Lightning

Usage

Development

API

  1. Docs
  2. Overview
  3. Template Class AdjointJacobian

Template Class AdjointJacobian

Inheritance Relationships

Base Type

  • public AdjointJacobianBase< StateVectorT, AdjointJacobian< StateVectorT > >

Class Documentation

template<class StateVectorT>
class AdjointJacobian : public AdjointJacobianBase<StateVectorT, AdjointJacobian<StateVectorT>>

Adjoint Jacobian evaluator following the method of arXiv:2009.02823.

Template Parameters:

StateVectorT – State vector type.

Public Functions

AdjointJacobian() = default
inline void adjointJacobianSingleObservable(std::span<PrecisionT> jac, const OpsData<StateVectorT> &ops, const std::vector<std::string> &ops_name, const std::vector<std::size_t> &tp, const Observable<StateVectorT> &ob, StateVectorLQubitManaged<PrecisionT> &lambda)

Helper function to calculate the Jacobian for a single observale.

This method is differnet from adjointJacobian in that it is optimized for the case where there is only a single observable in the circuit via the use of only two statevectors rather than a vector of statevectors.

Parameters:

  • jac – Preallocated vector for Jacobian data results.

  • ops – The operations from the PennyLane tape

  • ops_name – The operation names from the PennyLane tape

  • tp – The trainable parameters retrieved from the JacobianData

  • ob – The (single) observable to apply

  • lambda – The final statevector of the forward pass of the circuit prior to calculation.

inline void adjointJacobian(std::span<PrecisionT> jac, const JacobianData<StateVectorT> &jd, const StateVectorT &ref_data = {0}, bool apply_operations = false)

Calculates the Jacobian for the statevector for the selected set of parametric gates.

For the statevector data associated with psi of length num_elements, we make internal copies, one per required observable. The operations will be applied to the internal statevector copies, with the operation indices participating in the gradient calculations given in trainableParams, and the overall number of parameters for the gradient calculation provided within num_params. The resulting row-major ordered jac matrix representation will be of size jd.getSizeStateVec() * jd.getObservables().size(). OpenMP is used to enable independent operations to be offloaded to threads.

Note

Only gates with pre-defined generators can be differentiated. For example, QubitUnitary is not differentiable as there is no generator defined for this gate.

Note

For circuits with single observable, this method dispatches to adjointJacobianSingleObservable for higher performance.

Parameters:

  • jac – Preallocated vector for Jacobian data results.

  • jd – JacobianData represents the QuantumTape to differentiate.

  • apply_operations – Indicate whether to apply operations to tape.psi prior to calculation.

Previous
Next