Catalyst v0.8.1

New features

• The catalyst.mitigate_with_zne error mitigation compilation pass now supports the option to fold gates locally as well as the existing method of globally. (#1006) (#1129)

  While global folding applies the scale factor by forming the inverse of the entire quantum circuit (without measurements) and repeating the circuit with its inverse, local folding instead inserts per-gate folding sequences directly in place of each gate in the original circuit.

  For example,

  import jax
  import pennylane as qml
  from catalyst import qjit, mitigate_with_zne
  from pennylane.transforms import exponential_extrapolate
  
  dev = qml.device("lightning.qubit", wires=4, shots=5)
  
  @qml.qnode(dev)
  def circuit():
    qml.Hadamard(wires=0)
    qml.CNOT(wires=[0, 1])
    return qml.expval(qml.PauliY(wires=0))
  
  @qjit(keep_intermediate=True)
  def mitigated_circuit():
    s = jax.numpy.array([1, 2, 3])
    return mitigate_with_zne(
      circuit,
      scale_factors=s,
      extrapolate=exponential_extrapolate,
      folding="local-all" # "local-all" for local on all gates or "global" for the original method (default being "global")
    )()

  >>> circuit()
  >>> mitigated_circuit()

Improvements

• Fixes an issue where certain JAX linear algebra functions from jax.scipy.linalg gave incorrect results when invoked from within a qjit block, and adds full support for other jax.scipy.linalg functions. (#1097)

  The supported linear algebra functions include, but are not limited to:

  • jax.scipy.linalg.cholesky
  • jax.scipy.linalg.expm
  • jax.scipy.linalg.funm
  • jax.scipy.linalg.hessenberg
  • jax.scipy.linalg.lu
  • jax.scipy.linalg.lu_solve
  • jax.scipy.linalg.polar
  • jax.scipy.linalg.qr
  • jax.scipy.linalg.schur
  • jax.scipy.linalg.solve
  • jax.scipy.linalg.sqrtm
  • jax.scipy.linalg.svd

Breaking changes

• The argument scale_factors of mitigate_with_zne function now follows the proper literature definition. It now needs to be a list of positive odd integers, as we don't support the fractional part. (#1120)

Bug fixes

• Those functions calling the gather_p primitive (like jax.scipy.linalg.expm) can now be used in multiple qjits in a single program. (#1096)

Contributors

This release contains contributions from (in alphabetical order):

Joey Carter,
Alessandro Cosentino,
Paul Haochen Wang,
David Ittah,
Romain Moyard,
Daniel Strano,
Raul Torres.