Solver Parameters¶

Setting Parameters¶

Use a Python alias or its native key with any of these interfaces:

m.setParam("TimeLimit", 60)
m.setParams(OptimalityTol=1e-6, FeasibilityTol=1e-6)
m.Params.LogLevel = 0
m.Params["time_sec_limit"] = 120
print(m.getParam("TimeLimit"))

Unknown names and invalid values raise an error. Backend selection and CPU thread configuration are documented under devices.

Parameter Reference¶

Execution and Logging¶

Parameter	Native key	Type	Default	Description
`Threads`	`num_threads`	int	0	CPU thread limit
`LogLevel`	`verbose`	int	1	0: silent; 1: summary; 2: iteration details
`Presolve`	`presolve`	bool	False	Enable PreFOS when built with presolve support

Verbosity, OutputFlag, and LogToConsole are also aliases for verbose.

Termination Criteria¶

Parameter	Native key	Type	Default	Description
`TimeLimit`	`time_sec_limit`	float	3600.0	Time limit in seconds
`IterationLimit`	`iteration_limit`	int	2147483647	Maximum iteration count
`OptimalityTol`	`eps_optimal_relative`	float	1e-4	Relative optimality tolerance
`FeasibilityTol`	`eps_feasible_relative`	float	1e-4	Relative feasibility tolerance
`InfeasibleTol`	`eps_infeasible`	float	1e-10	Infeasibility certificate tolerance
`TermCheckFreq`	`termination_evaluation_frequency`	int	200	Iterations between termination checks
`OptimalityNorm`	`optimality_norm`	str	"linf"	Residual norm: "l2" or "linf"
—	`feasibility_polishing`	bool	False	Enable feasibility polishing
—	`eps_feas_polish_relative`	float	1e-6	Relative polishing tolerance

Scaling and Quadratic Updates¶

Parameter	Native key	Type	Default	Description
`CurtisReidIters`	`curtis_reid_iterations`	int	0	Curtis–Reid scaling iterations; 0 disables it
`RuizIters`	`l_inf_ruiz_iterations`	int	10	L-inf Ruiz scaling iterations
`PCAlpha`	`pock_chambolle_alpha`	float	1.0	Pock–Chambolle scaling exponent
`UsePCAlpha`	`has_pock_chambolle_alpha`	bool	True	Enable Pock–Chambolle scaling
`BoundObjRescaling`	`bound_objective_rescaling`	bool	True	Rescale bounds and objective
`UseConePreservingScaling`	`use_cone_preserving_scaling`	bool	True	Use one scaling value per cone block
`ReflectionCoeff`	`reflection_coefficient`	float	1.0	Reflection coefficient
`NonDiagonalQuadraticMode`	`non_diagonal_quadratic_mode`	str	"inner"	Non-diagonal Q update: "inner" or "linearized"

Restarts¶

Parameter	Native key	Type	Default	Description
`RestartArtificialThresh`	`artificial_restart_threshold`	float	0.36	Artificial restart threshold
`RestartSufficientReduction`	`sufficient_reduction_for_restart`	float	0.2	Sufficient reduction threshold
`RestartNecessaryReduction`	`necessary_reduction_for_restart`	float	0.8	Necessary reduction threshold
`RestartKp`	`k_p`	float	0.99	Proportional coefficient for primal-weight updates

Inner Solver¶

Parameter	Native key	Type	Default	Description
`InnerIterLimit`	`inner_iter_limit`	int	1000	Maximum inner iteration count
`InnerInitTol`	`inner_init_tol`	float	1e-3	Initial inner tolerance
`InnerMinTol`	`inner_min_tol`	float	1e-9	Minimum inner tolerance
`DiagJacobiPrecond`	`diag_jacobi_precond`	bool	True	Jacobi diagonal preconditioner

Singular Value Estimation¶

Parameter	Native key	Type	Default	Description
`SVMaxIter`	`sv_max_iter`	int	5000	Maximum power-method iteration count
`SVTol`	`sv_tol`	float	1e-4	Power-method tolerance

Cone specifications are model data, not solver parameters. The lower-level Python API accepts native keys in its params dictionary.