Optimization problem solver for quantum optimal control. More...

#include <optimproblem.hpp>

Collaboration diagram for OptimProblem:

[legend]

Public Member Functions
	OptimProblem (const Config &config, OptimTarget optim_target_, TimeStepper timestepper_, MasterEq mastereq_, MPI_Comm comm_init_, MPI_Comm comm_optim, Output output_, bool quietmode=false)
	Constructor for optimization problem.

	~OptimProblem ()

int	getNdesign ()

double	getObjective ()

double	getCostT ()

double	getRegul ()

double	getPenaltyLeakage ()

double	getPenaltyWeightedCost ()

double	getPenaltyDpDm ()

double	getPenaltyVariation ()

double	getPenaltyEnergy ()

double	getFidelity ()

double	getTolFinalCost ()

double	getTolGradAbs ()

double	getTolInfidelity ()

int	getMPIrank_world ()

int	getMaxIter ()

int	getOutputOptimizationStride ()

Output *	getOutput ()

TimeStepper *	getTimeStepper ()

double	evalF (const Vec x)
	Evaluates the objective function F(x).

void	evalGradF (const Vec x, Vec G)
	Evaluates the gradient of the objective function with respect to the control parameters.

void	solve (Vec xinit)
	Runs the optimization solver.

void	getStartingPoint (Vec x)
	Computes initial guess for optimization variables.

void	getSolution (Vec *opt)
	Retrieves the optimization solution and prints summary information.

Public Attributes
Vec	xlower

Vec	xupper
	Lower and upper bounds for optimization variables.

Vec	xprev
	Design vector at previous iteration.

Vec	xinit
	Initial design vector.

Protected Attributes
size_t	ninit
	Number of initial conditions to be considered (N^2, N, or 1)

int	ninit_local
	Local number of initial conditions on this processor.

Vec	rho_t0
	Storage for initial condition of the ODE.

Vec	rho_t0_bar
	Storage for adjoint initial condition of the adjoint ODE (aka the terminal condition)

OptimTarget *	optim_target
	Pointer to the optimization target (gate or state)

MPI_Comm	comm_init
	MPI communicator for initial condition parallelization.

MPI_Comm	comm_optim
	MPI communicator for optimization parallelization, currently not used (size 1)

int	mpirank_optim

int	mpisize_optim
	MPI rank and size for optimization communicator.

int	mpirank_petsc

int	mpisize_petsc
	MPI rank and size for spatial parallelization (PETSc)

int	mpirank_world

int	mpisize_world
	MPI rank and size for global communicator.

int	mpirank_init

int	mpisize_init
	MPI rank and size for initial condition communicator.

bool	quietmode
	Flag for quiet mode operation.

std::vector< double >	obj_weights
	Weights for averaging objective over initial conditions.

int	ndesign
	Number of global design (optimization) parameters.

double	objective
	Current objective function value (sum over final-time cost, regularization terms and penalty terms)

double	obj_cost
	Final-time measure J(T) in objective.

double	obj_regul
	Regularization term in objective.

double	obj_penal_leakage
	Penalty term for leakage into guard levels.

double	obj_penal_weightedcost
	Penalty term for weighted running cost.

double	obj_penal_dpdm
	Penalty term second-order state derivatives (penalizes variations of the state evolution)

double	obj_penal_variation
	Penalty term for variation of control parameters.

double	obj_penal_energy
	Energy penalty term in objective.

double	fidelity
	Final-time fidelity: 1/ninit sum_i Tr(rho_target^dag rho(T)) for Lindblad, \|1/ninit sum_i phi_target^dag phi\|^2 for Schrodinger.

double	gnorm
	Current norm of gradient.

double	gamma_tikhonov
	Parameter for Tikhonov regularization.

bool	tikhonov_use_x0
	Switch to use \|\|x - x0\|\|^2 for Tikhonov regularization instead of \|\|x\|\|^2.

double	gamma_penalty_leakage
	Parameter multiplying integral leakage term.

double	gamma_penalty_weightedcost
	Parameter multiplying integral weighted cost function.

double	gamma_penalty_dpdm
	Parameter multiplying integral penalty term for 2nd derivative of state variation.

double	gamma_penalty_energy
	Parameter multiplying energy penalty.

double	gamma_penalty_variation
	Parameter multiplying finite-difference squared regularization term.

double	tol_grad_abs
	Stopping criterion based on absolute gradient norm.

double	tol_grad_rel
	Stopping criterion based on relative gradient norm.

double	tol_final_cost
	Stopping criterion based on objective function value.

double	tol_infidelity
	Stopping criterion based on infidelity.

int	maxiter
	Stopping criterion based on maximum number of iterations.

Tao	tao
	PETSc's TAO optimization solver.

double *	mygrad
	Auxiliary gradient storage.

Vec	xtmp
	Temporary vector storage.

int	output_optimization_stride
	Write output files every N optimization iterations.

TimeStepper *	timestepper
	Pointer to time-stepping scheme.

Output *	output
	Pointer to output handler.

MasterEq *	mastereq
	Pointer to master equation solver.

Detailed Description

Optimization problem solver for quantum optimal control.

This class manages the optimization of quantum control pulses using PETSc's TAO optimization library. It handles objective function evaluation, by propagating initial states forward in time solving the dynamical equation and computing the final-time objective cost function and integral penalty terms, as well as the gradient computation by backpropagating the adjoint terminal states backwards in time solving the adjoint dynamical equation and collecting gradient contributions. It further defines the interface functions for PETSc's TAO optimization via L-BFGS, including a callback function to monitor optimization progress.

Main functionality:

evalF evaluates the objective function by calling TimeStepper::solveODE to evolve initial states to final time T and summing up the objective function measure over each target state
evalGradF evaluates the objective function and its gradient with respect to the optimization parameters by calling TimeStepper::solveODE and TimeStepper::solveAdjointODE to propagate initial states forward and backward through the time domain while accumulating objective and gradient information.

This class contains references to:

TimeStepper for handling the forward and backward time stepping process
OptimTarget for evaluating the final-time cost for each initial condition
Output for writing monitored optimization convergence to file

Constructor & Destructor Documentation

◆ OptimProblem()

OptimProblem::OptimProblem	(	const Config &	config,
		OptimTarget *	optim_target_,
		TimeStepper *	timestepper_,
		MasterEq *	mastereq_,
		MPI_Comm	comm_init_,
		MPI_Comm	comm_optim,
		Output *	output_,
		bool	quietmode = `false`
	)

Constructor for optimization problem.

Parameters

config	Configuration parameters from input file
optim_target_	Pointer to optimization target
timestepper_	Pointer to time-stepping scheme
mastereq_	Pointer to master equation solver
comm_init_	MPI communicator for initial condition parallelization
comm_optim	MPI communicator for optimization parallelization
output_	Pointer to output handler
quietmode	Flag for quiet operation (default: false)

◆ ~OptimProblem()

OptimProblem::~OptimProblem ( )

Member Function Documentation

◆ evalF()

double OptimProblem::evalF ( const Vec x )

Evaluates the objective function F(x).

Performs forward simulations for each initial conditions and evaluates the objective function.

Parameters

x	Design vector

Returns: double Objective function value

◆ evalGradF()

void OptimProblem::evalGradF	(	const Vec	x,
		Vec	G
	)

Evaluates the gradient of the objective function with respect to the control parameters.

Parameters

x	Design (optimization) vector
G	Gradient vector to store result

◆ getCostT()

double OptimProblem::getCostT ( )

inline

◆ getFidelity()

double OptimProblem::getFidelity ( )

inline

◆ getMaxIter()

int OptimProblem::getMaxIter ( )

inline

◆ getMPIrank_world()

int OptimProblem::getMPIrank_world ( )

inline

◆ getNdesign()

int OptimProblem::getNdesign ( )

inline

◆ getObjective()

double OptimProblem::getObjective ( )

inline

◆ getOutput()

Output * OptimProblem::getOutput ( )

inline

◆ getOutputOptimizationStride()

int OptimProblem::getOutputOptimizationStride ( )

inline

◆ getPenaltyDpDm()

double OptimProblem::getPenaltyDpDm ( )

inline

◆ getPenaltyEnergy()

double OptimProblem::getPenaltyEnergy ( )

inline

◆ getPenaltyLeakage()

double OptimProblem::getPenaltyLeakage ( )

inline

◆ getPenaltyVariation()

double OptimProblem::getPenaltyVariation ( )

inline

◆ getPenaltyWeightedCost()

double OptimProblem::getPenaltyWeightedCost ( )

inline

◆ getRegul()

double OptimProblem::getRegul ( )

inline

◆ getSolution()

void OptimProblem::getSolution ( Vec * opt )

Retrieves the optimization solution and prints summary information.

This method should be called after TaoSolve() has finished.

Parameters

opt	Pointer to vector to store the optimal solution

◆ getStartingPoint()

void OptimProblem::getStartingPoint ( Vec x )

Computes initial guess for optimization variables.

Parameters

x	Vector to store the initial guess

◆ getTimeStepper()

TimeStepper * OptimProblem::getTimeStepper ( )

inline

◆ getTolFinalCost()

double OptimProblem::getTolFinalCost ( )

inline

◆ getTolGradAbs()

double OptimProblem::getTolGradAbs ( )

inline

◆ getTolInfidelity()

double OptimProblem::getTolInfidelity ( )

inline

◆ solve()

void OptimProblem::solve ( Vec xinit )

Runs the optimization solver.

Parameters

xinit Initial guess for design variables

Member Data Documentation

◆ comm_init

MPI_Comm OptimProblem::comm_init

protected

MPI communicator for initial condition parallelization.

◆ comm_optim

MPI_Comm OptimProblem::comm_optim

protected

MPI communicator for optimization parallelization, currently not used (size 1)

◆ fidelity

double OptimProblem::fidelity

protected

Final-time fidelity: 1/ninit sum_i Tr(rho_target^dag rho(T)) for Lindblad, |1/ninit sum_i phi_target^dag phi|^2 for Schrodinger.

◆ gamma_penalty_dpdm

double OptimProblem::gamma_penalty_dpdm

protected

Parameter multiplying integral penalty term for 2nd derivative of state variation.

◆ gamma_penalty_energy

double OptimProblem::gamma_penalty_energy

protected

Parameter multiplying energy penalty.

◆ gamma_penalty_leakage

double OptimProblem::gamma_penalty_leakage

protected

Parameter multiplying integral leakage term.

◆ gamma_penalty_variation

double OptimProblem::gamma_penalty_variation

protected

Parameter multiplying finite-difference squared regularization term.

◆ gamma_penalty_weightedcost

double OptimProblem::gamma_penalty_weightedcost

protected

Parameter multiplying integral weighted cost function.

◆ gamma_tikhonov

double OptimProblem::gamma_tikhonov

protected

Parameter for Tikhonov regularization.

◆ gnorm

double OptimProblem::gnorm

protected

Current norm of gradient.

◆ mastereq

MasterEq* OptimProblem::mastereq

protected

Pointer to master equation solver.

◆ maxiter

int OptimProblem::maxiter

protected

Stopping criterion based on maximum number of iterations.

◆ mpirank_init

int OptimProblem::mpirank_init

protected

◆ mpirank_optim

int OptimProblem::mpirank_optim

protected

◆ mpirank_petsc

int OptimProblem::mpirank_petsc

protected

◆ mpirank_world

int OptimProblem::mpirank_world

protected

◆ mpisize_init

int OptimProblem::mpisize_init

protected

MPI rank and size for initial condition communicator.

◆ mpisize_optim

int OptimProblem::mpisize_optim

protected

MPI rank and size for optimization communicator.

◆ mpisize_petsc

int OptimProblem::mpisize_petsc

protected

MPI rank and size for spatial parallelization (PETSc)

◆ mpisize_world

int OptimProblem::mpisize_world

protected

MPI rank and size for global communicator.

◆ mygrad

double* OptimProblem::mygrad

protected

Auxiliary gradient storage.

◆ ndesign

int OptimProblem::ndesign

protected

Number of global design (optimization) parameters.

◆ ninit

size_t OptimProblem::ninit

protected

Number of initial conditions to be considered (N^2, N, or 1)

◆ ninit_local

int OptimProblem::ninit_local

protected

Local number of initial conditions on this processor.

◆ obj_cost

double OptimProblem::obj_cost

protected

Final-time measure J(T) in objective.

◆ obj_penal_dpdm

double OptimProblem::obj_penal_dpdm

protected

Penalty term second-order state derivatives (penalizes variations of the state evolution)

◆ obj_penal_energy

double OptimProblem::obj_penal_energy

protected

Energy penalty term in objective.

◆ obj_penal_leakage

double OptimProblem::obj_penal_leakage

protected

Penalty term for leakage into guard levels.

◆ obj_penal_variation

double OptimProblem::obj_penal_variation

protected

Penalty term for variation of control parameters.

◆ obj_penal_weightedcost

double OptimProblem::obj_penal_weightedcost

protected

Penalty term for weighted running cost.

◆ obj_regul

double OptimProblem::obj_regul

protected

Regularization term in objective.

◆ obj_weights

std::vector<double> OptimProblem::obj_weights

protected

Weights for averaging objective over initial conditions.

◆ objective

double OptimProblem::objective

protected

Current objective function value (sum over final-time cost, regularization terms and penalty terms)

◆ optim_target

OptimTarget* OptimProblem::optim_target

protected

Pointer to the optimization target (gate or state)

◆ output

Output* OptimProblem::output

protected

Pointer to output handler.

◆ output_optimization_stride

int OptimProblem::output_optimization_stride

protected

Write output files every N optimization iterations.

◆ quietmode

bool OptimProblem::quietmode

protected

Flag for quiet mode operation.

◆ rho_t0

Vec OptimProblem::rho_t0

protected

Storage for initial condition of the ODE.

◆ rho_t0_bar

Vec OptimProblem::rho_t0_bar

protected

Storage for adjoint initial condition of the adjoint ODE (aka the terminal condition)

◆ tao

Tao OptimProblem::tao

protected

PETSc's TAO optimization solver.

◆ tikhonov_use_x0

bool OptimProblem::tikhonov_use_x0

protected

Switch to use ||x - x0||^2 for Tikhonov regularization instead of ||x||^2.

◆ timestepper

TimeStepper* OptimProblem::timestepper

protected

Pointer to time-stepping scheme.

◆ tol_final_cost

double OptimProblem::tol_final_cost

protected

Stopping criterion based on objective function value.

◆ tol_grad_abs

double OptimProblem::tol_grad_abs

protected

Stopping criterion based on absolute gradient norm.

◆ tol_grad_rel

double OptimProblem::tol_grad_rel

protected

Stopping criterion based on relative gradient norm.

◆ tol_infidelity

double OptimProblem::tol_infidelity

protected

Stopping criterion based on infidelity.

◆ xinit

Vec OptimProblem::xinit

Initial design vector.

◆ xlower

Vec OptimProblem::xlower

◆ xprev

Vec OptimProblem::xprev

Design vector at previous iteration.

◆ xtmp

Vec OptimProblem::xtmp

protected

Temporary vector storage.

◆ xupper

Vec OptimProblem::xupper

Lower and upper bounds for optimization variables.

The documentation for this class was generated from the following files:

/home/runner/work/quandary/quandary/include/optimproblem.hpp
/home/runner/work/quandary/quandary/src/optimproblem.cpp

Public Member Functions

Public Attributes

Protected Attributes

Detailed Description

Constructor & Destructor Documentation

◆ OptimProblem()

◆ ~OptimProblem()

Member Function Documentation

◆ evalF()

◆ evalGradF()

◆ getCostT()

◆ getFidelity()

◆ getMaxIter()

◆ getMPIrank_world()

◆ getNdesign()

◆ getObjective()

◆ getOutput()

◆ getOutputOptimizationStride()

◆ getPenaltyDpDm()

◆ getPenaltyEnergy()

◆ getPenaltyLeakage()

◆ getPenaltyVariation()

◆ getPenaltyWeightedCost()

◆ getRegul()

◆ getSolution()

◆ getStartingPoint()

◆ getTimeStepper()

◆ getTolFinalCost()

◆ getTolGradAbs()

◆ getTolInfidelity()

◆ solve()

Member Data Documentation

◆ comm_init

◆ comm_optim

◆ fidelity

◆ gamma_penalty_dpdm

◆ gamma_penalty_energy

◆ gamma_penalty_leakage

◆ gamma_penalty_variation

◆ gamma_penalty_weightedcost

◆ gamma_tikhonov

◆ gnorm

◆ mastereq

◆ maxiter

◆ mpirank_init

◆ mpirank_optim

◆ mpirank_petsc

◆ mpirank_world

◆ mpisize_init

◆ mpisize_optim

◆ mpisize_petsc

◆ mpisize_world

◆ mygrad

◆ ndesign

◆ ninit

◆ ninit_local

◆ obj_cost

◆ obj_penal_dpdm

◆ obj_penal_energy

◆ obj_penal_leakage

◆ obj_penal_variation

◆ obj_penal_weightedcost

◆ obj_regul

◆ obj_weights

◆ objective

◆ optim_target

◆ output

◆ output_optimization_stride

◆ quietmode

◆ rho_t0

◆ rho_t0_bar

◆ tao

◆ tikhonov_use_x0

◆ timestepper

◆ tol_final_cost

◆ tol_grad_abs

◆ tol_grad_rel

◆ tol_infidelity

◆ xinit

◆ xlower