Release notes

This page contains the release notes for PennyLane.

orphan

Release 0.35.0 (current release)

New features since last release

Qiskit 1.0 integration 🔌

  • This version of PennyLane makes it easier to import circuits from Qiskit. (#5218) (#5168)

    The qml.from_qiskit function converts a Qiskit QuantumCircuit into a PennyLane quantum function. Although qml.from_qiskit already exists in PennyLane, we have made a number of improvements to make importing from Qiskit easier. And yes — qml.from_qiskit functionality is compatible with both Qiskit 1.0 and earlier versions! Here’s a comprehensive list of the improvements:

    • You can now append PennyLane measurements onto the quantum function returned by qml.from_qiskit. Consider this simple Qiskit circuit:

      import pennylane as qml
      from qiskit import QuantumCircuit
      
      qc = QuantumCircuit(2)
      qc.rx(0.785, 0)
      qc.ry(1.57, 1)
      

      We can convert it into a PennyLane QNode in just a few lines, with PennyLane measurements easily included:

      >>> dev = qml.device("default.qubit")
      >>> measurements = qml.expval(qml.Z(0) @ qml.Z(1))
      >>> qfunc = qml.from_qiskit(qc, measurements=measurements)
      >>> qnode = qml.QNode(qfunc, dev)
      >>> qnode()
      tensor(0.00056331, requires_grad=True)
      
    • Quantum circuits that already contain Qiskit-side measurements can be faithfully converted with qml.from_qiskit. Consider this example Qiskit circuit:

      qc = QuantumCircuit(3, 2)  # Teleportation
      
      qc.rx(0.9, 0)  # Prepare input state on qubit 0
      
      qc.h(1)  # Prepare Bell state on qubits 1 and 2
      qc.cx(1, 2)
      
      qc.cx(0, 1)  # Perform teleportation
      qc.h(0)
      qc.measure(0, 0)
      qc.measure(1, 1)
      
      with qc.if_test((1, 1)):  # Perform first conditional
          qc.x(2)
      

      This circuit can be converted into PennyLane with the Qiskit measurements still accessible. For example, we can use those results as inputs to a mid-circuit measurement in PennyLane:

      @qml.qnode(dev)
      def teleport():
          m0, m1 = qml.from_qiskit(qc)()
          qml.cond(m0, qml.Z)(2)
          return qml.density_matrix(2)
      
      >>> teleport()
      tensor([[0.81080498+0.j        , 0.        +0.39166345j],
              [0.        -0.39166345j, 0.18919502+0.j        ]], requires_grad=True)
      
    • It is now more intuitive to handle and differentiate parametrized Qiskit circuits. Consider the following circuit:

      from qiskit.circuit import Parameter
      from pennylane import numpy as np
      
      angle0 = Parameter("x")
      angle1 = Parameter("y")
      
      qc = QuantumCircuit(2, 2)
      qc.rx(angle0, 0)
      qc.ry(angle1, 1)
      qc.cx(1, 0)
      

      We can convert this circuit into a QNode with two arguments, corresponding to x and y:

      measurements = qml.expval(qml.PauliZ(0))
      qfunc = qml.from_qiskit(qc, measurements)
      qnode = qml.QNode(qfunc, dev)
      

      The QNode can be evaluated and differentiated:

      >>> x, y = np.array([0.4, 0.5], requires_grad=True)
      >>> qnode(x, y)
      tensor(0.80830707, requires_grad=True)
      >>> qml.grad(qnode)(x, y)
      (tensor(-0.34174675, requires_grad=True),
       tensor(-0.44158016, requires_grad=True))
      

      This shows how easy it is to make a Qiskit circuit differentiable with PennyLane.

  • In addition to circuits, it is also possible to convert operators from Qiskit to PennyLane with a new function called qml.from_qiskit_op. (#5251)

    A Qiskit SparsePauliOp can be converted to a PennyLane operator using qml.from_qiskit_op:

    >>> from qiskit.quantum_info import SparsePauliOp
    >>> qiskit_op = SparsePauliOp(["II", "XY"])
    >>> qiskit_op
    SparsePauliOp(['II', 'XY'],
                  coeffs=[1.+0.j, 1.+0.j])
    >>> pl_op = qml.from_qiskit_op(qiskit_op)
    >>> pl_op
    I(0) + X(1) @ Y(0)
    

    Combined with qml.from_qiskit, it becomes easy to quickly calculate quantities like expectation values by converting the whole workflow to PennyLane:

    qc = QuantumCircuit(2)  # Create circuit
    qc.rx(0.785, 0)
    qc.ry(1.57, 1)
    
    measurements = qml.expval(pl_op)  # Create QNode
    qfunc = qml.from_qiskit(qc, measurements)
    qnode = qml.QNode(qfunc, dev)
    
    >>> qnode()  # Evaluate!
    tensor(0.29317504, requires_grad=True)
    

Native mid-circuit measurements on Default Qubit 💡

  • Mid-circuit measurements can now be more scalable and efficient in finite-shots mode with default.qubit by simulating them in a similar way to what happens on quantum hardware. (#5088) (#5120)

    Previously, mid-circuit measurements (MCMs) would be automatically replaced with an additional qubit using the @qml.defer_measurements transform. The circuit below would have required thousands of qubits to simulate.

    Now, MCMs are performed in a similar way to quantum hardware with finite shots on default.qubit. For each shot and each time an MCM is encountered, the device evaluates the probability of projecting onto |0> or |1> and makes a random choice to collapse the circuit state. This approach works well when there are a lot of MCMs and the number of shots is not too high.

    import pennylane as qml
    
    dev = qml.device("default.qubit", shots=10)
    
    @qml.qnode(dev)
    def f():
        for i in range(1967):
            qml.Hadamard(0)
            qml.measure(0)
        return qml.sample(qml.PauliX(0))
    
    >>> f()
    tensor([-1, -1, -1,  1,  1, -1,  1, -1,  1, -1], requires_grad=True)
    

Work easily and efficiently with operators 🔧

  • Over the past few releases, PennyLane’s approach to operator arithmetic has been in the process of being overhauled. We have a few objectives:

    1. To make it as easy to work with PennyLane operators as it would be with pen and paper.

    2. To improve the efficiency of operator arithmetic.

    The updated operator arithmetic functionality is still being finalized, but can be activated using qml.operation.enable_new_opmath(). In the next release, the new behaviour will become the default, so we recommend enabling now to become familiar with the new system!

    The following updates have been made in this version of PennyLane:

    • You can now easily access Pauli operators via I, X, Y, and Z: (#5116)

      >>> from pennylane import I, X, Y, Z
      >>> X(0)
      X(0)
      

      The original long-form names Identity, PauliX, PauliY, and PauliZ remain available, but use of the short-form names is now recommended.

    • A new qml.commutator function is now available that allows you to compute commutators between PennyLane operators. (#5051) (#5052) (#5098)

      >>> qml.commutator(X(0), Y(0))
      2j * Z(0)
      
    • Operators in PennyLane can have a backend Pauli representation, which can be used to perform faster operator arithmetic. Now, the Pauli representation will be automatically used for calculations when available. (#4989) (#5001) (#5003) (#5017) (#5027)

      The Pauli representation can be optionally accessed via op.pauli_rep:

      >>> qml.operation.enable_new_opmath()
      >>> op = X(0) + Y(0)
      >>> op.pauli_rep
      1.0 * X(0)
      + 1.0 * Y(0)
      
    • Extensive improvements have been made to the string representations of PennyLane operators, making them shorter and possible to copy-paste as valid PennyLane code. (#5116) (#5138)

      >>> 0.5 * X(0)
      0.5 * X(0)
      >>> 0.5 * (X(0) + Y(1))
      0.5 * (X(0) + Y(1))
      

      Sums with many terms are broken up into multiple lines, but can still be copied back as valid code:

      >>> 0.5 * (X(0) @ X(1)) + 0.7 * (X(1) @ X(2)) + 0.8 * (X(2) @ X(3))
      (
          0.5 * (X(0) @ X(1))
        + 0.7 * (X(1) @ X(2))
        + 0.8 * (X(2) @ X(3))
      )
      
    • Linear combinations of operators and operator multiplication via Sum and Prod, respectively, have been updated to reach feature parity with Hamiltonian and Tensor, respectively. This should minimize the effort to port over any existing code. (#5070) (#5132) (#5133)

      Updates include support for grouping via the pauli module:

      >>> obs = [X(0) @ Y(1), Z(0), Y(0) @ Z(1), Y(1)]
      >>> qml.pauli.group_observables(obs)
      [[Y(0) @ Z(1)], [X(0) @ Y(1), Y(1)], [Z(0)]]
      

New Clifford device 🦾

  • A new default.clifford device enables efficient simulation of large-scale Clifford circuits defined in PennyLane through the use of stim as a backend. (#4936) (#4954) (#5144)

    Given a circuit with only Clifford gates, one can use this device to obtain the usual range of PennyLane measurements as well as the state represented in the Tableau form of Aaronson & Gottesman (2004):

    import pennylane as qml
    
    dev = qml.device("default.clifford", tableau=True)
    @qml.qnode(dev)
    def circuit():
        qml.CNOT(wires=[0, 1])
        qml.PauliX(wires=[1])
        qml.ISWAP(wires=[0, 1])
        qml.Hadamard(wires=[0])
        return qml.state()
    
    >>> circuit()
    array([[0, 1, 1, 0, 0],
          [1, 0, 1, 1, 1],
          [0, 0, 0, 1, 0],
          [1, 0, 0, 1, 1]])
    

    The default.clifford device also supports the PauliError, DepolarizingChannel, BitFlip and PhaseFlip noise channels when operating in finite-shot mode.

Improvements 🛠

Faster gradients with VJPs and other performance improvements

  • Vector-Jacobian products (VJPs) can result in faster computations when the output of your quantum Node has a low dimension. They can be enabled by setting device_vjp=True when loading a QNode. In the next release of PennyLane, VJPs are planned to be used by default, when available.

    In this release, we have unlocked:

    • Adjoint device VJPs can be used with jax.jacobian, meaning that device_vjp=True is always faster when using JAX with default.qubit. (#4963)

    • PennyLane can now use lightning-provided VJPs. (#4914)

    • VJPs can be used with TensorFlow, though support has not yet been added for tf.Function and Tensorflow Autograph. (#4676)

  • Measuring qml.probs is now faster due to an optimization in converting samples to counts. (#5145)

  • The performance of circuit-cutting workloads with large numbers of generated tapes has been improved. (#5005)

  • Queueing (AnnotatedQueue) has been removed from qml.cut_circuit and qml.cut_circuit_mc to improve performance for large workflows. (#5108)

Community contributions 🥳

  • A new function called qml.fermi.parity_transform has been added for parity mapping of a fermionic Hamiltonian. (#4928)

    It is now possible to transform a fermionic Hamiltonian to a qubit Hamiltonian with parity mapping.

    import pennylane as qml
    fermi_ham = qml.fermi.FermiWord({(0, 0) : '+', (1, 1) : '-'})
    
    qubit_ham = qml.fermi.parity_transform(fermi_ham, n=6)
    
    >>> print(qubit_ham)
    -0.25j * Y(0) + (-0.25+0j) * (X(0) @ Z(1)) + (0.25+0j) * X(0) + 0.25j * (Y(0) @ Z(1))
    
  • The transform split_non_commuting now accepts measurements of type probs, sample, and counts, which accept both wires and observables. (#4972)

  • The efficiency of matrix calculations when an operator is symmetric over a given set of wires has been improved. (#3601)

  • The pennylane/math/quantum.py module now has support for computing the minimum entropy of a density matrix. (#3959)

    >>> x = [1, 0, 0, 1] / np.sqrt(2)
    >>> x = qml.math.dm_from_state_vector(x)
    >>> qml.math.min_entropy(x, indices=[0])
    0.6931471805599455
    
  • A function called apply_operation that applies operations to device-compatible states has been added to the new qutrit_mixed module found in qml.devices. (#5032)

  • A function called measure has been added to the new qutrit_mixed module found in qml.devices that measures device-compatible states for a collection of measurement processes. (#5049)

  • A partial_trace function has been added to qml.math for taking the partial trace of matrices. (#5152)

Other operator arithmetic improvements

  • The following capabilities have been added for Pauli arithmetic: (#4989) (#5001) (#5003) (#5017) (#5027) (#5018)

    • You can now multiply PauliWord and PauliSentence instances by scalars (e.g., 0.5 * PauliWord({0: "X"}) or 0.5 * PauliSentence({PauliWord({0: "X"}): 1.})).

    • You can now intuitively add and subtract PauliWord and PauliSentence instances and scalars together (scalars are treated implicitly as multiples of the identity, I). For example, ps1 + pw1 + 1. for some Pauli word pw1 = PauliWord({0: "X", 1: "Y"}) and Pauli sentence ps1 = PauliSentence({pw1: 3.}).

    • You can now element-wise multiply PauliWord, PauliSentence, and operators together with qml.dot (e.g., qml.dot([0.5, -1.5, 2], [pw1, ps1, id_word]) with id_word = PauliWord({})).

    • qml.matrix now accepts PauliWord and PauliSentence instances (e.g., qml.matrix(PauliWord({0: "X"}))).

    • It is now possible to compute commutators with Pauli operators natively with the new commutator method.

      >>> op1 = PauliWord({0: "X", 1: "X"})
      >>> op2 = PauliWord({0: "Y"}) + PauliWord({1: "Y"})
      >>> op1.commutator(op2)
      2j * Z(0) @ X(1)
      + 2j * X(0) @ Z(1)
      
  • Composite operations (e.g., those made with qml.prod and qml.sum) and scalar-product operations convert Hamiltonian and Tensor operands to Sum and Prod types, respectively. This helps avoid the mixing of incompatible operator types. (#5031) (#5063)

  • qml.Identity() can be initialized without wires. Measuring it is currently not possible, though. (#5106)

  • qml.dot now returns a Sum class even when all the coefficients match. (#5143)

  • qml.pauli.group_observables now supports grouping Prod and SProd operators. (#5070)

  • The performance of converting a PauliSentence to a Sum has been improved. (#5141) (#5150)

  • Akin to qml.Hamiltonian features, the coefficients and operators that make up composite operators formed via Sum or Prod can now be accessed with the terms() method. (#5132) (#5133) (#5164)

    >>> qml.operation.enable_new_opmath()
    >>> op = X(0) @ (0.5 * X(1) + X(2))
    >>> op.terms()
    ([0.5, 1.0],
     [X(1) @ X(0),
      X(2) @ X(0)])
    
  • String representations of ParametrizedHamiltonian have been updated to match the style of other PL operators. (#5215)

Other improvements

  • The pl-device-test suite is now compatible with the qml.devices.Device interface. (#5229)

  • The QSVT operation now determines its data from the block encoding and projector operator data. (#5226) (#5248)

  • The BlockEncode operator is now JIT-compatible with JAX. (#5110)

  • The qml.qsvt function uses qml.GlobalPhase instead of qml.exp to define a global phase. (#5105)

  • The tests/ops/functions/conftest.py test has been updated to ensure that all operator types are tested for validity. (#4978)

  • A new pennylane.workflow module has been added. This module now contains qnode.py, execution.py, set_shots.py, jacobian_products.py, and the submodule interfaces. (#5023)

  • A more informative error is now raised when calling adjoint_jacobian with trainable state-prep operations. (#5026)

  • qml.workflow.get_transform_program and qml.workflow.construct_batch have been added to inspect the transform program and batch of tapes at different stages. (#5084)

  • All custom controlled operations such as CRX, CZ, CNOT, ControlledPhaseShift now inherit from ControlledOp, giving them additional properties such as control_wire and control_values. Calling qml.ctrl on RX, RY, RZ, Rot, and PhaseShift with a single control wire will return gates of types CRX, CRY, etc. as opposed to a general Controlled operator. (#5069) (#5199)

  • The CI will now fail if coverage data fails to upload to codecov. Previously, it would silently pass and the codecov check itself would never execute. (#5101)

  • qml.ctrl called on operators with custom controlled versions will now return instances of the custom class, and it will flatten nested controlled operators to a single multi-controlled operation. For PauliX, CNOT, Toffoli, and MultiControlledX, calling qml.ctrl will always resolve to the best option in CNOT, Toffoli, or MultiControlledX depending on the number of control wires and control values. (#5125)

  • Unwanted warning filters have been removed from tests and no PennyLaneDeprecationWarnings are being raised unexpectedly. (#5122)

  • New error tracking and propagation functionality has been added (#5115) (#5121)

  • The method map_batch_transform has been replaced with the method _batch_transform implemented in TransformDispatcher. (#5212)

  • TransformDispatcher can now dispatch onto a batch of tapes, making it easier to compose transforms when working in the tape paradigm. (#5163)

  • qml.ctrl is now a simple wrapper that either calls PennyLane’s built in create_controlled_op or uses the Catalyst implementation. (#5247)

  • Controlled composite operations can now be decomposed using ZYZ rotations. (#5242)

  • New functions called qml.devices.modifiers.simulator_tracking and qml.devices.modifiers.single_tape_support have been added to add basic default behavior onto a device class. (#5200)

Breaking changes 💔

  • Passing additional arguments to a transform that decorates a QNode must now be done through the use of functools.partial. (#5046)

  • qml.ExpvalCost has been removed. Users should use qml.expval() moving forward. (#5097)

  • Caching of executions is now turned off by default when max_diff == 1, as the classical overhead cost outweighs the probability that duplicate circuits exists. (#5243)

  • The entry point convention registering compilers with PennyLane has changed. (#5140)

    To allow for packages to register multiple compilers with PennyLane, the entry_points convention under the designated group name pennylane.compilers has been modified.

    Previously, compilers would register qjit (JIT decorator), ops (compiler-specific operations), and context (for tracing and program capture).

    Now, compilers must register compiler_name.qjit, compiler_name.ops, and compiler_name.context, where compiler_name is replaced by the name of the provided compiler.

    For more information, please see the documentation on adding compilers.

  • PennyLane source code is now compatible with the latest version of black. (#5112) (#5119)

  • gradient_analysis_and_validation has been renamed to find_and_validate_gradient_methods. Instead of returning a list, it now returns a dictionary of gradient methods for each parameter index, and no longer mutates the tape. (#5035)

  • Multiplying two PauliWord instances no longer returns a tuple (new_word, coeff) but instead PauliSentence({new_word: coeff}). The old behavior is still available with the private method PauliWord._matmul(other) for faster processing. (#5045)

  • Observable.return_type has been removed. Instead, you should inspect the type of the surrounding measurement process. (#5044)

  • ClassicalShadow.entropy() no longer needs an atol keyword as a better method to estimate entropies from approximate density matrix reconstructions (with potentially negative eigenvalues). (#5048)

  • Controlled operators with a custom controlled version decompose like how their controlled counterpart decomposes as opposed to decomposing into their controlled version. (#5069) (#5125)

    For example:

    >>> qml.ctrl(qml.RX(0.123, wires=1), control=0).decomposition()
    [
      RZ(1.5707963267948966, wires=[1]),
      RY(0.0615, wires=[1]),
      CNOT(wires=[0, 1]),
      RY(-0.0615, wires=[1]),
      CNOT(wires=[0, 1]),
      RZ(-1.5707963267948966, wires=[1])
    ]
    
  • QuantumScript.is_sampled and QuantumScript.all_sampled have been removed. Users should now validate these properties manually. (#5072)

  • qml.transforms.one_qubit_decomposition and qml.transforms.two_qubit_decomposition have been removed. Instead, you should use qml.ops.one_qubit_decomposition and qml.ops.two_qubit_decomposition. (#5091)

Deprecations 👋

  • Calling qml.matrix without providing a wire_order on objects where the wire order could be ambiguous now raises a warning. In the future, the wire_order argument will be required in these cases. (#5039)

  • Operator.validate_subspace(subspace) has been relocated to the qml.ops.qutrit.parametric_ops module and will be removed from the Operator class in an upcoming release. (#5067)

  • Matrix and tensor products between PauliWord and PauliSentence instances are done using the @ operator, * will be used only for scalar multiplication. Note also the breaking change that the product of two PauliWord instances now returns a PauliSentence instead of a tuple (new_word, coeff). (#4989) (#5054)

  • MeasurementProcess.name and MeasurementProcess.data are now deprecated, as they contain dummy values that are no longer needed. (#5047) (#5071) (#5076) (#5122)

  • qml.pauli.pauli_mult and qml.pauli.pauli_mult_with_phase are now deprecated. Instead, you should use qml.simplify(qml.prod(pauli_1, pauli_2)) to get the reduced operator. (#5057)

  • The private functions _pauli_mult, _binary_matrix and _get_pauli_map from the pauli module have been deprecated, as they are no longer used anywhere and the same functionality can be achieved using newer features in the pauli module. (#5057)

  • Sum.ops, Sum.coeffs, Prod.ops and Prod.coeffs will be deprecated in the future. (#5164)

Documentation 📝

  • The module documentation for pennylane.tape now explains the difference between QuantumTape and QuantumScript. (#5065)

  • A typo in a code example in the qml.transforms API has been fixed. (#5014)

  • Documentation for qml.data has been updated and now mentions a way to access the same dataset simultaneously from multiple environments. (#5029)

  • A clarification for the definition of argnum added to gradient methods has been made. (#5035)

  • A typo in the code example for qml.qchem.dipole_of has been fixed. (#5036)

  • A development guide on deprecations and removals has been added. (#5083)

  • A note about the eigenspectrum of second-quantized Hamiltonians has been added to qml.eigvals. (#5095)

  • A warning about two mathematically equivalent Hamiltonians undergoing different time evolutions has been added to qml.TrotterProduct and qml.ApproxTimeEvolution. (#5137)

  • A reference to the paper that provides the image of the qml.QAOAEmbedding template has been added. (#5130)

  • The docstring of qml.sample has been updated to advise the use of single-shot expectations instead when differentiating a circuit. (#5237)

  • A quick start page has been added called “Importing Circuits”. This explains how to import quantum circuits and operations defined outside of PennyLane. (#5281)

Bug fixes 🐛

  • QubitChannel can now be used with jitting. (#5288)

  • Fixed a bug in the matplotlib drawer where the colour of Barrier did not match the requested style. (#5276)

  • qml.draw and qml.draw_mpl now apply all applied transforms before drawing. (#5277)

  • ctrl_decomp_zyz is now differentiable. (#5198)

  • qml.ops.Pow.matrix() is now differentiable with TensorFlow with integer exponents. (#5178)

  • The qml.MottonenStatePreparation template has been updated to include a global phase operation. (#5166)

  • Fixed a queuing bug when using qml.prod with a quantum function that queues a single operator. (#5170)

  • The qml.TrotterProduct template has been updated to accept scalar products of operators as an input Hamiltonian. (#5073)

  • Fixed a bug where caching together with JIT compilation and broadcasted tapes yielded wrong results Operator.hash now depends on the memory location, id, of a JAX tracer instead of its string representation. (#3917)

  • qml.transforms.undo_swaps can now work with operators with hyperparameters or nesting. (#5081)

  • qml.transforms.split_non_commuting will now pass the original shots along. (#5081)

  • If argnum is provided to a gradient transform, only the parameters specified in argnum will have their gradient methods validated. (#5035)

  • StatePrep operations expanded onto more wires are now compatible with backprop. (#5028)

  • qml.equal works well with qml.Sum operators when wire labels are a mix of integers and strings. (#5037)

  • The return value of Controlled.generator now contains a projector that projects onto the correct subspace based on the control value specified. (#5068)

  • CosineWindow no longer raises an unexpected error when used on a subset of wires at the beginning of a circuit. (#5080)

  • tf.function now works with TensorSpec(shape=None) by skipping batch size computation. (#5089)

  • PauliSentence.wires no longer imposes a false order. (#5041)

  • qml.qchem.import_state now applies the chemist-to-physicist sign convention when initializing a PennyLane state vector from classically pre-computed wavefunctions. That is, it interleaves spin-up/spin-down operators for the same spatial orbital index, as standard in PennyLane (instead of commuting all spin-up operators to the left, as is standard in quantum chemistry). (#5114)

  • Multi-wire controlled CNOT and PhaseShift are now be decomposed correctly. (#5125) (#5148)

  • draw_mpl no longer raises an error when drawing a circuit containing an adjoint of a controlled operation. (#5149)

  • default.mixed no longer throws ValueError when applying a state vector that is not of type complex128 when used with tensorflow. (#5155)

  • ctrl_decomp_zyz no longer raises a TypeError if the rotation parameters are of type torch.Tensor (#5183)

  • Comparing Prod and Sum objects now works regardless of nested structure with qml.equal if the operators have a valid pauli_rep property. (#5177)

  • Controlled GlobalPhase with non-zero control wires no longer throws an error. (#5194)

  • A QNode transformed with mitigate_with_zne now accepts batch parameters. (#5195)

  • The matrix of an empty PauliSentence instance is now correct (all-zeros). Further, matrices of empty PauliWord and PauliSentence instances can now be turned into matrices. (#5188)

  • PauliSentence instances can handle matrix multiplication with PauliWord instances. (#5208)

  • CompositeOp.eigendecomposition is now JIT-compatible. (#5207)

  • QubitDensityMatrix now works with JAX-JIT on the default.mixed device. (#5203) (#5236)

  • When a QNode specifies diff_method="adjoint", default.qubit no longer tries to decompose non-trainable operations with non-scalar parameters such as QubitUnitary. (#5233)

  • The overwriting of the class names of I, X, Y, and Z no longer happens in the initialization after causing problems with datasets. This now happens globally. (#5252)

  • The adjoint_metric_tensor transform now works with jax. (#5271)

Contributors ✍️

This release contains contributions from (in alphabetical order):

Abhishek Abhishek, Mikhail Andrenkov, Utkarsh Azad, Trenten Babcock, Gabriel Bottrill, Thomas Bromley, Astral Cai, Skylar Chan, Isaac De Vlugt, Diksha Dhawan, Lillian Frederiksen, Pietropaolo Frisoni, Eugenio Gigante, Diego Guala, David Ittah, Soran Jahangiri, Jacky Jiang, Korbinian Kottmann, Christina Lee, Xiaoran Li, Vincent Michaud-Rioux, Romain Moyard, Pablo Antonio Moreno Casares, Erick Ochoa Lopez, Lee J. O’Riordan, Mudit Pandey, Alex Preciado, Matthew Silverman, Jay Soni.

orphan

Release 0.34.0

New features since last release

Statistics and drawing for mid-circuit measurements 🎨

  • It is now possible to return statistics of composite mid-circuit measurements. (#4888)

    Mid-circuit measurement results can be composed using basic arithmetic operations and then statistics can be calculated by putting the result within a PennyLane measurement like qml.expval(). For example:

    import pennylane as qml
    
    dev = qml.device("default.qubit")
    
    @qml.qnode(dev)
    def circuit(phi, theta):
        qml.RX(phi, wires=0)
        m0 = qml.measure(wires=0)
        qml.RY(theta, wires=1)
        m1 = qml.measure(wires=1)
        return qml.expval(~m0 + m1)
    
    print(circuit(1.23, 4.56))
    
    1.2430187928114291
    

    Another option, for ease-of-use when using qml.sample(), qml.probs(), or qml.counts(), is to provide a simple list of mid-circuit measurement results:

    dev = qml.device("default.qubit")
    
    @qml.qnode(dev)
    def circuit(phi, theta):
        qml.RX(phi, wires=0)
        m0 = qml.measure(wires=0)
        qml.RY(theta, wires=1)
        m1 = qml.measure(wires=1)
        return qml.sample(op=[m0, m1])
    
    print(circuit(1.23, 4.56, shots=5))
    
    [[0 1]
     [0 1]
     [0 0]
     [1 0]
     [0 1]]
    

    Composite mid-circuit measurement statistics are supported on default.qubit and default.mixed. To learn more about which measurements and arithmetic operators are supported, refer to the measurements page and the documentation for qml.measure.

  • Mid-circuit measurements can now be visualized with the text-based qml.draw() and the graphical qml.draw_mpl() methods. (#4775) (#4803) (#4832) (#4901) (#4850) (#4917) (#4930) (#4957)

    Drawing of mid-circuit measurement capabilities including qubit reuse and reset, postselection, conditioning, and collecting statistics is now supported. Here is an all-encompassing example:

    def circuit():
        m0 = qml.measure(0, reset=True)
        m1 = qml.measure(1, postselect=1)
        qml.cond(m0 - m1 == 0, qml.S)(0)
        m2 = qml.measure(1)
        qml.cond(m0 + m1 == 2, qml.T)(0)
        qml.cond(m2, qml.PauliX)(1)
    

    The text-based drawer outputs:

    >>> print(qml.draw(circuit)())
    0: ──┤↗│  │0⟩────────S───────T────┤
    1: ───║────────┤↗₁├──║──┤↗├──║──X─┤
          ╚═════════║════╬═══║═══╣  ║
                    ╚════╩═══║═══╝  ║
                             ╚══════╝
    

    The graphical drawer outputs:

    >>> print(qml.draw_mpl(circuit)())
    

Catalyst is seamlessly integrated with PennyLane ⚗️

  • Catalyst, our next-generation compilation framework, is now accessible within PennyLane, allowing you to more easily benefit from hybrid just-in-time (JIT) compilation.

    To access these features, simply install pennylane-catalyst:

    pip install pennylane-catalyst
    

    The qml.compiler module provides support for hybrid quantum-classical compilation. (#4692) (#4979)

    Through the use of the qml.qjit decorator, entire workflows can be JIT compiled — including both quantum and classical processing — down to a machine binary on first-function execution. Subsequent calls to the compiled function will execute the previously-compiled binary, resulting in significant performance improvements.

    import pennylane as qml
    
    dev = qml.device("lightning.qubit", wires=2)
    
    @qml.qjit
    @qml.qnode(dev)
    def circuit(theta):
        qml.Hadamard(wires=0)
        qml.RX(theta, wires=1)
        qml.CNOT(wires=[0,1])
        return qml.expval(qml.PauliZ(wires=1))
    
    >>> circuit(0.5)  # the first call, compilation occurs here
    array(0.)
    >>> circuit(0.5)  # the precompiled quantum function is called
    array(0.)
    

    Currently, PennyLane supports the Catalyst hybrid compiler with the qml.qjit decorator. A significant benefit of Catalyst is the ability to preserve complex control flow around quantum operations — such as if statements and for loops, and including measurement feedback — during compilation, while continuing to support end-to-end autodifferentiation.

  • The following functions can now be used with the qml.qjit decorator: qml.grad, qml.jacobian, qml.vjp, qml.jvp, and qml.adjoint. (#4709) (#4724) (#4725) (#4726)

    When qml.grad or qml.jacobian are used with @qml.qjit, they are patched to catalyst.grad and catalyst.jacobian, respectively.

    dev = qml.device("lightning.qubit", wires=1)
    
    @qml.qjit
    def workflow(x):
    
        @qml.qnode(dev)
        def circuit(x):
            qml.RX(np.pi * x[0], wires=0)
            qml.RY(x[1], wires=0)
            return qml.probs()
    
        g = qml.jacobian(circuit)
    
        return g(x)
    
    >>> workflow(np.array([2.0, 1.0]))
    array([[ 3.48786850e-16, -4.20735492e-01],
           [-8.71967125e-17,  4.20735492e-01]])
    
  • JIT-compatible functionality for control flow has been added via qml.for_loop, qml.while_loop, and qml.cond. (#4698) (#5006)

    qml.for_loop and qml.while_loop can be deployed as decorators on functions that are the body of the loop. The arguments to both follow typical conventions:

    @qml.for_loop(lower_bound, upper_bound, step)
    
    @qml.while_loop(cond_function)
    

    Here is a concrete example with qml.for_loop:

    qml.for_loop and qml.while_loop can be deployed as decorators on functions that are the body of the loop. The arguments to both follow typical conventions:

    @qml.for_loop(lower_bound, upper_bound, step)
    
    @qml.while_loop(cond_function)
    

    Here is a concrete example with qml.for_loop:

    dev = qml.device("lightning.qubit", wires=1)
    
    @qml.qjit
    @qml.qnode(dev)
    def circuit(n: int, x: float):
    
        @qml.for_loop(0, n, 1)
        def loop_rx(i, x):
            # perform some work and update (some of) the arguments
            qml.RX(x, wires=0)
    
            # update the value of x for the next iteration
            return jnp.sin(x)
    
        # apply the for loop
        final_x = loop_rx(x)
    
        return qml.expval(qml.PauliZ(0)), final_x
    
    >>> circuit(7, 1.6)
    (array(0.97926626), array(0.55395718))
    

Decompose circuits into the Clifford+T gateset 🧩

  • The new qml.clifford_t_decomposition() transform provides an approximate breakdown of an input circuit into the Clifford+T gateset. Behind the scenes, this decomposition is enacted via the sk_decomposition() function using the Solovay-Kitaev algorithm. (#4801) (#4802)

    The Solovay-Kitaev algorithm approximately decomposes a quantum circuit into the Clifford+T gateset. To account for this, a desired total circuit decomposition error, epsilon, must be specified when using qml.clifford_t_decomposition:

    dev = qml.device("default.qubit")
    
    @qml.qnode(dev)
    def circuit():
        qml.RX(1.1, 0)
        return qml.state()
    
    circuit = qml.clifford_t_decomposition(circuit, epsilon=0.1)
    
    >>> print(qml.draw(circuit)())
    0: ──T†──H──T†──H──T──H──T──H──T──H──T──H──T†──H──T†──T†──H──T†──H──T──H──T──H──T──H──T──H──T†──H
    
    ───T†──H──T──H──GlobalPhase(0.39)─┤
    

    The resource requirements of this circuit can also be evaluated:

    >>> with qml.Tracker(dev) as tracker:
    ...     circuit()
    >>> resources_lst = tracker.history["resources"]
    >>> resources_lst[0]
    wires: 1
    gates: 34
    depth: 34
    shots: Shots(total=None)
    gate_types:
    {'Adjoint(T)': 8, 'Hadamard': 16, 'T': 9, 'GlobalPhase': 1}
    gate_sizes:
    {1: 33, 0: 1}
    

Use an iterative approach for quantum phase estimation 🔄

  • Iterative Quantum Phase Estimation is now available with qml.iterative_qpe. (#4804)

    The subroutine can be used similarly to mid-circuit measurements:

    import pennylane as qml
    
    dev = qml.device("default.qubit", shots=5)
    
    @qml.qnode(dev)
    def circuit():
    
      # Initial state
      qml.PauliX(wires=[0])
    
      # Iterative QPE
      measurements = qml.iterative_qpe(qml.RZ(2., wires=[0]), ancilla=[1], iters=3)
    
      return [qml.sample(op=meas) for meas in measurements]
    
    >>> print(circuit())
    [array([0, 0, 0, 0, 0]), array([1, 0, 0, 0, 0]), array([0, 1, 1, 1, 1])]
    

    The \(i\)-th element in the list refers to the 5 samples generated by the \(i\)-th measurement of the algorithm.

Improvements 🛠

Community contributions 🥳

  • The += operand can now be used with a PauliSentence, which has also provides a performance boost. (#4662)

  • The Approximate Quantum Fourier Transform (AQFT) is now available with qml.AQFT. (#4715)

  • qml.draw and qml.draw_mpl now render operator IDs. (#4749)

    The ID can be specified as a keyword argument when instantiating an operator:

    >>> def circuit():
    ...     qml.RX(0.123, id="data", wires=0)
    >>> print(qml.draw(circuit)())
    0: ──RX(0.12,"data")─┤
    
  • Non-parametric operators such as Barrier, Snapshot, and Wirecut have been grouped together and moved to pennylane/ops/meta.py. Additionally, the relevant tests have been organized and placed in a new file, tests/ops/test_meta.py. (#4789)

  • The TRX, TRY, and TRZ operators are now differentiable via backpropagation on default.qutrit. (#4790)

  • The function qml.equal now supports ControlledSequence operators. (#4829)

  • XZX decomposition has been added to the list of supported single-qubit unitary decompositions. (#4862)

  • == and != operands can now be used with TransformProgram and TransformContainers instances. (#4858)

  • A qutrit_mixed module has been added to qml.devices to store helper functions for a future qutrit mixed-state device. A function called create_initial_state has been added to this module that creates device-compatible initial states. (#4861)

  • The function qml.Snapshot now supports arbitrary state-based measurements (i.e., measurements of type StateMeasurement). (#4876)

  • qml.equal now supports the comparison of QuantumScript and BasisRotation objects. (#4902) (#4919)

  • The function qml.draw_mpl now accept a keyword argument fig to specify the output figure window. (#4956)

Better support for batching

  • qml.AmplitudeEmbedding now supports batching when used with Tensorflow. (#4818)

  • default.qubit can now evolve already batched states with qml.pulse.ParametrizedEvolution. (#4863)

  • qml.ArbitraryUnitary now supports batching. (#4745)

  • Operator and tape batch sizes are evaluated lazily, helping run expensive computations less frequently and an issue with Tensorflow pre-computing batch sizes. (#4911)

Performance improvements and benchmarking

  • Autograd, PyTorch, and JAX can now use vector-Jacobian products (VJPs) provided by the device from the new device API. If a device provides a VJP, this can be selected by providing device_vjp=True to a QNode or qml.execute. (#4935) (#4557) (#4654) (#4878) (#4841)

    >>> dev = qml.device('default.qubit')
    >>> @qml.qnode(dev, diff_method="adjoint", device_vjp=True)
    >>> def circuit(x):
    ...     qml.RX(x, wires=0)
    ...     return qml.expval(qml.PauliZ(0))
    >>> with dev.tracker:
    ...     g = qml.grad(circuit)(qml.numpy.array(0.1))
    >>> dev.tracker.totals
    {'batches': 1, 'simulations': 1, 'executions': 1, 'vjp_batches': 1, 'vjps': 1}
    >>> g
    -0.09983341664682815
    
  • qml.expval with large Hamiltonian objects is now faster and has a significantly lower memory footprint (and constant with respect to the number of Hamiltonian terms) when the Hamiltonian is a PauliSentence. This is due to the introduction of a specialized dot method in the PauliSentence class which performs PauliSentence-state products. (#4839)

  • default.qubit no longer uses a dense matrix for MultiControlledX for more than 8 operation wires. (#4673)

  • Some relevant Pytests have been updated to enable its use as a suite of benchmarks. (#4703)

  • default.qubit now applies GroverOperator faster by not using its matrix representation but a custom rule for apply_operation. Also, the matrix representation of GroverOperator now runs faster. (#4666)

  • A new pipeline to run benchmarks and plot graphs comparing with a fixed reference has been added. This pipeline will run on a schedule and can be activated on a PR with the label ci:run_benchmarks. (#4741)

  • default.qubit now supports adjoint differentiation for arbitrary diagonal state-based measurements. (#4865)

  • The benchmarks pipeline has been expanded to export all benchmark data to a single JSON file and a CSV file with runtimes. This includes all references and local benchmarks. (#4873)

Final phase of updates to transforms

  • qml.quantum_monte_carlo and qml.simplify now use the new transform system. (#4708) (#4949)

  • The formal requirement that type hinting be provided when using the qml.transform decorator has been removed. Type hinting can still be used, but is now optional. Please use a type checker such as mypy if you wish to ensure types are being passed correctly. (#4942)

Other improvements

  • Add PyTree-serialization interface for the Wires class. (#4998)

  • PennyLane now supports Python 3.12. (#4985)

  • SampleMeasurement now has an optional method process_counts for computing the measurement results from a counts dictionary. (#4941)

  • A new function called ops.functions.assert_valid has been added for checking if an Operator class is defined correctly. (#4764)

  • Shots objects can now be multiplied by scalar values. (#4913)

  • GlobalPhase now decomposes to nothing in case devices do not support global phases. (#4855)

  • Custom operations can now provide their matrix directly through the Operator.matrix() method without needing to update the has_matrix property. has_matrix will now automatically be True if Operator.matrix is overridden, even if Operator.compute_matrix is not. (#4844)

  • The logic for re-arranging states before returning them has been improved. (#4817)

  • When multiplying SparseHamiltonians by a scalar value, the result now stays as a SparseHamiltonian. (#4828)

  • trainable_params can now be set upon initialization of a QuantumScript instead of having to set the parameter after initialization. (#4877)

  • default.qubit now calculates the expectation value of Hermitian operators in a differentiable manner. (#4866)

  • The rot decomposition now has support for returning a global phase. (#4869)

  • The "pennylane_sketch" MPL-drawer style has been added. This is the same as the "pennylane" style, but with sketch-style lines. (#4880)

  • Operators now define a pauli_rep property, an instance of PauliSentence, defaulting to None if the operator has not defined it (or has no definition in the Pauli basis). (#4915)

  • qml.ShotAdaptiveOptimizer can now use a multinomial distribution for spreading shots across the terms of a Hamiltonian measured in a QNode. Note that this is equivalent to what can be done with qml.ExpvalCost, but this is the preferred method because ExpvalCost is deprecated. (#4896)

  • Decomposition of qml.PhaseShift now uses qml.GlobalPhase for retaining the global phase information. (#4657) (#4947)

  • qml.equal for Controlled operators no longer returns False when equivalent but differently-ordered sets of control wires and control values are compared. (#4944)

  • All PennyLane Operator subclasses are automatically tested by ops.functions.assert_valid to ensure that they follow PennyLane Operator standards. (#4922)

  • Probability measurements can now be calculated from a counts dictionary with the addition of a process_counts method in the ProbabilityMP class. (#4952)

  • ClassicalShadow.entropy now uses the algorithm outlined in 1106.5458 to project the approximate density matrix (with potentially negative eigenvalues) onto the closest valid density matrix. (#4959)

  • The ControlledSequence.compute_decomposition default now decomposes the Pow operators, improving compatibility with machine learning interfaces. (#4995)

Breaking changes 💔

  • The function qml.transforms.classical_jacobian has been moved to the gradients module and is now accessible as qml.gradients.classical_jacobian. (#4900)

  • The transforms submodule qml.transforms.qcut is now its own module: qml.qcut. (#4819)

  • The decomposition of GroverOperator now has an additional global phase operation. (#4666)

  • qml.cond and the Conditional operation have been moved from the transforms folder to the ops/op_math folder. qml.transforms.Conditional will now be available as qml.ops.Conditional. (#4860)

  • The prep keyword argument has been removed from QuantumScript and QuantumTape. StatePrepBase operations should be placed at the beginning of the ops list instead. (#4756)

  • qml.gradients.pulse_generator is now named qml.gradients.pulse_odegen to adhere to paper naming conventions. (#4769)

  • Specifying control_values passed to qml.ctrl as a string is no longer supported. (#4816)

  • The rot decomposition will now normalize its rotation angles to the range [0, 4pi] for consistency (#4869)

  • QuantumScript.graph is now built using tape.measurements instead of tape.observables because it depended on the now-deprecated Observable.return_type property. (#4762)

  • The "pennylane" MPL-drawer style now draws straight lines instead of sketch-style lines. (#4880)

  • The default value for the term_sampling argument of ShotAdaptiveOptimizer is now None instead of "weighted_random_sampling". (#4896)

Deprecations 👋

  • single_tape_transform, batch_transform, qfunc_transform, and op_transform are deprecated. Use the new qml.transform function instead. (#4774)

  • Observable.return_type is deprecated. Instead, you should inspect the type of the surrounding measurement process. (#4762) (#4798)

  • All deprecations now raise a qml.PennyLaneDeprecationWarning instead of a UserWarning. (#4814)

  • QuantumScript.is_sampled and QuantumScript.all_sampled are deprecated. Users should now validate these properties manually. (#4773)

  • With an algorithmic improvement to ClassicalShadow.entropy, the keyword atol becomes obsolete and will be removed in v0.35. (#4959)

Documentation 📝

  • Documentation for unitaries and operations’ decompositions has been moved from qml.transforms to qml.ops.ops_math. (#4906)

  • Documentation for qml.metric_tensor and qml.adjoint_metric_tensor and qml.transforms.classical_jacobian is now accessible via the gradients API page qml.gradients in the documentation. (#4900)

  • Documentation for qml.specs has been moved to the resource module. (#4904)

  • Documentation for QCut has been moved to its own API page: qml.qcut. (#4819)

  • The documentation page for qml.measurements now links top-level accessible functions (e.g., qml.expval) to their top-level pages rather than their module-level pages (e.g., qml.measurements.expval). (#4750)

  • Information for the documentation of qml.matrix about wire ordering has been added for using qml.matrix on a QNode which uses a device with device.wires=None. (#4874)

Bug fixes 🐛

  • TransformDispatcher now stops queuing when performing the transform when applying it to a qfunc. Only the output of the transform will be queued. (#4983)

  • qml.map_wires now works properly with qml.cond and qml.measure. (#4884)

  • Pow operators are now picklable. (#4966)

  • Finite differences and SPSA can now be used with tensorflow-autograph on setups that were seeing a bus error. (#4961)

  • qml.cond no longer incorrectly queues operators used arguments. (#4948)

  • Attribute objects now return False instead of raising a TypeError when checking if an object is inside the set. (#4933)

  • Fixed a bug where the parameter-shift rule of qml.ctrl(op) was wrong if op had a generator that has two or more eigenvalues and is stored as a SparseHamiltonian. (#4899)

  • Fixed a bug where trainable parameters in the post-processing of finite-differences were incorrect for JAX when applying the transform directly on a QNode. (#4879)

  • qml.grad and qml.jacobian now explicitly raise errors if trainable parameters are integers. (#4836)

  • JAX-JIT now works with shot vectors. (#4772)

  • JAX can now differentiate a batch of circuits where one tape does not have trainable parameters. (#4837)

  • The decomposition of GroverOperator now has the same global phase as its matrix. (#4666)

  • The tape.to_openqasm method no longer mistakenly includes interface information in the parameter string when converting tapes using non-NumPy interfaces. (#4849)

  • qml.defer_measurements now correctly transforms circuits when terminal measurements include wires used in mid-circuit measurements. (#4787)

  • Fixed a bug where the adjoint differentiation method would fail if an operation that has a parameter with grad_method=None is present. (#4820)

  • MottonenStatePreparation and BasisStatePreparation now raise an error when decomposing a broadcasted state vector. (#4767)

  • Gradient transforms now work with overridden shot vectors and default.qubit. (#4795)

  • Any ScalarSymbolicOp, like Evolution, now states that it has a matrix if the target is a Hamiltonian. (#4768)

  • In default.qubit, initial states are now initialized with the simulator’s wire order, not the circuit’s wire order. (#4781)

  • qml.compile will now always decompose to expand_depth, even if a target basis set is not specified. (#4800)

  • qml.transforms.transpile can now handle measurements that are broadcasted onto all wires. (#4793)

  • Parametrized circuits whose operators do not act on all wires return PennyLane tensors instead of NumPy arrays, as expected. (#4811) (#4817)

  • qml.transforms.merge_amplitude_embedding no longer depends on queuing, allowing it to work as expected with QNodes. (#4831)

  • qml.pow(op) and qml.QubitUnitary.pow() now also work with Tensorflow data raised to an integer power. (#4827)

  • The text drawer has been fixed to correctly label qml.qinfo measurements, as well as qml.classical_shadow qml.shadow_expval. (#4803)

  • Removed an implicit assumption that an empty PauliSentence gets treated as identity under multiplication. (#4887)

  • Using a CNOT or PauliZ operation with large batched states and the Tensorflow interface no longer raises an unexpected error. (#4889)

  • qml.map_wires no longer fails when mapping nested quantum tapes. (#4901)

  • Conversion of circuits to openqasm now decomposes to a depth of 10, allowing support for operators requiring more than 2 iterations of decomposition, such as the ApproxTimeEvolution gate. (#4951)

  • MPLDrawer does not add the bonus space for classical wires when no classical wires are present. (#4987)

  • Projector now works with parameter-broadcasting. (#4993)

  • The jax-jit interface can now be used with float32 mode. (#4990)

  • Keras models with a qnn.KerasLayer no longer fail to save and load weights properly when they are named “weights”. (#5008)

Contributors ✍️

This release contains contributions from (in alphabetical order):

Guillermo Alonso, Ali Asadi, Utkarsh Azad, Gabriel Bottrill, Thomas Bromley, Astral Cai, Minh Chau, Isaac De Vlugt, Amintor Dusko, Pieter Eendebak, Lillian Frederiksen, Pietropaolo Frisoni, Josh Izaac, Juan Giraldo, Emiliano Godinez Ramirez, Ankit Khandelwal, Korbinian Kottmann, Christina Lee, Vincent Michaud-Rioux, Anurav Modak, Romain Moyard, Mudit Pandey, Matthew Silverman, Jay Soni, David Wierichs, Justin Woodring, Sergei Mironov.

orphan

Release 0.33.1

Bug fixes 🐛

  • Fix gradient performance regression due to expansion of VJP products. (#4806)

  • qml.defer_measurements now correctly transforms circuits when terminal measurements include wires used in mid-circuit measurements. (#4787)

  • Any ScalarSymbolicOp, like Evolution, now states that it has a matrix if the target is a Hamiltonian. (#4768)

  • In default.qubit, initial states are now initialized with the simulator’s wire order, not the circuit’s wire order. (#4781)

Contributors ✍️

This release contains contributions from (in alphabetical order):

Christina Lee, Lee James O’Riordan, Mudit Pandey

orphan

Release 0.33.0

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 #4634

    Check 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 HardwareHamiltonians 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 #4521

  • The 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 SProds 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 AmplitudeEmbeddings 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

    (#2845) (#2846) (#2847)

  • 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

    (#4081) (#4082)

  • 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

    (#4234)

  • The following functions in qml.math now support parameter broadcasting:

    • purity

    • vn_entropy

    • mutual_info

    • fidelity

    • relative_entropy

    • max_entropy

    • sqrt_matrix

    (#4186)

  • 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 HardwareHamiltonians 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 termqml.pulse.rydberg_drive — and an interaction termqml.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. #3864

  • Ensured 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 QNodes. (#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 calledqml.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)

  • The collections module has been deprecated. (#3686) (#3687)

  • 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. (#3306)

  • 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 from

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

    to

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

    (#3421)

  • 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 with

    • arbitrary 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.

    (#3049)

    >>> 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 errors if 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.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:

    1. The custom device inherits from DefaultQubit, not QubitDevice.

    2. The device implements custom methods in the simulation pipeline that are incompatible with broadcasting (for example expval, apply_operation or analytic_probability).

    3. 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_recipes 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 or

    • multiple 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 Observables, but now extends to all Operators: (#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 Operators 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 as

    symbols = ["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 rounding

    • a show_matrices keyword for controlling display of matrices

    • a different algorithm for determining positions

    • deprecation of the charset keyword

    • additional 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 overview were rewritten to reflect the past and planned changes of the operator refactor. (#2066)

  • Links to the Strawberry Fields documentation for information on the CV model. (#2259)

  • Fixes the documentation example for qml.QFT. (#2232)

  • Fixes the documentation example for using qml.sample with jax.jit. (#2196)

  • The qml.numpy subpackage is now included in the PennyLane API documentation. (#2179)

  • Improves the documentation of RotosolveOptimizer regarding the usage of the passed substep_optimizer and its keyword arguments. (#2160)

  • Ensures that signatures of @qml.qfunc_transform decorated functions display correctly in the docs. (#2286)

  • Docstring examples now display using the updated text-based circuit drawer. (#2252)

  • Add docstring to OrbitalRotation.grad_recipe. (#2193)

Contributors

This release contains contributions from (in alphabetical order):

Catalina Albornoz, Jack Y. Araz, Juan Miguel Arrazola, Ali Asadi, Utkarsh Azad, Sam Banning, Thomas Bromley, Olivia Di Matteo, Christian Gogolin, Diego Guala, Anthony Hayes, David Ittah, Josh Izaac, Soran Jahangiri, Nathan Killoran, Christina Lee, Angus Lowe, Maria Fernanda Morris, Romain Moyard, Zeyue Niu, Lee James O’Riordan, Chae-Yeun Park, Maria Schuld, Jay Soni, Antal Száva, David Wierichs.

orphan

Release 0.21.0

New features since last release

Reduce qubit requirements of simulating Hamiltonians ⚛️

  • Functions for tapering qubits based on molecular symmetries have been added, following results from Setia et al. (#1966) (#1974) (#2041) (#2042)

    With this functionality, a molecular Hamiltonian and the corresponding Hartree-Fock (HF) state can be transformed to a new Hamiltonian and HF state that acts on a reduced number of qubits, respectively.

    # molecular geometry
    symbols = ["He", "H"]
    geometry = np.array([[0.0, 0.0, 0.0], [0.0, 0.0, 1.4588684632]])
    mol = qml.hf.Molecule(symbols, geometry, charge=1)
    
    # generate the qubit Hamiltonian
    H = qml.hf.generate_hamiltonian(mol)(geometry)
    
    # determine Hamiltonian symmetries
    generators, paulix_ops = qml.hf.generate_symmetries(H, len(H.wires))
    opt_sector = qml.hf.optimal_sector(H, generators, mol.n_electrons)
    
    # taper the Hamiltonian
    H_tapered = qml.hf.transform_hamiltonian(H, generators, paulix_ops, opt_sector)
    

    We can compare the number of qubits required by the original Hamiltonian and the tapered Hamiltonian:

    >>> len(H.wires)
    4
    >>> len(H_tapered.wires)
    2
    

    For quantum chemistry algorithms, the Hartree-Fock state can also be tapered:

    n_elec = mol.n_electrons
    n_qubits = mol.n_orbitals * 2
    
    hf_tapered = qml.hf.transform_hf(
        generators, paulix_ops, opt_sector, n_elec, n_qubits
    )
    
    >>> hf_tapered
    tensor([1, 1], requires_grad=True)
    

New tensor network templates 🪢

  • Quantum circuits with the shape of a matrix product state tensor network can now be easily implemented using the new qml.MPS template, based on the work arXiv:1803.11537. (#1871)

    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
    template_weights = np.array([[0.1, -0.3], [0.4, 0.2], [-0.15, 0.5]], requires_grad=True)
    
    dev = qml.device("default.qubit", wires=range(n_wires))
    
    @qml.qnode(dev)
    def circuit(weights):
        qml.MPS(range(n_wires), n_block_wires, block, n_params_block, weights)
        return qml.expval(qml.PauliZ(wires=n_wires - 1))
    

    The resulting circuit is:

    >>> print(qml.draw(circuit, expansion_strategy="device")(template_weights))
    0: ──╭C──RY(0.1)───────────────────────────────┤
    1: ──╰X──RY(-0.3)──╭C──RY(0.4)─────────────────┤
    2: ────────────────╰X──RY(0.2)──╭C──RY(-0.15)──┤
    3: ─────────────────────────────╰X──RY(0.5)────┤ ⟨Z⟩
    
  • Added a template for tree tensor networks, qml.TTN. (#2043)

    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.MPS.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.TTN(range(n_wires), n_block_wires, block, n_params_block, template_weights)
        return qml.expval(qml.PauliZ(wires=n_wires - 1))
    

    The resulting circuit is:

    >>> print(qml.draw(circuit, expansion_strategy="device")(template_weights))
    0: ──╭C──RY(0.1)─────────────────┤
    1: ──╰X──RY(-0.3)──╭C──RY(0.1)───┤
    2: ──╭C──RY(0.1)───│─────────────┤
    3: ──╰X──RY(-0.3)──╰X──RY(-0.3)──┤ ⟨Z⟩
    

Generalized RotosolveOptmizer 📉

  • The RotosolveOptimizer has been generalized to arbitrary frequency spectra in the cost function. Also note the changes in behaviour listed under Breaking changes. (#2081)

    Previously, the RotosolveOptimizer only supported variational circuits using special gates such as single-qubit Pauli rotations. Now, circuits with arbitrary gates are supported natively without decomposition, as long as the frequencies of the gate parameters are known. This new generalization extends the Rotosolve optimization method to a larger class of circuits, and can reduce the cost of the optimization compared to decomposing all gates to single-qubit rotations.

    Consider the QNode

    dev = qml.device("default.qubit", wires=2)
    
    @qml.qnode(dev)
    def qnode(x, Y):
        qml.RX(2.5 * x, wires=0)
        qml.CNOT(wires=[0, 1])
        qml.RZ(0.3 * Y[0], wires=0)
        qml.CRY(1.1 * Y[1], wires=[1, 0])
        return qml.expval(qml.PauliX(0) @ qml.PauliZ(1))
    
    x = np.array(0.8, requires_grad=True)
    Y = np.array([-0.2, 1.5], requires_grad=True)
    

    Its frequency spectra can be easily obtained via qml.fourier.qnode_spectrum:

    >>> spectra = qml.fourier.qnode_spectrum(qnode)(x, Y)
    >>> spectra
    {'x': {(): [-2.5, 0.0, 2.5]},
     'Y': {(0,): [-0.3, 0.0, 0.3], (1,): [-1.1, -0.55, 0.0, 0.55, 1.1]}}
    

    We may then initialize the RotosolveOptimizer and minimize the QNode cost function by providing this information about the frequency spectra. We also compare the cost at each step to the initial cost.

    >>> cost_init = qnode(x, Y)
    >>> opt = qml.RotosolveOptimizer()
    >>> for _ in range(2):
    ...     x, Y = opt.step(qnode, x, Y, spectra=spectra)
    ...     print(f"New cost: {np.round(qnode(x, Y), 3)}; Initial cost: {np.round(cost_init, 3)}")
    New cost: 0.0; Initial cost: 0.706
    New cost: -1.0; Initial cost: 0.706
    

    The optimization with RotosolveOptimizer is performed in substeps. The minimal cost of these substeps can be retrieved by setting full_output=True.

    >>> x = np.array(0.8, requires_grad=True)
    >>> Y = np.array([-0.2, 1.5], requires_grad=True)
    >>> opt = qml.RotosolveOptimizer()
    >>> for _ in range(2):
    ...     (x, Y), history = opt.step(qnode, x, Y, spectra=spectra, full_output=True)
    ...     print(f"New cost: {np.round(qnode(x, Y), 3)} reached via substeps {np.round(history, 3)}")
    New cost: 0.0 reached via substeps [-0.  0.  0.]
    New cost: -1.0 reached via substeps [-1. -1. -1.]
    

    However, note that these intermediate minimal values are evaluations of the reconstructions that Rotosolve creates and uses internally for the optimization, and not of the original objective function. For noisy cost functions, these intermediate evaluations may differ significantly from evaluations of the original cost function.

Improved JAX support 💻

  • The JAX interface now supports evaluating vector-valued QNodes. (#2110)

    Vector-valued QNodes include those with:

    • qml.probs;

    • qml.state;

    • qml.sample or

    • multiple 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)
    
    @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])
    

    The QNode can be evaluated and its jacobian can be computed:

    >>> circuit(x, y)
    Array([0.8397495 , 0.16025047], dtype=float32)
    >>> jax.jacobian(circuit, argnums=[0, 1])(x, y)
    (Array([-0.2050439,  0.2050439], dtype=float32, weak_type=True),
     Array([ 0.26043, -0.26043], dtype=float32, weak_type=True))
    

    Note that jax.jit is not yet supported for vector-valued QNodes.

Speedier quantum natural gradient ⚡

  • A new function for computing the metric tensor on simulators, qml.adjoint_metric_tensor, has been added, that uses classically efficient methods to massively improve performance. (#1992)

    This method, detailed in Jones (2020), computes the metric tensor using four copies of the state vector and a number of operations that scales quadratically in the number of trainable parameters.

    Note that as it makes use of state cloning, it is inherently classical and can only be used with statevector simulators and shots=None.

    It is particularly useful for larger circuits for which backpropagation requires inconvenient or even unfeasible amounts of storage, but is slower. Furthermore, the adjoint method is only available for analytic computation, not for measurements simulation with shots!=None.

    dev = qml.device("default.qubit", wires=3)
    
    @qml.qnode(dev)
    def circuit(x, y):
        qml.Rot(*x[0], wires=0)
        qml.Rot(*x[1], wires=1)
        qml.Rot(*x[2], wires=2)
        qml.CNOT(wires=[0, 1])
        qml.CNOT(wires=[1, 2])
        qml.CNOT(wires=[2, 0])
        qml.RY(y[0], wires=0)
        qml.RY(y[1], wires=1)
        qml.RY(y[0], wires=2)
        return qml.expval(qml.PauliZ(0) @ qml.PauliZ(1)), qml.expval(qml.PauliY(1))
    
    x = np.array([[0.2, 0.4, -0.1], [-2.1, 0.5, -0.2], [0.1, 0.7, -0.6]], requires_grad=False)
    y = np.array([1.3, 0.2], requires_grad=True)
    
    >>> qml.adjoint_metric_tensor(circuit)(x, y)
    tensor([[ 0.25495723, -0.07086695],
            [-0.07086695,  0.24945606]], requires_grad=True)
    

    Computational cost

    The adjoint method uses \(2P^2+4P+1\) gates and state cloning operations if the circuit is composed only of trainable gates, where \(P\) is the number of trainable operations. If non-trainable gates are included, each of them is applied about \(n^2-n\) times, where \(n\) is the number of trainable operations that follow after the respective non-trainable operation in the circuit. This means that non-trainable gates later in the circuit are executed less often, making the adjoint method a bit cheaper if such gates appear later. The adjoint method requires memory for 4 independent state vectors, which corresponds roughly to storing a state vector of a system with 2 additional qubits.

Compute the Hessian on hardware ⬆️

  • A new gradient transform qml.gradients.param_shift_hessian has been added to directly compute the Hessian (2nd order partial derivative matrix) of QNodes on hardware. (#1884)

    The function generates parameter-shifted tapes which allow the Hessian to be computed analytically on hardware and software devices. Compared to using an auto-differentiation framework to compute the Hessian via parameter shifts, this function will use fewer device invocations and can be used to inspect the parameter-shifted “Hessian tapes” directly. The function remains fully differentiable on all supported PennyLane interfaces.

    Additionally, the parameter-shift Hessian comes with a new batch transform decorator @qml.gradients.hessian_transform, which can be used to create custom Hessian functions.

    The following code demonstrates how to use the parameter-shift Hessian:

    dev = qml.device("default.qubit", wires=2)
    
    @qml.qnode(dev)
    def circuit(x):
        qml.RX(x[0], wires=0)
        qml.RY(x[1], wires=0)
        return qml.expval(qml.PauliZ(0))
    
    x = np.array([0.1, 0.2], requires_grad=True)
    
    hessian = qml.gradients.param_shift_hessian(circuit)(x)
    
    >>> hessian
    tensor([[-0.97517033,  0.01983384],
            [ 0.01983384, -0.97517033]], requires_grad=True)
    

Improvements

  • The qml.transforms.insert transform now supports adding operation after or before certain specific gates. (#1980)

  • Added a modified version of the simplify function to the hf module. (#2103)

    This function combines redundant terms in a Hamiltonian and eliminates terms with a coefficient smaller than a cutoff value. The new function makes construction of molecular Hamiltonians more efficient. For LiH, as an example, the time to construct the Hamiltonian is reduced roughly by a factor of 20.

  • The QAOA module now accepts both NetworkX and RetworkX graphs as function inputs. (#1791)

  • The CircuitGraph, used to represent circuits via directed acyclic graphs, now uses RetworkX for its internal representation. This results in significant speedup for algorithms that rely on a directed acyclic graph representation. (#1791)

  • For subclasses of Operator where the number of parameters is known before instantiation, the num_params is reverted back to being a static property. This allows to programmatically know the number of parameters before an operator is instantiated without changing the user interface. A test was added to ensure that different ways of defining num_params work as expected. (#2101) (#2135)

  • A WireCut operator has been added for manual wire cut placement when constructing a QNode. (#2093)

  • The new function qml.drawer.tape_text produces a string drawing of a tape. This function differs in implementation and minor stylistic details from the old string circuit drawing infrastructure. (#1885)

  • The RotosolveOptimizer now raises an error if no trainable arguments are detected, instead of silently skipping update steps for all arguments. (#2109)

  • The function qml.math.safe_squeeze is introduced and gradient_transform allows for QNode argument axes of size 1. (#2080)

    qml.math.safe_squeeze wraps qml.math.squeeze, with slight modifications:

    • When provided the axis keyword argument, axes that do not have size 1 will be ignored, instead of raising an error.

    • The keyword argument exclude_axis allows to explicitly exclude axes from the squeezing.

  • The adjoint transform now raises and error whenever the object it is applied to is not callable. (#2060)

    An example is a list of operations to which one might apply qml.adjoint:

    dev = qml.device("default.qubit", wires=2)
    @qml.qnode(dev)
    def circuit_wrong(params):
        # Note the difference:                  v                         v
        qml.adjoint(qml.templates.AngleEmbedding(params, wires=dev.wires))
        return qml.state()
    
    @qml.qnode(dev)
    def circuit_correct(params):
        # Note the difference:                  v                         v
        qml.adjoint(qml.templates.AngleEmbedding)(params, wires=dev.wires)
        return qml.state()
    
    params = list(range(1, 3))
    

    Evaluating circuit_wrong(params) now raises a ValueError and if we apply qml.adjoint correctly, we get

    >>> circuit_correct(params)
    [ 0.47415988+0.j          0.         0.73846026j  0.         0.25903472j
     -0.40342268+0.j        ]
    
  • A precision argument has been added to the tape’s to_openqasm function to control the precision of parameters. (#2071)

  • Interferometer now has a shape method. (#1946)

  • The Barrier and Identity operations now support the adjoint method. (#2062) (#2063)

  • qml.BasisStatePreparation now supports the batch_params decorator. (#2091)

  • Added a new multi_dispatch decorator that helps ease the definition of new functions inside PennyLane. The decorator is used throughout the math module, demonstrating use cases. (#2082) (#2096)

    We can decorate a function, indicating the arguments that are tensors handled by the interface:

    >>> @qml.math.multi_dispatch(argnum=[0, 1])
    ... def some_function(tensor1, tensor2, option, like):
    ...     # the interface string is stored in ``like``.
    ...     ...
    

    Previously, this was done using the private utility function _multi_dispatch.

    >>> def some_function(tensor1, tensor2, option):
    ...     interface = qml.math._multi_dispatch([tensor1, tensor2])
    ...     ...
    
  • The IsingZZ gate was added to the diagonal_in_z_basis attribute. For this an explicit _eigvals method was added. (#2113)

  • The IsingXX, IsingYY and IsingZZ gates were added to the composable_rotations attribute. (#2113)

Breaking changes

  • QNode arguments will no longer be considered trainable by default when using the Autograd interface. In order to obtain derivatives with respect to a parameter, it should be instantiated via PennyLane’s NumPy wrapper using the requires_grad=True attribute. The previous behaviour was deprecated in version v0.19.0 of PennyLane. (#2116) (#2125) (#2139) (#2148) (#2156)

    from pennylane import numpy as np
    
    @qml.qnode(qml.device("default.qubit", wires=2))
    def circuit(x):
      ...
    
    x = np.array([0.1, 0.2], requires_grad=True)
    qml.grad(circuit)(x)
    

    For the qml.grad and qml.jacobian functions, trainability can alternatively be indicated via the argnum keyword:

    import numpy as np
    
    @qml.qnode(qml.device("default.qubit", wires=2))
    def circuit(hyperparam, param):
      ...
    
    x = np.array([0.1, 0.2])
    qml.grad(circuit, argnum=1)(0.5, x)
    
  • qml.jacobian now follows a different convention regarding its output shape. (#2059)

    Previously, qml.jacobian would attempt to stack the Jacobian for multiple QNode arguments, which succeeded whenever the arguments have the same shape. In this case, the stacked Jacobian would also be transposed, leading to the output shape (*reverse_QNode_args_shape, *reverse_output_shape, num_QNode_args)

    If no stacking and transposing occurs, the output shape instead is a tuple where each entry corresponds to one QNode argument and has the shape (*output_shape, *QNode_arg_shape).

    This breaking change alters the behaviour in the first case and removes the attempt to stack and transpose, so that the output always has the shape of the second type.

    Note that the behaviour is unchanged — that is, the Jacobian tuple is unpacked into a single Jacobian — if argnum=None and there is only one QNode argument with respect to which the differentiation takes place, or if an integer is provided as argnum.

    A workaround that allowed qml.jacobian to differentiate multiple QNode arguments will no longer support higher-order derivatives. In such cases, combining multiple arguments into a single array is recommended.

  • qml.metric_tensor, qml.adjoint_metric_tensor and qml.transforms.classical_jacobian now follow a different convention regarding their output shape when being used with the Autograd interface (#2059)

    See the previous entry for details. This breaking change immediately follows from the change in qml.jacobian whenever hybrid=True is used in the above methods.

  • The behaviour of RotosolveOptimizer has been changed regarding its keyword arguments. (#2081)

    The keyword arguments optimizer and optimizer_kwargs for the RotosolveOptimizer have been renamed to substep_optimizer and substep_kwargs, respectively. Furthermore they have been moved from step and step_and_cost to the initialization __init__.

    The keyword argument num_freqs has been renamed to nums_frequency and is expected to take a different shape now: Previously, it was expected to be an int or a list of entries, with each entry in turn being either an int or a list of int entries. Now the expected structure is a nested dictionary, matching the formatting expected by qml.fourier.reconstruct This also matches the expected formatting of the new keyword arguments spectra and shifts.

    For more details, see the RotosolveOptimizer documentation.

Deprecations

  • Deprecates the caching ability provided by QubitDevice. (#2154)

    Going forward, the preferred way is to use the caching abilities of the QNode:

    dev = qml.device("default.qubit", wires=2)
    
    cache = {}
    
    @qml.qnode(dev, diff_method='parameter-shift', cache=cache)
    def circuit():
        qml.RY(0.345, wires=0)
        return qml.expval(qml.PauliZ(0))
    
    >>> for _ in range(10):
    ...    circuit()
    >>> dev.num_executions
    1
    

Bug fixes

  • Fixes a bug where an incorrect number of executions are recorded by a QNode using a custom cache with diff_method="backprop". (#2171)

  • Fixes a bug where the default.qubit.jax device can’t be used with diff_method=None and jitting. (#2136)

  • Fixes a bug where the Torch interface was not properly unwrapping Torch tensors to NumPy arrays before executing gradient tapes on devices. (#2117)

  • Fixes a bug for the TensorFlow interface where the dtype of input tensors was not cast. (#2120)

  • Fixes a bug where batch transformed QNodes would fail to apply batch transforms provided by the underlying device. (#2111)

  • An error is now raised during QNode creation if backpropagation is requested on a device with finite shots specified. (#2114)

  • Pytest now ignores any DeprecationWarning raised within autograd’s numpy_wrapper module. Other assorted minor test warnings are fixed. (#2007)

  • Fixes a bug where the QNode was not correctly diagonalizing qubit-wise commuting observables. (#2097)

  • Fixes a bug in gradient_transform where the hybrid differentiation of circuits with a single parametrized gate failed and QNode argument axes of size 1 where removed from the output gradient. (#2080)

  • The available diff_method options for QNodes has been corrected in both the error messages and the documentation. (#2078)

  • Fixes a bug in DefaultQubit where the second derivative of QNodes at positions corresponding to vanishing state vector amplitudes is wrong. (#2057)

  • Fixes a bug where PennyLane didn’t require v0.20.0 of PennyLane-Lightning, but raised an error with versions of Lightning earlier than v0.20.0 due to the new batch execution pipeline. (#2033)

  • Fixes a bug in classical_jacobian when used with Torch, where the Jacobian of the preprocessing was also computed for non-trainable parameters. (#2020)

  • Fixes a bug in queueing of the two_qubit_decomposition method that originally led to circuits with >3 two-qubit unitaries failing when passed through the unitary_to_rot optimization transform. (#2015)

  • Fixes a bug which allows using jax.jit to be compatible with circuits which return qml.probs when the default.qubit.jax is provided with a custom shot vector. (#2028)

  • Updated the adjoint() method for non-parametric qubit operations to solve a bug where repeated adjoint() calls don’t return the correct operator. (#2133)

  • Fixed a bug in insert() which prevented operations that inherited from multiple classes to be inserted. (#2172)

Documentation

  • Fixes an error in the signs of equations in the DoubleExcitation page. (#2072)

  • Extends the interfaces description page to explicitly mention device compatibility. (#2031)

Contributors

This release contains contributions from (in alphabetical order):

Juan Miguel Arrazola, Ali Asadi, Utkarsh Azad, Sam Banning, Thomas Bromley, Esther Cruz, Olivia Di Matteo, Christian Gogolin, Diego Guala, Anthony Hayes, David Ittah, Josh Izaac, Soran Jahangiri, Edward Jiang, Ankit Khandelwal, Nathan Killoran, Korbinian Kottmann, Christina Lee, Romain Moyard, Lee James O’Riordan, Maria Schuld, Jay Soni, Antal Száva, David Wierichs, Shaoming Zhang.

orphan

Release 0.20.0

New features since last release

Shiny new circuit drawer!🎨🖌️

  • PennyLane now supports drawing a QNode with matplotlib! (#1803) (#1811) (#1931) (#1954)

    dev = qml.device("default.qubit", wires=4)
    
    @qml.qnode(dev)
    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))
    fig, ax = qml.draw_mpl(circuit)(1.2345, 1.2345)
    fig.show()
    

New and improved quantum-aware optimizers

  • Added qml.LieAlgebraOptimizer, a new quantum-aware Lie Algebra optimizer that allows one to perform gradient descent on the special unitary group. (#1911)

    dev = qml.device("default.qubit", wires=2)
    H = -1.0 * qml.PauliX(0) - qml.PauliZ(1) - qml.PauliY(0) @ qml.PauliX(1)
    
    @qml.qnode(dev)
    def circuit():
        qml.RX(0.1, wires=[0])
        qml.RY(0.5, wires=[1])
        qml.CNOT(wires=[0,1])
        qml.RY(0.6, wires=[0])
        return qml.expval(H)
    opt = qml.LieAlgebraOptimizer(circuit=circuit, stepsize=0.1)
    

    Note that, unlike other optimizers, the LieAlgebraOptimizer accepts a QNode with no parameters, and instead grows the circuit by appending operations during the optimization:

    >>> circuit()
    tensor(-1.3351865, requires_grad=True)
    >>> circuit1, cost = opt.step_and_cost()
    >>> circuit1()
    tensor(-1.99378872, requires_grad=True)
    

    For more details, see the LieAlgebraOptimizer documentation.

  • The qml.metric_tensor transform can now be used to compute the full tensor, beyond the block diagonal approximation. (#1725)

    This is performed using Hadamard tests, and requires an additional wire on the device to execute the circuits produced by the transform, as compared to the number of wires required by the original circuit. The transform defaults to computing the full tensor, which can be controlled by the approx keyword argument.

    As an example, consider the QNode

    dev = qml.device("default.qubit", wires=3)
    
    @qml.qnode(dev)
    def circuit(weights):
        qml.RX(weights[0], wires=0)
        qml.RY(weights[1], wires=0)
        qml.CNOT(wires=[0, 1])
        qml.RZ(weights[2], wires=1)
        return qml.expval(qml.PauliZ(0) @ qml.PauliZ(1))
    
    weights = np.array([0.2, 1.2, -0.9], requires_grad=True)
    

    Then we can compute the (block) diagonal metric tensor as before, now using the approx="block-diag" keyword:

    >>> qml.metric_tensor(circuit, approx="block-diag")(weights)
    [[0.25       0.         0.        ]
     [0.         0.24013262 0.        ]
     [0.         0.         0.21846983]]
    

    Instead, we now can also compute the full metric tensor, using Hadamard tests on the additional wire of the device:

    >>> qml.metric_tensor(circuit)(weights)
    [[ 0.25        0.         -0.23300977]
     [ 0.          0.24013262  0.01763859]
     [-0.23300977  0.01763859  0.21846983]]
    

    See the metric tensor documentation. for more information and usage details.

Faster performance with optimized quantum workflows

  • The QNode has been re-written to support batch execution across the board, custom gradients, better decomposition strategies, and higher-order derivatives. (#1807) (#1969)

    • Internally, if multiple circuits are generated for simultaneous execution, they will be packaged into a single job for execution on the device. This can lead to significant performance improvement when executing the QNode on remote quantum hardware or simulator devices with parallelization capabilities.

    • Custom gradient transforms can be specified as the differentiation method:

      @qml.gradients.gradient_transform
      def my_gradient_transform(tape):
          ...
          return tapes, processing_fn
      
      @qml.qnode(dev, diff_method=my_gradient_transform)
      def circuit():
      

    For breaking changes related to the use of the new QNode, refer to the Breaking Changes section.

    Note that the old QNode remains accessible at @qml.qnode_old.qnode, however this will be removed in the next release.

  • Custom decompositions can now be applied to operations at the device level. (#1900)

    For example, suppose we would like to implement the following QNode:

    def circuit(weights):
        qml.BasicEntanglerLayers(weights, wires=[0, 1, 2])
        return qml.expval(qml.PauliZ(0))
    
    original_dev = qml.device("default.qubit", wires=3)
    original_qnode = qml.QNode(circuit, original_dev)
    
    >>> weights = np.array([[0.4, 0.5, 0.6]])
    >>> print(qml.draw(original_qnode, expansion_strategy="device")(weights))
     0: ──RX(0.4)──╭C──────╭X──┤ ⟨Z⟩
     1: ──RX(0.5)──╰X──╭C──│───┤
     2: ──RX(0.6)──────╰X──╰C──┤
    

    Now, let’s swap out the decomposition of the CNOT gate into CZ and Hadamard, and furthermore the decomposition of Hadamard into RZ and RY rather than the decomposition already available in PennyLane. We define the two decompositions like so, and pass them to a device:

    def custom_cnot(wires):
        return [
            qml.Hadamard(wires=wires[1]),
            qml.CZ(wires=[wires[0], wires[1]]),
            qml.Hadamard(wires=wires[1])
        ]
    
    def custom_hadamard(wires):
        return [
            qml.RZ(np.pi, wires=wires),
            qml.RY(np.pi / 2, wires=wires)
        ]
    
    # Can pass the operation itself, or a string
    custom_decomps = {qml.CNOT : custom_cnot, "Hadamard" : custom_hadamard}
    
    decomp_dev = qml.device("default.qubit", wires=3, custom_decomps=custom_decomps)
    decomp_qnode = qml.QNode(circuit, decomp_dev)
    

    Now when we draw or run a QNode on this device, the gates will be expanded according to our specifications:

    >>> print(qml.draw(decomp_qnode, expansion_strategy="device")(weights))
     0: ──RX(0.4)──────────────────────╭C──RZ(3.14)──RY(1.57)──────────────────────────╭Z──RZ(3.14)──RY(1.57)──┤ ⟨Z⟩
     1: ──RX(0.5)──RZ(3.14)──RY(1.57)──╰Z──RZ(3.14)──RY(1.57)──╭C──────────────────────│───────────────────────┤
     2: ──RX(0.6)──RZ(3.14)──RY(1.57)──────────────────────────╰Z──RZ(3.14)──RY(1.57)──╰C──────────────────────┤
    

    A separate context manager, set_decomposition, has also been implemented to enable application of custom decompositions on devices that have already been created.

    >>> with qml.transforms.set_decomposition(custom_decomps, original_dev):
    ...     print(qml.draw(original_qnode, expansion_strategy="device")(weights))
     0: ──RX(0.4)──────────────────────╭C──RZ(3.14)──RY(1.57)──────────────────────────╭Z──RZ(3.14)──RY(1.57)──┤ ⟨Z⟩
     1: ──RX(0.5)──RZ(3.14)──RY(1.57)──╰Z──RZ(3.14)──RY(1.57)──╭C──────────────────────│───────────────────────┤
     2: ──RX(0.6)──RZ(3.14)──RY(1.57)──────────────────────────╰Z──RZ(3.14)──RY(1.57)──╰C──────────────────────┤
    
  • Given an operator of the form \(U=e^{iHt}\), where \(H\) has commuting terms and known eigenvalues, qml.gradients.generate_shift_rule computes the generalized parameter shift rules for determining the gradient of the expectation value \(f(t) = \langle 0|U(t)^\dagger \hat{O} U(t)|0\rangle\) on hardware. (#1788) (#1932)

    Given

    \[H = \sum_i a_i h_i,\]

    where the eigenvalues of \(H\) are known and all \(h_i\) commute, we can compute the frequencies (the unique positive differences of any two eigenvalues) using qml.gradients.eigvals_to_frequencies.

    qml.gradients.generate_shift_rule can then be used to compute the parameter shift rules to compute \(f'(t)\) using 2R shifted cost function evaluations. This becomes cheaper than the standard application of the chain rule and two-term shift rule when R is less than the number of Pauli words in the generator.

    For example, consider the case where \(H\) has eigenspectrum (-1, 0, 1):

    >>> frequencies = qml.gradients.eigvals_to_frequencies((-1, 0, 1))
    >>> frequencies
    (1, 2)
    >>> coeffs, shifts = qml.gradients.generate_shift_rule(frequencies)
    >>> coeffs
    array([ 0.85355339, -0.85355339, -0.14644661,  0.14644661])
    >>> shifts
    array([ 0.78539816, -0.78539816,  2.35619449, -2.35619449])
    

    As we can see, generate_shift_rule returns four coefficients \(c_i\) and shifts \(s_i\) corresponding to a four term parameter shift rule. The gradient can then be reconstructed via:

    \[\frac{\partial}{\partial\phi}f = \sum_{i} c_i f(\phi + s_i),\]

    where \(f(\phi) = \langle 0|U(\phi)^\dagger \hat{O} U(\phi)|0\rangle\) for some observable \(\hat{O}\) and the unitary \(U(\phi)=e^{iH\phi}\).

Support for TensorFlow AutoGraph mode with quantum hardware

  • It is now possible to use TensorFlow’s AutoGraph mode with QNodes on all devices and with arbitrary differentiation methods. Previously, AutoGraph mode only support diff_method="backprop". This will result in significantly more performant model execution, at the cost of a more expensive initial compilation. (#1866)

    Use AutoGraph to convert your QNodes or cost functions into TensorFlow graphs by decorating them with @tf.function:

    dev = qml.device("lightning.qubit", wires=2)
    
    @qml.qnode(dev, diff_method="adjoint", interface="tf", max_diff=1)
    def circuit(x):
        qml.RX(x[0], wires=0)
        qml.RY(x[1], wires=1)
        return qml.expval(qml.PauliZ(0) @ qml.PauliZ(1)), qml.expval(qml.PauliZ(0))
    
    @tf.function
    def cost(x):
        return tf.reduce_sum(circuit(x))
    
    x = tf.Variable([0.5, 0.7], dtype=tf.float64)
    
    with tf.GradientTape() as tape:
        loss = cost(x)
    
    grad = tape.gradient(loss, x)
    

    The initial execution may take slightly longer than when executing the circuit in eager mode; this is because TensorFlow is tracing the function to create the graph. Subsequent executions will be much more performant.

    Note that using AutoGraph with backprop-enabled devices, such as default.qubit, will yield the best performance.

    For more details, please see the TensorFlow AutoGraph documentation.

Characterize your quantum models with classical QNode reconstruction

  • The qml.fourier.reconstruct function is added. It can be used to reconstruct QNodes outputting expectation values along a specified parameter dimension, with a minimal number of calls to the original QNode. The returned reconstruction is exact and purely classical, and can be evaluated without any quantum executions. (#1864)

    The reconstruction technique differs for functions with equidistant frequencies that are reconstructed using the function value at equidistant sampling points, and for functions with arbitrary frequencies reconstructed using arbitrary sampling points.

    As an example, consider the following QNode:

    dev = qml.device("default.qubit", wires=2)
    
    @qml.qnode(dev)
    def circuit(x, Y, f=1.0):
        qml.RX(f * x, wires=0)
        qml.RY(Y[0], wires=0)
        qml.RY(Y[1], wires=1)
        qml.CNOT(wires=[0, 1])
        qml.RY(3 * Y[1], wires=1)
        return qml.expval(qml.PauliZ(0) @ qml.PauliZ(1))
    

    It has three variational parameters overall: A scalar input x and an array-valued input Y with two entries. Additionally, we can tune the dependence on x with the frequency f. We then can reconstruct the QNode output function with respect to x via

    >>> x = 0.3
    >>> Y = np.array([0.1, -0.9])
    >>> rec = qml.fourier.reconstruct(circuit, ids="x", nums_frequency={"x": {0: 1}})(x, Y)
    >>> rec
    {'x': {0: <function pennylane.fourier.reconstruct._reconstruct_equ.<locals>._reconstruction(x)>}}
    

    As we can see, we get a nested dictionary in the format of the input nums_frequency with functions as values. These functions are simple float-to-float callables:

    >>> univariate = rec["x"][0]
    >>> univariate(x)
    -0.880208251507
    

    For more details on usage, reconstruction cost and differentiability support, please see the fourier.reconstruct docstring.

State-of-the-art operations and templates

  • A circuit template for time evolution under a commuting Hamiltonian utilizing generalized parameter shift rules for cost function gradients is available as qml.CommutingEvolution. (#1788)

    If the template is handed a frequency spectrum during its instantiation, then generate_shift_rule is internally called to obtain the general parameter shift rules with respect to CommutingEvolution‘s \(t\) parameter, otherwise the shift rule for a decomposition of CommutingEvolution will be used.

    The template can be initialized within QNode as:

    import pennylane as qml
    
    n_wires = 2
    dev = qml.device('default.qubit', wires=n_wires)
    
    coeffs = [1, -1]
    obs = [qml.PauliX(0) @ qml.PauliY(1), qml.PauliY(0) @ qml.PauliX(1)]
    hamiltonian = qml.Hamiltonian(coeffs, obs)
    frequencies = (2,4)
    
    @qml.qnode(dev)
    def circuit(time):
        qml.PauliX(0)
        qml.CommutingEvolution(hamiltonian, time, frequencies)
        return qml.expval(qml.PauliZ(0))
    

    Note that there is no internal validation that 1) the input qml.Hamiltonian is fully commuting and 2) the eigenvalue frequency spectrum is correct, since these checks become prohibitively expensive for large Hamiltonians.

  • The qml.Barrier() operator has been added. With it we can separate blocks in compilation or use it as a visual tool. (#1844)

  • Added the identity observable to be an operator. Now we can explicitly call the identity operation on our quantum circuits for both qubit and CV devices. (#1829)

  • Added the qml.QubitDensityMatrix initialization gate for mixed state simulation. (#1850)

  • A thermal relaxation channel is added to the Noisy channels. The channel description can be found on the supplementary information of Quantum classifier with tailored quantum kernels. (#1766)

  • Added a new qml.PauliError channel that allows the application of an arbitrary number of Pauli operators on an arbitrary number of wires. (#1781)

Manipulate QNodes to your ❤️s content with new transforms

  • The merge_amplitude_embedding transformation has been created to automatically merge all gates of this type into one. (#1933)

    from pennylane.transforms import merge_amplitude_embedding
    
    dev = qml.device("default.qubit", wires = 3)
    
    @qml.qnode(dev)
    @merge_amplitude_embedding
    def qfunc():
        qml.AmplitudeEmbedding([0,1,0,0], wires = [0,1])
        qml.AmplitudeEmbedding([0,1], wires = 2)
        return qml.expval(qml.PauliZ(wires = 0))
    
    >>> print(qml.draw(qnode)())
     0: ──╭AmplitudeEmbedding(M0)──┤ ⟨Z⟩
     1: ──├AmplitudeEmbedding(M0)──┤
     2: ──╰AmplitudeEmbedding(M0)──┤
     M0 =
     [0.+0.j 0.+0.j 0.+0.j 1.+0.j 0.+0.j 0.+0.j 0.+0.j 0.+0.j]
    
  • The undo_swaps transformation has been created to automatically remove all swaps of a circuit. (#1960)

    dev = qml.device('default.qubit', wires=3)
    
    @qml.qnode(dev)
    @qml.transforms.undo_swaps
    def qfunc():
        qml.Hadamard(wires=0)
        qml.PauliX(wires=1)
        qml.SWAP(wires=[0,1])
        qml.SWAP(wires=[0,2])
        qml.PauliY(wires=0)
        return qml.expval(qml.PauliZ(0))
    
    >>> print(qml.draw(qfunc)())
     0: ──Y──┤ ⟨Z⟩
     1: ──H──┤
     2: ──X──┤
    

Improvements

  • Added functions for computing the values of atomic and molecular orbitals at a given position. (#1867)

    The functions atomic_orbital and molecular_orbital can be used, as shown in the following codeblock, to evaluate the orbitals. By generating values of the orbitals at different positions, one can plot the spatial shape of a desired orbital.

    symbols  = ['H', 'H']
    geometry = np.array([[0.0, 0.0, 0.0], [0.0, 0.0, 1.0]], requires_grad = False)
    mol = hf.Molecule(symbols, geometry)
    hf.generate_scf(mol)()
    
    ao = mol.atomic_orbital(0)
    mo = mol.molecular_orbital(1)
    
    >>> print(ao(0.0, 0.0, 0.0))
    >>> print(mo(0.0, 0.0, 0.0))
    0.6282468778183719
    0.018251285973461928
    
  • Added support for Python 3.10. (#1964)

  • The execution of QNodes that have

    • multiple return types;

    • a return type other than Variance and Expectation

    now raises a descriptive error message when using the JAX interface. (#2011)

  • The PennyLane qchem package is now lazily imported; it will only be imported the first time it is accessed. (#1962)

  • qml.math.scatter_element_add now supports adding multiple values at multiple indices with a single function call, in all interfaces (#1864)

    For example, we may set five values of a three-dimensional tensor in the following way:

    >>> X = tf.zeros((3, 2, 9), dtype=tf.float64)
    >>> indices = [(0, 0, 1, 2, 2), (0, 0, 0, 0, 1), (1, 3, 8, 6, 7)]
    >>> values = [1 * i for i in range(1,6)]
    >>> qml.math.scatter_element_add(X, indices, values)
    <tf.Tensor: shape=(3, 2, 9), dtype=float64, numpy=
    array([[[0., 1., 0., 2., 0., 0., 0., 0., 0.],
            [0., 0., 0., 0., 0., 0., 0., 0., 0.]],
    
           [[0., 0., 0., 0., 0., 0., 0., 0., 3.],
            [0., 0., 0., 0., 0., 0., 0., 0., 0.]],
    
           [[0., 0., 0., 0., 0., 0., 4., 0., 0.],
            [0., 0., 0., 0., 0., 0., 0., 5., 0.]]])>
    
  • All instances of str.format have been replace with f-strings. (#1970)

  • Tests do not loop over automatically imported and instantiated operations any more, which was opaque and created unnecessarily many tests. (#1895)

  • A decompose() method has been added to the Operator class such that we can obtain (and queue) decompositions directly from instances of operations. (#1873)

    >>> op = qml.PhaseShift(0.3, wires=0)
    >>> op.decompose()
    [RZ(0.3, wires=[0])]
    
  • qml.circuit_drawer.tape_mpl produces a matplotlib figure and axes given a tape. (#1787)

  • The AngleEmbedding, BasicEntanglerLayers and MottonenStatePreparation templates now support parameters with batch dimension when using the @qml.batch_params decorator. (#1812) (#1883) (#1893)

  • qml.draw now supports a max_length argument to help prevent text overflows when printing circuits. (#1892)

  • Identity operation is now part of both the ops.qubit and ops.cv modules. (#1956)

Breaking changes

  • The QNode has been re-written to support batch execution across the board, custom gradients, better decomposition strategies, and higher-order derivatives. (#1807) (#1969)

    • Arbitrary \(n\)-th order derivatives are supported on hardware using gradient transforms such as the parameter-shift rule. To specify that an \(n\)-th order derivative of a QNode will be computed, the max_diff argument should be set. By default, this is set to 1 (first-order derivatives only). Increasing this value allows for higher order derivatives to be extracted, at the cost of additional (classical) computational overhead during the backwards pass.

    • When decomposing the circuit, the default decomposition strategy expansion_strategy="gradient" will prioritize decompositions that result in the smallest number of parametrized operations required to satisfy the differentiation method. While this may lead to a slight increase in classical processing, it significantly reduces the number of circuit evaluations needed to compute gradients of complicated unitaries.

      To return to the old behaviour, expansion_strategy="device" can be specified.

    Note that the old QNode remains accessible at @qml.qnode_old.qnode, however this will be removed in the next release.

  • Certain features deprecated in v0.19.0 have been removed: (#1981) (#1963)

    • The qml.template decorator (use a ` QuantumTape <https://pennylane.readthedocs.io/en/stable/code/api/pennylane.tape.QuantumTape.html>`_ as a context manager to record operations and its operations attribute to return them, see the linked page for examples);

    • The default.tensor and default.tensor.tf experimental devices;

    • The qml.fourier.spectrum function (use the qml.fourier.circuit_spectrum or qml.fourier.qnode_spectrum functions instead);

    • The diag_approx keyword argument of qml.metric_tensor and qml.QNGOptimizer (pass approx='diag' instead).

  • The default behaviour of the qml.metric_tensor transform has been modified. By default, the full metric tensor is computed, leading to higher cost than the previous default of computing the block diagonal only. At the same time, the Hadamard tests for the full metric tensor require an additional wire on the device, so that

    >>> qml.metric_tensor(some_qnode)(weights)
    

    will revert back to the block diagonal restriction and raise a warning if the used device does not have an additional wire. (#1725)

  • The circuit_drawer module has been renamed drawer. (#1949)

  • The par_domain attribute in the operator class has been removed. (#1907)

  • The mutable keyword argument has been removed from the QNode, due to underlying bugs that result in incorrect results being returned from immutable QNodes. This functionality will return in an upcoming release. (#1807)

  • The reversible QNode differentiation method has been removed; the adjoint differentiation method is preferred instead (diff_method='adjoint'). (#1807)

  • QuantumTape.trainable_params now is a list instead of a set. This means that tape.trainable_params will return a list unlike before, but setting the trainable_params with a set works exactly as before. (#1904)

  • The num_params attribute in the operator class is now dynamic. This makes it easier to define operator subclasses with a flexible number of parameters. (#1898) (#1909)

  • The static method decomposition(), formerly in the Operation class, has been moved to the base Operator class. (#1873)

  • DiagonalOperation is not a separate subclass any more. (#1889)

    Instead, devices can check for the diagonal property using attributes:

    from pennylane.ops.qubit.attributes import diagonal_in_z_basis
    
    if op in diagonal_in_z_basis:
        # do something
    

    Custom operations can be added to this attribute at runtime via diagonal_in_z_basis.add("MyCustomOp").

Bug fixes

  • Fixes a bug with qml.probs when using default.qubit.jax. (#1998)

  • Fixes a bug where output tensors of a QNode would always be put on the default GPU with default.qubit.torch. (#1982)

  • Device test suite doesn’t use empty circuits so that it can also test the IonQ plugin, and it checks if operations are supported in more places. (#1979)

  • Fixes a bug where the metric tensor was computed incorrectly when using gates with gate.inverse=True. (#1987)

  • Corrects the documentation of qml.transforms.classical_jacobian for the Autograd interface (and improves test coverage). (#1978)

  • Fixes a bug where differentiating a QNode with qml.state using the JAX interface raised an error. (#1906)

  • Fixes a bug with the adjoint of qml.QFT. (#1955)

  • Fixes a bug where the ApproxTimeEvolution template was not correctly computing the operation wires from the input Hamiltonian. This did not affect computation with the ApproxTimeEvolution template, but did cause circuit drawing to fail. (#1952)

  • Fixes a bug where the classical preprocessing Jacobian computed by qml.transforms.classical_jacobian with JAX returned a reduced submatrix of the Jacobian. (#1948)

  • Fixes a bug where the operations are not accessed in the correct order in qml.fourier.qnode_spectrum, leading to wrong outputs. (#1935)

  • Fixes several Pylint errors. (#1951)

  • Fixes a bug where the device test suite wasn’t testing certain operations. (#1943)

  • Fixes a bug where batch transforms would mutate a QNodes execution options. (#1934)

  • qml.draw now supports arbitrary templates with matrix parameters. (#1917)

  • QuantumTape.trainable_params now is a list instead of a set, making it more stable in very rare edge cases. (#1904)

  • ExpvalCost now returns corrects results shape when optimize=True with shots batch. (#1897)

  • qml.circuit_drawer.MPLDrawer was slightly modified to work with matplotlib version 3.5. (#1899)

  • qml.CSWAP and qml.CRot now define control_wires, and qml.SWAP returns the default empty wires object. (#1830)

  • The requires_grad attribute of qml.numpy.tensor objects is now preserved when pickling/unpickling the object. (#1856)

  • Device tests no longer throw warnings about the requires_grad attribute of variational parameters. (#1913)

  • AdamOptimizer and AdagradOptimizer had small fixes to their optimization step updates. (#1929)

  • Fixes a bug where differentiating a QNode with multiple array arguments via qml.gradients.param_shift throws an error. (#1989)

  • AmplitudeEmbedding template no longer produces a ComplexWarning when the features parameter is batched and provided as a 2D array. (#1990)

  • qml.circuit_drawer.CircuitDrawer no longer produces an error when attempting to draw tapes inside of circuits (e.g. from decomposition of an operation or manual placement). (#1994)

  • Fixes a bug where using SciPy sparse matrices with the new QNode could lead to a warning being raised about prioritizing the TensorFlow and PyTorch interfaces. (#2001)

  • Fixed a bug where the QueueContext was not empty when first importing PennyLane. (#1957)

  • Fixed circuit drawing problem with Interferometer and CVNeuralNet. (#1953)

Documentation

  • Added examples in documentation for some operations. (#1902)

  • Improves the Developer’s Guide Testing document. (#1896)

  • Added documentation examples for AngleEmbedding, BasisEmbedding, StronglyEntanglingLayers, SqueezingEmbedding, DisplacementEmbedding, MottonenStatePreparation and Interferometer. (#1910) (#1908) (#1912) (#1920) (#1936) (#1937)

Contributors

This release contains contributions from (in alphabetical order):

Catalina Albornoz, Guillermo Alonso-Linaje, Juan Miguel Arrazola, Ali Asadi, Utkarsh Azad, Samuel Banning, Benjamin Cordier, Alain Delgado, Olivia Di Matteo, Anthony Hayes, David Ittah, Josh Izaac, Soran Jahangiri, Jalani Kanem, Ankit Khandelwal, Nathan Killoran, Shumpei Kobayashi, Robert Lang, Christina Lee, Cedric Lin, Alejandro Montanez, Romain Moyard, Lee James O’Riordan, Chae-Yeun Park, Isidor Schoch, Maria Schuld, Jay Soni, Antal Száva, Rodrigo Vargas, David Wierichs, Roeland Wiersema, Moritz Willmann.

orphan

Release 0.19.1

Bug fixes

  • Fixes several bugs when using parametric operations with the default.qubit.tensor device on GPU. The device takes the torch_device argument once again to allow running non-parametric QNodes on the GPU. (#1927)

  • Fixes a bug where using JAX’s jit function on certain QNodes that contain the qml.QubitStateVector operation raised an error with earlier JAX versions (e.g., jax==0.2.10 and jaxlib==0.1.64). (#1924)

Contributors

This release contains contributions from (in alphabetical order):

Josh Izaac, Christina Lee, Romain Moyard, Lee James O’Riordan, Antal Száva.

orphan

Release 0.19.0

New features since last release

Differentiable Hartree-Fock solver

  • A differentiable Hartree-Fock (HF) solver has been added. It can be used to construct molecular Hamiltonians that can be differentiated with respect to nuclear coordinates and basis-set parameters. (#1610)

    The HF solver computes the integrals over basis functions, constructs the relevant matrices, and performs self-consistent-field iterations to obtain a set of optimized molecular orbital coefficients. These coefficients and the computed integrals over basis functions are used to construct the one- and two-body electron integrals in the molecular orbital basis which can be used to generate a differentiable second-quantized Hamiltonian in the fermionic and qubit basis.

    The following code shows the construction of the Hamiltonian for the hydrogen molecule where the geometry of the molecule is differentiable.

    symbols = ["H", "H"]
    geometry = np.array([[0.0000000000, 0.0000000000, -0.6943528941],
                         [0.0000000000, 0.0000000000,  0.6943528941]], requires_grad=True)
    
    mol = qml.hf.Molecule(symbols, geometry)
    args_mol = [geometry]
    
    hamiltonian = qml.hf.generate_hamiltonian(mol)(*args_mol)
    
    >>> hamiltonian.coeffs
    tensor([-0.09041082+0.j,  0.17220382+0.j,  0.17220382+0.j,
             0.16893367+0.j,  0.04523101+0.j, -0.04523101+0.j,
            -0.04523101+0.j,  0.04523101+0.j, -0.22581352+0.j,
             0.12092003+0.j, -0.22581352+0.j,  0.16615103+0.j,
             0.16615103+0.j,  0.12092003+0.j,  0.17464937+0.j], requires_grad=True)
    

    The generated Hamiltonian can be used in a circuit where the atomic coordinates and circuit parameters are optimized simultaneously.

    symbols = ["H", "H"]
    geometry = np.array([[0.0000000000, 0.0000000000, 0.0],
                         [0.0000000000, 0.0000000000, 2.0]], requires_grad=True)
    
    mol = qml.hf.Molecule(symbols, geometry)
    
    dev = qml.device("default.qubit", wires=4)
    params = [np.array([0.0], requires_grad=True)]
    
    def generate_circuit(mol):
        @qml.qnode(dev)
        def circuit(*args):
            qml.BasisState(np.array([1, 1, 0, 0]), wires=[0, 1, 2, 3])
            qml.DoubleExcitation(*args[0][0], wires=[0, 1, 2, 3])
            return qml.expval(qml.hf.generate_hamiltonian(mol)(*args[1:]))
        return circuit
    
    for n in range(25):
    
        mol = qml.hf.Molecule(symbols, geometry)
        args = [params, geometry] # initial values of the differentiable parameters
    
        g_params = qml.grad(generate_circuit(mol), argnum = 0)(*args)
        params = params - 0.5 * g_params[0]
    
        forces = qml.grad(generate_circuit(mol), argnum = 1)(*args)
        geometry = geometry - 0.5 * forces
    
        print(f'Step: {n}, Energy: {generate_circuit(mol)(*args)}, Maximum Force: {forces.max()}')
    

    In addition, the new Hartree-Fock solver can further be used to optimize the basis set parameters. For details, please refer to the differentiable Hartree-Fock solver documentation.

Integration with Mitiq

  • Error mitigation using the zero-noise extrapolation method is now available through the transforms.mitigate_with_zne transform. This transform can integrate with the Mitiq package for unitary folding and extrapolation functionality. (#1813)

    Consider the following noisy device:

    noise_strength = 0.05
    dev = qml.device("default.mixed", wires=2)
    dev = qml.transforms.insert(qml.AmplitudeDamping, noise_strength)(dev)
    

    We can mitigate the effects of this noise for circuits run on this device by using the added transform:

    from mitiq.zne.scaling import fold_global
    from mitiq.zne.inference import RichardsonFactory
    
    n_wires = 2
    n_layers = 2
    
    shapes = qml.SimplifiedTwoDesign.shape(n_wires, n_layers)
    np.random.seed(0)
    w1, w2 = [np.random.random(s) for s in shapes]
    
    @qml.transforms.mitigate_with_zne([1, 2, 3], fold_global, RichardsonFactory.extrapolate)
    @qml.beta.qnode(dev)
    def circuit(w1, w2):
        qml.SimplifiedTwoDesign(w1, w2, wires=range(2))
        return qml.expval(qml.PauliZ(0))
    

    Now, when we execute circuit, errors will be automatically mitigated:

    >>> circuit(w1, w2)
    0.19113067083636542
    

Powerful new transforms

  • The unitary matrix corresponding to a quantum circuit can now be generated using the new get_unitary_matrix() transform. (#1609) (#1786)

    This transform is fully differentiable across all supported PennyLane autodiff frameworks.

    def circuit(theta):
        qml.RX(theta, wires=1)
        qml.PauliZ(wires=0)
        qml.CNOT(wires=[0, 1])
    
    >>> theta = torch.tensor(0.3, requires_grad=True)
    >>> matrix = qml.transforms.get_unitary_matrix(circuit)(theta)
    >>> print(matrix)
    tensor([[ 0.9888+0.0000j,  0.0000+0.0000j,  0.0000-0.1494j,  0.0000+0.0000j],
          [ 0.0000+0.0000j,  0.0000+0.1494j,  0.0000+0.0000j, -0.9888+0.0000j],
          [ 0.0000-0.1494j,  0.0000+0.0000j,  0.9888+0.0000j,  0.0000+0.0000j],
          [ 0.0000+0.0000j, -0.9888+0.0000j,  0.0000+0.0000j,  0.0000+0.1494j]],
         grad_fn=<MmBackward>)
    >>> loss = torch.real(torch.trace(matrix))
    >>> loss.backward()
    >>> theta.grad
    tensor(-0.1494)
    
  • Arbitrary two-qubit unitaries can now be decomposed into elementary gates. This functionality has been incorporated into the qml.transforms.unitary_to_rot transform, and is available separately as qml.transforms.two_qubit_decomposition. (#1552)

    As an example, consider the following randomly-generated matrix and circuit that uses it:

    U = np.array([
        [-0.03053706-0.03662692j,  0.01313778+0.38162226j, 0.4101526 -0.81893687j, -0.03864617+0.10743148j],
        [-0.17171136-0.24851809j,  0.06046239+0.1929145j, -0.04813084-0.01748555j, -0.29544883-0.88202604j],
        [ 0.39634931-0.78959795j, -0.25521689-0.17045233j, -0.1391033 -0.09670952j, -0.25043606+0.18393466j],
        [ 0.29599198-0.19573188j,  0.55605806+0.64025769j, 0.06140516+0.35499559j,  0.02674726+0.1563311j ]
    ])
    
    dev = qml.device('default.qubit', wires=2)
    
    @qml.qnode(dev)
    @qml.transforms.unitary_to_rot
    def circuit(x, y):
        qml.QubitUnitary(U, wires=[0, 1])
        return qml.expval(qml.PauliZ(wires=0))
    

    If we run the circuit, we can see the new decomposition:

    >>> circuit(0.3, 0.4)
    tensor(-0.81295986, requires_grad=True)
    >>> print(qml.draw(circuit)(0.3, 0.4))
    0: ──Rot(2.78, 0.242, -2.28)──╭X──RZ(0.176)───╭C─────────────╭X──Rot(-3.87, 0.321, -2.09)──┤ ⟨Z⟩
    1: ──Rot(4.64, 2.69, -1.56)───╰C──RY(-0.883)──╰X──RY(-1.47)──╰C──Rot(1.68, 0.337, 0.587)───┤
    
  • A new transform, @qml.batch_params, has been added, that makes QNodes handle a batch dimension in trainable parameters. (#1710) (#1761)

    This transform will create multiple circuits, one per batch dimension. As a result, it is both simulator and hardware compatible.

    @qml.batch_params
    @qml.beta.qnode(dev)
    def circuit(x, weights):
        qml.RX(x, wires=0)
        qml.RY(0.2, wires=1)
        qml.templates.StronglyEntanglingLayers(weights, wires=[0, 1, 2])
        return qml.expval(qml.Hadamard(0))
    

    The qml.batch_params decorator allows us to pass arguments x and weights that have a batch dimension. For example,

    >>> batch_size = 3
    >>> x = np.linspace(0.1, 0.5, batch_size)
    >>> weights = np.random.random((batch_size, 10, 3, 3))
    

    If we evaluate the QNode with these inputs, we will get an output of shape (batch_size,):

    >>> circuit(x, weights)
    tensor([0.08569816, 0.12619101, 0.21122004], requires_grad=True)
    
  • The insert transform has now been added, providing a way to insert single-qubit operations into a quantum circuit. The transform can apply to quantum functions, tapes, and devices. (#1795)

    The following QNode can be transformed to add noise to the circuit:

    dev = qml.device("default.mixed", wires=2)
    
    @qml.qnode(dev)
    @qml.transforms.insert(qml.AmplitudeDamping, 0.2, position="end")
    def f(w, x, y, z):
        qml.RX(w, wires=0)
        qml.RY(x, wires=1)
        qml.CNOT(wires=[0, 1])
        qml.RY(y, wires=0)
        qml.RX(z, wires=1)
        return qml.expval(qml.PauliZ(0) @ qml.PauliZ(1))
    

    Executions of this circuit will differ from the noise-free value:

    >>> f(0.9, 0.4, 0.5, 0.6)
    tensor(0.754847, requires_grad=True)
    >>> print(qml.draw(f)(0.9, 0.4, 0.5, 0.6))
     0: ──RX(0.9)──╭C──RY(0.5)──AmplitudeDamping(0.2)──╭┤ ⟨Z ⊗ Z⟩
     1: ──RY(0.4)──╰X──RX(0.6)──AmplitudeDamping(0.2)──╰┤ ⟨Z ⊗ Z⟩
    
  • Common tape expansion functions are now available in qml.transforms, alongside a new create_expand_fn function for easily creating expansion functions from stopping criteria. (#1734) (#1760)

    create_expand_fn takes the default depth to which the expansion function should expand a tape, a stopping criterion, an optional device, and a docstring to be set for the created function. The stopping criterion must take a queuable object and return a boolean.

    For example, to create an expansion function that decomposes all trainable, multi-parameter operations:

    >>> stop_at = ~(qml.operation.has_multipar & qml.operation.is_trainable)
    >>> expand_fn = qml.transforms.create_expand_fn(depth=5, stop_at=stop_at)
    

    The created expansion function can be used within a custom transform. Devices can also be provided, producing expansion functions that decompose tapes to support the native gate set of the device.

Batch execution of circuits

  • A new, experimental QNode has been added, that adds support for batch execution of circuits, custom quantum gradient support, and arbitrary order derivatives. This QNode is available via qml.beta.QNode, and @qml.beta.qnode. (#1642) (#1646) (#1651) (#1804)

    It differs from the standard QNode in several ways:

    • Custom gradient transforms can be specified as the differentiation method:

      @qml.gradients.gradient_transform
      def my_gradient_transform(tape):
          ...
          return tapes, processing_fn
      
      @qml.beta.qnode(dev, diff_method=my_gradient_transform)
      def circuit():
      
    • Arbitrary \(n\)-th order derivatives are supported on hardware using gradient transforms such as the parameter-shift rule. To specify that an \(n\)-th order derivative of a QNode will be computed, the max_diff argument should be set. By default, this is set to 1 (first-order derivatives only).

    • Internally, if multiple circuits are generated for execution simultaneously, they will be packaged into a single job for execution on the device. This can lead to significant performance improvement when executing the QNode on remote quantum hardware.

    • When decomposing the circuit, the default decomposition strategy will prioritize decompositions that result in the smallest number of parametrized operations required to satisfy the differentiation method. Additional decompositions required to satisfy the native gate set of the quantum device will be performed later, by the device at execution time. While this may lead to a slight increase in classical processing, it significantly reduces the number of circuit evaluations needed to compute gradients of complex unitaries.

    In an upcoming release, this QNode will replace the existing one. If you come across any bugs while using this QNode, please let us know via a bug report on our GitHub bug tracker.

    Currently, this beta QNode does not support the following features:

    • Non-mutability via the mutable keyword argument

    • The reversible QNode differentiation method

    • The ability to specify a dtype when using PyTorch and TensorFlow.

    It is also not tested with the qml.qnn module.

New operations and templates

  • Added a new operation OrbitalRotation, which implements the spin-adapted spatial orbital rotation gate. (#1665)

    An example circuit that uses OrbitalRotation operation is:

    dev = qml.device('default.qubit', wires=4)
    
    @qml.qnode(dev)
    def circuit(phi):
        qml.BasisState(np.array([1, 1, 0, 0]), wires=[0, 1, 2, 3])
        qml.OrbitalRotation(phi, wires=[0, 1, 2, 3])
        return qml.state()
    

    If we run this circuit, we will get the following output

    >>> circuit(0.1)
    array([ 0.        +0.j,  0.        +0.j,  0.        +0.j,
            0.00249792+0.j,  0.        +0.j,  0.        +0.j,
            -0.04991671+0.j,  0.        +0.j,  0.        +0.j,
            -0.04991671+0.j,  0.        +0.j,  0.        +0.j,
            0.99750208+0.j,  0.        +0.j,  0.        +0.j,
            0.        +0.j])
    
  • Added a new template GateFabric, which implements a local, expressive, quantum-number-preserving ansatz proposed by Anselmetti et al. in arXiv:2104.05692. (#1687)

    An example of a circuit using GateFabric template is:

    coordinates = np.array([0.0, 0.0, -0.6614, 0.0, 0.0, 0.6614])
    H, qubits = qml.qchem.molecular_hamiltonian(["H", "H"], coordinates)
    ref_state = qml.qchem.hf_state(electrons=2, orbitals=qubits)
    
    dev = qml.device('default.qubit', wires=qubits)
    
    @qml.qnode(dev)
    def ansatz(weights):
        qml.templates.GateFabric(weights, wires=[0,1,2,3],
                                    init_state=ref_state, include_pi=True)
        return qml.expval(H)
    

    For more details, see the GateFabric documentation.

  • Added a new template kUpCCGSD, which implements a unitary coupled cluster ansatz with generalized singles and pair doubles excitation operators, proposed by Joonho Lee et al. in arXiv:1810.02327. (#1743)

    An example of a circuit using kUpCCGSD template is:

    coordinates = np.array([0.0, 0.0, -0.6614, 0.0, 0.0, 0.6614])
    H, qubits = qml.qchem.molecular_hamiltonian(["H", "H"], coordinates)
    ref_state = qml.qchem.hf_state(electrons=2, orbitals=qubits)
    
    dev = qml.device('default.qubit', wires=qubits)
    
    @qml.qnode(dev)
    def ansatz(weights):
        qml.templates.kUpCCGSD(weights, wires=[0,1,2,3], k=0, delta_sz=0,
                                    init_state=ref_state)
        return qml.expval(H)
    

Improved utilities for quantum compilation and characterization

  • The new qml.fourier.qnode_spectrum function extends the former qml.fourier.spectrum function and takes classical processing of QNode arguments into account. The frequencies are computed per (requested) QNode argument instead of per gate id. The gate ids are ignored. (#1681) (#1720)

    Consider the following example, which uses non-trainable inputs x, y and z as well as trainable parameters w as arguments to the QNode.

    import pennylane as qml
    import numpy as np
    
    n_qubits = 3
    dev = qml.device("default.qubit", wires=n_qubits)
    
    @qml.qnode(dev)
    def circuit(x, y, z, w):
        for i in range(n_qubits):
            qml.RX(0.5*x[i], wires=i)
            qml.Rot(w[0,i,0], w[0,i,1], w[0,i,2], wires=i)
            qml.RY(2.3*y[i], wires=i)
            qml.Rot(w[1,i,0], w[1,i,1], w[1,i,2], wires=i)
            qml.RX(z, wires=i)
        return qml.expval(qml.PauliZ(wires=0))
    
    x = np.array([1., 2., 3.])
    y = np.array([0.1, 0.3, 0.5])
    z = -1.8
    w = np.random.random((2, n_qubits, 3))
    

    This circuit looks as follows:

    >>> print(qml.draw(circuit)(x, y, z, w))
    0: ──RX(0.5)──Rot(0.598, 0.949, 0.346)───RY(0.23)──Rot(0.693, 0.0738, 0.246)──RX(-1.8)──┤ ⟨Z⟩
    1: ──RX(1)────Rot(0.0711, 0.701, 0.445)──RY(0.69)──Rot(0.32, 0.0482, 0.437)───RX(-1.8)──┤
    2: ──RX(1.5)──Rot(0.401, 0.0795, 0.731)──RY(1.15)──Rot(0.756, 0.38, 0.38)─────RX(-1.8)──┤
    

    Applying the qml.fourier.qnode_spectrum function to the circuit for the non-trainable parameters, we obtain:

    >>> spec = qml.fourier.qnode_spectrum(circuit, encoding_args={"x", "y", "z"})(x, y, z, w)
    >>> for inp, freqs in spec.items():
    ...     print(f"{inp}: {freqs}")
    "x": {(0,): [-0.5, 0.0, 0.5], (1,): [-0.5, 0.0, 0.5], (2,): [-0.5, 0.0, 0.5]}
    "y": {(0,): [-2.3, 0.0, 2.3], (1,): [-2.3, 0.0, 2.3], (2,): [-2.3, 0.0, 2.3]}
    "z": {(): [-3.0, -2.0, -1.0, 0.0, 1.0, 2.0, 3.0]}
    

    We can see that all three parameters in the QNode arguments x and y contribute the spectrum of a Pauli rotation [-1.0, 0.0, 1.0], rescaled with the prefactor of the respective parameter in the circuit. The three RX rotations using the parameter z accumulate, yielding a more complex frequency spectrum.

    For details on how to control for which parameters the spectrum is computed, a comparison to qml.fourier.circuit_spectrum, and other usage details, please see the fourier.qnode_spectrum docstring.

  • Two new methods were added to the Device API, allowing PennyLane devices increased control over circuit decompositions. (#1651)

    • Device.expand_fn(tape) -> tape: expands a tape such that it is supported by the device. By default, performs the standard device-specific gate set decomposition done in the default QNode. Devices may overwrite this method in order to define their own decomposition logic.

      Note that the numerical result after applying this method should remain unchanged; PennyLane will assume that the expanded tape returns exactly the same value as the original tape when executed.

    • Device.batch_transform(tape) -> (tapes, processing_fn): preprocesses the tape in the case where the device needs to generate multiple circuits to execute from the input circuit. The requirement of a post-processing function makes this distinct to the expand_fn method above.

      By default, this method applies the transform

      \[\left\langle \sum_i c_i h_i\right\rangle → \sum_i c_i \left\langle h_i \right\rangle\]

      if expval(H) is present on devices that do not natively support Hamiltonians with non-commuting terms.

  • A new class has been added to store operator attributes, such as self_inverses, and composable_rotation, as a list of operation names. (#1763)

    A number of such attributes, for the purpose of compilation transforms, can be found in ops/qubit/attributes.py, but the class can also be used to create your own. For example, we can create a new Attribute, pauli_ops, like so:

    >>> from pennylane.ops.qubit.attributes import Attribute
    >>> pauli_ops = Attribute(["PauliX", "PauliY", "PauliZ"])
    

    We can check either a string or an Operation for inclusion in this set:

    >>> qml.PauliX(0) in pauli_ops
    True
    >>> "Hadamard" in pauli_ops
    False
    

    We can also dynamically add operators to the sets at runtime. This is useful for adding custom operations to the attributes such as composable_rotations and self_inverses that are used in compilation transforms. For example, suppose you have created a new Operation, MyGate, which you know to be its own inverse. Adding it to the set, like so

    >>> from pennylane.ops.qubit.attributes import self_inverses
    >>> self_inverses.add("MyGate")
    

    will enable the gate to be considered by the cancel_inverses compilation transform if two such gates are adjacent in a circuit.

Improvements

  • The qml.metric_tensor transform has been improved with regards to both function and performance. (#1638) (#1721)

    • If the underlying device supports batch execution of circuits, the quantum circuits required to compute the metric tensor elements will be automatically submitted as a batched job. This can lead to significant performance improvements for devices with a non-trivial job submission overhead.

    • Previously, the transform would only return the metric tensor with respect to gate arguments, and ignore any classical processing inside the QNode, even very trivial classical processing such as parameter permutation. The metric tensor now takes into account classical processing, and returns the metric tensor with respect to QNode arguments, not simply gate arguments:

      >>> @qml.qnode(dev)
      ... def circuit(x):
      ...     qml.Hadamard(wires=1)
      ...     qml.RX(x[0], wires=0)
      ...     qml.CNOT(wires=[0, 1])
      ...     qml.RY(x[1] ** 2, wires=1)
      ...     qml.RY(x[1], wires=0)
      ...     return qml.expval(qml.PauliZ(0))
      >>> x = np.array([0.1, 0.2], requires_grad=True)
      >>> qml.metric_tensor(circuit)(x)
      array([[0.25      , 0.        ],
             [0.        , 0.28750832]])
      

      To revert to the previous behaviour of returning the metric tensor with respect to gate arguments, qml.metric_tensor(qnode, hybrid=False) can be passed.

      >>> qml.metric_tensor(circuit, hybrid=False)(x)
      array([[0.25      , 0.        , 0.        ],
             [0.        , 0.25      , 0.        ],
             [0.        , 0.        , 0.24750832]])
      
    • The metric tensor transform now works with a larger set of operations. In particular, all operations that have a single variational parameter and define a generator are now supported. In addition to a reduction in decomposition overhead, the change also results in fewer circuit evaluations.

  • The expansion rule in the qml.metric_tensor transform has been changed. (#1721)

    If hybrid=False, the changed expansion rule might lead to a changed output.

  • The ApproxTimeEvolution template can now be used with Hamiltonians that have trainable coefficients. (#1789)

    Resulting QNodes can be differentiated with respect to both the time parameter and the Hamiltonian coefficients.

    dev = qml.device('default.qubit', wires=2)
    obs = [qml.PauliX(0) @ qml.PauliY(1), qml.PauliY(0) @ qml.PauliX(1)]
    
    @qml.qnode(dev)
    def circuit(coeffs, t):
        H = qml.Hamiltonian(coeffs, obs)
        qml.templates.ApproxTimeEvolution(H, t, 2)
        return qml.expval(qml.PauliZ(0))
    
    >>> t = np.array(0.54, requires_grad=True)
    >>> coeffs = np.array([-0.6, 2.0], requires_grad=True)
    >>> qml.grad(circuit)(coeffs, t)
    (array([-1.07813375, -1.07813375]), array(-2.79516158))
    

    All differentiation methods, including backpropagation and the parameter-shift rule, are supported.

  • Quantum function transforms and batch transforms can now be applied to devices. Once applied to a device, any quantum function executed on the modified device will be transformed prior to execution. (#1809) (#1810)

    dev = qml.device("default.mixed", wires=1)
    dev = qml.transforms.merge_rotations()(dev)
    
    @qml.beta.qnode(dev)
    def f(w, x, y, z):
        qml.RX(w, wires=0)
        qml.RX(x, wires=0)
        qml.RX(y, wires=0)
        qml.RX(z, wires=0)
        return qml.expval(qml.PauliZ(0))
    
    >>> print(f(0.9, 0.4, 0.5, 0.6))
     -0.7373937155412453
    >>> print(qml.draw(f, expansion_strategy="device")(0.9, 0.4, 0.5, 0.6))
     0: ──RX(2.4)──┤ ⟨Z⟩
    
  • It is now possible to draw QNodes that have been transformed by a ‘batch transform’; that is, a transform that maps a single QNode into multiple circuits under the hood. Examples of batch transforms include @qml.metric_tensor and @qml.gradients. (#1762)

    For example, consider the parameter-shift rule, which generates two circuits per parameter; one circuit that has the parameter shifted forward, and another that has the parameter shifted backwards:

    dev = qml.device("default.qubit", wires=2)
    
    @qml.gradients.param_shift
    @qml.beta.qnode(dev)
    def circuit(x):
        qml.RX(x, wires=0)
        qml.CNOT(wires=[0, 1])
        return qml.expval(qml.PauliZ(wires=0))
    
    >>> print(qml.draw(circuit)(0.6))
     0: ──RX(2.17)──╭C──┤ ⟨Z⟩
     1: ────────────╰X──┤
    
     0: ──RX(-0.971)──╭C──┤ ⟨Z⟩
     1: ──────────────╰X──┤
    
  • Support for differentiable execution of batches of circuits has been extended to the JAX interface for scalar functions, via the beta pennylane.interfaces.batch module. (#1634) (#1685)

    For example using the execute function from the pennylane.interfaces.batch module:

    from pennylane.interfaces.batch import execute
    
    def cost_fn(x):
        with qml.tape.JacobianTape() as tape1:
            qml.RX(x[0], wires=[0])
            qml.RY(x[1], wires=[1])
            qml.CNOT(wires=[0, 1])
            qml.var(qml.PauliZ(0) @ qml.PauliX(1))
    
        with qml.tape.JacobianTape() as tape2:
            qml.RX(x[0], wires=0)
            qml.RY(x[0], wires=1)
            qml.CNOT(wires=[0, 1])
            qml.probs(wires=1)
    
        result = execute(
          [tape1, tape2], dev,
          gradient_fn=qml.gradients.param_shift,
          interface="autograd"
        )
        return (result[0] + result[1][0, 0])[0]
    
    res = jax.grad(cost_fn)(params)
    
  • All qubit operations have been re-written to use the qml.math framework for internal classical processing and the generation of their matrix representations. As a result these representations are now fully differentiable, and the framework-specific device classes no longer need to maintain framework-specific versions of these matrices. (#1749) (#1802)

  • The use of expval(H), where H is a cost Hamiltonian generated by the qaoa module, has been sped up. This was achieved by making PennyLane decompose a circuit with an expval(H) measurement into subcircuits if the Hamiltonian.grouping_indices attribute is set, and setting this attribute in the relevant qaoa module functions. (#1718)

  • Operations can now have gradient recipes that depend on the state of the operation. (#1674)

    For example, this allows for gradient recipes that are parameter dependent:

    class RX(qml.RX):
    
        @property
        def grad_recipe(self):
            # The gradient is given by [f(2x) - f(0)] / (2 sin(x)), by subsituting
            # shift = x into the two term parameter-shift rule.
            x = self.data[0]
            c = 0.5 / np.sin(x)
            return ([[c, 0.0, 2 * x], [-c, 0.0, 0.0]],)
    
  • Shots can now be passed as a runtime argument to transforms that execute circuits in batches, similarly to QNodes. (#1707)

    An example of such a transform are the gradient transforms in the qml.gradients module. As a result, we can now call gradient transforms (such as qml.gradients.param_shift) and set the number of shots at runtime.

    >>> dev = qml.device("default.qubit", wires=1, shots=1000)
    >>> @qml.beta.qnode(dev)
    ... def circuit(x):
    ...     qml.RX(x, wires=0)
    ...     return qml.expval(qml.PauliZ(0))
    >>> grad_fn = qml.gradients.param_shift(circuit)
    >>> param = np.array(0.564, requires_grad=True)
    >>> grad_fn(param, shots=[(1, 10)]).T
    array([[-1., -1., -1., -1., -1.,  0., -1.,  0., -1.,  0.]])
    >>> param2 = np.array(0.1233, requires_grad=True)
    >>> grad_fn(param2, shots=None)
    array([[-0.12298782]])
    
  • Templates are now top level imported and can be used directly e.g. qml.QFT(wires=0). (#1779)

  • qml.probs now accepts an attribute op that allows to rotate the computational basis and get the probabilities in the rotated basis. (#1692)

  • Refactored the expand_fn functionality in the Device class to avoid any edge cases leading to failures with plugins. (#1838)

  • Updated the qml.QNGOptimizer.step_and_cost method to avoid the use of deprecated functionality. (#1834)

  • Added a custom torch.to_numpy implementation to pennylane/math/single_dispatch.py to ensure compabilitity with PyTorch 1.10. (#1824) (#1825)

  • The default for an Operation‘s control_wires attribute is now an empty Wires object instead of the attribute raising a NonImplementedError. (#1821)

  • qml.circuit_drawer.MPLDrawer will now automatically rotate and resize text to fit inside the rectangle created by the box_gate method. (#1764)

  • Operators now have a label method to determine how they are drawn. This will eventually override the RepresentationResolver class. (#1678)

  • The operation label method now supports string variables. (#1815)

  • A new utility class qml.BooleanFn is introduced. It wraps a function that takes a single argument and returns a Boolean. (#1734)

    After wrapping, qml.BooleanFn can be called like the wrapped function, and multiple instances can be manipulated and combined with the bitwise operators &, | and ~.

  • There is a new utility function qml.math.is_independent that checks whether a callable is independent of its arguments. (#1700)

    This function is experimental and might behave differently than expected.

    Note that the test relies on both numerical and analytical checks, except when using the PyTorch interface which only performs a numerical check. It is known that there are edge cases on which this test will yield wrong results, in particular non-smooth functions may be problematic. For details, please refer to the is_indpendent docstring.

  • The qml.beta.QNode now supports the qml.qnn module. (#1748)

  • @qml.beta.QNode now supports the qml.specs transform. (#1739)

  • qml.circuit_drawer.drawable_layers and qml.circuit_drawer.drawable_grid process a list of operations to layer positions for drawing. (#1639)

  • qml.transforms.batch_transform now accepts expand_fns that take additional arguments and keyword arguments. In fact, expand_fn and transform_fn now must have the same signature. (#1721)

  • The qml.batch_transform decorator is now ignored during Sphinx builds, allowing the correct signature to display in the built documentation. (#1733)

  • The tests for qubit operations are split into multiple files. (#1661)

  • The transform for the Jacobian of the classical preprocessing within a QNode, qml.transforms.classical_jacobian, now takes a keyword argument argnum to specify the QNode argument indices with respect to which the Jacobian is computed. (#1645)

    An example for the usage of argnum is

    @qml.qnode(dev)
    def circuit(x, y, z):
        qml.RX(qml.math.sin(x), wires=0)
        qml.CNOT(wires=[0, 1])
        qml.RY(y ** 2, wires=1)
        qml.RZ(1 / z, wires=1)
        return qml.expval(qml.PauliZ(0))
    
    jac_fn = qml.transforms.classical_jacobian(circuit, argnum=[1, 2])
    

    The Jacobian can then be computed at specified parameters.

    >>> x, y, z = np.array([0.1, -2.5, 0.71])
    >>> jac_fn(x, y, z)
    (array([-0., -5., -0.]), array([-0.        , -0.        , -1.98373339]))
    

    The returned arrays are the derivatives of the three parametrized gates in the circuit with respect to y and z respectively.

    There also are explicit tests for classical_jacobian now, which previously was tested implicitly via its use in the metric_tensor transform.

    For more usage details, please see the classical Jacobian docstring.

  • A new utility function qml.math.is_abstract(tensor) has been added. This function returns True if the tensor is abstract; that is, it has no value or shape. This can occur if within a function that has been just-in-time compiled. (#1845)

  • qml.circuit_drawer.CircuitDrawer can accept a string for the charset keyword, instead of a CharSet object. (#1640)

  • qml.math.sort will now return only the sorted torch tensor and not the corresponding indices, making sort consistent across interfaces. (#1691)

  • Specific QNode execution options are now re-used by batch transforms to execute transformed QNodes. (#1708)

  • To standardize across all optimizers, qml.optimize.AdamOptimizer now also uses accumulation (in form of collections.namedtuple) to keep track of running quantities. Before it used three variables fm, sm and t. (#1757)

Breaking changes

  • The operator attributes has_unitary_generator, is_composable_rotation, is_self_inverse, is_symmetric_over_all_wires, and is_symmetric_over_control_wires have been removed as attributes from the base class. They have been replaced by the sets that store the names of operations with similar properties in ops/qubit/attributes.py. (#1763)

  • The qml.inv function has been removed, qml.adjoint should be used instead. (#1778)

  • The input signature of an expand_fn used in a batch_transform now must have the same signature as the provided transform_fn, and vice versa. (#1721)

  • The default.qubit.torch device automatically determines if computations should be run on a CPU or a GPU and doesn’t take a torch_device argument anymore. (#1705)

  • The utility function qml.math.requires_grad now returns True when using Autograd if and only if the requires_grad=True attribute is set on the NumPy array. Previously, this function would return True for all NumPy arrays and Python floats, unless requires_grad=False was explicitly set. (#1638)

  • The operation qml.Interferometer has been renamed qml.InterferometerUnitary in order to distinguish it from the template qml.templates.Interferometer. (#1714)

  • The qml.transforms.invisible decorator has been replaced with qml.tape.stop_recording, which may act as a context manager as well as a decorator to ensure that contained logic is non-recordable or non-queueable within a QNode or quantum tape context. (#1754)

  • Templates SingleExcitationUnitary and DoubleExcitationUnitary have been renamed to FermionicSingleExcitation and FermionicDoubleExcitation, respectively. (#1822)

Deprecations

  • Allowing cost functions to be differentiated using qml.grad or qml.jacobian without explicitly marking parameters as trainable is being deprecated, and will be removed in an upcoming release. Please specify the requires_grad attribute for every argument, or specify argnum when using qml.grad or qml.jacobian. (#1773)

    The following raises a warning in v0.19.0 and will raise an error in an upcoming release:

    import pennylane as qml
    
    dev = qml.device('default.qubit', wires=1)
    
    @qml.qnode(dev)
    def test(x):
        qml.RY(x, wires=[0])
        return qml.expval(qml.PauliZ(0))
    
    par = 0.3
    qml.grad(test)(par)
    

    Preferred approaches include specifying the requires_grad attribute:

    import pennylane as qml
    from pennylane import numpy as np
    
    dev = qml.device('default.qubit', wires=1)
    
    @qml.qnode(dev)
    def test(x):
        qml.RY(x, wires=[0])
        return qml.expval(qml.PauliZ(0))
    
    par = np.array(0.3, requires_grad=True)
    qml.grad(test)(par)
    

    Or specifying the argnum argument when using qml.grad or qml.jacobian:

    import pennylane as qml
    
    dev = qml.device('default.qubit', wires=1)
    
    @qml.qnode(dev)
    def test(x):
        qml.RY(x, wires=[0])
        return qml.expval(qml.PauliZ(0))
    
    par = 0.3
    qml.grad(test, argnum=0)(par)
    

  • The default.tensor device from the beta folder is no longer maintained and has been deprecated. It will be removed in future releases. (#1851)

  • The qml.metric_tensor and qml.QNGOptimizer keyword argument diag_approx is deprecated. Approximations can be controlled with the more fine-grained approx keyword argument, with approx="block-diag" (the default) reproducing the old behaviour. (#1721) (#1834)

  • The template decorator is now deprecated with a warning message and will be removed in release v0.20.0. It has been removed from different PennyLane functions. (#1794) (#1808)

  • The qml.fourier.spectrum function has been renamed to qml.fourier.circuit_spectrum, in order to clearly separate the new qnode_spectrum function from this one. qml.fourier.spectrum is now an alias for circuit_spectrum but is flagged for deprecation and will be removed soon. (#1681)

  • The init module, which contains functions to generate random parameter tensors for templates, is flagged for deprecation and will be removed in the next release cycle. Instead, the templates’ shape method can be used to get the desired shape of the tensor, which can then be generated manually. (#1689)

    To generate the parameter tensors, the np.random.normal and np.random.uniform functions can be used (just like in the init module). Considering the default arguments of these functions as of NumPy v1.21, some non-default options were used by the init module:

    • All functions generating normally distributed parameters used np.random.normal by passing scale=0.1;

    • Most functions generating uniformly distributed parameters (except for certain CVQNN initializers) used np.random.uniform by passing high=2*math.pi;

    • The cvqnn_layers_r_uniform, cvqnn_layers_a_uniform, cvqnn_layers_kappa_uniform functions used np.random.uniform by passing high=0.1.

  • The QNode.draw method has been deprecated, and will be removed in an upcoming release. Please use the qml.draw transform instead. (#1746)

  • The QNode.metric_tensor method has been deprecated, and will be removed in an upcoming release. Please use the qml.metric_tensor transform instead. (#1638)

  • The pad parameter of the qml.AmplitudeEmbedding template has been removed. It has instead been renamed to the pad_with parameter. (#1805)

Bug fixes

  • Fixes a bug where qml.math.dot failed to work with @tf.function autograph mode. (#1842)

  • Fixes a bug where in rare instances the parameters of a tape are returned unsorted by Tape.get_parameters. (#1836)

  • Fixes a bug with the arrow width in the measure of qml.circuit_drawer.MPLDrawer. (#1823)

  • The helper functions qml.math.block_diag and qml.math.scatter_element_add now are entirely differentiable when using Autograd. Previously only indexed entries of the block diagonal could be differentiated, while the derivative w.r.t to the second argument of qml.math.scatter_element_add dispatched to NumPy instead of Autograd. (#1816) (#1818)

  • Fixes a bug such that the original shot vector information of a device is preserved, so that outside the context manager the device remains unchanged. (#1792)

  • Modifies qml.math.take to be compatible with a breaking change released in JAX 0.2.24 and ensure that PennyLane supports this JAX version. (#1769)

  • Fixes a bug where the GPU cannot be used with qml.qnn.TorchLayer. (#1705)

  • Fix a bug where the devices cache the same result for different observables return types. (#1719)

  • Fixed a bug of the default circuit drawer where having more measurements compared to the number of measurements on any wire raised a KeyError. (#1702)

  • Fix a bug where it was not possible to use jax.jit on a QNode when using QubitStateVector. (#1683)

  • The device suite tests can now execute successfully if no shots configuration variable is given. (#1641)

  • Fixes a bug where the qml.gradients.param_shift transform would raise an error while attempting to compute the variance of a QNode with ragged output. (#1646)

  • Fixes a bug in default.mixed, to ensure that returned probabilities are always non-negative. (#1680)

  • Fixes a bug where gradient transforms would fail to apply to QNodes containing classical processing. (#1699)

  • Fixes a bug where the the parameter-shift method was not correctly using the fallback gradient function when all circuit parameters required the fallback. (#1782)

Documentation

  • Adds a link to https://pennylane.ai/qml/demonstrations.html in the navbar. (#1624)

  • Corrects the docstring of ExpvalCost by adding wires to the signature of the ansatz argument. (#1715)

  • Updated docstring examples using the qchem.molecular_hamiltonian function. (#1724)

  • Updates the ‘Gradients and training’ quickstart guide to provide information on gradient transforms. (#1751)

  • All instances of qnode.draw() have been updated to instead use the transform qml.draw(qnode). (#1750)

  • Add the jax interface in QNode Documentation. (#1755)

  • Reorganized all the templates related to quantum chemistry under a common header Quantum Chemistry templates. (#1822)

Contributors

This release contains contributions from (in alphabetical order):

Catalina Albornoz, Juan Miguel Arrazola, Utkarsh Azad, Akash Narayanan B, Sam Banning, Thomas Bromley, Jack Ceroni, Alain Delgado, Olivia Di Matteo, Andrew Gardhouse, Anthony Hayes, Theodor Isacsson, David Ittah, Josh Izaac, Soran Jahangiri, Nathan Killoran, Christina Lee, Guillermo Alonso-Linaje, Romain Moyard, Lee James O’Riordan, Carrie-Anne Rubidge, Maria Schuld, Rishabh Singh, Jay Soni, Ingrid Strandberg, Antal Száva, Teresa Tamayo-Mendoza, Rodrigo Vargas, Cody Wang, David Wierichs, Moritz Willmann.

orphan

Release 0.18.0

New features since last release

PennyLane now comes packaged with lightning.qubit

  • The C++-based lightning.qubit device is now included with installations of PennyLane. (#1663)

    The lightning.qubit device is a fast state-vector simulator equipped with the efficient adjoint method for differentiating quantum circuits, check out the plugin release notes for more details! The device can be accessed in the following way:

    import pennylane as qml
    
    wires = 3
    layers = 2
    dev = qml.device("lightning.qubit", wires=wires)
    
    @qml.qnode(dev, diff_method="adjoint")
    def circuit(weights):
        qml.templates.StronglyEntanglingLayers(weights, wires=range(wires))
        return qml.expval(qml.PauliZ(0))
    
    weights = qml.init.strong_ent_layers_normal(layers, wires, seed=1967)
    

    Evaluating circuits and their gradients on the device can be achieved using the standard approach:

    >>> print(f"Circuit evaluated: {circuit(weights)}")
    Circuit evaluated: 0.9801286266677633
    >>> print(f"Circuit gradient:\n{qml.grad(circuit)(weights)}")
    Circuit gradient:
    [[[-9.35301749e-17 -1.63051504e-01 -4.14810501e-04]
      [-7.88816484e-17 -1.50136528e-04 -1.77922957e-04]
      [-5.20670796e-17 -3.92874550e-02  8.14523075e-05]]
    
     [[-1.14472273e-04  3.85963953e-02 -9.39190132e-18]
      [-5.76791765e-05 -9.78478343e-02  0.00000000e+00]
      [ 0.00000000e+00  0.00000000e+00  0.00000000e+00]]]
    

    The adjoint method operates after a forward pass by iteratively applying inverse gates to scan backwards through the circuit. The method is already available in PennyLane’s default.qubit device, but the version provided by lightning.qubit integrates with the C++ backend and is more performant, as shown in the plot below:

Support for native backpropagation using PyTorch

  • The built-in PennyLane simulator default.qubit now supports backpropogation with PyTorch. (#1360) (#1598)

    As a result, default.qubit can now use end-to-end classical backpropagation as a means to compute gradients. End-to-end backpropagation can be faster than the parameter-shift rule for computing quantum gradients when the number of parameters to be optimized is large. This is now the default differentiation method when using default.qubit with PyTorch.

    Using this method, the created QNode is a ‘white-box’ that is tightly integrated with your PyTorch computation, including TorchScript and GPU support.

    x = torch.tensor(0.43316321, dtype=torch.float64, requires_grad=True)
    y = torch.tensor(0.2162158, dtype=torch.float64, requires_grad=True)
    z = torch.tensor(0.75110998, dtype=torch.float64, requires_grad=True)
    
    p = torch.tensor([x, y, z], requires_grad=True)
    dev = qml.device("default.qubit", wires=1)
    
    @qml.qnode(dev, interface="torch", diff_method="backprop")
    def circuit(x):
        qml.Rot(x[0], x[1], x[2], wires=0)
        return qml.expval(qml.PauliZ(0))
    
    res = circuit(p)
    res.backward()
    
    >>> res = circuit(p)
    >>> res.backward()
    >>> print(p.grad)
    tensor([-9.1798e-17, -2.1454e-01, -1.0511e-16], dtype=torch.float64)
    

Improved quantum optimization methods

  • The RotosolveOptimizer now can tackle general parametrized circuits, and is no longer restricted to single-qubit Pauli rotations. (#1489)

    This includes:

    • layers of gates controlled by the same parameter,

    • controlled variants of parametrized gates, and

    • Hamiltonian time evolution.

    Note that the eigenvalue spectrum of the gate generator needs to be known to use RotosolveOptimizer for a general gate, and it is required to produce equidistant frequencies. For details see Vidal and Theis, 2018 and Wierichs, Izaac, Wang, Lin 2021.

    Consider a circuit with a mixture of Pauli rotation gates, controlled Pauli rotations, and single-parameter layers of Pauli rotations:

    dev = qml.device('default.qubit', wires=3, shots=None)
    
    @qml.qnode(dev)
    def cost_function(rot_param, layer_par, crot_param):
        for i, par in enumerate(rot_param):
            qml.RX(par, wires=i)
        for w in dev.wires:
            qml.RX(layer_par, wires=w)
        for i, par in enumerate(crot_param):
            qml.CRY(par, wires=[i, (i+1) % 3])
    
        return qml.expval(qml.PauliZ(0) @ qml.PauliZ(1) @ qml.PauliZ(2))
    

    This cost function has one frequency for each of the first RX rotation angles, three frequencies for the layer of RX gates that depend on layer_par, and two frequencies for each of the CRY gate parameters. Rotosolve can then be used to minimize the cost_function:

    # Initial parameters
    init_param = [
        np.array([0.3, 0.2, 0.67], requires_grad=True),
        np.array(1.1, requires_grad=True),
        np.array([-0.2, 0.1, -2.5], requires_grad=True),
    ]
    # Numbers of frequencies per parameter
    num_freqs = [[1, 1, 1], 3, [2, 2, 2]]
    
    opt = qml.RotosolveOptimizer()
    param = init_param.copy()
    

    In addition, the optimization technique for the Rotosolve substeps can be chosen via the optimizer and optimizer_kwargs keyword arguments and the minimized cost of the intermediate univariate reconstructions can be read out via full_output, including the cost after the full Rotosolve step:

    for step in range(3):
        param, cost, sub_cost = opt.step_and_cost(
            cost_function,
            *param,
            num_freqs=num_freqs,
            full_output=True,
            optimizer="brute",
        )
        print(f"Cost before step: {cost}")
        print(f"Minimization substeps: {np.round(sub_cost, 6)}")
    
    Cost before step: 0.042008210392535605
    Minimization substeps: [-0.230905 -0.863336 -0.980072 -0.980072 -1.       -1.       -1.      ]
    Cost before step: -0.999999999068121
    Minimization substeps: [-1. -1. -1. -1. -1. -1. -1.]
    Cost before step: -1.0
    Minimization substeps: [-1. -1. -1. -1. -1. -1. -1.]
    

    For usage details please consider the docstring of the optimizer.

Faster, trainable, Hamiltonian simulations

  • Hamiltonians are now trainable with respect to their coefficients. (#1483)

    from pennylane import numpy as np
    
    dev = qml.device("default.qubit", wires=2)
    @qml.qnode(dev)
    def circuit(coeffs, param):
        qml.RX(param, wires=0)
        qml.RY(param, wires=0)
        return qml.expval(
            qml.Hamiltonian(coeffs, [qml.PauliX(0), qml.PauliZ(0)], simplify=True)
        )
    
    coeffs = np.array([-0.05, 0.17])
    param = np.array(1.7)
    grad_fn = qml.grad(circuit)
    
    >>> grad_fn(coeffs, param)
    (array([-0.12777055,  0.0166009 ]), array(0.0917819))
    

    Furthermore, a gradient recipe for Hamiltonian coefficients has been added. This makes it possible to compute parameter-shift gradients of these coefficients on devices that natively support Hamiltonians. (#1551)

  • Hamiltonians are now natively supported on the default.qubit device if shots=None. This makes VQE workflows a lot faster in some cases. (#1551) (#1596)

  • The Hamiltonian can now store grouping information, which can be accessed by a device to speed up computations of the expectation value of a Hamiltonian. (#1515)

    obs = [qml.PauliX(0), qml.PauliX(1), qml.PauliZ(0)]
    coeffs = np.array([1., 2., 3.])
    H = qml.Hamiltonian(coeffs, obs, grouping_type='qwc')
    

    Initialization with a grouping_type other than None stores the indices required to make groups of commuting observables and their coefficients.

    >>> H.grouping_indices
    [[0, 1], [2]]
    

Create multi-circuit quantum transforms and custom gradient rules

  • Custom gradient transforms can now be created using the new @qml.gradients.gradient_transform decorator on a batch-tape transform. (#1589)

    Quantum gradient transforms are a specific case of qml.batch_transform.

    Supported gradient transforms must be of the following form:

    @qml.gradients.gradient_transform
    def my_custom_gradient(tape, argnum=None, **kwargs):
        ...
        return gradient_tapes, processing_fn
    

    Various built-in quantum gradient transforms are provided within the qml.gradients module, including qml.gradients.param_shift. Once defined, quantum gradient transforms can be applied directly to QNodes:

    >>> @qml.qnode(dev)
    ... def circuit(x):
    ...     qml.RX(x, wires=0)
    ...     qml.CNOT(wires=[0, 1])
    ...     return qml.expval(qml.PauliZ(0))
    >>> circuit(0.3)
    tensor(0.95533649, requires_grad=True)
    >>> qml.gradients.param_shift(circuit)(0.5)
    array([[-0.47942554]])
    

    Quantum gradient transforms are fully differentiable, allowing higher order derivatives to be accessed:

    >>> qml.grad(qml.gradients.param_shift(circuit))(0.5)
    tensor(-0.87758256, requires_grad=True)
    

    Refer to the page of quantum gradient transforms for more details.

  • The ability to define batch transforms has been added via the new @qml.batch_transform decorator. (#1493)

    A batch transform is a transform that takes a single tape or QNode as input, and executes multiple tapes or QNodes independently. The results may then be post-processed before being returned.

    For example, consider the following batch transform:

    @qml.batch_transform
    def my_transform(tape, a, b):
        """Generates two tapes, one with all RX replaced with RY,
        and the other with all RX replaced with RZ."""
        tape1 = qml.tape.JacobianTape()
        tape2 = qml.tape.JacobianTape()
    
        # loop through all operations on the input tape
        for op in tape.operations + tape.measurements:
            if op.name == "RX":
                with tape1:
                    qml.RY(a * qml.math.abs(op.parameters[0]), wires=op.wires)
                with tape2:
                    qml.RZ(b * qml.math.abs(op.parameters[0]), wires=op.wires)
            else:
                for t in [tape1, tape2]:
                    with t:
                        qml.apply(op)
    
        def processing_fn(results):
            return qml.math.sum(qml.math.stack(results))
    
        return [tape1, tape2], processing_fn
    

    We can transform a QNode directly using decorator syntax:

    >>> @my_transform(0.65, 2.5)
    ... @qml.qnode(dev)
    ... def circuit(x):
    ...     qml.Hadamard(wires=0)
    ...     qml.RX(x, wires=0)
    ...     return qml.expval(qml.PauliX(0))
    >>> print(circuit(-0.5))
    1.2629730888100839
    

    Batch tape transforms are fully differentiable:

    >>> gradient = qml.grad(circuit)(-0.5)
    >>> print(gradient)
    2.5800122591960153
    

    Batch transforms can also be applied to existing QNodes,

    >>> new_qnode = my_transform(existing_qnode, *transform_weights)
    >>> new_qnode(weights)
    

    or to tapes (in which case, the processed tapes and classical post-processing functions are returned):

    >>> tapes, fn = my_transform(tape, 0.65, 2.5)
    >>> from pennylane.interfaces.batch import execute
    >>> dev = qml.device("default.qubit", wires=1)
    >>> res = execute(tapes, dev, interface="autograd", gradient_fn=qml.gradients.param_shift)
    1.2629730888100839
    
  • Vector-Jacobian product transforms have been added to the qml.gradients package. (#1494)

    The new transforms include:

    • qml.gradients.vjp

    • qml.gradients.batch_vjp

  • Support for differentiable execution of batches of circuits has been added, via the beta pennylane.interfaces.batch module. (#1501) (#1508) (#1542) (#1549) (#1608) (#1618) (#1637)

    For now, this is a low-level feature, and will be integrated into the QNode in a future release. For example:

    from pennylane.interfaces.batch import execute
    
    def cost_fn(x):
        with qml.tape.JacobianTape() as tape1:
            qml.RX(x[0], wires=[0])
            qml.RY(x[1], wires=[1])
            qml.CNOT(wires=[0, 1])
            qml.var(qml.PauliZ(0) @ qml.PauliX(1))
    
        with qml.tape.JacobianTape() as tape2:
            qml.RX(x[0], wires=0)
            qml.RY(x[0], wires=1)
            qml.CNOT(wires=[0, 1])
            qml.probs(wires=1)
    
        result = execute(
            [tape1, tape2], dev,
            gradient_fn=qml.gradients.param_shift,
            interface="autograd"
        )
        return result[0] + result[1][0, 0]
    
    res = qml.grad(cost_fn)(params)
    

Improvements

  • A new operation qml.SISWAP has been added, the square-root of the qml.ISWAP operation. (#1563)

  • The frobenius_inner_product function has been moved to the qml.math module, and is now differentiable using all autodiff frameworks. (#1388)

  • A warning is raised to inform the user that specifying a list of shots is only supported for QubitDevice based devices. (#1659)

  • The qml.circuit_drawer.MPLDrawer class provides manual circuit drawing functionality using Matplotlib. While not yet integrated with automatic circuit drawing, this class provides customization and control. (#1484)

    from pennylane.circuit_drawer import MPLDrawer
    
    drawer = MPLDrawer(n_wires=3, n_layers=3)
    
    drawer.label([r"$|\Psi\rangle$", r"$|\theta\rangle$", "aux"])
    
    drawer.box_gate(layer=0, wires=[0, 1, 2], text="Entangling Layers", text_options={'rotation': 'vertical'})
    drawer.box_gate(layer=1, wires=[0, 1], text="U(θ)")
    
    drawer.CNOT(layer=2, wires=[1, 2])
    drawer.measure(layer=3, wires=2)
    
    drawer.fig.suptitle('My Circuit', fontsize='xx-large')
    

  • The slowest tests, more than 1.5 seconds, now have the pytest mark slow, and can be selected or deselected during local execution of tests. (#1633)

  • The device test suite has been expanded to cover more qubit operations and observables. (#1510)

  • The MultiControlledX class now inherits from Operation instead of ControlledQubitUnitary which makes the MultiControlledX gate a non-parameterized gate. (#1557)

  • The utils.sparse_hamiltonian function can now deal with non-integer wire labels, and it throws an error for the edge case of observables that are created from multi-qubit operations. (#1550)

  • Added the matrix attribute to qml.templates.subroutines.GroverOperator (#1553)

  • The tape.to_openqasm() method now has a measure_all argument that specifies whether the serialized OpenQASM script includes computational basis measurements on all of the qubits or just those specified by the tape. (#1559)

  • An error is now raised when no arguments are passed to an observable, to inform that wires have not been supplied. (#1547)

  • The group_observables transform is now differentiable. (#1483)

    For example:

    import jax
    from jax import numpy as jnp
    
    coeffs = jnp.array([1., 2., 3.])
    obs = [PauliX(wires=0), PauliX(wires=1), PauliZ(wires=1)]
    
    def group(coeffs, select=None):
      _, grouped_coeffs = qml.grouping.group_observables(obs, coeffs)
      # in this example, grouped_coeffs is a list of two jax tensors
      # [Array([1., 2.], dtype=float32), Array([3.], dtype=float32)]
      return grouped_coeffs[select]
    
    jac_fn = jax.jacobian(group)
    
    >>> jac_fn(coeffs, select=0)
    [[1. 0. 0.]
    [0. 1. 0.]]
    
    >>> jac_fn(coeffs, select=1)
    [[0., 0., 1.]]
    
  • The tape does not verify any more that all Observables have owners in the annotated queue. (#1505)

    This allows manipulation of Observables inside a tape context. An example is expval(Tensor(qml.PauliX(0), qml.Identity(1)).prune()) which makes the expval an owner of the pruned tensor and its constituent observables, but leaves the original tensor in the queue without an owner.

  • The qml.ResetError is now supported for default.mixed device. (#1541)

  • QNode.diff_method will now reflect which method was selected from diff_method="best". (#1568)

  • QNodes now support diff_method=None. This works the same as interface=None. Such QNodes accept floats, ints, lists and NumPy arrays and return NumPy output but can not be differentiated. (#1585)

  • QNodes now include validation to warn users if a supplied keyword argument is not one of the recognized arguments. (#1591)

Breaking changes

  • The QFT operation has been moved, and is now accessible via pennylane.templates.QFT. (#1548)

  • Specifying shots=None with qml.sample was previously deprecated. From this release onwards, setting shots=None when sampling will raise an error also for default.qubit.jax. (#1629)

  • An error is raised during QNode creation when a user requests backpropagation on a device with finite-shots. (#1588)

  • The class qml.Interferometer is deprecated and will be renamed qml.InterferometerUnitary after one release cycle. (#1546)

  • All optimizers except for Rotosolve and Rotoselect now have a public attribute stepsize. Temporary backward compatibility has been added to support the use of _stepsize for one release cycle. update_stepsize method is deprecated. (#1625)

Bug fixes

  • Fixed a bug with shot vectors and Device base class. (#1666)

  • Fixed a bug where @jax.jit would fail on a QNode that used qml.QubitStateVector. (#1649)

  • Fixed a bug related to an edge case of single-qubit zyz_decomposition when only off-diagonal elements are present. (#1643)

  • MottonenStatepreparation can now be run with a single wire label not in a list. (#1620)

  • Fixed the circuit representation of CY gates to align with CNOT and CZ gates when calling the circuit drawer. (#1504)

  • Dask and CVXPY dependent tests are skipped if those packages are not installed. (#1617)

  • The qml.layer template now works with tensorflow variables. (#1615)

  • Remove QFT from possible operations in default.qubit and default.mixed. (#1600)

  • Fixed a bug when computing expectations of Hamiltonians using TensorFlow. (#1586)

  • Fixed a bug when computing the specs of a circuit with a Hamiltonian observable. (#1533)

Documentation

  • The qml.Identity operation is placed under the sections Qubit observables and CV observables. (#1576)

  • Updated the documentation of qml.grouping, qml.kernels and qml.qaoa modules to present the list of functions first followed by the technical details of the module. (#1581)

  • Recategorized Qubit operations into new and existing categories so that code for each operation is easier to locate. (#1566)

Contributors

This release contains contributions from (in alphabetical order):

Vishnu Ajith, Akash Narayanan B, Thomas Bromley, Olivia Di Matteo, Sahaj Dhamija, Tanya Garg, Anthony Hayes, Theodor Isacsson, Josh Izaac, Prateek Jain, Ankit Khandelwal, Nathan Killoran, Christina Lee, Ian McLean, Johannes Jakob Meyer, Romain Moyard, Lee James O’Riordan, Esteban Payares, Pratul Saini, Maria Schuld, Arshpreet Singh, Jay Soni, Ingrid Strandberg, Antal Száva, Slimane Thabet, David Wierichs, Vincent Wong.

orphan

Release 0.17.0

New features since the last release

Circuit optimization

  • PennyLane can now perform quantum circuit optimization using the top-level transform qml.compile. The compile transform allows you to chain together sequences of tape and quantum function transforms into custom circuit optimization pipelines. (#1475)

    For example, take the following decorated quantum function:

    dev = qml.device('default.qubit', wires=[0, 1, 2])
    
    @qml.qnode(dev)
    @qml.compile()
    def qfunc(x, y, z):
        qml.Hadamard(wires=0)
        qml.Hadamard(wires=1)
        qml.Hadamard(wires=2)
        qml.RZ(z, wires=2)
        qml.CNOT(wires=[2, 1])
        qml.RX(z, wires=0)
        qml.CNOT(wires=[1, 0])
        qml.RX(x, wires=0)
        qml.CNOT(wires=[1, 0])
        qml.RZ(-z, wires=2)
        qml.RX(y, wires=2)
        qml.PauliY(wires=2)
        qml.CZ(wires=[1, 2])
        return qml.expval(qml.PauliZ(wires=0))
    

    The default behaviour of qml.compile is to apply a sequence of three transforms: commute_controlled, cancel_inverses, and then merge_rotations.

    >>> print(qml.draw(qfunc)(0.2, 0.3, 0.4))
     0: ──H───RX(0.6)──────────────────┤ ⟨Z⟩
     1: ──H──╭X────────────────────╭C──┤
     2: ──H──╰C────────RX(0.3)──Y──╰Z──┤
    

    The qml.compile transform is flexible and accepts a custom pipeline of tape and quantum function transforms (you can even write your own!). For example, if we wanted to only push single-qubit gates through controlled gates and cancel adjacent inverses, we could do:

    from pennylane.transforms import commute_controlled, cancel_inverses
    pipeline = [commute_controlled, cancel_inverses]
    
    @qml.qnode(dev)
    @qml.compile(pipeline=pipeline)
    def qfunc(x, y, z):
        qml.Hadamard(wires=0)
        qml.Hadamard(wires=1)
        qml.Hadamard(wires=2)
        qml.RZ(z, wires=2)
        qml.CNOT(wires=[2, 1])
        qml.RX(z, wires=0)
        qml.CNOT(wires=[1, 0])
        qml.RX(x, wires=0)
        qml.CNOT(wires=[1, 0])
        qml.RZ(-z, wires=2)
        qml.RX(y, wires=2)
        qml.PauliY(wires=2)
        qml.CZ(wires=[1, 2])
        return qml.expval(qml.PauliZ(wires=0))
    
    >>> print(qml.draw(qfunc)(0.2, 0.3, 0.4))
     0: ──H───RX(0.4)──RX(0.2)────────────────────────────┤ ⟨Z⟩
     1: ──H──╭X───────────────────────────────────────╭C──┤
     2: ──H──╰C────────RZ(0.4)──RZ(-0.4)──RX(0.3)──Y──╰Z──┤
    

    The following compilation transforms have been added and are also available to use, either independently, or within a qml.compile pipeline:

    • commute_controlled: push commuting single-qubit gates through controlled operations. (#1464)

    • cancel_inverses: removes adjacent pairs of operations that cancel out. (#1455)

    • merge_rotations: combines adjacent rotation gates of the same type into a single gate, including controlled rotations. (#1455)

    • single_qubit_fusion: acts on all sequences of single-qubit operations in a quantum function, and converts each sequence to a single Rot gate. (#1458)

    For more details on qml.compile and the available compilation transforms, see the compilation documentation.

QNodes are even more powerful

  • Computational basis samples directly from the underlying device can now be returned directly from QNodes via qml.sample(). (#1441)

    dev = qml.device("default.qubit", wires=3, shots=5)
    
    @qml.qnode(dev)
    def circuit_1():
        qml.Hadamard(wires=0)
        qml.Hadamard(wires=1)
        return qml.sample()
    
    @qml.qnode(dev)
    def circuit_2():
        qml.Hadamard(wires=0)
        qml.Hadamard(wires=1)
        return qml.sample(wires=[0,2])    # no observable provided and wires specified
    
    >>> print(circuit_1())
    [[1, 0, 0],
     [1, 1, 0],
     [1, 0, 0],
     [0, 0, 0],
     [0, 1, 0]]
    
    >>> print(circuit_2())
    [[1, 0],
     [1, 0],
     [1, 0],
     [0, 0],
     [0, 0]]
    
    >>> print(qml.draw(circuit_2)())
     0: ──H──╭┤ Sample[basis]
     1: ──H──│┤
     2: ─────╰┤ Sample[basis]
    
  • The new qml.apply function can be used to add operations that might have already been instantiated elsewhere to the QNode and other queuing contexts: (#1433)

    op = qml.RX(0.4, wires=0)
    dev = qml.device("default.qubit", wires=2)
    
    @qml.qnode(dev)
    def circuit(x):
        qml.RY(x, wires=0)
        qml.apply(op)
        return qml.expval(qml.PauliZ(0))
    
    >>> print(qml.draw(circuit)(0.6))
    0: ──RY(0.6)──RX(0.4)──┤ ⟨Z⟩
    

    Previously instantiated measurements can also be applied to QNodes.

Device Resource Tracker

  • The new Device Tracker capabilities allows for flexible and versatile tracking of executions, even inside parameter-shift gradients. This functionality will improve the ease of monitoring large batches and remote jobs. (#1355)

    dev = qml.device('default.qubit', wires=1, shots=100)
    
    @qml.qnode(dev, diff_method="parameter-shift")
    def circuit(x):
        qml.RX(x, wires=0)
        return qml.expval(qml.PauliZ(0))
    
    x = np.array(0.1)
    
    with qml.Tracker(circuit.device) as tracker:
        qml.grad(circuit)(x)
    
    >>> tracker.totals
    {'executions': 3, 'shots': 300, 'batches': 1, 'batch_len': 2}
    >>> tracker.history
    {'executions': [1, 1, 1],
     'shots': [100, 100, 100],
     'batches': [1],
     'batch_len': [2]}
    >>> tracker.latest
    {'batches': 1, 'batch_len': 2}
    

    Users can also provide a custom function to the callback keyword that gets called each time the information is updated. This functionality allows users to monitor remote jobs or large parameter-shift batches.

    >>> def shots_info(totals, history, latest):
    ...     print("Total shots: ", totals['shots'])
    >>> with qml.Tracker(circuit.device, callback=shots_info) as tracker:
    ...     qml.grad(circuit)(0.1)
    Total shots:  100
    Total shots:  200
    Total shots:  300
    Total shots:  300
    

Containerization support

  • Docker support for building PennyLane with support for all interfaces (TensorFlow, Torch, and Jax), as well as device plugins and QChem, for GPUs and CPUs, has been added. (#1391)

    The build process using Docker and make requires that the repository source code is cloned or downloaded from GitHub. Visit the the detailed description for an extended list of options.

Improved Hamiltonian simulations

  • Added a sparse Hamiltonian observable and the functionality to support computing its expectation value with default.qubit. (#1398)

    For example, the following QNode returns the expectation value of a sparse Hamiltonian:

    dev = qml.device("default.qubit", wires=2)
    
    @qml.qnode(dev, diff_method="parameter-shift")
    def circuit(param, H):
        qml.PauliX(0)
        qml.SingleExcitation(param, wires=[0, 1])
        return qml.expval(qml.SparseHamiltonian(H, [0, 1]))
    

    We can execute this QNode, passing in a sparse identity matrix:

    >>> print(circuit([0.5], scipy.sparse.eye(4).tocoo()))
    0.9999999999999999
    

    The expectation value of the sparse Hamiltonian is computed directly, which leads to executions that are faster by orders of magnitude. Note that “parameter-shift” is the only differentiation method that is currently supported when the observable is a sparse Hamiltonian.

  • VQE problems can now be intuitively set up by passing the Hamiltonian as an observable. (#1474)

    dev = qml.device("default.qubit", wires=2)
    H = qml.Hamiltonian([1., 2., 3.],  [qml.PauliZ(0), qml.PauliY(0), qml.PauliZ(1)])
    w = qml.init.strong_ent_layers_uniform(1, 2, seed=1967)
    
    @qml.qnode(dev)
    def circuit(w):
        qml.templates.StronglyEntanglingLayers(w, wires=range(2))
        return qml.expval(H)
    
    >>> print(circuit(w))
    -1.5133943637878295
    >>> print(qml.grad(circuit)(w))
    [[[-8.32667268e-17  1.39122955e+00 -9.12462052e-02]
    [ 1.02348685e-16 -7.77143238e-01 -1.74708049e-01]]]
    

    Note that other measurement types like var(H) or sample(H), as well as multiple expectations like expval(H1), expval(H2) are not supported.

  • Added functionality to compute the sparse matrix representation of a qml.Hamiltonian object. (#1394)

New gradients module

  • A new gradients module qml.gradients has been added, which provides differentiable quantum gradient transforms. (#1476) (#1479) (#1486)

    Available quantum gradient transforms include:

    • qml.gradients.finite_diff

    • qml.gradients.param_shift

    • qml.gradients.param_shift_cv

    For example,

    >>> params = np.array([0.3,0.4,0.5], requires_grad=True)
    >>> with qml.tape.JacobianTape() as tape:
    ...     qml.RX(params[0], wires=0)
    ...     qml.RY(params[1], wires=0)
    ...     qml.RX(params[2], wires=0)
    ...     qml.expval(qml.PauliZ(0))
    ...     qml.var(qml.PauliZ(0))
    >>> tape.trainable_params = {0, 1, 2}
    >>> gradient_tapes, fn = qml.gradients.finite_diff(tape)
    >>> res = dev.batch_execute(gradient_tapes)
    >>> fn(res)
    array([[-0.69688381, -0.32648317, -0.68120105],
           [ 0.8788057 ,  0.41171179,  0.85902895]])
    

Even more new operations and templates

  • Grover Diffusion Operator template added. (#1442)

    For example, if we have an oracle that marks the “all ones” state with a negative sign:

    n_wires = 3
    wires = list(range(n_wires))
    
    def oracle():
        qml.Hadamard(wires[-1])
        qml.Toffoli(wires=wires)
        qml.Hadamard(wires[-1])
    

    We can perform Grover’s Search Algorithm:

    dev = qml.device('default.qubit', wires=wires)
    
    @qml.qnode(dev)
    def GroverSearch(num_iterations=1):
        for wire in wires:
            qml.Hadamard(wire)
    
        for _ in range(num_iterations):
            oracle()
            qml.templates.GroverOperator(wires=wires)
    
        return qml.probs(wires)
    

    We can see this circuit yields the marked state with high probability:

    >>> GroverSearch(num_iterations=1)
    tensor([0.03125, 0.03125, 0.03125, 0.03125, 0.03125, 0.03125, 0.03125,
            0.78125], requires_grad=True)
    >>> GroverSearch(num_iterations=2)
    tensor([0.0078125, 0.0078125, 0.0078125, 0.0078125, 0.0078125, 0.0078125,
        0.0078125, 0.9453125], requires_grad=True)
    
  • A decomposition has been added to QubitUnitary that makes the single-qubit case fully differentiable in all interfaces. Furthermore, a quantum function transform, unitary_to_rot(), has been added to decompose all single-qubit instances of QubitUnitary in a quantum circuit. (#1427)

    Instances of QubitUnitary may now be decomposed directly to Rot operations, or RZ operations if the input matrix is diagonal. For example, let

    >>> U = np.array([
        [-0.28829348-0.78829734j,  0.30364367+0.45085995j],
        [ 0.53396245-0.10177564j,  0.76279558-0.35024096j]
    ])
    

    Then, we can compute the decomposition as:

    >>> qml.QubitUnitary.decomposition(U, wires=0)
    [Rot(-0.24209530281458358, 1.1493817777199102, 1.733058145303424, wires=[0])]
    

    We can also apply the transform directly to a quantum function, and compute the gradients of parameters used to construct the unitary matrices.

    def qfunc_with_qubit_unitary(angles):
        z, x = angles[0], angles[1]
    
        Z_mat = np.array([[np.exp(-1j * z / 2), 0.0], [0.0, np.exp(1j * z / 2)]])
    
        c = np.cos(x / 2)
        s = np.sin(x / 2) * 1j
        X_mat = np.array([[c, -s], [-s, c]])
    
        qml.Hadamard(wires="a")
        qml.QubitUnitary(Z_mat, wires="a")
        qml.QubitUnitary(X_mat, wires="b")
        qml.CNOT(wires=["b", "a"])
        return qml.expval(qml.PauliX(wires="a"))
    
    >>> dev = qml.device("default.qubit", wires=["a", "b"])
    >>> transformed_qfunc = qml.transforms.unitary_to_rot(qfunc_with_qubit_unitary)
    >>> transformed_qnode = qml.QNode(transformed_qfunc, dev)
    >>> input = np.array([0.3, 0.4], requires_grad=True)
    >>> transformed_qnode(input)
    tensor(0.95533649, requires_grad=True)
    >>> qml.grad(transformed_qnode)(input)
    array([-0.29552021,  0.        ])
    
  • Ising YY gate functionality added. (#1358)

Improvements

  • The tape does not verify any more that all Observables have owners in the annotated queue. (#1505)

    This allows manipulation of Observables inside a tape context. An example is expval(Tensor(qml.PauliX(0), qml.Identity(1)).prune()) which makes the expval an owner of the pruned tensor and its constituent observables, but leaves the original tensor in the queue without an owner.

  • The step and step_and_cost methods of QNGOptimizer now accept a custom grad_fn keyword argument to use for gradient computations. (#1487)

  • The precision used by default.qubit.jax now matches the float precision indicated by

    from jax.config import config
    config.read('jax_enable_x64')
    

    where True means float64/complex128 and False means float32/complex64. (#1485)

  • The ./pennylane/ops/qubit.py file is broken up into a folder of six separate files. (#1467)

  • Changed to using commas as the separator of wires in the string representation of qml.Hamiltonian objects for multi-qubit terms. (#1465)

  • Changed to using np.object_ instead of np.object as per the NumPy deprecations starting version 1.20. (#1466)

  • Change the order of the covariance matrix and the vector of means internally in default.gaussian. (#1331)

  • Added the id attribute to templates. (#1438)

  • The qml.math module, for framework-agnostic tensor manipulation, has two new functions available: (#1490)

    • qml.math.get_trainable_indices(sequence_of_tensors): returns the indices corresponding to trainable tensors in the input sequence.

    • qml.math.unwrap(sequence_of_tensors): unwraps a sequence of tensor-like objects to NumPy arrays.

    In addition, the behaviour of qml.math.requires_grad has been improved in order to correctly determine trainability during Autograd and JAX backwards passes.

  • A new tape method, tape.unwrap() is added. This method is a context manager; inside the context, the tape’s parameters are unwrapped to NumPy arrays and floats, and the trainable parameter indices are set. (#1491)

    These changes are temporary, and reverted on exiting the context.

    >>> with tf.GradientTape():
    ...     with qml.tape.QuantumTape() as tape:
    ...         qml.RX(tf.Variable(0.1), wires=0)
    ...         qml.RY(tf.constant(0.2), wires=0)
    ...         qml.RZ(tf.Variable(0.3), wires=0)
    ...     with tape.unwrap():
    ...         print("Trainable params:", tape.trainable_params)
    ...         print("Unwrapped params:", tape.get_parameters())
    Trainable params: {0, 2}
    Unwrapped params: [0.1, 0.3]
    >>> print("Original parameters:", tape.get_parameters())
    Original parameters: [<tf.Variable 'Variable:0' shape=() dtype=float32, numpy=0.1>,
      <tf.Variable 'Variable:0' shape=() dtype=float32, numpy=0.3>]
    

    In addition, qml.tape.Unwrap is a context manager that unwraps multiple tapes:

    >>> with qml.tape.Unwrap(tape1, tape2):
    

Breaking changes

  • Removed the deprecated tape methods get_resources and get_depth as they are superseded by the specs tape attribute. (#1522)

  • Specifying shots=None with qml.sample was previously deprecated. From this release onwards, setting shots=None when sampling will raise an error. (#1522)

  • The existing pennylane.collections.apply function is no longer accessible via qml.apply, and needs to be imported directly from the collections package. (#1358)

Bug fixes

  • Fixes a bug in qml.adjoint and qml.ctrl where the adjoint of operations outside of a QNode or a QuantumTape could not be obtained. (#1532)

  • Fixes a bug in GradientDescentOptimizer and NesterovMomentumOptimizer where a cost function with one trainable parameter and non-trainable parameters raised an error. (#1495)

  • Fixed an example in the documentation’s introduction to numpy gradients, where the wires were a non-differentiable argument to the QNode. (#1499)

  • Fixed a bug where the adjoint of qml.QFT when using the qml.adjoint function was not correctly computed. (#1451)

  • Fixed the differentiability of the operationIsingYY for Autograd, Jax and Tensorflow. (#1425)

  • Fixed a bug in the torch interface that prevented gradients from being computed on a GPU. (#1426)

  • Quantum function transforms now preserve the format of the measurement results, so that a single measurement returns a single value rather than an array with a single element. (#1434)

  • Fixed a bug in the parameter-shift Hessian implementation, which resulted in the incorrect Hessian being returned for a cost function that performed post-processing on a vector-valued QNode. (#1436)

  • Fixed a bug in the initialization of QubitUnitary where the size of the matrix was not checked against the number of wires. (#1439)

Documentation

  • Improved Contribution Guide and Pull Requests Guide. (#1461)

  • Examples have been added to clarify use of the continuous-variable FockStateVector operation in the multi-mode case. (#1472)

Contributors

This release contains contributions from (in alphabetical order):

Juan Miguel Arrazola, Olivia Di Matteo, Anthony Hayes, Theodor Isacsson, Josh Izaac, Soran Jahangiri, Nathan Killoran, Arshpreet Singh Khangura, Leonhard Kunczik, Christina Lee, Romain Moyard, Lee James O’Riordan, Ashish Panigrahi, Nahum Sá, Maria Schuld, Jay Soni, Antal Száva, David Wierichs.

orphan

Release 0.16.0

First class support for quantum kernels

  • The new qml.kernels module provides basic functionalities for working with quantum kernels as well as post-processing methods to mitigate sampling errors and device noise: (#1102)

    num_wires = 6
    wires = range(num_wires)
    dev = qml.device('default.qubit', wires=wires)
    
    @qml.qnode(dev)
    def kernel_circuit(x1, x2):
        qml.templates.AngleEmbedding(x1, wires=wires)
        qml.adjoint(qml.templates.AngleEmbedding)(x2, wires=wires)
        return qml.probs(wires)
    
    kernel = lambda x1, x2: kernel_circuit(x1, x2)[0]
    X_train = np.random.random((10, 6))
    X_test = np.random.random((5, 6))
    
    # Create symmetric square kernel matrix (for training)
    K = qml.kernels.square_kernel_matrix(X_train, kernel)
    
    # Compute kernel between test and training data.
    K_test = qml.kernels.kernel_matrix(X_train, X_test, kernel)
    K1 = qml.kernels.mitigate_depolarizing_noise(K, num_wires, method='single')
    

Extract the fourier representation of quantum circuits

  • PennyLane now has a fourier module, which hosts a growing library of methods that help with investigating the Fourier representation of functions implemented by quantum circuits. The Fourier representation can be used to examine and characterize the expressivity of the quantum circuit. (#1160) (#1378)

    For example, one can plot distributions over Fourier series coefficients like this one:

Seamless support for working with the Pauli group

  • Added functionality for constructing and manipulating the Pauli group (#1181).

    The function qml.grouping.pauli_group provides a generator to easily loop over the group, or construct and store it in its entirety. For example, we can construct the single-qubit Pauli group like so:

    >>> from pennylane.grouping import pauli_group
    >>> pauli_group_1_qubit = list(pauli_group(1))
    >>> pauli_group_1_qubit
    [Identity(wires=[0]), PauliZ(wires=[0]), PauliX(wires=[0]), PauliY(wires=[0])]
    

    We can multiply together its members at the level of Pauli words using the pauli_mult and pauli_multi_with_phase functions. This can be done on arbitrarily-labeled wires as well, by defining a wire map.

    >>> from pennylane.grouping import pauli_group, pauli_mult
    >>> wire_map = {'a' : 0, 'b' : 1, 'c' : 2}
    >>> pg = list(pauli_group(3, wire_map=wire_map))
    >>> pg[3]
    PauliZ(wires=['b']) @ PauliZ(wires=['c'])
    >>> pg[55]
    PauliY(wires=['a']) @ PauliY(wires=['b']) @ PauliZ(wires=['c'])
    >>> pauli_mult(pg[3], pg[55], wire_map=wire_map)
    PauliY(wires=['a']) @ PauliX(wires=['b'])
    

    Functions for conversion of Pauli observables to strings (and back), are included.

    >>> from pennylane.grouping import pauli_word_to_string, string_to_pauli_word
    >>> pauli_word_to_string(pg[55], wire_map=wire_map)
    'YYZ'
    >>> string_to_pauli_word('ZXY', wire_map=wire_map)
    PauliZ(wires=['a']) @ PauliX(wires=['b']) @ PauliY(wires=['c'])
    

    Calculation of the matrix representation for arbitrary Paulis and wire maps is now also supported.

    >>> from pennylane.grouping import pauli_word_to_matrix
    >>> wire_map = {'a' : 0, 'b' : 1}
    >>> pauli_word = qml.PauliZ('b')  # corresponds to Pauli 'IZ'
    >>> pauli_word_to_matrix(pauli_word, wire_map=wire_map)
    array([[ 1.,  0.,  0.,  0.],
           [ 0., -1.,  0., -0.],
           [ 0.,  0.,  1.,  0.],
           [ 0., -0.,  0., -1.]])
    

New transforms

  • The qml.specs QNode transform creates a function that returns specifications or details about the QNode, including depth, number of gates, and number of gradient executions required. (#1245)

    For example:

    dev = qml.device('default.qubit', wires=4)
    
    @qml.qnode(dev, diff_method='parameter-shift')
    def circuit(x, y):
        qml.RX(x[0], wires=0)
        qml.Toffoli(wires=(0, 1, 2))
        qml.CRY(x[1], wires=(0, 1))
        qml.Rot(x[2], x[3], y, wires=0)
        return qml.expval(qml.PauliZ(0)), qml.expval(qml.PauliX(1))
    

    We can now use the qml.specs transform to generate a function that returns details and resource information:

    >>> x = np.array([0.05, 0.1, 0.2, 0.3], requires_grad=True)
    >>> y = np.array(0.4, requires_grad=False)
    >>> specs_func = qml.specs(circuit)
    >>> specs_func(x, y)
    {'gate_sizes': defaultdict(int, {1: 2, 3: 1, 2: 1}),
     'gate_types': defaultdict(int, {'RX': 1, 'Toffoli': 1, 'CRY': 1, 'Rot': 1}),
     'num_operations': 4,
     'num_observables': 2,
     'num_diagonalizing_gates': 1,
     'num_used_wires': 3,
     'depth': 4,
     'num_trainable_params': 4,
     'num_parameter_shift_executions': 11,
     'num_device_wires': 4,
     'device_name': 'default.qubit',
     'diff_method': 'parameter-shift'}
    

    The tape methods get_resources and get_depth are superseded by specs and will be deprecated after one release cycle.

  • Adds a decorator @qml.qfunc_transform to easily create a transformation that modifies the behaviour of a quantum function. (#1315)

    For example, consider the following transform, which scales the parameter of all RX gates by \(x \rightarrow \sin(a) \sqrt{x}\), and the parameters of all RY gates by \(y \rightarrow \cos(a * b) y\):

    @qml.qfunc_transform
    def my_transform(tape, a, b):
        for op in tape.operations + tape.measurements:
            if op.name == "RX":
                x = op.parameters[0]
                qml.RX(qml.math.sin(a) * qml.math.sqrt(x), wires=op.wires)
            elif op.name == "RY":
                y = op.parameters[0]
                qml.RX(qml.math.cos(a * b) * y, wires=op.wires)
            else:
                op.queue()
    

    We can now apply this transform to any quantum function:

    dev = qml.device("default.qubit", wires=2)
    
    def ansatz(x):
        qml.Hadamard(wires=0)
        qml.RX(x[0], wires=0)
        qml.RY(x[1], wires=1)
        qml.CNOT(wires=[0, 1])
    
    @qml.qnode(dev)
    def circuit(params, transform_weights):
        qml.RX(0.1, wires=0)
    
        # apply the transform to the ansatz
        my_transform(*transform_weights)(ansatz)(params)
    
        return qml.expval(qml.PauliZ(1))
    

    We can print this QNode to show that the qfunc transform is taking place:

    >>> x = np.array([0.5, 0.3], requires_grad=True)
    >>> transform_weights = np.array([0.1, 0.6], requires_grad=True)
    >>> print(qml.draw(circuit)(x, transform_weights))
     0: ──RX(0.1)────H──RX(0.0706)──╭C──┤
     1: ──RX(0.299)─────────────────╰X──┤ ⟨Z⟩
    

    Evaluating the QNode, as well as the derivative, with respect to the gate parameter and the transform weights:

    >>> circuit(x, transform_weights)
    tensor(0.00672829, requires_grad=True)
    >>> qml.grad(circuit)(x, transform_weights)
    (array([ 0.00671711, -0.00207359]), array([6.69695008e-02, 3.73694364e-06]))
    
  • Adds a hamiltonian_expand tape transform. This takes a tape ending in qml.expval(H), where H is a Hamiltonian, and maps it to a collection of tapes which can be executed and passed into a post-processing function yielding the expectation value. (#1142)

    Example use:

    H = qml.PauliZ(0) + 3 * qml.PauliZ(0) @ qml.PauliX(1)
    
    with qml.tape.QuantumTape() as tape:
        qml.Hadamard(wires=1)
        qml.expval(H)
    
    tapes, fn = qml.transforms.hamiltonian_expand(tape)
    

    We can now evaluate the transformed tapes, and apply the post-processing function:

    >>> dev = qml.device("default.qubit", wires=3)
    >>> res = dev.batch_execute(tapes)
    >>> fn(res)
    3.999999999999999
    
  • The quantum_monte_carlo transform has been added, allowing an input circuit to be transformed into the full quantum Monte Carlo algorithm. (#1316)

    Suppose we want to measure the expectation value of the sine squared function according to a standard normal distribution. We can calculate the expectation value analytically as 0.432332, but we can also estimate using the quantum Monte Carlo algorithm. The first step is to discretize the problem:

    from scipy.stats import norm
    
    m = 5
    M = 2 ** m
    
    xmax = np.pi  # bound to region [-pi, pi]
    xs = np.linspace(-xmax, xmax, M)
    
    probs = np.array([norm().pdf(x) for x in xs])
    probs /= np.sum(probs)
    
    func = lambda i: np.sin(xs[i]) ** 2
    r_rotations = np.array([2 * np.arcsin(np.sqrt(func(i))) for i in range(M)])
    

    The quantum_monte_carlo transform can then be used:

    from pennylane.templates.state_preparations.mottonen import (
        _uniform_rotation_dagger as r_unitary,
    )
    
    n = 6
    N = 2 ** n
    
    a_wires = range(m)
    wires = range(m + 1)
    target_wire = m
    estimation_wires = range(m + 1, n + m + 1)
    
    dev = qml.device("default.qubit", wires=(n + m + 1))
    
    def fn():
        qml.templates.MottonenStatePreparation(np.sqrt(probs), wires=a_wires)
        r_unitary(qml.RY, r_rotations, control_wires=a_wires[::-1], target_wire=target_wire)
    
    @qml.qnode(dev)
    def qmc():
        qml.quantum_monte_carlo(fn, wires, target_wire, estimation_wires)()
        return qml.probs(estimation_wires)
    
    phase_estimated = np.argmax(qmc()[:int(N / 2)]) / N
    

    The estimated value can be retrieved using:

    >>> (1 - np.cos(np.pi * phase_estimated)) / 2
    0.42663476277231915
    

    The resources required to perform the quantum Monte Carlo algorithm can also be inspected using the specs transform.

Extended QAOA module

  • Functionality to support solving the maximum-weighted cycle problem has been added to the qaoa module. (#1207) (#1209) (#1251) (#1213) (#1220) (#1214) (#1283) (#1297) (#1396) (#1403)

    The max_weight_cycle function returns the appropriate cost and mixer Hamiltonians:

    >>> a = np.random.random((3, 3))
    >>> np.fill_diagonal(a, 0)
    >>> g = nx.DiGraph(a)  # create a random directed graph
    >>> cost, mixer, mapping = qml.qaoa.max_weight_cycle(g)
    >>> print(cost)
      (-0.9775906842165344) [Z2]
    + (-0.9027248603361988) [Z3]
    + (-0.8722207409852838) [Z0]
    + (-0.6426184210832898) [Z5]
    + (-0.2832594164291379) [Z1]
    + (-0.0778133996933755) [Z4]
    >>> print(mapping)
    {0: (0, 1), 1: (0, 2), 2: (1, 0), 3: (1, 2), 4: (2, 0), 5: (2, 1)}
    

    Additional functionality can be found in the qml.qaoa.cycle module.

Extended operations and templates

  • Added functionality to compute the sparse matrix representation of a qml.Hamiltonian object. (#1394)

    coeffs = [1, -0.45]
    obs = [qml.PauliZ(0) @ qml.PauliZ(1), qml.PauliY(0) @ qml.PauliZ(1)]
    H = qml.Hamiltonian(coeffs, obs)
    H_sparse = qml.utils.sparse_hamiltonian(H)
    

    The resulting matrix is a sparse matrix in scipy coordinate list (COO) format:

    >>> H_sparse
    <4x4 sparse matrix of type '<class 'numpy.complex128'>'
        with 8 stored elements in COOrdinate format>
    

    The sparse matrix can be converted to an array as:

    >>> H_sparse.toarray()
    array([[ 1.+0.j  ,  0.+0.j  ,  0.+0.45j,  0.+0.j  ],
           [ 0.+0.j  , -1.+0.j  ,  0.+0.j  ,  0.-0.45j],
           [ 0.-0.45j,  0.+0.j  , -1.+0.j  ,  0.+0.j  ],
           [ 0.+0.j  ,  0.+0.45j,  0.+0.j  ,  1.+0.j  ]])
    
  • Adds the new template AllSinglesDoubles to prepare quantum states of molecules using the SingleExcitation and DoubleExcitation operations. The new template reduces significantly the number of operations and the depth of the quantum circuit with respect to the traditional UCCSD unitary. (#1383)

    For example, consider the case of two particles and four qubits. First, we define the Hartree-Fock initial state and generate all possible single and double excitations.

    import pennylane as qml
    from pennylane import numpy as np
    
    electrons = 2
    qubits = 4
    
    hf_state = qml.qchem.hf_state(electrons, qubits)
    singles, doubles = qml.qchem.excitations(electrons, qubits)
    

    Now we can use the template AllSinglesDoubles to define the quantum circuit,

    from pennylane.templates import AllSinglesDoubles
    
    wires = range(qubits)
    
    dev = qml.device('default.qubit', wires=wires)
    
    @qml.qnode(dev)
    def circuit(weights, hf_state, singles, doubles):
        AllSinglesDoubles(weights, wires, hf_state, singles, doubles)
        return qml.expval(qml.PauliZ(0))
    
    params = np.random.normal(0, np.pi, len(singles) + len(doubles))
    

    and execute it:

    >>> circuit(params, hf_state, singles=singles, doubles=doubles)
    tensor(-0.73772194, requires_grad=True)
    
  • Adds QubitCarry and QubitSum operations for basic arithmetic. (#1169)

    The following example adds two 1-bit numbers, returning a 2-bit answer:

    dev = qml.device('default.qubit', wires = 4)
    a = 0
    b = 1
    
    @qml.qnode(dev)
    def circuit():
        qml.BasisState(np.array([a, b]), wires=[1, 2])
        qml.QubitCarry(wires=[0, 1, 2, 3])
        qml.CNOT(wires=[1, 2])
        qml.QubitSum(wires=[0, 1, 2])
        return qml.probs(wires=[3, 2])
    
    probs = circuit()
    bitstrings = tuple(itertools.product([0, 1], repeat = 2))
    indx = np.argwhere(probs == 1).flatten()[0]
    output = bitstrings[indx]
    
    >>> print(output)
    (0, 1)
    
  • Added the qml.Projector observable, which is available on all devices inheriting from the QubitDevice class. (#1356) (#1368)

    Using qml.Projector, we can define the basis state projectors to use when computing expectation values. Let us take for example a circuit that prepares Bell states:

    dev = qml.device("default.qubit", wires=2)
    
    @qml.qnode(dev)
    def circuit(basis_state):
        qml.Hadamard(wires=[0])
        qml.CNOT(wires=[0, 1])
        return qml.expval(qml.Projector(basis_state, wires=[0, 1]))
    

    We can then specify the |00> basis state to construct the |00><00| projector and compute the expectation value:

    >>> basis_state = [0, 0]
    >>> circuit(basis_state)
    tensor(0.5, requires_grad=True)
    

    As expected, we get similar results when specifying the |11> basis state:

    >>> basis_state = [1, 1]
    >>> circuit(basis_state)
    tensor(0.5, requires_grad=True)
    
  • The following new operations have been added:

    • The IsingXX gate qml.IsingXX (#1194)

    • The IsingZZ gate qml.IsingZZ (#1199)

    • The ISWAP gate qml.ISWAP (#1298)

    • The reset error noise channel qml.ResetError (#1321)

Improvements

  • The argnum keyword argument can now be specified for a QNode to define a subset of trainable parameters used to estimate the Jacobian. (#1371)

    For example, consider two trainable parameters and a quantum function:

    dev = qml.device("default.qubit", wires=2)
    
    x = np.array(0.543, requires_grad=True)
    y = np.array(-0.654, requires_grad=True)
    
    def circuit(x,y):
        qml.RX(x, wires=[0])
        qml.RY(y, wires=[1])
        qml.CNOT(wires=[0, 1])
        return qml.expval(qml.PauliZ(0) @ qml.PauliX(1))
    

    When computing the gradient of the QNode, we can specify the trainable parameters to consider by passing the argnum keyword argument:

    >>> qnode1 = qml.QNode(circuit, dev, diff_method="parameter-shift", argnum=[0,1])
    >>> print(qml.grad(qnode1)(x,y))
    (array(0.31434679), array(0.67949903))
    

    Specifying a proper subset of the trainable parameters will estimate the Jacobian:

    >>> qnode2 = qml.QNode(circuit, dev, diff_method="parameter-shift", argnum=[0])
    >>> print(qml.grad(qnode2)(x,y))
    (array(0.31434679), array(0.))
    
  • Allows creating differentiable observables that return custom objects such that the observable is supported by devices. (1291)

    As an example, first we define NewObservable class:

    from pennylane.devices import DefaultQubit
    
    class NewObservable(qml.operation.Observable):
        """NewObservable"""
    
        num_wires = qml.operation.AnyWires
        num_params = 0
        par_domain = None
    
        def diagonalizing_gates(self):
            """Diagonalizing gates"""
            return []
    

    Once we have this new observable class, we define a SpecialObject class that can be used to encode data in an observable and a new device that supports our new observable and returns a SpecialObject as the expectation value (the code is shortened for brevity, the extended example can be found as a test in the previously referenced pull request):

    class SpecialObject:
    
        def __init__(self, val):
            self.val = val
    
        def __mul__(self, other):
            new = SpecialObject(self.val)
            new *= other
            return new
    
        ...
    
    class DeviceSupportingNewObservable(DefaultQubit):
        name = "Device supporting NewObservable"
        short_name = "default.qubit.newobservable"
        observables = DefaultQubit.observables.union({"NewObservable"})
    
        def expval(self, observable, **kwargs):
            if self.shots is None and isinstance(observable, NewObservable):
                val = super().expval(qml.PauliZ(wires=0), **kwargs)
                return SpecialObject(val)
    
            return super().expval(observable, **kwargs)
    

    At this point, we can create a device that will support the differentiation of a NewObservable object:

    dev = DeviceSupportingNewObservable(wires=1, shots=None)
    
    @qml.qnode(dev, diff_method="parameter-shift")
    def qnode(x):
        qml.RY(x, wires=0)
        return qml.expval(NewObservable(wires=0))
    

    We can then compute the jacobian of this object:

    >>> result = qml.jacobian(qnode)(0.2)
    >>> print(result)
    <__main__.SpecialObject object at 0x7fd2c54721f0>
    >>> print(result.item().val)
    -0.19866933079506116
    
  • PennyLane NumPy now includes the random module’s Generator objects, the recommended way of random number generation. This allows for random number generation using a local, rather than global seed. (#1267)

    from pennylane import numpy as np
    
    rng = np.random.default_rng()
    random_mat1 = rng.random((3,2))
    random_mat2 = rng.standard_normal(3, requires_grad=False)
    
  • The performance of adjoint jacobian differentiation was significantly improved as the method now reuses the state computed on the forward pass. This can be turned off to save memory with the Torch and TensorFlow interfaces by passing adjoint_cache=False during QNode creation. (#1341)

  • The Operator (and by inheritance, the Operation and Observable class and their children) now have an id attribute, which can mark an operator in a circuit, for example to identify it on the tape by a tape transform. (#1377)

  • The benchmark module was deleted, since it was outdated and is superseded by the new separate benchmark repository. (#1343)

  • Decompositions in terms of elementary gates has been added for:

    • qml.CSWAP (#1306)

    • qml.SWAP (#1329)

    • qml.SingleExcitation (#1303)

    • qml.SingleExcitationPlus and qml.SingleExcitationMinus (#1278)

    • qml.DoubleExcitation (#1303)

    • qml.Toffoli (#1320)

    • qml.MultiControlledX. (#1287) When controlling on three or more wires, an ancilla register of worker wires is required to support the decomposition.

      ctrl_wires = [f"c{i}" for i in range(5)]
      work_wires = [f"w{i}" for i in range(3)]
      target_wires = ["t0"]
      all_wires = ctrl_wires + work_wires + target_wires
      
      dev = qml.device("default.qubit", wires=all_wires)
      
      with qml.tape.QuantumTape() as tape:
          qml.MultiControlledX(control_wires=ctrl_wires, wires=target_wires, work_wires=work_wires)
      
      >>> tape = tape.expand(depth=1)
      >>> print(tape.draw(wire_order=qml.wires.Wires(all_wires)))
      
       c0: ──────────────╭C──────────────────────╭C──────────┤
       c1: ──────────────├C──────────────────────├C──────────┤
       c2: ──────────╭C──│───╭C──────────────╭C──│───╭C──────┤
       c3: ──────╭C──│───│───│───╭C──────╭C──│───│───│───╭C──┤
       c4: ──╭C──│───│───│───│───│───╭C──│───│───│───│───│───┤
       w0: ──│───│───├C──╰X──├C──│───│───│───├C──╰X──├C──│───┤
       w1: ──│───├C──╰X──────╰X──├C──│───├C──╰X──────╰X──├C──┤
       w2: ──├C──╰X──────────────╰X──├C──╰X──────────────╰X──┤
       t0: ──╰X──────────────────────╰X──────────────────────┤
      
  • Added qml.CPhase as an alias for the existing qml.ControlledPhaseShift operation. (#1319).

  • The Device class now uses caching when mapping wires. (#1270)

  • The Wires class now uses caching for computing its hash. (#1270)

  • Added custom gate application for Toffoli in default.qubit. (#1249)

  • Added validation for noise channel parameters. Invalid noise parameters now raise a ValueError. (#1357)

  • The device test suite now provides test cases for checking gates by comparing expectation values. (#1212)

  • PennyLane’s test suite is now code-formatted using black -l 100. (#1222)

  • PennyLane’s qchem package and tests are now code-formatted using black -l 100. (#1311)

Breaking changes

  • The qml.inv() function is now deprecated with a warning to use the more general qml.adjoint(). (#1325)

  • Removes support for Python 3.6 and adds support for Python 3.9. (#1228)

  • The tape methods get_resources and get_depth are superseded by specs and will be deprecated after one release cycle. (#1245)

  • Using the qml.sample() measurement on devices with shots=None continue to raise a warning with this functionality being fully deprecated and raising an error after one release cycle. (#1079) (#1196)

Bug fixes

  • QNodes now display readable information when in interactive environments or when printed. (#1359).

  • Fixes a bug with qml.math.cast where the MottonenStatePreparation operation expected a float type instead of double. (#1400)

  • Fixes a bug where a copy of qml.ControlledQubitUnitary was non-functional as it did not have all the necessary information. (#1411)

  • Warns when adjoint or reversible differentiation specified or called on a device with finite shots. (#1406)

  • Fixes the differentiability of the operations IsingXX and IsingZZ for Autograd, Jax and Tensorflow. (#1390)

  • Fixes a bug where multiple identical Hamiltonian terms will produce a different result with optimize=True using ExpvalCost. (#1405)

  • Fixes bug where shots=None was not reset when changing shots temporarily in a QNode call like circuit(0.1, shots=3). (#1392)

  • Fixes floating point errors with diff_method="finite-diff" and order=1 when parameters are float32. (#1381)

  • Fixes a bug where qml.ctrl would fail to transform gates that had no control defined and no decomposition defined. (#1376)

  • Copying the JacobianTape now correctly also copies the jacobian_options attribute. This fixes a bug allowing the JAX interface to support adjoint differentiation. (#1349)

  • Fixes drawing QNodes that contain multiple measurements on a single wire. (#1353)

  • Fixes drawing QNodes with no operations. (#1354)

  • Fixes incorrect wires in the decomposition of the ControlledPhaseShift operation. (#1338)

  • Fixed tests for the Permute operation that used a QNode and hence expanded tapes twice instead of once due to QNode tape expansion and an explicit tape expansion call. (#1318).

  • Prevent Hamiltonians that share wires from being multiplied together. (#1273)

  • Fixed a bug where the custom range sequences could not be passed to the StronglyEntanglingLayers template. (#1332)

  • Fixed a bug where qml.sum() and qml.dot() do not support the JAX interface. (#1380)

Documentation

  • Math present in the QubitParamShiftTape class docstring now renders correctly. (#1402)

  • Fix typo in the documentation of qml.StronglyEntanglingLayers. (#1367)

  • Fixed typo in TensorFlow interface documentation (#1312)

  • Fixed typos in the mathematical expressions in documentation of qml.DoubleExcitation. (#1278)

  • Remove unsupported None option from the qml.QNode docstrings. (#1271)

  • Updated the docstring of qml.PolyXP to reference the new location of internal usage. (#1262)

  • Removes occurrences of the deprecated device argument analytic from the documentation. (#1261)

  • Updated PyTorch and TensorFlow interface introductions. (#1333)

  • Updates the quantum chemistry quickstart to reflect recent changes to the qchem module. (#1227)

Contributors

This release contains contributions from (in alphabetical order):

Marius Aglitoiu, Vishnu Ajith, Juan Miguel Arrazola, Thomas Bromley, Jack Ceroni, Alaric Cheng, Miruna Daian, Olivia Di Matteo, Tanya Garg, Christian Gogolin, Alain Delgado Gran, Diego Guala, Anthony Hayes, Ryan Hill, Theodor Isacsson, Josh Izaac, Soran Jahangiri, Pavan Jayasinha, Nathan Killoran, Christina Lee, Ryan Levy, Alberto Maldonado, Johannes Jakob Meyer, Romain Moyard, Ashish Panigrahi, Nahum Sá, Maria Schuld, Brian Shi, Antal Száva, David Wierichs, Vincent Wong.

orphan

Release 0.15.1

Bug fixes

  • Fixes two bugs in the parameter-shift Hessian. (#1260)

    • Fixes a bug where having an unused parameter in the Autograd interface would result in an indexing error during backpropagation.

    • The parameter-shift Hessian only supports the two-term parameter-shift rule currently, so raises an error if asked to differentiate any unsupported gates (such as the controlled rotation gates).

  • A bug which resulted in qml.adjoint() and qml.inv() failing to work with templates has been fixed. (#1243)

  • Deprecation warning instances in PennyLane have been changed to UserWarning, to account for recent changes to how Python warnings are filtered in PEP565. (#1211)

Documentation

  • Updated the order of the parameters to the GaussianState operation to match the way that the PennyLane-SF plugin uses them. (#1255)

Contributors

This release contains contributions from (in alphabetical order):

Thomas Bromley, Olivia Di Matteo, Diego Guala, Anthony Hayes, Ryan Hill, Josh Izaac, Christina Lee, Maria Schuld, Antal Száva.

orphan

Release 0.15.0

New features since last release

Better and more flexible shot control

  • Adds a new optimizer qml.ShotAdaptiveOptimizer, a gradient-descent optimizer where the shot rate is adaptively calculated using the variances of the parameter-shift gradient. (#1139)

    By keeping a running average of the parameter-shift gradient and the variance of the parameter-shift gradient, this optimizer frugally distributes a shot budget across the partial derivatives of each parameter.

    In addition, if computing the expectation value of a Hamiltonian, weighted random sampling can be used to further distribute the shot budget across the local terms from which the Hamiltonian is constructed.

    This optimizer is based on both the iCANS1 and Rosalin shot-adaptive optimizers.

    Once constructed, the cost function can be passed directly to the optimizer’s step method. The attribute opt.total_shots_used can be used to track the number of shots per iteration.

    >>> coeffs = [2, 4, -1, 5, 2]
    >>> obs = [
    ...   qml.PauliX(1),
    ...   qml.PauliZ(1),
    ...   qml.PauliX(0) @ qml.PauliX(1),
    ...   qml.PauliY(0) @ qml.PauliY(1),
    ...   qml.PauliZ(0) @ qml.PauliZ(1)
    ... ]
    >>> H = qml.Hamiltonian(coeffs, obs)
    >>> dev = qml.device("default.qubit", wires=2, shots=100)
    >>> cost = qml.ExpvalCost(qml.templates.StronglyEntanglingLayers, H, dev)
    >>> params = qml.init.strong_ent_layers_uniform(n_layers=2, n_wires=2)
    >>> opt = qml.ShotAdaptiveOptimizer(min_shots=10)
    >>> for i in range(5):
    ...    params = opt.step(cost, params)
    ...    print(f"Step {i}: cost = {cost(params):.2f}, shots_used = {opt.total_shots_used}")
    Step 0: cost = -5.68, shots_used = 240
    Step 1: cost = -2.98, shots_used = 336
    Step 2: cost = -4.97, shots_used = 624
    Step 3: cost = -5.53, shots_used = 1054
    Step 4: cost = -6.50, shots_used = 1798
    
  • Batches of shots can now be specified as a list, allowing measurement statistics to be course-grained with a single QNode evaluation. (#1103)

    >>> shots_list = [5, 10, 1000]
    >>> dev = qml.device("default.qubit", wires=2, shots=shots_list)
    

    When QNodes are executed on this device, a single execution of 1015 shots will be submitted. However, three sets of measurement statistics will be returned; using the first 5 shots, second set of 10 shots, and final 1000 shots, separately.

    For example, executing a circuit with two outputs will lead to a result of shape (3, 2):

    >>> @qml.qnode(dev)
    ... def circuit(x):
    ...     qml.RX(x, wires=0)
    ...     qml.CNOT(wires=[0, 1])
    ...     return qml.expval(qml.PauliZ(0) @ qml.PauliX(1)), qml.expval(qml.PauliZ(0))
    >>> circuit(0.5)
    [[0.33333333 1.        ]
     [0.2        1.        ]
     [0.012      0.868     ]]
    

    This output remains fully differentiable.

  • The number of shots can now be specified on a per-call basis when evaluating a QNode. (#1075).

    For this, the qnode should be called with an additional shots keyword argument:

    >>> dev = qml.device('default.qubit', wires=1, shots=10) # default is 10
    >>> @qml.qnode(dev)
    ... def circuit(a):
    ...     qml.RX(a, wires=0)
    ...     return qml.sample(qml.PauliZ(wires=0))
    >>> circuit(0.8)
    [ 1  1  1 -1 -1  1  1  1  1  1]
    >>> circuit(0.8, shots=3)
    [ 1  1  1]
    >>> circuit(0.8)
    [ 1  1  1 -1 -1  1  1  1  1  1]
    

New differentiable quantum transforms

A new module is available, qml.transforms, which contains differentiable quantum transforms. These are functions that act on QNodes, quantum functions, devices, and tapes, transforming them while remaining fully differentiable.

  • A new adjoint transform has been added. (#1111) (#1135)

    This new method allows users to apply the adjoint of an arbitrary sequence of operations.

    def subroutine(wire):
        qml.RX(0.123, wires=wire)
        qml.RY(0.456, wires=wire)
    
    dev = qml.device('default.qubit', wires=1)
    @qml.qnode(dev)
    def circuit():
        subroutine(0)
        qml.adjoint(subroutine)(0)
        return qml.expval(qml.PauliZ(0))
    

    This creates the following circuit:

    >>> print(qml.draw(circuit)())
    0: --RX(0.123)--RY(0.456)--RY(-0.456)--RX(-0.123)--| <Z>
    

    Directly applying to a gate also works as expected.

    qml.adjoint(qml.RX)(0.123, wires=0) # applies RX(-0.123)
    
  • A new transform qml.ctrl is now available that adds control wires to subroutines. (#1157)

    def my_ansatz(params):
       qml.RX(params[0], wires=0)
       qml.RZ(params[1], wires=1)
    
    # Create a new operation that applies `my_ansatz`
    # controlled by the "2" wire.
    my_ansatz2 = qml.ctrl(my_ansatz, control=2)
    
    @qml.qnode(dev)
    def circuit(params):
        my_ansatz2(params)
        return qml.state()
    

    This is equivalent to:

    @qml.qnode(...)
    def circuit(params):
        qml.CRX(params[0], wires=[2, 0])
        qml.CRZ(params[1], wires=[2, 1])
        return qml.state()
    
  • The qml.transforms.classical_jacobian transform has been added. (#1186)

    This transform returns a function to extract the Jacobian matrix of the classical part of a QNode, allowing the classical dependence between the QNode arguments and the quantum gate arguments to be extracted.

    For example, given the following QNode:

    >>> @qml.qnode(dev)
    ... def circuit(weights):
    ...     qml.RX(weights[0], wires=0)
    ...     qml.RY(weights[0], wires=1)
    ...     qml.RZ(weights[2] ** 2, wires=1)
    ...     return qml.expval(qml.PauliZ(0))
    

    We can use this transform to extract the relationship \(f: \mathbb{R}^n \rightarrow\mathbb{R}^m\) between the input QNode arguments \(w\) and the gate arguments \(g\), for a given value of the QNode arguments:

    >>> cjac_fn = qml.transforms.classical_jacobian(circuit)
    >>> weights = np.array([1., 1., 1.], requires_grad=True)
    >>> cjac = cjac_fn(weights)
    >>> print(cjac)
    [[1. 0. 0.]
     [1. 0. 0.]
     [0. 0. 2.]]
    

    The returned Jacobian has rows corresponding to gate arguments, and columns corresponding to QNode arguments; that is, \(J_{ij} = \frac{\partial}{\partial g_i} f(w_j)\).

More operations and templates

  • Added the SingleExcitation two-qubit operation, which is useful for quantum chemistry applications. (#1121)

    It can be used to perform an SO(2) rotation in the subspace spanned by the states \(|01\rangle\) and \(|10\rangle\). For example, the following circuit performs the transformation \(|10\rangle \rightarrow \cos(\phi/2)|10\rangle - \sin(\phi/2)|01\rangle\):

    dev = qml.device('default.qubit', wires=2)
    
    @qml.qnode(dev)
    def circuit(phi):
        qml.PauliX(wires=0)
        qml.SingleExcitation(phi, wires=[0, 1])
    

    The SingleExcitation operation supports analytic gradients on hardware using only four expectation value calculations, following results from Kottmann et al.

  • Added the DoubleExcitation four-qubit operation, which is useful for quantum chemistry applications. (#1123)

    It can be used to perform an SO(2) rotation in the subspace spanned by the states \(|1100\rangle\) and \(|0011\rangle\). For example, the following circuit performs the transformation \(|1100\rangle\rightarrow \cos(\phi/2)|1100\rangle - \sin(\phi/2)|0011\rangle\):

    dev = qml.device('default.qubit', wires=2)
    
    @qml.qnode(dev)
    def circuit(phi):
        qml.PauliX(wires=0)
        qml.PauliX(wires=1)
        qml.DoubleExcitation(phi, wires=[0, 1, 2, 3])
    

    The DoubleExcitation operation supports analytic gradients on hardware using only four expectation value calculations, following results from Kottmann et al..

  • Added the QuantumMonteCarlo template for performing quantum Monte Carlo estimation of an expectation value on simulator. (#1130)

    The following example shows how the expectation value of sine squared over a standard normal distribution can be approximated:

    from scipy.stats import norm
    
    m = 5
    M = 2 ** m
    n = 10
    N = 2 ** n
    target_wires = range(m + 1)
    estimation_wires = range(m + 1, n + m + 1)
    
    xmax = np.pi  # bound to region [-pi, pi]
    xs = np.linspace(-xmax, xmax, M)
    
    probs = np.array([norm().pdf(x) for x in xs])
    probs /= np.sum(probs)
    
    func = lambda i: np.sin(xs[i]) ** 2
    
    dev = qml.device("default.qubit", wires=(n + m + 1))
    
    @qml.qnode(dev)
    def circuit():
        qml.templates.QuantumMonteCarlo(
            probs,
            func,
            target_wires=target_wires,
            estimation_wires=estimation_wires,
        )
        return qml.probs(estimation_wires)
    
    phase_estimated = np.argmax(circuit()[:int(N / 2)]) / N
    expectation_estimated = (1 - np.cos(np.pi * phase_estimated)) / 2
    
  • Added the QuantumPhaseEstimation template for performing quantum phase estimation for an input unitary matrix. (#1095)

    Consider the matrix corresponding to a rotation from an RX gate:

    >>> phase = 5
    >>> target_wires = [0]
    >>> unitary = qml.RX(phase, wires=0).matrix
    

    The phase parameter can be estimated using QuantumPhaseEstimation. For example, using five phase-estimation qubits:

    n_estimation_wires = 5
    estimation_wires = range(1, n_estimation_wires + 1)
    
    dev = qml.device("default.qubit", wires=n_estimation_wires + 1)
    
    @qml.qnode(dev)
    def circuit():
        # Start in the |+> eigenstate of the unitary
        qml.Hadamard(wires=target_wires)
    
        QuantumPhaseEstimation(
            unitary,
            target_wires=target_wires,
            estimation_wires=estimation_wires,
        )
    
        return qml.probs(estimation_wires)
    
    phase_estimated = np.argmax(circuit()) / 2 ** n_estimation_wires
    
    # Need to rescale phase due to convention of RX gate
    phase_estimated = 4 * np.pi * (1 - phase)
    
  • Added the ControlledPhaseShift gate as well as the QFT operation for applying quantum Fourier transforms. (#1064)

    @qml.qnode(dev)
    def circuit_qft(basis_state):
        qml.BasisState(basis_state, wires=range(3))
        qml.templates.QFT(wires=range(3))
        return qml.state()
    
  • Added the ControlledQubitUnitary operation. This enables implementation of multi-qubit gates with a variable number of control qubits. It is also possible to specify a different state for the control qubits using the control_values argument (also known as a mixed-polarity multi-controlled operation). (#1069) (#1104)

    For example, we can create a multi-controlled T gate using:

    T = qml.T._matrix()
    qml.ControlledQubitUnitary(T, control_wires=[0, 1, 3], wires=2, control_values="110")
    

    Here, the T gate will be applied to wire 2 if control wires 0 and 1 are in state 1, and control wire 3 is in state 0. If no value is passed to control_values, the gate will be applied if all control wires are in the 1 state.

  • Added MultiControlledX for multi-controlled NOT gates. This is a special case of ControlledQubitUnitary that applies a Pauli X gate conditioned on the state of an arbitrary number of control qubits. (#1104)

Support for higher-order derivatives on hardware

  • Computing second derivatives and Hessians of QNodes is now supported with the parameter-shift differentiation method, on all machine learning interfaces. (#1130) (#1129) (#1110)

    Hessians are computed using the parameter-shift rule, and can be evaluated on both hardware and simulator devices.

    dev = qml.device('default.qubit', wires=1)
    
    @qml.qnode(dev, diff_method="parameter-shift")
    def circuit(p):
        qml.RY(p[0], wires=0)
        qml.RX(p[1], wires=0)
        return qml.expval(qml.PauliZ(0))
    
    x = np.array([1.0, 2.0], requires_grad=True)
    
    >>> hessian_fn = qml.jacobian(qml.grad(circuit))
    >>> hessian_fn(x)
    [[0.2248451 0.7651474]
     [0.7651474 0.2248451]]
    
  • Added the function finite_diff() to compute finite-difference approximations to the gradient and the second-order derivatives of arbitrary callable functions. (#1090)

    This is useful to compute the derivative of parametrized pennylane.Hamiltonian observables with respect to their parameters.

    For example, in quantum chemistry simulations it can be used to evaluate the derivatives of the electronic Hamiltonian with respect to the nuclear coordinates:

    >>> def H(x):
    ...    return qml.qchem.molecular_hamiltonian(['H', 'H'], x)[0]
    >>> x = np.array([0., 0., -0.66140414, 0., 0., 0.66140414])
    >>> grad_fn = qml.finite_diff(H, N=1)
    >>> grad = grad_fn(x)
    >>> deriv2_fn = qml.finite_diff(H, N=2, idx=[0, 1])
    >>> deriv2_fn(x)
    
  • The JAX interface now supports all devices, including hardware devices, via the parameter-shift differentiation method. (#1076)

    For example, using the JAX interface with Cirq:

    dev = qml.device('cirq.simulator', wires=1)
    @qml.qnode(dev, interface="jax", diff_method="parameter-shift")
    def circuit(x):
        qml.RX(x[1], wires=0)
        qml.Rot(x[0], x[1], x[2], wires=0)
        return qml.expval(qml.PauliZ(0))
    weights = jnp.array([0.2, 0.5, 0.1])
    print(circuit(weights))
    

    Currently, when used with the parameter-shift differentiation method, only a single returned expectation value or variance is supported. Multiple expectations/variances, as well as probability and state returns, are not currently allowed.

Improvements

   dev = qml.device("default.qubit", wires=2)

   inputstate = [np.sqrt(0.2), np.sqrt(0.3), np.sqrt(0.4), np.sqrt(0.1)]

   @qml.qnode(dev)
   def circuit():
       mottonen.MottonenStatePreparation(inputstate,wires=[0, 1])
       return qml.expval(qml.PauliZ(0))

Previously returned:
   >>> print(qml.draw(circuit)())
   0: ──RY(1.57)──╭C─────────────╭C──╭C──╭C──┤ ⟨Z⟩
   1: ──RY(1.35)──╰X──RY(0.422)──╰X──╰X──╰X──┤

In this release, it now returns:
>>> print(qml.draw(circuit)())
0: ──RY(1.57)──╭C─────────────╭C──┤ ⟨Z⟩
1: ──RY(1.35)──╰X──RY(0.422)──╰X──┤
  • The templates are now classes inheriting from Operation, and define the ansatz in their expand() method. This change does not affect the user interface. (#1138) (#1156) (#1163) (#1192)

    For convenience, some templates have a new method that returns the expected shape of the trainable parameter tensor, which can be used to create random tensors.

    shape = qml.templates.BasicEntanglerLayers.shape(n_layers=2, n_wires=4)
    weights = np.random.random(shape)
    qml.templates.BasicEntanglerLayers(weights, wires=range(4))
    
  • QubitUnitary now validates to ensure the input matrix is two dimensional. (#1128)

  • Most layers in Pytorch or Keras accept arbitrary dimension inputs, where each dimension barring the last (in the case where the actual weight function of the layer operates on one-dimensional vectors) is broadcast over. This is now also supported by KerasLayer and TorchLayer. (#1062).

    Example use:

    dev = qml.device("default.qubit", wires=4)
    x = tf.ones((5, 4, 4))
    
    @qml.qnode(dev)
    def layer(weights, inputs):
        qml.templates.AngleEmbedding(inputs, wires=range(4))
        qml.templates.StronglyEntanglingLayers(weights, wires=range(4))
        return [qml.expval(qml.PauliZ(i)) for i in range(4)]
    
    qlayer = qml.qnn.KerasLayer(layer, {"weights": (4, 4, 3)}, output_dim=4)
    out = qlayer(x)
    

    The output tensor has the following shape:

    >>> out.shape
    (5, 4, 4)
    
  • If only one argument to the function qml.grad has the requires_grad attribute set to True, then the returned gradient will be a NumPy array, rather than a tuple of length 1. (#1067) (#1081)

  • An improvement has been made to how QubitDevice generates and post-processess samples, allowing QNode measurement statistics to work on devices with more than 32 qubits. (#1088)

  • Due to the addition of density_matrix() as a return type from a QNode, tuples are now supported by the output_dim parameter in qnn.KerasLayer. (#1070)

  • Two new utility methods are provided for working with quantum tapes. (#1175)

    • qml.tape.get_active_tape() gets the currently recording tape.

    • tape.stop_recording() is a context manager that temporarily stops the currently recording tape from recording additional tapes or quantum operations.

    For example:

    >>> with qml.tape.QuantumTape():
    ...     qml.RX(0, wires=0)
    ...     current_tape = qml.tape.get_active_tape()
    ...     with current_tape.stop_recording():
    ...         qml.RY(1.0, wires=1)
    ...     qml.RZ(2, wires=1)
    >>> current_tape.operations
    [RX(0, wires=[0]), RZ(2, wires=[1])]
    
  • When printing qml.Hamiltonian objects, the terms are sorted by number of wires followed by coefficients. (#981)

  • Adds qml.math.conj to the PennyLane math module. (#1143)

    This new method will do elementwise conjugation to the given tensor-like object, correctly dispatching to the required tensor-manipulation framework to preserve differentiability.

    >>> a = np.array([1.0 + 2.0j])
    >>> qml.math.conj(a)
    array([1.0 - 2.0j])
    
  • The four-term parameter-shift rule, as used by the controlled rotation operations, has been updated to use coefficients that minimize the variance as per https://arxiv.org/abs/2104.05695. (#1206)

  • A new transform qml.transforms.invisible has been added, to make it easier to transform QNodes. (#1175)

Breaking changes

  • Devices do not have an analytic argument or attribute anymore. Instead, shots is the source of truth for whether a simulator estimates return values from a finite number of shots, or whether it returns analytic results (shots=None). (#1079) (#1196)

    dev_analytic = qml.device('default.qubit', wires=1, shots=None)
    dev_finite_shots = qml.device('default.qubit', wires=1, shots=1000)
    
    def circuit():
        qml.Hadamard(wires=0)
        return qml.expval(qml.PauliZ(wires=0))
    
    circuit_analytic = qml.QNode(circuit, dev_analytic)
    circuit_finite_shots = qml.QNode(circuit, dev_finite_shots)
    

    Devices with shots=None return deterministic, exact results:

    >>> circuit_analytic()
    0.0
    >>> circuit_analytic()
    0.0
    

    Devices with shots > 0 return stochastic results estimated from samples in each run:

    >>> circuit_finite_shots()
    -0.062
    >>> circuit_finite_shots()
    0.034
    

    The qml.sample() measurement can only be used on devices on which the number of shots is set explicitly.

  • If creating a QNode from a quantum function with an argument named shots, a UserWarning is raised, warning the user that this is a reserved argument to change the number of shots on a per-call basis. (#1075)

  • For devices inheriting from QubitDevice, the methods expval, var, sample accept two new keyword arguments — shot_range and bin_size. (#1103)

    These new arguments allow for the statistics to be performed on only a subset of device samples. This finer level of control is accessible from the main UI by instantiating a device with a batch of shots.

    For example, consider the following device:

    >>> dev = qml.device("my_device", shots=[5, (10, 3), 100])
    

    This device will execute QNodes using 135 shots, however measurement statistics will be course grained across these 135 shots:

    • All measurement statistics will first be computed using the first 5 shots — that is, shots_range=[0, 5], bin_size=5.

    • Next, the tuple (10, 3) indicates 10 shots, repeated 3 times. This will use shot_range=[5, 35], performing the expectation value in bins of size 10 (bin_size=10).

    • Finally, we repeat the measurement statistics for the final 100 shots, shot_range=[35, 135], bin_size=100.

  • The old PennyLane core has been removed, including the following modules: (#1100)

    • pennylane.variables

    • pennylane.qnodes

    As part of this change, the location of the new core within the Python module has been moved:

    • Moves pennylane.tape.interfacespennylane.interfaces

    • Merges pennylane.CircuitGraph and pennylane.TapeCircuitGraphpennylane.CircuitGraph

    • Merges pennylane.OperationRecorder and pennylane.TapeOperationRecorder

    • pennylane.tape.operation_recorder

    • Merges pennylane.measure and pennylane.tape.measurepennylane.measure

    • Merges pennylane.operation and pennylane.tape.operationpennylane.operation

    • Merges pennylane._queuing and pennylane.tape.queuingpennylane.queuing

    This has no affect on import location.

    In addition,

    • All tape-mode functions have been removed (qml.enable_tape(), qml.tape_mode_active()),

    • All tape fixtures have been deleted,

    • Tests specifically for non-tape mode have been deleted.

  • The device test suite no longer accepts the analytic keyword. (#1216)

Bug fixes

  • Fixes a bug where using the circuit drawer with a ControlledQubitUnitary operation raised an error. (#1174)

  • Fixes a bug and a test where the QuantumTape.is_sampled attribute was not being updated. (#1126)

  • Fixes a bug where BasisEmbedding would not accept inputs whose bits are all ones or all zeros. (#1114)

  • The ExpvalCost class raises an error if instantiated with non-expectation measurement statistics. (#1106)

  • Fixes a bug where decompositions would reset the differentiation method of a QNode. (#1117)

  • Fixes a bug where the second-order CV parameter-shift rule would error if attempting to compute the gradient of a QNode with more than one second-order observable. (#1197)

  • Fixes a bug where repeated Torch interface applications after expansion caused an error. (#1223)

  • Sampling works correctly with batches of shots specified as a list. (#1232)

Documentation

  • Updated the diagram used in the Architectural overview page of the Development guide such that it doesn’t mention Variables. (#1235)

  • Typos addressed in templates documentation. (#1094)

  • Upgraded the documentation to use Sphinx 3.5.3 and the new m2r2 package. (#1186)

  • Added flaky as dependency for running tests in the documentation. (#1113)

Contributors

This release contains contributions from (in alphabetical order):

Shahnawaz Ahmed, Juan Miguel Arrazola, Thomas Bromley, Olivia Di Matteo, Alain Delgado Gran, Kyle Godbey, Diego Guala, Theodor Isacsson, Josh Izaac, Soran Jahangiri, Nathan Killoran, Christina Lee, Daniel Polatajko, Chase Roberts, Sankalp Sanand, Pritish Sehzpaul, Maria Schuld, Antal Száva, David Wierichs.

orphan

Release 0.14.1

Bug fixes

  • Fixes a testing bug where tests that required JAX would fail if JAX was not installed. The tests will now instead be skipped if JAX can not be imported. (#1066)

  • Fixes a bug where inverse operations could not be differentiated using backpropagation on default.qubit. (#1072)

  • The QNode has a new keyword argument, max_expansion, that determines the maximum number of times the internal circuit should be expanded when executed on a device. In addition, the default number of max expansions has been increased from 2 to 10, allowing devices that require more than two operator decompositions to be supported. (#1074)

  • Fixes a bug where Hamiltonian objects created with non-list arguments raised an error for arithmetic operations. (#1082)

  • Fixes a bug where Hamiltonian objects with no coefficients or operations would return a faulty result when used with ExpvalCost. (#1082)

Documentation

  • Updates mentions of generate_hamiltonian to molecular_hamiltonian in the docstrings of the ExpvalCost and Hamiltonian classes. (#1077)

Contributors

This release contains contributions from (in alphabetical order):

Thomas Bromley, Josh Izaac, Antal Száva.

orphan

Release 0.14.0

New features since last release

Perform quantum machine learning with JAX

  • QNodes created with default.qubit now support a JAX interface, allowing JAX to be used to create, differentiate, and optimize hybrid quantum-classical models. (#947)

    This is supported internally via a new default.qubit.jax device. This device runs end to end in JAX, meaning that it supports all of the awesome JAX transformations (jax.vmap, jax.jit, jax.hessian, etc).

    Here is an example of how to use the new JAX interface:

    dev = qml.device("default.qubit", wires=1)
    @qml.qnode(dev, interface="jax", diff_method="backprop")
    def circuit(x):
        qml.RX(x[1], wires=0)
        qml.Rot(x[0], x[1], x[2], wires=0)
        return qml.expval(qml.PauliZ(0))
    
    weights = jnp.array([0.2, 0.5, 0.1])
    grad_fn = jax.grad(circuit)
    print(grad_fn(weights))
    

    Currently, only diff_method="backprop" is supported, with plans to support more in the future.

New, faster, quantum gradient methods

  • A new differentiation method has been added for use with simulators. The "adjoint" method operates after a forward pass by iteratively applying inverse gates to scan backwards through the circuit. (#1032)

    This method is similar to the reversible method, but has a lower time overhead and a similar memory overhead. It follows the approach provided by Jones and Gacon. This method is only compatible with certain statevector-based devices such as default.qubit.

    Example use:

    import pennylane as qml
    
    wires = 1
    device = qml.device("default.qubit", wires=wires)
    
    @qml.qnode(device, diff_method="adjoint")
    def f(params):
        qml.RX(0.1, wires=0)
        qml.Rot(*params, wires=0)
        qml.RX(-0.3, wires=0)
        return qml.expval(qml.PauliZ(0))
    
    params = [0.1, 0.2, 0.3]
    qml.grad(f)(params)
    
  • The default logic for choosing the ‘best’ differentiation method has been altered to improve performance. (#1008)

    • If the quantum device provides its own gradient, this is now the preferred differentiation method.

    • If the quantum device natively supports classical backpropagation, this is now preferred over the parameter-shift rule.

      This will lead to marked speed improvement during optimization when using default.qubit, with a sight penalty on the forward-pass evaluation.

    More details are available below in the ‘Improvements’ section for plugin developers.

  • PennyLane now supports analytical quantum gradients for noisy channels, in addition to its existing support for unitary operations. The noisy channels BitFlip, PhaseFlip, and DepolarizingChannel all support analytic gradients out of the box. (#968)

  • A method has been added for calculating the Hessian of quantum circuits using the second-order parameter shift formula. (#961)

    The following example shows the calculation of the Hessian:

    n_wires = 5
    weights = [2.73943676, 0.16289932, 3.4536312, 2.73521126, 2.6412488]
    
    dev = qml.device("default.qubit", wires=n_wires)
    
    with qml.tape.QubitParamShiftTape() as tape:
        for i in range(n_wires):
            qml.RX(weights[i], wires=i)
    
        qml.CNOT(wires=[0, 1])
        qml.CNOT(wires=[2, 1])
        qml.CNOT(wires=[3, 1])
        qml.CNOT(wires=[4, 3])
    
        qml.expval(qml.PauliZ(1))
    
    print(tape.hessian(dev))
    

    The Hessian is not yet supported via classical machine learning interfaces, but will be added in a future release.

More operations and templates

  • Two new error channels, BitFlip and PhaseFlip have been added. (#954)

    They can be used in the same manner as existing error channels:

    dev = qml.device("default.mixed", wires=2)
    
    @qml.qnode(dev)
    def circuit():
        qml.RX(0.3, wires=0)
        qml.RY(0.5, wires=1)
        qml.BitFlip(0.01, wires=0)
        qml.PhaseFlip(0.01, wires=1)
        return qml.expval(qml.PauliZ(0))
    
  • Apply permutations to wires using the Permute subroutine. (#952)

    import pennylane as qml
    dev = qml.device('default.qubit', wires=5)
    
    @qml.qnode(dev)
    def apply_perm():
        # Send contents of wire 4 to wire 0, of wire 2 to wire 1, etc.
        qml.templates.Permute([4, 2, 0, 1, 3], wires=dev.wires)
        return qml.expval(qml.PauliZ(0))
    

QNode transformations

  • The qml.metric_tensor function transforms a QNode to produce the Fubini-Study metric tensor with full autodifferentiation support—even on hardware. (#1014)

    Consider the following QNode:

    dev = qml.device("default.qubit", wires=3)
    
    @qml.qnode(dev, interface="autograd")
    def circuit(weights):
        # layer 1
        qml.RX(weights[0, 0], wires=0)
        qml.RX(weights[0, 1], wires=1)
    
        qml.CNOT(wires=[0, 1])
        qml.CNOT(wires=[1, 2])
    
        # layer 2
        qml.RZ(weights[1, 0], wires=0)
        qml.RZ(weights[1, 1], wires=2)
    
        qml.CNOT(wires=[0, 1])
        qml.CNOT(wires=[1, 2])
        return qml.expval(qml.PauliZ(0) @ qml.PauliZ(1)), qml.expval(qml.PauliY(2))
    

    We can use the metric_tensor function to generate a new function, that returns the metric tensor of this QNode:

    >>> met_fn = qml.metric_tensor(circuit)
    >>> weights = np.array([[0.1, 0.2, 0.3], [0.4, 0.5, 0.6]], requires_grad=True)
    >>> met_fn(weights)
    tensor([[0.25  , 0.    , 0.    , 0.    ],
            [0.    , 0.25  , 0.    , 0.    ],
            [0.    , 0.    , 0.0025, 0.0024],
            [0.    , 0.    , 0.0024, 0.0123]], requires_grad=True)
    

    The returned metric tensor is also fully differentiable, in all interfaces. For example, differentiating the (3, 2) element:

    >>> grad_fn = qml.grad(lambda x: met_fn(x)[3, 2])
    >>> grad_fn(weights)
    array([[ 0.04867729, -0.00049502,  0.        ],
           [ 0.        ,  0.        ,  0.        ]])
    

    Differentiation is also supported using Torch, Jax, and TensorFlow.

  • Adds the new function qml.math.cov_matrix(). This function accepts a list of commuting observables, and the probability distribution in the shared observable eigenbasis after the application of an ansatz. It uses these to construct the covariance matrix in a framework independent manner, such that the output covariance matrix is autodifferentiable. (#1012)

    For example, consider the following ansatz and observable list:

    obs_list = [qml.PauliX(0) @ qml.PauliZ(1), qml.PauliY(2)]
    ansatz = qml.templates.StronglyEntanglingLayers
    

    We can construct a QNode to output the probability distribution in the shared eigenbasis of the observables:

    dev = qml.device("default.qubit", wires=3)
    
    @qml.qnode(dev, interface="autograd")
    def circuit(weights):
        ansatz(weights, wires=[0, 1, 2])
        # rotate into the basis of the observables
        for o in obs_list:
            o.diagonalizing_gates()
        return qml.probs(wires=[0, 1, 2])
    

    We can now compute the covariance matrix:

    >>> weights = qml.init.strong_ent_layers_normal(n_layers=2, n_wires=3)
    >>> cov = qml.math.cov_matrix(circuit(weights), obs_list)
    >>> cov
    array([[0.98707611, 0.03665537],
           [0.03665537, 0.99998377]])
    

    Autodifferentiation is fully supported using all interfaces:

    >>> cost_fn = lambda weights: qml.math.cov_matrix(circuit(weights), obs_list)[0, 1]
    >>> qml.grad(cost_fn)(weights)[0]
    array([[[ 4.94240914e-17, -2.33786398e-01, -1.54193959e-01],
            [-3.05414996e-17,  8.40072236e-04,  5.57884080e-04],
            [ 3.01859411e-17,  8.60411436e-03,  6.15745204e-04]],
    
           [[ 6.80309533e-04, -1.23162742e-03,  1.08729813e-03],
            [-1.53863193e-01, -1.38700657e-02, -1.36243323e-01],
            [-1.54665054e-01, -1.89018172e-02, -1.56415558e-01]]])
    
  • A new qml.draw function is available, allowing QNodes to be easily drawn without execution by providing example input. (#962)

    @qml.qnode(dev)
    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))
    

    The QNode circuit structure may depend on the input arguments; this is taken into account by passing example QNode arguments to the qml.draw() drawing function:

    >>> drawer = qml.draw(circuit)
    >>> result = drawer(a=2.3, w=[1.2, 3.2, 0.7])
    >>> print(result)
    0: ──H──╭C────────────────────────────╭C─────────╭┤ ⟨Z ⊗ Z⟩
    1: ─────╰RX(2.3)──Rot(1.2, 3.2, 0.7)──╰RX(-2.3)──╰┤ ⟨Z ⊗ Z⟩
    

A faster, leaner, and more flexible core

  • The new core of PennyLane, rewritten from the ground up and developed over the last few release cycles, has achieved feature parity and has been made the new default in PennyLane v0.14. The old core has been marked as deprecated, and will be removed in an upcoming release. (#1046) (#1040) (#1034) (#1035) (#1027) (#1026) (#1021) (#1054) (#1049)

    While high-level PennyLane code and tutorials remain unchanged, the new core provides several advantages and improvements:

    • Faster and more optimized: The new core provides various performance optimizations, reducing pre- and post-processing overhead, and reduces the number of quantum evaluations in certain cases.

    • Support for in-QNode classical processing: this allows for differentiable classical processing within the QNode.

      dev = qml.device("default.qubit", wires=1)
      
      @qml.qnode(dev, interface="tf")
      def circuit(p):
          qml.RX(tf.sin(p[0])**2 + p[1], wires=0)
          return qml.expval(qml.PauliZ(0))
      

      The classical processing functions used within the QNode must match the QNode interface. Here, we use TensorFlow:

      >>> params = tf.Variable([0.5, 0.1], dtype=tf.float64)
      >>> with tf.GradientTape() as tape:
      ...     res = circuit(params)
      >>> grad = tape.gradient(res, params)
      >>> print(res)
      tf.Tensor(0.9460913127754935, shape=(), dtype=float64)
      >>> print(grad)
      tf.Tensor([-0.27255248 -0.32390003], shape=(2,), dtype=float64)
      

      As a result of this change, quantum decompositions that require classical processing are fully supported and end-to-end differentiable in tape mode.

    • No more Variable wrapping: QNode arguments no longer become Variable objects within the QNode.

      dev = qml.device("default.qubit", wires=1)
      
      @qml.qnode(dev)
      def circuit(x):
          print("Parameter value:", x)
          qml.RX(x, wires=0)
          return qml.expval(qml.PauliZ(0))
      

      Internal QNode parameters can be easily inspected, printed, and manipulated:

      >>> circuit(0.5)
      Parameter value: 0.5
      tensor(0.87758256, requires_grad=True)
      
    • Less restrictive QNode signatures: There is no longer any restriction on the QNode signature; the QNode can be defined and called following the same rules as standard Python functions.

      For example, the following QNode uses positional, named, and variable keyword arguments:

      x = torch.tensor(0.1, requires_grad=True)
      y = torch.tensor([0.2, 0.3], requires_grad=True)
      z = torch.tensor(0.4, requires_grad=True)
      
      @qml.qnode(dev, interface="torch")
      def circuit(p1, p2=y, **kwargs):
          qml.RX(p1, wires=0)
          qml.RY(p2[0] * p2[1], wires=0)
          qml.RX(kwargs["p3"], wires=0)
          return qml.var(qml.PauliZ(0))
      

      When we call the QNode, we may pass the arguments by name even if defined positionally; any argument not provided will use the default value.

      >>> res = circuit(p1=x, p3=z)
      >>> print(res)
      tensor(0.2327, dtype=torch.float64, grad_fn=<SelectBackward>)
      >>> res.backward()
      >>> print(x.grad, y.grad, z.grad)
      tensor(0.8396) tensor([0.0289, 0.0193]) tensor(0.8387)
      

      This extends to the qnn module, where KerasLayer and TorchLayer modules can be created from QNodes with unrestricted signatures.

    • Smarter measurements: QNodes can now measure wires more than once, as long as all observables are commuting:

      @qml.qnode(dev)
      def circuit(x):
          qml.RX(x, wires=0)
          return [
              qml.expval(qml.PauliZ(0)),
              qml.expval(qml.PauliZ(0) @ qml.PauliZ(1))
          ]
      

      Further, the qml.ExpvalCost() function allows for optimizing measurements to reduce the number of quantum evaluations required.

    With the new PennyLane core, there are a few small breaking changes, detailed below in the ‘Breaking Changes’ section.

Improvements

  • The built-in PennyLane optimizers allow more flexible cost functions. The cost function passed to most optimizers may accept any combination of trainable arguments, non-trainable arguments, and keyword arguments. (#959) (#1053)

    The full changes apply to:

    • AdagradOptimizer

    • AdamOptimizer

    • GradientDescentOptimizer

    • MomentumOptimizer

    • NesterovMomentumOptimizer

    • RMSPropOptimizer

    • RotosolveOptimizer

    The requires_grad=False property must mark any non-trainable constant argument. The RotoselectOptimizer allows passing only keyword arguments.

    Example use:

    def cost(x, y, data, scale=1.0):
        return scale * (x[0]-data)**2 + scale * (y-data)**2
    
    x = np.array([1.], requires_grad=True)
    y = np.array([1.0])
    data = np.array([2.], requires_grad=False)
    
    opt = qml.GradientDescentOptimizer()
    
    # the optimizer step and step_and_cost methods can
    # now update multiple parameters at once
    x_new, y_new, data = opt.step(cost, x, y, data, scale=0.5)
    (x_new, y_new, data), value = opt.step_and_cost(cost, x, y, data, scale=0.5)
    
    # list and tuple unpacking is also supported
    params = (x, y, data)
    params = opt.step(cost, *params)
    
  • The circuit drawer has been updated to support the inclusion of unused or inactive wires, by passing the show_all_wires argument. (#1033)

    dev = qml.device('default.qubit', wires=[-1, "a", "q2", 0])
    
    @qml.qnode(dev)
    def circuit():
        qml.Hadamard(wires=-1)
        qml.CNOT(wires=[-1, "q2"])
        return qml.expval(qml.PauliX(wires="q2"))
    
    >>> print(qml.draw(circuit, show_all_wires=True)())
    >>>
     -1: ──H──╭C──┤
      a: ─────│───┤
     q2: ─────╰X──┤ ⟨X⟩
      0: ─────────┤
    
  • The logic for choosing the ‘best’ differentiation method has been altered to improve performance. (#1008)

    • If the device provides its own gradient, this is now the preferred differentiation method.

    • If a device provides additional interface-specific versions that natively support classical backpropagation, this is now preferred over the parameter-shift rule.

      Devices define additional interface-specific devices via their capabilities() dictionary. For example, default.qubit supports supplementary devices for TensorFlow, Autograd, and JAX:

      {
        "passthru_devices": {
            "tf": "default.qubit.tf",
            "autograd": "default.qubit.autograd",
            "jax": "default.qubit.jax",
        },
      }
      

    As a result of this change, if the QNode diff_method is not explicitly provided, it is possible that the QNode will run on a supplementary device of the device that was specifically provided:

    dev = qml.device("default.qubit", wires=2)
    qml.QNode(dev) # will default to backprop on default.qubit.autograd
    qml.QNode(dev, interface="tf") # will default to backprop on default.qubit.tf
    qml.QNode(dev, interface="jax") # will default to backprop on default.qubit.jax
    
  • The default.qubit device has been updated so that internally it applies operations in a more functional style, i.e., by accepting an input state and returning an evolved state. (#1025)

  • A new test series, pennylane/devices/tests/test_compare_default_qubit.py, has been added, allowing to test if a chosen device gives the same result as default.qubit. (#897)

    Three tests are added:

    • test_hermitian_expectation,

    • test_pauliz_expectation_analytic, and

    • test_random_circuit.

  • Adds the following agnostic tensor manipulation functions to the qml.math module: abs, angle, arcsin, concatenate, dot, squeeze, sqrt, sum, take, where. These functions are required to fully support end-to-end differentiable Mottonen and Amplitude embedding. (#922) (#1011)

  • The qml.math module now supports JAX. (#985)

  • Several improvements have been made to the Wires class to reduce overhead and simplify the logic of how wire labels are interpreted: (#1019) (#1010) (#1005) (#983) (#967)

    • If the input wires to a wires class instantiation Wires(wires) can be iterated over, its elements are interpreted as wire labels. Otherwise, wires is interpreted as a single wire label. The only exception to this are strings, which are always interpreted as a single wire label, so users can address wires with labels such as "ancilla".

    • Any type can now be a wire label as long as it is hashable. The hash is used to establish the uniqueness of two labels.

    • Indexing wires objects now returns a label, instead of a new Wires object. For example:

      >>> w = Wires([0, 1, 2])
      >>> w[1]
      >>> 1
      
    • The check for uniqueness of wires moved from Wires instantiation to the qml.wires._process function in order to reduce overhead from repeated creation of Wires instances.

    • Calls to the Wires class are substantially reduced, for example by avoiding to call Wires on Wires instances on Operation instantiation, and by using labels instead of Wires objects inside the default qubit device.

  • Adds the PauliRot generator to the qml.operation module. This generator is required to construct the metric tensor. (#963)

  • The templates are modified to make use of the new qml.math module, for framework-agnostic tensor manipulation. This allows the template library to be differentiable in backpropagation mode (diff_method="backprop"). (#873)

  • The circuit drawer now allows for the wire order to be (optionally) modified: (#992)

    >>> dev = qml.device('default.qubit', wires=["a", -1, "q2"])
    >>> @qml.qnode(dev)
    ... def circuit():
    ...     qml.Hadamard(wires=-1)
    ...     qml.CNOT(wires=["a", "q2"])
    ...     qml.RX(0.2, wires="a")
    ...     return qml.expval(qml.PauliX(wires="q2"))
    

    Printing with default wire order of the device:

    >>> print(circuit.draw())
      a: ─────╭C──RX(0.2)──┤
     -1: ──H──│────────────┤
     q2: ─────╰X───────────┤ ⟨X⟩
    

    Changing the wire order:

    >>> print(circuit.draw(wire_order=["q2", "a", -1]))
     q2: ──╭X───────────┤ ⟨X⟩
      a: ──╰C──RX(0.2)──┤
     -1: ───H───────────┤
    

Breaking changes

  • QNodes using the new PennyLane core will no longer accept ragged arrays as inputs.

  • When using the new PennyLane core and the Autograd interface, non-differentiable data passed as a QNode argument or a gate must have the requires_grad property set to False:

    @qml.qnode(dev)
    def circuit(weights, data):
        basis_state = np.array([1, 0, 1, 1], requires_grad=False)
        qml.BasisState(basis_state, wires=[0, 1, 2, 3])
        qml.templates.AmplitudeEmbedding(data, wires=[0, 1, 2, 3])
        qml.templates.BasicEntanglerLayers(weights, wires=[0, 1, 2, 3])
        return qml.probs(wires=0)
    
    data = np.array(data, requires_grad=False)
    weights = np.array(weights, requires_grad=True)
    circuit(weights, data)
    

Bug fixes

  • Fixes an issue where if the constituent observables of a tensor product do not exist in the queue, an error is raised. With this fix, they are first queued before annotation occurs. (#1038)

  • Fixes an issue with tape expansions where information about sampling (specifically the is_sampled tape attribute) was not preserved. (#1027)

  • Tape expansion was not properly taking into devices that supported inverse operations, causing inverse operations to be unnecessarily decomposed. The QNode tape expansion logic, as well as the Operation.expand() method, has been modified to fix this. (#956)

  • Fixes an issue where the Autograd interface was not unwrapping non-differentiable PennyLane tensors, which can cause issues on some devices. (#941)

  • qml.vqe.Hamiltonian prints any observable with any number of strings. (#987)

  • Fixes a bug where parameter-shift differentiation would fail if the QNode contained a single probability output. (#1007)

  • Fixes an issue when using trainable parameters that are lists/arrays with tape.vjp. (#1042)

  • The TensorN observable is updated to support being copied without any parameters or wires passed. (#1047)

  • Fixed deprecation warning when importing Sequence from collections instead of collections.abc in vqe/vqe.py. (#1051)

Contributors

This release contains contributions from (in alphabetical order):

Juan Miguel Arrazola, Thomas Bromley, Olivia Di Matteo, Theodor Isacsson, Josh Izaac, Christina Lee, Alejandro Montanez, Steven Oud, Chase Roberts, Sankalp Sanand, Maria Schuld, Antal Száva, David Wierichs, Jiahao Yao.

orphan

Release 0.13.0

New features since last release

Automatically optimize the number of measurements

  • QNodes in tape mode now support returning observables on the same wire whenever the observables are qubit-wise commuting Pauli words. Qubit-wise commuting observables can be evaluated with a single device run as they are diagonal in the same basis, via a shared set of single-qubit rotations. (#882)

    The following example shows a single QNode returning the expectation values of the qubit-wise commuting Pauli words XX and XI:

    qml.enable_tape()
    
    @qml.qnode(dev)
    def f(x):
        qml.Hadamard(wires=0)
        qml.Hadamard(wires=1)
        qml.CRot(0.1, 0.2, 0.3, wires=[1, 0])
        qml.RZ(x, wires=1)
        return qml.expval(qml.PauliX(0) @ qml.PauliX(1)), qml.expval(qml.PauliX(0))
    
    >>> f(0.4)
    tensor([0.89431013, 0.9510565 ], requires_grad=True)
    
  • The ExpvalCost class (previously VQECost) now provides observable optimization using the optimize argument, resulting in potentially fewer device executions. (#902)

    This is achieved by separating the observables composing the Hamiltonian into qubit-wise commuting groups and evaluating those groups on a single QNode using functionality from the qml.grouping module:

    qml.enable_tape()
    commuting_obs = [qml.PauliX(0), qml.PauliX(0) @ qml.PauliZ(1)]
    H = qml.vqe.Hamiltonian([1, 1], commuting_obs)
    
    dev = qml.device("default.qubit", wires=2)
    ansatz = qml.templates.StronglyEntanglingLayers
    
    cost_opt = qml.ExpvalCost(ansatz, H, dev, optimize=True)
    cost_no_opt = qml.ExpvalCost(ansatz, H, dev, optimize=False)
    
    params = qml.init.strong_ent_layers_uniform(3, 2)
    

    Grouping these commuting observables leads to fewer device executions:

    >>> cost_opt(params)
    >>> ex_opt = dev.num_executions
    >>> cost_no_opt(params)
    >>> ex_no_opt = dev.num_executions - ex_opt
    >>> print("Number of executions:", ex_no_opt)
    Number of executions: 2
    >>> print("Number of executions (optimized):", ex_opt)
    Number of executions (optimized): 1
    

New quantum gradient features

  • Compute the analytic gradient of quantum circuits in parallel on supported devices. (#840)

    This release introduces support for batch execution of circuits, via a new device API method Device.batch_execute(). Devices that implement this new API support submitting a batch of circuits for parallel evaluation simultaneously, which can significantly reduce the computation time.

    Furthermore, if using tape mode and a compatible device, gradient computations will automatically make use of the new batch API—providing a speedup during optimization.

  • Gradient recipes are now much more powerful, allowing for operations to define their gradient via an arbitrary linear combination of circuit evaluations. (#909) (#915)

    With this change, gradient recipes can now be of the form \(\frac{\partial}{\partial\phi_k}f(\phi_k) = \sum_{i} c_i f(a_i \phi_k + s_i )\), and are no longer restricted to two-term shifts with identical (but opposite in sign) shift values.

    As a result, PennyLane now supports native analytic quantum gradients for the controlled rotation operations CRX, CRY, CRZ, and CRot. This allows for parameter-shift analytic gradients on hardware, without decomposition.

    Note that this is a breaking change for developers; please see the Breaking Changes section for more details.

  • The qnn.KerasLayer class now supports differentiating the QNode through classical backpropagation in tape mode. (#869)

    qml.enable_tape()
    
    dev = qml.device("default.qubit.tf", wires=2)
    
    @qml.qnode(dev, interface="tf", diff_method="backprop")
    def f(inputs, weights):
        qml.templates.AngleEmbedding(inputs, wires=range(2))
        qml.templates.StronglyEntanglingLayers(weights, wires=range(2))
        return [qml.expval(qml.PauliZ(i)) for i in range(2)]
    
    weight_shapes = {"weights": (3, 2, 3)}
    
    qlayer = qml.qnn.KerasLayer(f, weight_shapes, output_dim=2)
    
    inputs = tf.constant(np.random.random((4, 2)), dtype=tf.float32)
    
    with tf.GradientTape() as tape:
        out = qlayer(inputs)
    
    tape.jacobian(out, qlayer.trainable_weights)
    

New operations, templates, and measurements

  • Adds the qml.density_matrix QNode return with partial trace capabilities. (#878)

    The density matrix over the provided wires is returned, with all other subsystems traced out. qml.density_matrix currently works for both the default.qubit and default.mixed devices.

    qml.enable_tape()
    dev = qml.device("default.qubit", wires=2)
    
    def circuit(x):
        qml.PauliY(wires=0)
        qml.Hadamard(wires=1)
        return qml.density_matrix(wires=[1])  # wire 0 is traced out
    
  • Adds the square-root X gate SX. (#871)

    dev = qml.device("default.qubit", wires=1)
    
    @qml.qnode(dev)
    def circuit():
        qml.SX(wires=[0])
        return qml.expval(qml.PauliZ(wires=[0]))
    
  • Two new hardware-efficient particle-conserving templates have been implemented to perform VQE-based quantum chemistry simulations. The new templates apply several layers of the particle-conserving entanglers proposed in Figs. 2a and 2b of Barkoutsos et al., arXiv:1805.04340 (#875) (#876)

Estimate and track resources

  • The QuantumTape class now contains basic resource estimation functionality. The method tape.get_resources() returns a dictionary with a list of the constituent operations and the number of times they appear in the circuit. Similarly, tape.get_depth() computes the circuit depth. (#862)

    >>> with qml.tape.QuantumTape() as tape:
    ...    qml.Hadamard(wires=0)
    ...    qml.RZ(0.26, wires=1)
    ...    qml.CNOT(wires=[1, 0])
    ...    qml.Rot(1.8, -2.7, 0.2, wires=0)
    ...    qml.Hadamard(wires=1)
    ...    qml.CNOT(wires=[0, 1])
    ...    qml.expval(qml.PauliZ(0) @ qml.PauliZ(1))
    >>> tape.get_resources()
    {'Hadamard': 2, 'RZ': 1, 'CNOT': 2, 'Rot': 1}
    >>> tape.get_depth()
    4
    
  • The number of device executions over a QNode’s lifetime can now be returned using num_executions. (#853)

    >>> dev = qml.device("default.qubit", wires=2)
    >>> @qml.qnode(dev)
    ... def circuit(x, y):
    ...    qml.RX(x, wires=[0])
    ...    qml.RY(y, wires=[1])
    ...    qml.CNOT(wires=[0, 1])
    ...    return qml.expval(qml.PauliZ(0) @ qml.PauliX(1))
    >>> for _ in range(10):
    ...    circuit(0.432, 0.12)
    >>> print(dev.num_executions)
    10
    

Improvements

  • Support for tape mode has improved across PennyLane. The following features now work in tape mode:

  • A new function, qml.refresh_devices(), has been added, allowing PennyLane to rescan installed PennyLane plugins and refresh the device list. In addition, the qml.device loader will attempt to refresh devices if the required plugin device cannot be found. This will result in an improved experience if installing PennyLane and plugins within a running Python session (for example, on Google Colab), and avoid the need to restart the kernel/runtime. (#907)

  • When using grad_fn = qml.grad(cost) to compute the gradient of a cost function with the Autograd interface, the value of the intermediate forward pass is now available via the grad_fn.forward property (#914):

    def cost_fn(x, y):
        return 2 * np.sin(x[0]) * np.exp(-x[1]) + x[0] ** 3 + np.cos(y)
    
    params = np.array([0.1, 0.5], requires_grad=True)
    data = np.array(0.65, requires_grad=False)
    grad_fn = qml.grad(cost_fn)
    
    grad_fn(params, data)  # perform backprop and evaluate the gradient
    grad_fn.forward  # the cost function value
    
  • Gradient-based optimizers now have a step_and_cost method that returns both the next step as well as the objective (cost) function output. (#916)

    >>> opt = qml.GradientDescentOptimizer()
    >>> params, cost = opt.step_and_cost(cost_fn, params)
    
  • PennyLane provides a new experimental module qml.proc which provides framework-agnostic processing functions for array and tensor manipulations. (#886)

    Given the input tensor-like object, the call is dispatched to the corresponding array manipulation framework, allowing for end-to-end differentiation to be preserved.

    >>> x = torch.tensor([1., 2.])
    >>> qml.proc.ones_like(x)
    tensor([1, 1])
    >>> y = tf.Variable([[0], [5]])
    >>> qml.proc.ones_like(y, dtype=np.complex128)
    <tf.Tensor: shape=(2, 1), dtype=complex128, numpy=
    array([[1.+0.j],
           [1.+0.j]])>
    

    Note that these functions are experimental, and only a subset of common functionality is supported. Furthermore, the names and behaviour of these functions may differ from similar functions in common frameworks; please refer to the function docstrings for more details.

  • The gradient methods in tape mode now fully separate the quantum and classical processing. Rather than returning the evaluated gradients directly, they now return a tuple containing the required quantum and classical processing steps. (#840)

    def gradient_method(idx, param, **options):
        # generate the quantum tapes that must be computed
        # to determine the quantum gradient
        tapes = quantum_gradient_tapes(self)
    
        def processing_fn(results):
            # perform classical processing on the evaluated tapes
            # returning the evaluated quantum gradient
            return classical_processing(results)
    
        return tapes, processing_fn
    

    The JacobianTape.jacobian() method has been similarly modified to accumulate all gradient quantum tapes and classical processing functions, evaluate all quantum tapes simultaneously, and then apply the post-processing functions to the evaluated tape results.

  • The MultiRZ gate now has a defined generator, allowing it to be used in quantum natural gradient optimization. (#912)

  • The CRot gate now has a decomposition method, which breaks the gate down into rotations and CNOT gates. This allows CRot to be used on devices that do not natively support it. (#908)

  • The classical processing in the MottonenStatePreparation template has been largely rewritten to use dense matrices and tensor manipulations wherever possible. This is in preparation to support differentiation through the template in the future. (#864)

  • Device-based caching has replaced QNode caching. Caching is now accessed by passing a cache argument to the device. (#851)

    The cache argument should be an integer specifying the size of the cache. For example, a cache of size 10 is created using:

    >>> dev = qml.device("default.qubit", wires=2, cache=10)
    
  • The Operation, Tensor, and MeasurementProcess classes now have the __copy__ special method defined. (#840)

    This allows us to ensure that, when a shallow copy is performed of an operation, the mutable list storing the operation parameters is also shallow copied. Both the old operation and the copied operation will continue to share the same parameter data,

    >>> import copy
    >>> op = qml.RX(0.2, wires=0)
    >>> op2 = copy.copy(op)
    >>> op.data[0] is op2.data[0]
    True
    

    however the list container is not a reference:

    >>> op.data is op2.data
    False
    

    This allows the parameters of the copied operation to be modified, without mutating the parameters of the original operation.

  • The QuantumTape.copy method has been tweaked so that (#840):

    • Optionally, the tape’s operations are shallow copied in addition to the tape by passing the copy_operations=True boolean flag. This allows the copied tape’s parameters to be mutated without affecting the original tape’s parameters. (Note: the two tapes will share parameter data until one of the tapes has their parameter list modified.)

    • Copied tapes can be cast to another QuantumTape subclass by passing the tape_cls keyword argument.

Breaking changes

  • Updated how parameter-shift gradient recipes are defined for operations, allowing for gradient recipes that are specified as an arbitrary number of terms. (#909)

    Previously, Operation.grad_recipe was restricted to two-term parameter-shift formulas. With this change, the gradient recipe now contains elements of the form \([c_i, a_i, s_i]\), resulting in a gradient recipe of \(\frac{\partial}{\partial\phi_k}f(\phi_k) = \sum_{i} c_i f(a_i \phi_k + s_i )\).

    As this is a breaking change, all custom operations with defined gradient recipes must be updated to continue working with PennyLane 0.13. Note though that if grad_recipe = None, the default gradient recipe remains unchanged, and corresponds to the two terms \([c_0, a_0, s_0]=[1/2, 1, \pi/2]\) and \([c_1, a_1, s_1]=[-1/2, 1, -\pi/2]\) for every parameter.

  • The VQECost class has been renamed to ExpvalCost to reflect its general applicability beyond VQE. Use of VQECost is still possible but will result in a deprecation warning. (#913)

Bug fixes

  • The default.qubit.tf device is updated to handle TensorFlow objects (e.g., tf.Variable) as gate parameters correctly when using the MultiRZ and CRot operations. (#921)

  • PennyLane tensor objects are now unwrapped in BaseQNode when passed as a keyword argument to the quantum function. (#903) (#893)

  • The new tape mode now prevents multiple observables from being evaluated on the same wire if the observables are not qubit-wise commuting Pauli words. (#882)

  • Fixes a bug in default.qubit whereby inverses of common gates were not being applied via efficient gate-specific methods, instead falling back to matrix-vector multiplication. The following gates were affected: PauliX, PauliY, PauliZ, Hadamard, SWAP, S, T, CNOT, CZ. (#872)

  • The PauliRot operation now gracefully handles single-qubit Paulis, and all-identity Paulis (#860).

  • Fixes a bug whereby binary Python operators were not properly propagating the requires_grad attribute to the output tensor. (#889)

  • Fixes a bug which prevents TorchLayer from doing backward when CUDA is enabled. (#899)

  • Fixes a bug where multi-threaded execution of QNodeCollection sometimes fails because of simultaneous queuing. This is fixed by adding thread locking during queuing. (#910)

  • Fixes a bug in QuantumTape.set_parameters(). The previous implementation assumed that the self.trainable_parms set would always be iterated over in increasing integer order. However, this is not guaranteed behaviour, and can lead to the incorrect tape parameters being set if this is not the case. (#923)

  • Fixes broken error message if a QNode is instantiated with an unknown exception. (#930)

Contributors

This release contains contributions from (in alphabetical order):

Juan Miguel Arrazola, Thomas Bromley, Christina Lee, Alain Delgado Gran, Olivia Di Matteo, Anthony Hayes, Theodor Isacsson, Josh Izaac, Soran Jahangiri, Nathan Killoran, Shumpei Kobayashi, Romain Moyard, Zeyue Niu, Maria Schuld, Antal Száva.

orphan

Release 0.12.0

New features since last release

New and improved simulators

  • PennyLane now supports a new device, default.mixed, designed for simulating mixed-state quantum computations. This enables native support for implementing noisy channels in a circuit, which generally map pure states to mixed states. (#794) (#807) (#819)

    The device can be initialized as

    >>> dev = qml.device("default.mixed", wires=1)
    

    This allows the construction of QNodes that include non-unitary operations, such as noisy channels:

    >>> @qml.qnode(dev)
    ... def circuit(params):
    ...     qml.RX(params[0], wires=0)
    ...     qml.RY(params[1], wires=0)
    ...     qml.AmplitudeDamping(0.5, wires=0)
    ...     return qml.expval(qml.PauliZ(0))
    >>> print(circuit([0.54, 0.12]))
    0.9257702929524184
    >>> print(circuit([0, np.pi]))
    0.0
    

New tools for optimizing measurements

  • The new grouping module provides functionality for grouping simultaneously measurable Pauli word observables. (#761) (#850) (#852)

    • The optimize_measurements function will take as input a list of Pauli word observables and their corresponding coefficients (if any), and will return the partitioned Pauli terms diagonalized in the measurement basis and the corresponding diagonalizing circuits.

      from pennylane.grouping import optimize_measurements
      h, nr_qubits = qml.qchem.molecular_hamiltonian("h2", "h2.xyz")
      rotations, grouped_ops, grouped_coeffs = optimize_measurements(h.ops, h.coeffs, grouping="qwc")
      

      The diagonalizing circuits of rotations correspond to the diagonalized Pauli word groupings of grouped_ops.

    • Pauli word partitioning utilities are performed by the PauliGroupingStrategy class. An input list of Pauli words can be partitioned into mutually commuting, qubit-wise-commuting, or anticommuting groupings.

      For example, partitioning Pauli words into anticommutative groupings by the Recursive Largest First (RLF) graph colouring heuristic:

      from pennylane import PauliX, PauliY, PauliZ, Identity
      from pennylane.grouping import group_observables
      pauli_words = [
          Identity('a') @ Identity('b'),
          Identity('a') @ PauliX('b'),
          Identity('a') @ PauliY('b'),
          PauliZ('a') @ PauliX('b'),
          PauliZ('a') @ PauliY('b'),
          PauliZ('a') @ PauliZ('b')
      ]
      groupings = group_observables(pauli_words, grouping_type='anticommuting', method='rlf')
      
    • Various utility functions are included for obtaining and manipulating Pauli words in the binary symplectic vector space representation.

      For instance, two Pauli words may be converted to their binary vector representation:

      >>> from pennylane.grouping import pauli_to_binary
      >>> from pennylane.wires import Wires
      >>> wire_map = {Wires('a'): 0, Wires('b'): 1}
      >>> pauli_vec_1 = pauli_to_binary(qml.PauliX('a') @ qml.PauliY('b'))
      >>> pauli_vec_2 = pauli_to_binary(qml.PauliZ('a') @ qml.PauliZ('b'))
      >>> pauli_vec_1
      [1. 1. 0. 1.]
      >>> pauli_vec_2
      [0. 0. 1. 1.]
      

      Their product up to a phase may be computed by taking the sum of their binary vector representations, and returned in the operator representation.

      >>> from pennylane.grouping import binary_to_pauli
      >>> binary_to_pauli((pauli_vec_1 + pauli_vec_2) % 2, wire_map)
      Tensor product ['PauliY', 'PauliX']: 0 params, wires ['a', 'b']
      

      For more details on the grouping module, see the grouping module documentation

Returning the quantum state from simulators

  • The quantum state of a QNode can now be returned using the qml.state() return function. (#818)

    import pennylane as qml
    
    dev = qml.device("default.qubit", wires=3)
    qml.enable_tape()
    
    @qml.qnode(dev)
    def qfunc(x, y):
        qml.RZ(x, wires=0)
        qml.CNOT(wires=[0, 1])
        qml.RY(y, wires=1)
        qml.CNOT(wires=[0, 2])
        return qml.state()
    
    >>> qfunc(0.56, 0.1)
    array([0.95985437-0.27601028j, 0.        +0.j        ,
           0.04803275-0.01381203j, 0.        +0.j        ,
           0.        +0.j        , 0.        +0.j        ,
           0.        +0.j        , 0.        +0.j        ])
    

    Differentiating the state is currently available when using the classical backpropagation differentiation method (diff_method="backprop") with a compatible device, and when using the new tape mode.

New operations and channels

  • PennyLane now includes standard channels such as the Amplitude-damping, Phase-damping, and Depolarizing channels, as well as the ability to make custom qubit channels. (#760) (#766) (#778)

  • The controlled-Y operation is now available via qml.CY. For devices that do not natively support the controlled-Y operation, it will be decomposed into qml.RY, qml.CNOT, and qml.S operations. (#806)

Preview the next-generation PennyLane QNode

  • The new PennyLane tape module provides a re-formulated QNode class, rewritten from the ground-up, that uses a new QuantumTape object to represent the QNode’s quantum circuit. Tape mode provides several advantages over the standard PennyLane QNode. (#785) (#792) (#796) (#800) (#803) (#804) (#805) (#808) (#810) (#811) (#815) (#820) (#823) (#824) (#829)

    • Support for in-QNode classical processing: Tape mode allows for differentiable classical processing within the QNode.

    • No more Variable wrapping: In tape mode, QNode arguments no longer become Variable objects within the QNode.

    • Less restrictive QNode signatures: There is no longer any restriction on the QNode signature; the QNode can be defined and called following the same rules as standard Python functions.

    • Unifying all QNodes: The tape-mode QNode merges all QNodes (including the JacobianQNode and the PassthruQNode) into a single unified QNode, with identical behaviour regardless of the differentiation type.

    • Optimizations: Tape mode provides various performance optimizations, reducing pre- and post-processing overhead, and reduces the number of quantum evaluations in certain cases.

    Note that tape mode is experimental, and does not currently have feature-parity with the existing QNode. Feedback and bug reports are encouraged and will help improve the new tape mode.

    Tape mode can be enabled globally via the qml.enable_tape function, without changing your PennyLane code:

    qml.enable_tape()
    dev = qml.device("default.qubit", wires=1)
    
    @qml.qnode(dev, interface="tf")
    def circuit(p):
        print("Parameter value:", p)
        qml.RX(tf.sin(p[0])**2 + p[1], wires=0)
        return qml.expval(qml.PauliZ(0))
    

    For more details, please see the tape mode documentation.

Improvements

  • QNode caching has been introduced, allowing the QNode to keep track of the results of previous device executions and reuse those results in subsequent calls. Note that QNode caching is only supported in the new and experimental tape-mode. (#817)

    Caching is available by passing a caching argument to the QNode:

    dev = qml.device("default.qubit", wires=2)
    qml.enable_tape()
    
    @qml.qnode(dev, caching=10)  # cache up to 10 evaluations
    def qfunc(x):
        qml.RX(x, wires=0)
        qml.RX(0.3, wires=1)
        qml.CNOT(wires=[0, 1])
        return qml.expval(qml.PauliZ(1))
    
    qfunc(0.1)  # first evaluation executes on the device
    qfunc(0.1)  # second evaluation accesses the cached result
    
  • Sped up the application of certain gates in default.qubit by using array/tensor manipulation tricks. The following gates are affected: PauliX, PauliY, PauliZ, Hadamard, SWAP, S, T, CNOT, CZ. (#772)

  • The computation of marginal probabilities has been made more efficient for devices with a large number of wires, achieving in some cases a 5x speedup. (#799)

  • Adds arithmetic operations (addition, tensor product, subtraction, and scalar multiplication) between Hamiltonian, Tensor, and Observable objects, and inline arithmetic operations between Hamiltonians and other observables. (#765)

    Hamiltonians can now easily be defined as sums of observables:

    >>> H = 3 * qml.PauliZ(0) - (qml.PauliX(0) @ qml.PauliX(1)) + qml.Hamiltonian([4], [qml.PauliZ(0)])
    >>> print(H)
    (7.0) [Z0] + (-1.0) [X0 X1]
    
  • Adds compare() method to Observable and Hamiltonian classes, which allows for comparison between observable quantities. (#765)

    >>> H = qml.Hamiltonian([1], [qml.PauliZ(0)])
    >>> obs = qml.PauliZ(0) @ qml.Identity(1)
    >>> print(H.compare(obs))
    True
    
    >>> H = qml.Hamiltonian([2], [qml.PauliZ(0)])
    >>> obs = qml.PauliZ(1) @ qml.Identity(0)
    >>> print(H.compare(obs))
    False
    
  • Adds simplify() method to the Hamiltonian class. (#765)

    >>> H = qml.Hamiltonian([1, 2], [qml.PauliZ(0), qml.PauliZ(0) @ qml.Identity(1)])
    >>> H.simplify()
    >>> print(H)
    (3.0) [Z0]
    
  • Added a new bit-flip mixer to the qml.qaoa module. (#774)

  • Summation of two Wires objects is now supported and will return a Wires object containing the set of all wires defined by the terms in the summation. (#812)

Breaking changes

  • The PennyLane NumPy module now returns scalar (zero-dimensional) arrays where Python scalars were previously returned. (#820) (#833)

    For example, this affects array element indexing, and summation:

    >>> x = np.array([1, 2, 3], requires_grad=False)
    >>> x[0]
    tensor(1, requires_grad=False)
    >>> np.sum(x)
    tensor(6, requires_grad=True)
    

    This may require small updates to user code. A convenience method, np.tensor.unwrap(), has been added to help ease the transition. This converts PennyLane NumPy tensors to standard NumPy arrays and Python scalars:

    >>> x = np.array(1.543, requires_grad=False)
    >>> x.unwrap()
    1.543
    

    Note, however, that information regarding array differentiability will be lost.

  • The device capabilities dictionary has been redesigned, for clarity and robustness. In particular, the capabilities dictionary is now inherited from the parent class, various keys have more expressive names, and all keys are now defined in the base device class. For more details, please refer to the developer documentation. (#781)

Bug fixes

  • Changed to use lists for storing variable values inside BaseQNode allowing complex matrices to be passed to QubitUnitary. (#773)

  • Fixed a bug within default.qubit, resulting in greater efficiency when applying a state vector to all wires on the device. (#849)

Documentation

  • Equations have been added to the qml.sample and qml.probs docstrings to clarify the mathematical foundation of the performed measurements. (#843)

Contributors

This release contains contributions from (in alphabetical order):

Aroosa Ijaz, Juan Miguel Arrazola, Thomas Bromley, Jack Ceroni, Alain Delgado Gran, Josh Izaac, Soran Jahangiri, Nathan Killoran, Robert Lang, Cedric Lin, Olivia Di Matteo, Nicolás Quesada, Maria Schuld, Antal Száva.

orphan

Release 0.11.0

New features since last release

New and improved simulators

  • Added a new device, default.qubit.autograd, a pure-state qubit simulator written using Autograd. This device supports classical backpropagation (diff_method="backprop"); this can be faster than the parameter-shift rule for computing quantum gradients when the number of parameters to be optimized is large. (#721)

    >>> dev = qml.device("default.qubit.autograd", wires=1)
    >>> @qml.qnode(dev, diff_method="backprop")
    ... def circuit(x):
    ...     qml.RX(x[1], wires=0)
    ...     qml.Rot(x[0], x[1], x[2], wires=0)
    ...     return qml.expval(qml.PauliZ(0))
    >>> weights = np.array([0.2, 0.5, 0.1])
    >>> grad_fn = qml.grad(circuit)
    >>> print(grad_fn(weights))
    array([-2.25267173e-01, -1.00864546e+00,  6.93889390e-18])
    

    See the device documentation for more details.

  • A new experimental C++ state-vector simulator device is now available, lightning.qubit. It uses the C++ Eigen library to perform fast linear algebra calculations for simulating quantum state-vector evolution.

    lightning.qubit is currently in beta; it can be installed via pip:

    $ pip install pennylane-lightning
    

    Once installed, it can be used as a PennyLane device:

    >>> dev = qml.device("lightning.qubit", wires=2)
    

    For more details, please see the lightning qubit documentation.

New algorithms and templates

  • Added built-in QAOA functionality via the new qml.qaoa module. (#712) (#718) (#741) (#720)

    This includes the following features:

    • New qml.qaoa.x_mixer and qml.qaoa.xy_mixer functions for defining Pauli-X and XY mixer Hamiltonians.

    • MaxCut: The qml.qaoa.maxcut function allows easy construction of the cost Hamiltonian and recommended mixer Hamiltonian for solving the MaxCut problem for a supplied graph.

    • Layers: qml.qaoa.cost_layer and qml.qaoa.mixer_layer take cost and mixer Hamiltonians, respectively, and apply the corresponding QAOA cost and mixer layers to the quantum circuit

    For example, using PennyLane to construct and solve a MaxCut problem with QAOA:

    wires = range(3)
    graph = Graph([(0, 1), (1, 2), (2, 0)])
    cost_h, mixer_h = qaoa.maxcut(graph)
    
    def qaoa_layer(gamma, alpha):
        qaoa.cost_layer(gamma, cost_h)
        qaoa.mixer_layer(alpha, mixer_h)
    
    def antatz(params, **kwargs):
    
        for w in wires:
            qml.Hadamard(wires=w)
    
        # repeat the QAOA layer two times
        qml.layer(qaoa_layer, 2, params[0], params[1])
    
    dev = qml.device('default.qubit', wires=len(wires))
    cost_function = qml.VQECost(ansatz, cost_h, dev)
    
  • Added an ApproxTimeEvolution template to the PennyLane templates module, which can be used to implement Trotterized time-evolution under a Hamiltonian. (#710)

  • Added a qml.layer template-constructing function, which takes a unitary, and repeatedly applies it on a set of wires to a given depth. (#723)

    def subroutine():
        qml.Hadamard(wires=[0])
        qml.CNOT(wires=[0, 1])
        qml.PauliX(wires=[1])
    
    dev = qml.device('default.qubit', wires=3)
    
    @qml.qnode(dev)
    def circuit():
        qml.layer(subroutine, 3)
        return [qml.expval(qml.PauliZ(0)), qml.expval(qml.PauliZ(1))]
    

    This creates the following circuit:

    >>> circuit()
    >>> print(circuit.draw())
    0: ──H──╭C──X──H──╭C──X──H──╭C──X──┤ ⟨Z⟩
    1: ─────╰X────────╰X────────╰X─────┤ ⟨Z⟩
    
  • Added the qml.utils.decompose_hamiltonian function. This function can be used to decompose a Hamiltonian into a linear combination of Pauli operators. (#671)

    >>> A = np.array(
    ... [[-2, -2+1j, -2, -2],
    ... [-2-1j,  0,  0, -1],
    ... [-2,  0, -2, -1],
    ... [-2, -1, -1,  0]])
    >>> coeffs, obs_list = decompose_hamiltonian(A)
    

New device features

  • It is now possible to specify custom wire labels, such as ['anc1', 'anc2', 0, 1, 3], where the labels can be strings or numbers. (#666)

    Custom wire labels are defined by passing a list to the wires argument when creating the device:

    >>> dev = qml.device("default.qubit", wires=['anc1', 'anc2', 0, 1, 3])
    

    Quantum operations should then be invoked with these custom wire labels:

    >>> @qml.qnode(dev)
    >>> def circuit():
    ...    qml.Hadamard(wires='anc2')
    ...    qml.CNOT(wires=['anc1', 3])
    ...    ...
    

    The existing behaviour, in which the number of wires is specified on device initialization, continues to work as usual. This gives a default behaviour where wires are labelled by consecutive integers.

    >>> dev = qml.device("default.qubit", wires=5)
    
  • An integrated device test suite has been added, which can be used to run basic integration tests on core or external devices. (#695) (#724) (#733)

    The test can be invoked against a particular device by calling the pl-device-test command line program:

    $ pl-device-test --device=default.qubit --shots=1234 --analytic=False
    

    If the tests are run on external devices, the device and its dependencies must be installed locally. For more details, please see the plugin test documentation.

Improvements

  • The functions implementing the quantum circuits building the Unitary Coupled-Cluster (UCCSD) VQE ansatz have been improved, with a more consistent naming convention and improved docstrings. (#748)

    The changes include:

    • The terms 1particle-1hole (ph) and 2particle-2hole (pphh) excitations were replaced with the names single and double excitations, respectively.

    • The non-differentiable arguments in the UCCSD template were renamed accordingly: phs_wires, pphhd_wires

    • The term virtual, previously used to refer the unoccupied orbitals, was discarded.

    • The Usage Details sections were updated and improved.

  • Added support for TensorFlow 2.3 and PyTorch 1.6. (#725)

  • Returning probabilities is now supported from photonic QNodes. As with qubit QNodes, photonic QNodes returning probabilities are end-to-end differentiable. (#699)

    >>> dev = qml.device("strawberryfields.fock", wires=2, cutoff_dim=5)
    >>> @qml.qnode(dev)
    ... def circuit(a):
    ...     qml.Displacement(a, 0, wires=0)
    ...     return qml.probs(wires=0)
    >>> print(circuit(0.5))
    [7.78800783e-01 1.94700196e-01 2.43375245e-02 2.02812704e-03 1.26757940e-04]
    

Breaking changes

  • The pennylane.plugins and pennylane.beta.plugins folders have been renamed to pennylane.devices and pennylane.beta.devices, to reflect their content better. (#726)

Bug fixes

  • The PennyLane interface conversion functions can now convert QNodes with pre-existing interfaces. (#707)

Documentation

  • The interfaces section of the documentation has been renamed to ‘Interfaces and training’, and updated with the latest variable handling details. (#753)

Contributors

This release contains contributions from (in alphabetical order):

Juan Miguel Arrazola, Thomas Bromley, Jack Ceroni, Alain Delgado Gran, Shadab Hussain, Theodor Isacsson, Josh Izaac, Nathan Killoran, Maria Schuld, Antal Száva, Nicola Vitucci.

orphan

Release 0.10.0

New features since last release

New and improved simulators

  • Added a new device, default.qubit.tf, a pure-state qubit simulator written using TensorFlow. As a result, it supports classical backpropagation as a means to compute the Jacobian. This can be faster than the parameter-shift rule for computing quantum gradients when the number of parameters to be optimized is large.

    default.qubit.tf is designed to be used with end-to-end classical backpropagation (diff_method="backprop") with the TensorFlow interface. This is the default method of differentiation when creating a QNode with this device.

    Using this method, the created QNode is a ‘white-box’ that is tightly integrated with your TensorFlow computation, including AutoGraph support:

    >>> dev = qml.device("default.qubit.tf", wires=1)
    >>> @tf.function
    ... @qml.qnode(dev, interface="tf", diff_method="backprop")
    ... def circuit(x):
    ...     qml.RX(x[1], wires=0)
    ...     qml.Rot(x[0], x[1], x[2], wires=0)
    ...     return qml.expval(qml.PauliZ(0))
    >>> weights = tf.Variable([0.2, 0.5, 0.1])
    >>> with tf.GradientTape() as tape:
    ...     res = circuit(weights)
    >>> print(tape.gradient(res, weights))
    tf.Tensor([-2.2526717e-01 -1.0086454e+00  1.3877788e-17], shape=(3,), dtype=float32)
    

    See the default.qubit.tf documentation for more details.

  • The default.tensor plugin has been significantly upgraded. It now allows two different tensor network representations to be used: "exact" and "mps". The former uses a exact factorized representation of quantum states, while the latter uses a matrix product state representation. (#572) (#599)

New machine learning functionality and integrations

  • PennyLane QNodes can now be converted into Torch layers, allowing for creation of quantum and hybrid models using the torch.nn API. (#588)

    A PennyLane QNode can be converted into a torch.nn layer using the qml.qnn.TorchLayer class:

    >>> @qml.qnode(dev)
    ... def qnode(inputs, weights_0, weight_1):
    ...    # define the circuit
    ...    # ...
    
    >>> weight_shapes = {"weights_0": 3, "weight_1": 1}
    >>> qlayer = qml.qnn.TorchLayer(qnode, weight_shapes)
    

    A hybrid model can then be easily constructed:

    >>> model = torch.nn.Sequential(qlayer, torch.nn.Linear(2, 2))
    
  • Added a new “reversible” differentiation method which can be used in simulators, but not hardware.

    The reversible approach is similar to backpropagation, but trades off extra computation for enhanced memory efficiency. Where backpropagation caches the state tensors at each step during a simulated evolution, the reversible method only caches the final pre-measurement state.

    Compared to the parameter-shift method, the reversible method can be faster or slower, depending on the density and location of parametrized gates in a circuit (circuits with higher density of parametrized gates near the end of the circuit will see a benefit). (#670)

    >>> dev = qml.device("default.qubit", wires=2)
    ... @qml.qnode(dev, diff_method="reversible")
    ... def circuit(x):
    ...     qml.RX(x, wires=0)
    ...     qml.RX(x, wires=0)
    ...     qml.CNOT(wires=[0,1])
    ...     return qml.expval(qml.PauliZ(0))
    >>> qml.grad(circuit)(0.5)
    (array(-0.47942554),)
    

New templates and cost functions

  • Added the new templates UCCSD, SingleExcitationUnitary, andDoubleExcitationUnitary, which together implement the Unitary Coupled-Cluster Singles and Doubles (UCCSD) ansatz to perform VQE-based quantum chemistry simulations using PennyLane-QChem. (#622) (#638) (#654) (#659) (#622)

  • Added module pennylane.qnn.cost with class SquaredErrorLoss. The module contains classes to calculate losses and cost functions on circuits with trainable parameters. (#642)

Improvements

  • Improves the wire management by making the Operator.wires attribute a wires object. (#666)

  • A significant improvement with respect to how QNodes and interfaces mark quantum function arguments as differentiable when using Autograd, designed to improve performance and make QNodes more intuitive. (#648) (#650)

    In particular, the following changes have been made:

    • A new ndarray subclass pennylane.numpy.tensor, which extends NumPy arrays with the keyword argument and attribute requires_grad. Tensors which have requires_grad=False are treated as non-differentiable by the Autograd interface.

    • A new subpackage pennylane.numpy, which wraps autograd.numpy such that NumPy functions accept the requires_grad keyword argument, and allows Autograd to differentiate pennylane.numpy.tensor objects.

    • The argnum argument to qml.grad is now optional; if not provided, arguments explicitly marked as requires_grad=False are excluded for the list of differentiable arguments. The ability to pass argnum has been retained for backwards compatibility, and if present the old behaviour persists.

  • The QNode Torch interface now inspects QNode positional arguments. If any argument does not have the attribute requires_grad=True, it is automatically excluded from quantum gradient computations. (#652) (#660)

  • The QNode TF interface now inspects QNode positional arguments. If any argument is not being watched by a tf.GradientTape(), it is automatically excluded from quantum gradient computations. (#655) (#660)

  • QNodes have two new public methods: QNode.set_trainable_args() and QNode.get_trainable_args(). These are designed to be called by interfaces, to specify to the QNode which of its input arguments are differentiable. Arguments which are non-differentiable will not be converted to PennyLane Variable objects within the QNode. (#660)

  • Added decomposition method to PauliX, PauliY, PauliZ, S, T, Hadamard, and PhaseShift gates, which decomposes each of these gates into rotation gates. (#668)

  • The CircuitGraph class now supports serializing contained circuit operations and measurement basis rotations to an OpenQASM2.0 script via the new CircuitGraph.to_openqasm() method. (#623)

Breaking changes

  • Removes support for Python 3.5. (#639)

Documentation

  • Various small typos were fixed.

Contributors

This release contains contributions from (in alphabetical order):

Thomas Bromley, Jack Ceroni, Alain Delgado Gran, Theodor Isacsson, Josh Izaac, Nathan Killoran, Maria Schuld, Antal Száva, Nicola Vitucci.

orphan

Release 0.9.0

New features since last release

New machine learning integrations

  • PennyLane QNodes can now be converted into Keras layers, allowing for creation of quantum and hybrid models using the Keras API. (#529)

    A PennyLane QNode can be converted into a Keras layer using the KerasLayer class:

    from pennylane.qnn import KerasLayer
    
    @qml.qnode(dev)
    def circuit(inputs, weights_0, weight_1):
       # define the circuit
       # ...
    
    weight_shapes = {"weights_0": 3, "weight_1": 1}
    qlayer = qml.qnn.KerasLayer(circuit, weight_shapes, output_dim=2)
    

    A hybrid model can then be easily constructed:

    model = tf.keras.models.Sequential([qlayer, tf.keras.layers.Dense(2)])
    
  • Added a new type of QNode, qml.qnodes.PassthruQNode. For simulators which are coded in an external library which supports automatic differentiation, PennyLane will treat a PassthruQNode as a “white box”, and rely on the external library to directly provide gradients via backpropagation. This can be more efficient than the using parameter-shift rule for a large number of parameters. (#488)

    Currently this behaviour is supported by PennyLane’s default.tensor.tf device backend, compatible with the 'tf' interface using TensorFlow 2:

    dev = qml.device('default.tensor.tf', wires=2)
    
    @qml.qnode(dev, diff_method="backprop")
    def circuit(params):
        qml.RX(params[0], wires=0)
        qml.RX(params[1], wires=1)
        qml.CNOT(wires=[0, 1])
        return qml.expval(qml.PauliZ(0))
    
    qnode = PassthruQNode(circuit, dev)
    params = tf.Variable([0.3, 0.1])
    
    with tf.GradientTape() as tape:
        tape.watch(params)
        res = qnode(params)
    
    grad = tape.gradient(res, params)
    

New optimizers

  • Added the qml.RotosolveOptimizer, a gradient-free optimizer that minimizes the quantum function by updating each parameter, one-by-one, via a closed-form expression while keeping other parameters fixed. (#636) (#539)

  • Added the qml.RotoselectOptimizer, which uses Rotosolve to minimizes a quantum function with respect to both the rotation operations applied and the rotation parameters. (#636) (#539)

    For example, given a quantum function f that accepts parameters x and a list of corresponding rotation operations generators, the Rotoselect optimizer will, at each step, update both the parameter values and the list of rotation gates to minimize the loss:

    >>> opt = qml.optimize.RotoselectOptimizer()
    >>> x = [0.3, 0.7]
    >>> generators = [qml.RX, qml.RY]
    >>> for _ in range(100):
    ...     x, generators = opt.step(f, x, generators)
    

New operations

  • Added the PauliRot gate, which performs an arbitrary Pauli rotation on multiple qubits, and the MultiRZ gate, which performs a rotation generated by a tensor product of Pauli Z operators. (#559)

    dev = qml.device('default.qubit', wires=4)
    
    @qml.qnode(dev)
    def circuit(angle):
        qml.PauliRot(angle, "IXYZ", wires=[0, 1, 2, 3])
        return [qml.expval(qml.PauliZ(wire)) for wire in [0, 1, 2, 3]]
    
    >>> circuit(0.4)
    [1.         0.92106099 0.92106099 1.        ]
    >>> print(circuit.draw())
     0: ──╭RI(0.4)──┤ ⟨Z⟩
     1: ──├RX(0.4)──┤ ⟨Z⟩
     2: ──├RY(0.4)──┤ ⟨Z⟩
     3: ──╰RZ(0.4)──┤ ⟨Z⟩
    

    If the PauliRot gate is not supported on the target device, it will be decomposed into Hadamard, RX and MultiRZ gates. Note that identity gates in the Pauli word result in untouched wires:

    >>> print(circuit.draw())
     0: ───────────────────────────────────┤ ⟨Z⟩
     1: ──H──────────╭RZ(0.4)──H───────────┤ ⟨Z⟩
     2: ──RX(1.571)──├RZ(0.4)──RX(-1.571)──┤ ⟨Z⟩
     3: ─────────────╰RZ(0.4)──────────────┤ ⟨Z⟩
    

    If the MultiRZ gate is not supported, it will be decomposed into CNOT and RZ gates:

    >>> print(circuit.draw())
     0: ──────────────────────────────────────────────────┤ ⟨Z⟩
     1: ──H──────────────╭X──RZ(0.4)──╭X──────H───────────┤ ⟨Z⟩
     2: ──RX(1.571)──╭X──╰C───────────╰C──╭X──RX(-1.571)──┤ ⟨Z⟩
     3: ─────────────╰C───────────────────╰C──────────────┤ ⟨Z⟩
    
  • PennyLane now provides DiagonalQubitUnitary for diagonal gates, that are e.g., encountered in IQP circuits. These kinds of gates can be evaluated much faster on a simulator device. (#567)

    The gate can be used, for example, to efficiently simulate oracles:

    dev = qml.device('default.qubit', wires=3)
    
    # Function as a bitstring
    f = np.array([1, 0, 0, 1, 1, 0, 1, 0])
    
    @qml.qnode(dev)
    def circuit(weights1, weights2):
        qml.templates.StronglyEntanglingLayers(weights1, wires=[0, 1, 2])
    
        # Implements the function as a phase-kickback oracle
        qml.DiagonalQubitUnitary((-1)**f, wires=[0, 1, 2])
    
        qml.templates.StronglyEntanglingLayers(weights2, wires=[0, 1, 2])
        return [qml.expval(qml.PauliZ(w)) for w in range(3)]
    
  • Added the TensorN CVObservable that can represent the tensor product of the NumberOperator on photonic backends. (#608)

New templates

  • Added the ArbitraryUnitary and ArbitraryStatePreparation templates, which use PauliRot gates to perform an arbitrary unitary and prepare an arbitrary basis state with the minimal number of parameters. (#590)

    dev = qml.device('default.qubit', wires=3)
    
    @qml.qnode(dev)
    def circuit(weights1, weights2):
          qml.templates.ArbitraryStatePreparation(weights1, wires=[0, 1, 2])
          qml.templates.ArbitraryUnitary(weights2, wires=[0, 1, 2])
          return qml.probs(wires=[0, 1, 2])
    
  • Added the IQPEmbedding template, which encodes inputs into the diagonal gates of an IQP circuit. (#605)

  • Added the SimplifiedTwoDesign template, which implements the circuit design of Cerezo et al. (2020). (#556)

  • Added the BasicEntanglerLayers template, which is a simple layer architecture of rotations and CNOT nearest-neighbour entanglers. (#555)

  • PennyLane now offers a broadcasting function to easily construct templates: qml.broadcast() takes single quantum operations or other templates and applies them to wires in a specific pattern. (#515) (#522) (#526) (#603)

    For example, we can use broadcast to repeat a custom template across multiple wires:

    from pennylane.templates import template
    
    @template
    def mytemplate(pars, wires):
        qml.Hadamard(wires=wires)
        qml.RY(pars, wires=wires)
    
    dev = qml.device('default.qubit', wires=3)
    
    @qml.qnode(dev)
    def circuit(pars):
        qml.broadcast(mytemplate, pattern="single", wires=[0,1,2], parameters=pars)
        return qml.expval(qml.PauliZ(0))
    
    >>> circuit([1, 1, 0.1])
    -0.841470984807896
    >>> print(circuit.draw())
     0: ──H──RY(1.0)──┤ ⟨Z⟩
     1: ──H──RY(1.0)──┤
     2: ──H──RY(0.1)──┤
    

    For other available patterns, see the broadcast function documentation.

Breaking changes

  • The QAOAEmbedding now uses the new MultiRZ gate as a ZZ entangler, which changes the convention. While previously, the ZZ gate in the embedding was implemented as

    CNOT(wires=[wires[0], wires[1]])
    RZ(2 * parameter, wires=wires[0])
    CNOT(wires=[wires[0], wires[1]])
    

    the MultiRZ corresponds to

    CNOT(wires=[wires[1], wires[0]])
    RZ(parameter, wires=wires[0])
    CNOT(wires=[wires[1], wires[0]])
    

    which differs in the factor of 2, and fixes a bug in the wires that the CNOT was applied to. (#609)

  • Probability methods are handled by QubitDevice and device method requirements are modified to simplify plugin development. (#573)

  • The internal variables All and Any to mark an Operation as acting on all or any wires have been renamed to AllWires and AnyWires. (#614)

Improvements

  • A new Wires class was introduced for the internal bookkeeping of wire indices. (#615)

  • Improvements to the speed/performance of the default.qubit device. (#567) (#559)

  • Added the "backprop" and "device" differentiation methods to the qnode decorator. (#552)

    • "backprop": Use classical backpropagation. Default on simulator devices that are classically end-to-end differentiable. The returned QNode can only be used with the same machine learning framework (e.g., default.tensor.tf simulator with the tensorflow interface).

    • "device": Queries the device directly for the gradient.

    Using the "backprop" differentiation method with the default.tensor.tf device, the created QNode is a ‘white-box’, and is tightly integrated with the overall TensorFlow computation:

    >>> dev = qml.device("default.tensor.tf", wires=1)
    >>> @qml.qnode(dev, interface="tf", diff_method="backprop")
    >>> def circuit(x):
    ...     qml.RX(x[1], wires=0)
    ...     qml.Rot(x[0], x[1], x[2], wires=0)
    ...     return qml.expval(qml.PauliZ(0))
    >>> vars = tf.Variable([0.2, 0.5, 0.1])
    >>> with tf.GradientTape() as tape:
    ...     res = circuit(vars)
    >>> tape.gradient(res, vars)
    <tf.Tensor: shape=(3,), dtype=float32, numpy=array([-2.2526717e-01, -1.0086454e+00,  1.3877788e-17], dtype=float32)>
    
  • The circuit drawer now displays inverted operations, as well as wires where probabilities are returned from the device: (#540)

    >>> @qml.qnode(dev)
    ... def circuit(theta):
    ...     qml.RX(theta, wires=0)
    ...     qml.CNOT(wires=[0, 1])
    ...     qml.S(wires=1).inv()
    ...     return qml.probs(wires=[0, 1])
    >>> circuit(0.2)
    array([0.99003329, 0.        , 0.        , 0.00996671])
    >>> print(circuit.draw())
    0: ──RX(0.2)──╭C───────╭┤ Probs
    1: ───────────╰X──S⁻¹──╰┤ Probs
    
  • You can now evaluate the metric tensor of a VQE Hamiltonian via the new VQECost.metric_tensor method. This allows VQECost objects to be directly optimized by the quantum natural gradient optimizer (qml.QNGOptimizer). (#618)

  • The input check functions in pennylane.templates.utils are now public and visible in the API documentation. (#566)

  • Added keyword arguments for step size and order to the qnode decorator, as well as the QNode and JacobianQNode classes. This enables the user to set the step size and order when using finite difference methods. These options are also exposed when creating QNode collections. (#530) (#585) (#587)

  • The decomposition for the CRY gate now uses the simpler form RY @ CNOT @ RY @ CNOT (#547)

  • The underlying queuing system was refactored, removing the qml._current_context property that held the currently active QNode or OperationRecorder. Now, all objects that expose a queue for operations inherit from QueuingContext and register their queue globally. (#548)

  • The PennyLane repository has a new benchmarking tool which supports the comparison of different git revisions. (#568) (#560) (#516)

Documentation

  • Updated the development section by creating a landing page with links to sub-pages containing specific guides. (#596)

  • Extended the developer’s guide by a section explaining how to add new templates. (#564)

Bug fixes

  • tf.GradientTape().jacobian() can now be evaluated on QNodes using the TensorFlow interface. (#626)

  • RandomLayers() is now compatible with the qiskit devices. (#597)

  • DefaultQubit.probability() now returns the correct probability when called with device.analytic=False. (#563)

  • Fixed a bug in the StronglyEntanglingLayers template, allowing it to work correctly when applied to a single wire. (544)

  • Fixed a bug when inverting operations with decompositions; operations marked as inverted are now correctly inverted when the fallback decomposition is called. (#543)

  • The QNode.print_applied() method now correctly displays wires where qml.prob() is being returned. #542

Contributors

This release contains contributions from (in alphabetical order):

Ville Bergholm, Lana Bozanic, Thomas Bromley, Theodor Isacsson, Josh Izaac, Nathan Killoran, Maggie Li, Johannes Jakob Meyer, Maria Schuld, Sukin Sim, Antal Száva.

orphan

Release 0.8.0

New features since last release

  • Added a quantum chemistry package, pennylane.qchem, which supports integration with OpenFermion, Psi4, PySCF, and OpenBabel. (#453)

    Features include:

    • Generate the qubit Hamiltonians directly starting with the atomic structure of the molecule.

    • Calculate the mean-field (Hartree-Fock) electronic structure of molecules.

    • Allow to define an active space based on the number of active electrons and active orbitals.

    • Perform the fermionic-to-qubit transformation of the electronic Hamiltonian by using different functions implemented in OpenFermion.

    • Convert OpenFermion’s QubitOperator to a Pennylane Hamiltonian class.

    • Perform a Variational Quantum Eigensolver (VQE) computation with this Hamiltonian in PennyLane.

    Check out the quantum chemistry quickstart, as well the quantum chemistry and VQE tutorials.

  • PennyLane now has some functions and classes for creating and solving VQE problems. (#467)

    • qml.Hamiltonian: a lightweight class for representing qubit Hamiltonians

    • qml.VQECost: a class for quickly constructing a differentiable cost function given a circuit ansatz, Hamiltonian, and one or more devices

      >>> H = qml.vqe.Hamiltonian(coeffs, obs)
      >>> cost = qml.VQECost(ansatz, hamiltonian, dev, interface="torch")
      >>> params = torch.rand([4, 3])
      >>> cost(params)
      tensor(0.0245, dtype=torch.float64)
      
  • Added a circuit drawing feature that provides a text-based representation of a QNode instance. It can be invoked via qnode.draw(). The user can specify to display variable names instead of variable values and choose either an ASCII or Unicode charset. (#446)

    Consider the following circuit as an example:

    @qml.qnode(dev)
    def qfunc(a, w):
        qml.Hadamard(0)
        qml.CRX(a, wires=[0, 1])
        qml.Rot(w[0], w[1], w[2], wires=[1])
        qml.CRX(-a, wires=[0, 1])
    
        return qml.expval(qml.PauliZ(0) @ qml.PauliZ(1))
    

    We can draw the circuit after it has been executed:

    >>> result = qfunc(2.3, [1.2, 3.2, 0.7])
    >>> print(qfunc.draw())
     0: ──H──╭C────────────────────────────╭C─────────╭┤ ⟨Z ⊗ Z⟩
     1: ─────╰RX(2.3)──Rot(1.2, 3.2, 0.7)──╰RX(-2.3)──╰┤ ⟨Z ⊗ Z⟩
    >>> print(qfunc.draw(charset="ascii"))
     0: --H--+C----------------------------+C---------+| <Z @ Z>
     1: -----+RX(2.3)--Rot(1.2, 3.2, 0.7)--+RX(-2.3)--+| <Z @ Z>
    >>> print(qfunc.draw(show_variable_names=True))
     0: ──H──╭C─────────────────────────────╭C─────────╭┤ ⟨Z ⊗ Z⟩
     1: ─────╰RX(a)──Rot(w[0], w[1], w[2])──╰RX(-1*a)──╰┤ ⟨Z ⊗ Z⟩
    
  • Added QAOAEmbedding and its parameter initialization as a new trainable template. (#442)

  • Added the qml.probs() measurement function, allowing QNodes to differentiate variational circuit probabilities on simulators and hardware. (#432)

    @qml.qnode(dev)
    def circuit(x):
        qml.Hadamard(wires=0)
        qml.RY(x, wires=0)
        qml.RX(x, wires=1)
        qml.CNOT(wires=[0, 1])
        return qml.probs(wires=[0])
    

    Executing this circuit gives the marginal probability of wire 1:

    >>> circuit(0.2)
    [0.40066533 0.59933467]
    

    QNodes that return probabilities fully support autodifferentiation.

  • Added the convenience load functions qml.from_pyquil, qml.from_quil and qml.from_quil_file that convert pyQuil objects and Quil code to PennyLane templates. This feature requires version 0.8 or above of the PennyLane-Forest plugin. (#459)

  • Added a qml.inv method that inverts templates and sequences of Operations. Added a @qml.template decorator that makes templates return the queued Operations. (#462)

    For example, using this function to invert a template inside a QNode:

    @qml.template
    def ansatz(weights, wires):
        for idx, wire in enumerate(wires):
            qml.RX(weights[idx], wires=[wire])
    
        for idx in range(len(wires) - 1):
            qml.CNOT(wires=[wires[idx], wires[idx + 1]])
    
    dev = qml.device('default.qubit', wires=2)
    
    @qml.qnode(dev)
    def circuit(weights):
        qml.inv(ansatz(weights, wires=[0, 1]))
        return qml.expval(qml.PauliZ(0) @ qml.PauliZ(1))
    
  • Added the QNodeCollection container class, that allows independent QNodes to be stored and evaluated simultaneously. Experimental support for asynchronous evaluation of contained QNodes is provided with the parallel=True keyword argument. (#466)

  • Added a high level qml.map function, that maps a quantum circuit template over a list of observables or devices, returning a QNodeCollection. (#466)

    For example:

    >>> def my_template(params, wires, **kwargs):
    >>>    qml.RX(params[0], wires=wires[0])
    >>>    qml.RX(params[1], wires=wires[1])
    >>>    qml.CNOT(wires=wires)
    
    >>> obs_list = [qml.PauliX(0) @ qml.PauliZ(1), qml.PauliZ(0) @ qml.PauliX(1)]
    >>> dev = qml.device("default.qubit", wires=2)
    >>> qnodes = qml.map(my_template, obs_list, dev, measure="expval")
    >>> qnodes([0.54, 0.12])
    array([-0.06154835  0.99280864])
    
  • Added high level qml.sum, qml.dot, qml.apply functions that act on QNode collections. (#466)

    qml.apply allows vectorized functions to act over the entire QNode collection:

    >>> qnodes = qml.map(my_template, obs_list, dev, measure="expval")
    >>> cost = qml.apply(np.sin, qnodes)
    >>> cost([0.54, 0.12])
    array([-0.0615095  0.83756375])
    

    qml.sum and qml.dot take the sum of a QNode collection, and a dot product of tensors/arrays/QNode collections, respectively.

Breaking changes

  • Deprecated the old-style QNode such that only the new-style QNode and its syntax can be used, moved all related files from the pennylane/beta folder to pennylane. (#440)

Improvements

  • Added the Tensor.prune() method and the Tensor.non_identity_obs property for extracting non-identity instances from the observables making up a Tensor instance. (#498)

  • Renamed the expt.tensornet and expt.tensornet.tf devices to default.tensor and default.tensor.tf. (#495)

  • Added a serialization method to the CircuitGraph class that is used to create a unique hash for each quantum circuit graph. (#470)

  • Added the Observable.eigvals method to return the eigenvalues of observables. (#449)

  • Added the Observable.diagonalizing_gates method to return the gates that diagonalize an observable in the computational basis. (#454)

  • Added the Operator.matrix method to return the matrix representation of an operator in the computational basis. (#454)

  • Added a QubitDevice class which implements common functionalities of plugin devices such that plugin devices can rely on these implementations. The new QubitDevice also includes a new execute method, which allows for more convenient plugin design. In addition, QubitDevice also unifies the way samples are generated on qubit-based devices. (#452) (#473)

  • Improved documentation of AmplitudeEmbedding and BasisEmbedding templates. (#441) (#439)

  • Codeblocks in the documentation now have a ‘copy’ button for easily copying examples. (#437)

Documentation

  • Update the developers plugin guide to use QubitDevice. (#483)

Bug fixes

  • Fixed a bug in CVQNode._pd_analytic, where non-descendant observables were not Heisenberg-transformed before evaluating the partial derivatives when using the order-2 parameter-shift method, resulting in an erroneous Jacobian for some circuits. (#433)

Contributors

This release contains contributions from (in alphabetical order):

Juan Miguel Arrazola, Ville Bergholm, Alain Delgado Gran, Olivia Di Matteo, Theodor Isacsson, Josh Izaac, Soran Jahangiri, Nathan Killoran, Johannes Jakob Meyer, Zeyue Niu, Maria Schuld, Antal Száva.

orphan

Release 0.7.0

New features since last release

  • Custom padding constant in AmplitudeEmbedding is supported (see ‘Breaking changes’.) (#419)

  • StronglyEntanglingLayer and RandomLayer now work with a single wire. (#409) (#413)

  • Added support for applying the inverse of an Operation within a circuit. (#377)

  • Added an OperationRecorder() context manager, that allows templates and quantum functions to be executed while recording events. The recorder can be used with and without QNodes as a debugging utility. (#388)

  • Operations can now specify a decomposition that is used when the desired operation is not supported on the target device. (#396)

  • The ability to load circuits from external frameworks as templates has been added via the new qml.load() function. This feature requires plugin support — this initial release provides support for Qiskit circuits and QASM files when pennylane-qiskit is installed, via the functions qml.from_qiskit and qml.from_qasm. (#418)

  • An experimental tensor network device has been added (#416) (#395) (#394) (#380)

  • An experimental tensor network device which uses TensorFlow for backpropagation has been added (#427)

  • Custom padding constant in AmplitudeEmbedding is supported (see ‘Breaking changes’.) (#419)

Breaking changes

  • The pad parameter in AmplitudeEmbedding() is now either None (no automatic padding), or a number that is used as the padding constant. (#419)

  • Initialization functions now return a single array of weights per function. Utilities for multi-weight templates Interferometer() and CVNeuralNetLayers() are provided. (#412)

  • The single layer templates RandomLayer(), CVNeuralNetLayer() and StronglyEntanglingLayer() have been turned into private functions _random_layer(), _cv_neural_net_layer() and _strongly_entangling_layer(). Recommended use is now via the corresponding Layers() templates. (#413)

Improvements

  • Added extensive input checks in templates. (#419)

  • Templates integration tests are rewritten - now cover keyword/positional argument passing, interfaces and combinations of templates. (#409) (#419)

  • State vector preparation operations in the default.qubit plugin can now be applied to subsets of wires, and are restricted to being the first operation in a circuit. (#346)

  • The QNode class is split into a hierarchy of simpler classes. (#354) (#398) (#415) (#417) (#425)

  • Added the gates U1, U2 and U3 parametrizing arbitrary unitaries on 1, 2 and 3 qubits and the Toffoli gate to the set of qubit operations. (#396)

  • Changes have been made to accomodate the movement of the main function in pytest._internal to pytest._internal.main in pip 19.3. (#404)

  • Added the templates BasisStatePreparation and MottonenStatePreparation that use gates to prepare a basis state and an arbitrary state respectively. (#336)

  • Added decompositions for BasisState and QubitStateVector based on state preparation templates. (#414)

  • Replaces the pseudo-inverse in the quantum natural gradient optimizer (which can be numerically unstable) with np.linalg.solve. (#428)

Contributors

This release contains contributions from (in alphabetical order):

Ville Bergholm, Josh Izaac, Nathan Killoran, Angus Lowe, Johannes Jakob Meyer, Oluwatobi Ogunbayo, Maria Schuld, Antal Száva.

orphan

Release 0.6.0

New features since last release

  • The devices default.qubit and default.gaussian have a new initialization parameter analytic that indicates if expectation values and variances should be calculated analytically and not be estimated from data. (#317)

  • Added C-SWAP gate to the set of qubit operations (#330)

  • The TensorFlow interface has been renamed from "tfe" to "tf", and now supports TensorFlow 2.0. (#337)

  • Added the S and T gates to the set of qubit operations. (#343)

  • Tensor observables are now supported within the expval, var, and sample functions, by using the @ operator. (#267)

Breaking changes

  • The argument n specifying the number of samples in the method Device.sample was removed. Instead, the method will always return Device.shots many samples. (#317)

Improvements

  • The number of shots / random samples used to estimate expectation values and variances, Device.shots, can now be changed after device creation. (#317)

  • Unified import shortcuts to be under qml in qnode.py and test_operation.py (#329)

  • The quantum natural gradient now uses scipy.linalg.pinvh which is more efficient for symmetric matrices than the previously used scipy.linalg.pinv. (#331)

  • The deprecated qml.expval.Observable syntax has been removed. (#267)

  • Remainder of the unittest-style tests were ported to pytest. (#310)

  • The do_queue argument for operations now only takes effect within QNodes. Outside of QNodes, operations can now be instantiated without needing to specify do_queue. (#359)

Documentation

  • The docs are rewritten and restructured to contain a code introduction section as well as an API section. (#314)

  • Added Ising model example to the tutorials (#319)

  • Added tutorial for QAOA on MaxCut problem (#328)

  • Added QGAN flow chart figure to its tutorial (#333)

  • Added missing figures for gallery thumbnails of state-preparation and QGAN tutorials (#326)

  • Fixed typos in the state preparation tutorial (#321)

  • Fixed bug in VQE tutorial 3D plots (#327)

Bug fixes

  • Fixed typo in measurement type error message in qnode.py (#341)

Contributors

This release contains contributions from (in alphabetical order):

Shahnawaz Ahmed, Ville Bergholm, Aroosa Ijaz, Josh Izaac, Nathan Killoran, Angus Lowe, Johannes Jakob Meyer, Maria Schuld, Antal Száva, Roeland Wiersema.

orphan

Release 0.5.0

New features since last release

  • Adds a new optimizer, qml.QNGOptimizer, which optimizes QNodes using quantum natural gradient descent. See https://arxiv.org/abs/1909.02108 for more details. (#295) (#311)

  • Adds a new QNode method, QNode.metric_tensor(), which returns the block-diagonal approximation to the Fubini-Study metric tensor evaluated on the attached device. (#295)

  • Sampling support: QNodes can now return a specified number of samples from a given observable via the top-level pennylane.sample() function. To support this on plugin devices, there is a new Device.sample method.

    Calculating gradients of QNodes that involve sampling is not possible. (#256)

  • default.qubit has been updated to provide support for sampling. (#256)

  • Added controlled rotation gates to PennyLane operations and default.qubit plugin. (#251)

Breaking changes

  • The method Device.supported was removed, and replaced with the methods Device.supports_observable and Device.supports_operation. Both methods can be called with string arguments (dev.supports_observable('PauliX')) and class arguments (dev.supports_observable(qml.PauliX)). (#276)

  • The following CV observables were renamed to comply with the new Operation/Observable scheme: MeanPhoton to NumberOperator, Homodyne to QuadOperator and NumberState to FockStateProjector. (#254)

Improvements

  • The AmplitudeEmbedding function now provides options to normalize and pad features to ensure a valid state vector is prepared. (#275)

  • Operations can now optionally specify generators, either as existing PennyLane operations, or by providing a NumPy array. (#295) (#313)

  • Adds a Device.parameters property, so that devices can view a dictionary mapping free parameters to operation parameters. This will allow plugin devices to take advantage of parametric compilation. (#283)

  • Introduces two enumerations: Any and All, representing any number of wires and all wires in the system respectively. They can be imported from pennylane.operation, and can be used when defining the Operation.num_wires class attribute of operations. (#277)

    As part of this change:

    • All is equivalent to the integer 0, for backwards compatibility with the existing test suite

    • Any is equivalent to the integer -1 to allow numeric comparison operators to continue working

    • An additional validation is now added to the Operation class, which will alert the user that an operation with num_wires = All is being incorrectly.

  • The one-qubit rotations in pennylane.plugins.default_qubit no longer depend on Scipy’s expm. Instead they are calculated with Euler’s formula. (#292)

  • Creates an ObservableReturnTypes enumeration class containing Sample, Variance and Expectation. These new values can be assigned to the return_type attribute of an Observable. (#290)

  • Changed the signature of the RandomLayer and RandomLayers templates to have a fixed seed by default. (#258)

  • setup.py has been cleaned up, removing the non-working shebang, and removing unused imports. (#262)

Documentation

  • A documentation refactor to simplify the tutorials and include Sphinx-Gallery. (#291)

    • Examples and tutorials previously split across the examples/ and doc/tutorials/ directories, in a mixture of ReST and Jupyter notebooks, have been rewritten as Python scripts with ReST comments in a single location, the examples/ folder.

    • Sphinx-Gallery is used to automatically build and run the tutorials. Rendered output is displayed in the Sphinx documentation.

    • Links are provided at the top of every tutorial page for downloading the tutorial as an executable python script, downloading the tutorial as a Jupyter notebook, or viewing the notebook on GitHub.

    • The tutorials table of contents have been moved to a single quick start page.

  • Fixed a typo in QubitStateVector. (#296)

  • Fixed a typo in the default_gaussian.gaussian_state function. (#293)

  • Fixed a typo in the gradient recipe within the RX, RY, RZ operation docstrings. (#248)

  • Fixed a broken link in the tutorial documentation, as a result of the qml.expval.Observable deprecation. (#246)

Bug fixes

  • Fixed a bug where a PolyXP observable would fail if applied to subsets of wires on default.gaussian. (#277)

Contributors

This release contains contributions from (in alphabetical order):

Simon Cross, Aroosa Ijaz, Josh Izaac, Nathan Killoran, Johannes Jakob Meyer, Rohit Midha, Nicolás Quesada, Maria Schuld, Antal Száva, Roeland Wiersema.

orphan

Release 0.4.0

New features since last release

  • pennylane.expval() is now a top-level function, and is no longer a package of classes. For now, the existing pennylane.expval.Observable interface continues to work, but will raise a deprecation warning. (#232)

  • Variance support: QNodes can now return the variance of observables, via the top-level pennylane.var() function. To support this on plugin devices, there is a new Device.var method.

    The following observables support analytic gradients of variances:

    • All qubit observables (requiring 3 circuit evaluations for involutory observables such as Identity, X, Y, Z; and 5 circuit evals for non-involutary observables, currently only qml.Hermitian)

    • First-order CV observables (requiring 5 circuit evaluations)

    Second-order CV observables support numerical variance gradients.

  • pennylane.about() function added, providing details on current PennyLane version, installed plugins, Python, platform, and NumPy versions (#186)

  • Removed the logic that allowed wires to be passed as a positional argument in quantum operations. This allows us to raise more useful error messages for the user if incorrect syntax is used. (#188)

  • Adds support for multi-qubit expectation values of the pennylane.Hermitian() observable (#192)

  • Adds support for multi-qubit expectation values in default.qubit. (#202)

  • Organize templates into submodules (#195). This included the following improvements:

    • Distinguish embedding templates from layer templates.

    • New random initialization functions supporting the templates available in the new submodule pennylane.init.

    • Added a random circuit template (RandomLayers()), in which rotations and 2-qubit gates are randomly distributed over the wires

    • Add various embedding strategies

Breaking changes

  • The Device methods expectations, pre_expval, and post_expval have been renamed to observables, pre_measure, and post_measure respectively. (#232)

Improvements

  • default.qubit plugin now uses np.tensordot when applying quantum operations and evaluating expectations, resulting in significant speedup (#239), (#241)

  • PennyLane now allows division of quantum operation parameters by a constant (#179)

  • Portions of the test suite are in the process of being ported to pytest. Note: this is still a work in progress.

    Ported tests include:

    • test_ops.py

    • test_about.py

    • test_classical_gradients.py

    • test_observables.py

    • test_measure.py

    • test_init.py

    • test_templates*.py

    • test_ops.py

    • test_variable.py

    • test_qnode.py (partial)

Bug fixes

  • Fixed a bug in Device.supported, which would incorrectly mark an operation as supported if it shared a name with an observable (#203)

  • Fixed a bug in Operation.wires, by explicitly casting the type of each wire to an integer (#206)

  • Removed code in PennyLane which configured the logger, as this would clash with users’ configurations (#208)

  • Fixed a bug in default.qubit, in which QubitStateVector operations were accidentally being cast to np.float instead of np.complex. (#211)

Contributors

This release contains contributions from:

Shahnawaz Ahmed, riveSunder, Aroosa Ijaz, Josh Izaac, Nathan Killoran, Maria Schuld.

orphan

Release 0.3.1

Bug fixes

  • Fixed a bug where the interfaces submodule was not correctly being packaged via setup.py

orphan

Release 0.3.0

New features since last release

  • PennyLane now includes a new interfaces submodule, which enables QNode integration with additional machine learning libraries.

  • Adds support for an experimental PyTorch interface for QNodes

  • Adds support for an experimental TensorFlow eager execution interface for QNodes

  • Adds a PyTorch+GPU+QPU tutorial to the documentation

  • Documentation now includes links and tutorials including the new PennyLane-Forest plugin.

Improvements

  • Printing a QNode object, via print(qnode) or in an interactive terminal, now displays more useful information regarding the QNode, including the device it runs on, the number of wires, it’s interface, and the quantum function it uses:

    >>> print(qnode)
    <QNode: device='default.qubit', func=circuit, wires=2, interface=PyTorch>
    

Contributors

This release contains contributions from:

Josh Izaac and Nathan Killoran.

orphan

Release 0.2.0

New features since last release

  • Added the Identity expectation value for both CV and qubit models (#135)

  • Added the templates.py submodule, containing some commonly used QML models to be used as ansatz in QNodes (#133)

  • Added the qml.Interferometer CV operation (#152)

  • Wires are now supported as free QNode parameters (#151)

  • Added ability to update stepsizes of the optimizers (#159)

Improvements

  • Removed use of hardcoded values in the optimizers, made them parameters (see #131 and #132)

  • Created the new PlaceholderExpectation, to be used when both CV and qubit expval modules contain expectations with the same name

  • Provide a way for plugins to view the operation queue before applying operations. This allows for on-the-fly modifications of the queue, allowing hardware-based plugins to support the full range of qubit expectation values. (#143)

  • QNode return values now support any form of sequence, such as lists, sets, etc. (#144)

  • CV analytic gradient calculation is now more robust, allowing for operations which may not themselves be differentiated, but have a well defined _heisenberg_rep method, and so may succeed operations that are analytically differentiable (#152)

Bug fixes

  • Fixed a bug where the variational classifier example was not batching when learning parity (see #128 and #129)

  • Fixed an inconsistency where some initial state operations were documented as accepting complex parameters - all operations now accept real values (#146)

Contributors

This release contains contributions from:

Christian Gogolin, Josh Izaac, Nathan Killoran, and Maria Schuld.

orphan

Release 0.1.0

Initial public release.

Contributors

This release contains contributions from:

Ville Bergholm, Josh Izaac, Maria Schuld, Christian Gogolin, and Nathan Killoran.