After setting Primal and Dual variables, retriving them returns all zeros

[Levi-Armstrong]

There seems to be a discrepancy between how these are stored and retrieved. I am using the setWarmStart to set both the initial primal and dual variable values but then I use the getPrimalVariable and getDualVariable methods to confirm they are set and it is returning all zeros. Should this be the case after setting the warm start values?

[S-Dafarra]

Hi @Levi-Armstrong, thanks for the issue! The osqp-eigen code is mainly an interface to the underlying osqp solver. The definition of the methods you mentioned is here:

osqp-eigen/include/OsqpEigen/Solver.tpp

Lines 300 to 455 in 19e1d8d

	template <typename T, int n, int m>
	bool OsqpEigen::Solver::setWarmStart(const Eigen::Matrix<T, n, 1>& primalVariable,
	const Eigen::Matrix<T, m, 1>& dualVariable)
	{
	if (!m_isSolverInitialized)
	{
	debugStream() << "[OsqpEigen::Solver::setWarmStart] The solver is not initialized"
	<< std::endl;
	return false;
	}
	if (primalVariable.rows() != getData()->n)
	{
	debugStream() << "[OsqpEigen::Solver::setWarmStart] The size of the primal variable vector "
	"has to be equal to "
	<< " the number of variables." << std::endl;
	return false;
	}
	if (dualVariable.rows() != getData()->m)
	{
	debugStream() << "[OsqpEigen::Solver::setWarmStart] The size of the dual variable vector "
	"has to be equal to "
	<< " the number of constraints." << std::endl;
	return false;
	}
	m_primalVariables = primalVariable.template cast<c_float>();
	m_dualVariables = dualVariable.template cast<c_float>();
	#ifdef OSQP_EIGEN_OSQP_IS_V1
	return (osqp_warm_start(m_solver.get(), m_primalVariables.data(), m_dualVariables.data()) == 0);
	#else
	return (osqp_warm_start(m_workspace.get(), m_primalVariables.data(), m_dualVariables.data())
	== 0);
	#endif
	}
	template <typename T, int n>
	bool OsqpEigen::Solver::setPrimalVariable(const Eigen::Matrix<T, n, 1>& primalVariable)
	{
	if (!m_isSolverInitialized)
	{
	debugStream() << "[OsqpEigen::Solver::setPrimalVariable] The solver is not initialized"
	<< std::endl;
	return false;
	}
	if (primalVariable.rows() != getData()->n)
	{
	debugStream() << "[OsqpEigen::Solver::setPrimalVariable] The size of the primal variable "
	"vector has to be equal to "
	<< " the number of variables." << std::endl;
	return false;
	}
	m_primalVariables = primalVariable.template cast<c_float>();
	#ifdef OSQP_EIGEN_OSQP_IS_V1
	return (osqp_warm_start(m_solver.get(), m_primalVariables.data(), nullptr) == 0);
	#else
	return (osqp_warm_start_x(m_workspace.get(), m_primalVariables.data()) == 0);
	#endif
	}
	template <typename T, int m>
	bool OsqpEigen::Solver::setDualVariable(const Eigen::Matrix<T, m, 1>& dualVariable)
	{
	if (dualVariable.rows() != getData()->m)
	{
	debugStream() << "[OsqpEigen::Solver::setDualVariable] The size of the dual variable "
	"vector has to be equal to "
	<< " the number of constraints." << std::endl;
	return false;
	}
	m_dualVariables = dualVariable.template cast<c_float>();
	#ifdef OSQP_EIGEN_OSQP_IS_V1
	return (osqp_warm_start(m_solver.get(), nullptr, m_dualVariables.data()) == 0);
	#else
	return (osqp_warm_start_y(m_workspace.get(), m_dualVariables.data()) == 0);
	#endif
	}
	template <typename T, int n>
	bool OsqpEigen::Solver::getPrimalVariable(Eigen::Matrix<T, n, 1>& primalVariable)
	{
	if (!m_isSolverInitialized)
	{
	debugStream() << "[OsqpEigen::Solver::getPrimalVariable] The solver is not initialized"
	<< std::endl;
	return false;
	}
	if (n == Eigen::Dynamic)
	{
	primalVariable.resize(getData()->n, 1);
	} else
	{
	if (n != getData()->n)
	{
	debugStream() << "[OsqpEigen::Solver::getPrimalVariable] The size of the vector has to "
	"be equal to the number of variables. (You can use an eigen dynamic "
	"vector)"
	<< std::endl;
	return false;
	}
	}
	#ifdef OSQP_EIGEN_OSQP_IS_V1
	primalVariable = Eigen::Map<Eigen::Matrix<c_float, n, 1>>(m_solver->solution->x, getData()->n)
	.template cast<T>();
	#else
	primalVariable
	= Eigen::Map<Eigen::Matrix<c_float, n, 1>>(m_workspace->x, getData()->n).template cast<T>();
	#endif
	return true;
	}
	template <typename T, int m>
	bool OsqpEigen::Solver::getDualVariable(Eigen::Matrix<T, m, 1>& dualVariable)
	{
	if (!m_isSolverInitialized)
	{
	debugStream() << "[OsqpEigen::Solver::getDualVariable] The solver is not initialized"
	<< std::endl;
	return false;
	}
	if (m == Eigen::Dynamic)
	{
	dualVariable.resize(getData()->m, 1);
	} else
	{
	if (m != getData()->m)
	{
	debugStream() << "[OsqpEigen::Solver::getDualVariable] The size of the vector has to "
	"be equal to the number of constraints. (You can use an eigen dynamic "
	"vector)"
	<< std::endl;
	return false;
	}
	}
	#ifdef OSQP_EIGEN_OSQP_IS_V1
	dualVariable = Eigen::Map<Eigen::Matrix<c_float, m, 1>>(m_solver->solution->y, getData()->m)
	.template cast<T>();
	#else
	dualVariable
	= Eigen::Map<Eigen::Matrix<c_float, m, 1>>(m_workspace->y, getData()->m).template cast<T>();
	#endif
	return true;
	}

For the primal variables in particular, they are get from the solution. Hence getPrimalVariable and getDualVariable should be meaningful only after performing the optimization.

So, short answer is no. Not sure if it is possible to get the warm start from osqp anyway.