Package de.labathome.trustconstr

Class MinimizeTrustConstr

java.lang.Object

de.labathome.trustconstr.MinimizeTrustConstr

public class MinimizeTrustConstr
extends Object

Java port of scipy.optimize.minimize(method='trust-constr')

Constructor Summary

Constructors

Constructor	Description
MinimizeTrustConstr()

Method Summary

Modifier and Type	Method	Description
static <C extends Constraint & Jacobian> OptimizeResult	minimize(Function<Matrix,Double> fun, Matrix x0, C constraint, int maxIter, double xtol, double gtol)	Most general convenience overload: only the objective function is supplied.
static <C extends Constraint & Jacobian> OptimizeResult	minimize(Function<Matrix,Double> fun, Function<Matrix,Matrix> grad, HessianUpdateStrategy strategy, Matrix x0, C constraint, int maxIter, double xtol, double gtol)	Convenience overload accepting any HessianUpdateStrategy (BFGS, SR1, custom).
static <C extends Constraint & Jacobian> OptimizeResult	minimize(Function<Matrix,Double> fun, Function<Matrix,Matrix> grad, Matrix x0, C constraint, int maxIter, double xtol, double gtol)	Convenience overload that approximates the objective Hessian via a fresh BFGS update strategy -- the scipy-default behaviour when hess is omitted from the user call.
static <C extends Constraint & Jacobian> OptimizeResult	minimize(Function<Matrix,Double> fun, Function<Matrix,Matrix> grad, Function<Matrix,Matrix> hess, Matrix x0, C constraint, int maxIter, double xtol, double gtol)	General-purpose entry point: routes to minimizeEqualityConstrained(java.util.function.Function<de.labathome.trustconstr.matrix.Matrix, java.lang.Double>, java.util.function.Function<de.labathome.trustconstr.matrix.Matrix, de.labathome.trustconstr.matrix.Matrix>, java.util.function.Function<de.labathome.trustconstr.matrix.Matrix, de.labathome.trustconstr.matrix.Matrix>, de.labathome.trustconstr.matrix.Matrix, C, int, double, double) when the constraint has only equality rows, and to a TrustRegionInteriorPoint-driven path when any inequality rows are present.
static OptimizeResult	minimize(Function<Matrix,Double> fun, Function<Matrix,Matrix> grad, Function<Matrix,Matrix> hess, Matrix x0, Object[] constraints, int maxIter, double xtol, double gtol)	Multi-constraint entry point: combines the supplied constraints into a CombinedConstraint and dispatches as in the single-constraint overload.
static <C extends Constraint & Jacobian> OptimizeResult	minimizeEqualityConstrained(Function<Matrix,Double> fun, Function<Matrix,Matrix> grad, Function<Matrix,Matrix> hess, Matrix x0, C eq, int maxIter, double xtol, double gtol)	Convenience entry point covering the equality-constrained slice of the trust-constr API: a scalar objective with an explicit gradient and Hessian, subject to optional pure-equality constraints from a single LinearConstraint or NonlinearConstraint.
static OptimizeResult	minimizeTrustConstr(ToDoubleBiFunction<Matrix,Object> fun, Matrix x0, Object args, BiFunction<Matrix,Object,Matrix> grad, BiFunction<Matrix,Object,Matrix> hess, HessianProduct hessp, Bounds bounds, Object constraints, double xTol, double gTol, double barrierTol, Optional<Boolean> sparseJacobian, IterationCallback callback, int maxIter, int verbose, Matrix finiteDifferenceRelStep, double initialConstraintPenalty, double initialTrustRadius, double initialBarrierParameter, double initialBarrierTolerance, ProjectionMethod factorizationMethod, boolean disp)	Minimize a scalar function subject to constraints.
static OptimizeResult	minimizeTrustConstr(ToDoubleBiFunction<Matrix,Object> fun, Matrix x0, Object args, BiFunction<Matrix,Object,Matrix> grad, BiFunction<Matrix,Object,Matrix> hess, HessianProduct hessp, Bounds bounds, Object constraints, double xTol, double gTol, double barrierTol, Optional<Boolean> sparseJacobian, IterationCallback callback, int maxIter, int verbose, Matrix finiteDifferenceRelStep, double initialConstraintPenalty, double initialTrustRadius, double initialBarrierParameter, double initialBarrierTolerance, ProjectionMethod factorizationMethod, boolean disp, ExecutorService workers)	Strict scipy _minimize_trustregion_constr(..., workers=N) counterpart.
static State	updateState(State state, Matrix x, boolean lastIterationFailed, ScalarFunction objective, PreparedConstraint[] preparedConstraints, long startTime, double trustRadius, double constraintPenalty, CGInfo cgInfo)	Refresh the equality-path State record after an outer iteration (the function value, gradient, optimality, constraint violation, trust-region and CG fields).
static StateIP	updateStateIP(StateIP state, Matrix x, boolean lastIterationFailed, ScalarFunction objective, PreparedConstraint[] preparedConstraints, long startTime, double trustRadius, double constraintPenalty, CGInfo cgInfo, double barrierParameter, double barrierTolerance)	Refresh the IP-path StateIP record after an outer iteration: delegates to updateState(de.labathome.trustconstr.records.State, de.labathome.trustconstr.matrix.Matrix, boolean, de.labathome.trustconstr.ScalarFunction, de.labathome.trustconstr.records.PreparedConstraint[], long, double, double, de.labathome.trustconstr.records.CGInfo) and then sets the barrier-specific fields.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Constructor Details

MinimizeTrustConstr

public MinimizeTrustConstr()

Method Details

updateState

public static State updateState(State state,
 Matrix x,
 boolean lastIterationFailed,
 ScalarFunction objective,
 PreparedConstraint[] preparedConstraints,
 long startTime,
 double trustRadius,
 double constraintPenalty,
 CGInfo cgInfo)

Refresh the equality-path State record after an outer iteration (the function value, gradient, optimality, constraint violation, trust-region and CG fields).

Parameters:

state - outer-iteration state to update

x - current iterate

lastIterationFailed - whether the previous step was rejected

objective - wrapped objective providing eval-count metadata

preparedConstraints - prepared constraint records

startTime - wall-clock System.nanoTime() at the start of the run

trustRadius - current trust-region radius

constraintPenalty - current merit-function penalty

cgInfo - info from the projected-CG inner solve

Returns:

the same state after the in-place mutation

updateStateIP

public static StateIP updateStateIP(StateIP state,
 Matrix x,
 boolean lastIterationFailed,
 ScalarFunction objective,
 PreparedConstraint[] preparedConstraints,
 long startTime,
 double trustRadius,
 double constraintPenalty,
 CGInfo cgInfo,
 double barrierParameter,
 double barrierTolerance)

Refresh the IP-path StateIP record after an outer iteration: delegates to updateState(de.labathome.trustconstr.records.State, de.labathome.trustconstr.matrix.Matrix, boolean, de.labathome.trustconstr.ScalarFunction, de.labathome.trustconstr.records.PreparedConstraint[], long, double, double, de.labathome.trustconstr.records.CGInfo) and then sets the barrier-specific fields.

Parameters:

state - outer-iteration IP state to update

x - current iterate

lastIterationFailed - whether the previous step was rejected

objective - wrapped objective providing eval-count metadata

preparedConstraints - prepared constraint records

startTime - wall-clock System.nanoTime() at the start of the run

trustRadius - current trust-region radius

constraintPenalty - current merit-function penalty

cgInfo - info from the projected-CG inner solve

barrierParameter - current log-barrier coefficient

barrierTolerance - current barrier-subproblem tolerance

Returns:

the same state after the in-place mutation

minimizeEqualityConstrained

public static <C extends Constraint & Jacobian>
OptimizeResult minimizeEqualityConstrained(Function<Matrix,Double> fun,
 Function<Matrix,Matrix> grad,
 Function<Matrix,Matrix> hess,
 Matrix x0,
 C eq,
 int maxIter,
 double xtol,
 double gtol)

Convenience entry point covering the equality-constrained slice of the trust-constr API: a scalar objective with an explicit gradient and Hessian, subject to optional pure-equality constraints from a single LinearConstraint or NonlinearConstraint.

This is a focused subset of scipy.optimize.minimize(method='trust-constr') intended for problems where every constraint row is an equality (lb[i] == ub[i]). It dispatches to EqualityConstrainedSQP.eqSQP(de.labathome.trustconstr.interfaces.IFunctionAndConstraint, de.labathome.trustconstr.interfaces.IGradientAndJacobian, de.labathome.trustconstr.interfaces.LagrangeHessian, de.labathome.trustconstr.matrix.Matrix, double, de.labathome.trustconstr.matrix.Matrix, de.labathome.trustconstr.matrix.Matrix, de.labathome.trustconstr.matrix.Matrix, de.labathome.trustconstr.interfaces.StoppingCriterion, de.labathome.trustconstr.records.State, double, double, de.labathome.trustconstr.enums.ProjectionMethod, de.labathome.trustconstr.matrix.Matrix, de.labathome.trustconstr.matrix.Matrix, de.labathome.trustconstr.interfaces.LinearOperator) directly without the full canonical-form constraint pipeline. Callers with inequalities, Bounds, multiple constraints, or finite-difference Hessians/Jacobians should wait for the full minimizeTrustConstr(java.util.function.ToDoubleBiFunction<de.labathome.trustconstr.matrix.Matrix, java.lang.Object>, de.labathome.trustconstr.matrix.Matrix, java.lang.Object, java.util.function.BiFunction<de.labathome.trustconstr.matrix.Matrix, java.lang.Object, de.labathome.trustconstr.matrix.Matrix>, java.util.function.BiFunction<de.labathome.trustconstr.matrix.Matrix, java.lang.Object, de.labathome.trustconstr.matrix.Matrix>, de.labathome.trustconstr.interfaces.HessianProduct, de.labathome.trustconstr.records.Bounds, java.lang.Object, double, double, double, java.util.Optional<java.lang.Boolean>, de.labathome.trustconstr.interfaces.IterationCallback, int, int, de.labathome.trustconstr.matrix.Matrix, double, double, double, double, de.labathome.trustconstr.enums.ProjectionMethod, boolean) orchestrator.

Type Parameters:

C - constraint type, must implement both Constraint and Jacobian

Parameters:

fun - objective function f : R^n -> R

grad - analytic gradient g : R^n -> R^n

hess - analytic Hessian H : R^n -> R^{n x n}

x0 - initial point as a column vector (n x 1)

eq - equality constraint, or null for unconstrained problems (the constraint object must satisfy nIneq() == 0)

maxIter - maximum number of outer iterations

xtol - stop when trustRadius < xtol

gtol - stop when optimality < gtol and constrViolation < gtol

Returns:

populated OptimizeResult

minimize

public static <C extends Constraint & Jacobian>
OptimizeResult minimize(Function<Matrix,Double> fun,
 Function<Matrix,Matrix> grad,
 Function<Matrix,Matrix> hess,
 Matrix x0,
 C constraint,
 int maxIter,
 double xtol,
 double gtol)

General-purpose entry point: routes to minimizeEqualityConstrained(java.util.function.Function<de.labathome.trustconstr.matrix.Matrix, java.lang.Double>, java.util.function.Function<de.labathome.trustconstr.matrix.Matrix, de.labathome.trustconstr.matrix.Matrix>, java.util.function.Function<de.labathome.trustconstr.matrix.Matrix, de.labathome.trustconstr.matrix.Matrix>, de.labathome.trustconstr.matrix.Matrix, C, int, double, double) when the constraint has only equality rows, and to a TrustRegionInteriorPoint-driven path when any inequality rows are present. Pure unconstrained problems route through the equality path with an empty constraint set.

Like minimizeEqualityConstrained(java.util.function.Function<de.labathome.trustconstr.matrix.Matrix, java.lang.Double>, java.util.function.Function<de.labathome.trustconstr.matrix.Matrix, de.labathome.trustconstr.matrix.Matrix>, java.util.function.Function<de.labathome.trustconstr.matrix.Matrix, de.labathome.trustconstr.matrix.Matrix>, de.labathome.trustconstr.matrix.Matrix, C, int, double, double), this is a focused subset of the full scipy.optimize.minimize(method='trust-constr') API: a single LinearConstraint or NonlinearConstraint (or null), no Bounds, no callback, no finite-difference Hessians/Jacobians.

Type Parameters:

C - constraint type, must implement both Constraint and Jacobian

Parameters:

fun - objective function f : R^n -> R

grad - analytic gradient g : R^n -> R^n

hess - analytic Hessian H : R^n -> R^{nxn}

x0 - initial point (n x 1)

constraint - single constraint or null for unconstrained

maxIter - maximum outer iterations

xtol - stop when trustRadius < xtol

gtol - stop when optimality < gtol and constrViolation < gtol

Returns:

populated OptimizeResult

minimize

public static <C extends Constraint & Jacobian>
OptimizeResult minimize(Function<Matrix,Double> fun,
 Matrix x0,
 C constraint,
 int maxIter,
 double xtol,
 double gtol)

Most general convenience overload: only the objective function is supplied. The gradient is approximated by 2-point finite differences and the Hessian by a fresh BFGS update strategy. Mirrors scipy's default behaviour when both jac and hess are omitted.

Type Parameters:

C - constraint type, must implement Constraint and Jacobian

Parameters:

fun - objective function f : R^n -> R

x0 - initial point (n x 1)

constraint - single constraint or null

maxIter - maximum outer iterations

xtol - stop when trustRadius < xtol

gtol - stop when optimality < gtol and constrViolation < gtol

Returns:

populated OptimizeResult

minimize

public static <C extends Constraint & Jacobian>
OptimizeResult minimize(Function<Matrix,Double> fun,
 Function<Matrix,Matrix> grad,
 Matrix x0,
 C constraint,
 int maxIter,
 double xtol,
 double gtol)

Convenience overload that approximates the objective Hessian via a fresh BFGS update strategy -- the scipy-default behaviour when hess is omitted from the user call.

The BFGS update is driven manually inside the lagrHess closure: every time the SQP / IP outer loop asks for a fresh Lagrangian Hessian (i.e. after a successful step), we compare the current x and gradient to the previous values and call HessianUpdateStrategy.update(de.labathome.trustconstr.matrix.Matrix, de.labathome.trustconstr.matrix.Matrix) to refresh the approximation.

Type Parameters:

C - constraint type, must implement Constraint and Jacobian

Parameters:

fun - objective function f : R^n -> R

grad - analytic gradient g : R^n -> R^n

x0 - initial point (n x 1)

constraint - single constraint or null

maxIter - maximum outer iterations

xtol - stop when trustRadius < xtol

gtol - stop when optimality < gtol and constrViolation < gtol

Returns:

populated OptimizeResult

minimize

public static <C extends Constraint & Jacobian>
OptimizeResult minimize(Function<Matrix,Double> fun,
 Function<Matrix,Matrix> grad,
 HessianUpdateStrategy strategy,
 Matrix x0,
 C constraint,
 int maxIter,
 double xtol,
 double gtol)

Convenience overload accepting any HessianUpdateStrategy (BFGS, SR1, custom). The strategy is updated by the orchestrator after each successful iteration; users typically pass BFGS.FACTORY.build() or SR1.FACTORY.build().

Type Parameters:

C - constraint type, must implement Constraint and Jacobian

Parameters:

fun - objective function f : R^n -> R

grad - analytic gradient

strategy - Hessian-update strategy (BFGS, SR1, ...)

x0 - initial point (n x 1)

constraint - single constraint or null

maxIter - maximum outer iterations

xtol - stop when trustRadius < xtol

gtol - stop when optimality < gtol and constrViolation < gtol

Returns:

populated OptimizeResult

minimize

public static OptimizeResult minimize(Function<Matrix,Double> fun,
 Function<Matrix,Matrix> grad,
 Function<Matrix,Matrix> hess,
 Matrix x0,
 Object[] constraints,
 int maxIter,
 double xtol,
 double gtol)

Multi-constraint entry point: combines the supplied constraints into a CombinedConstraint and dispatches as in the single-constraint overload. Pass null or an empty array for unconstrained problems.

Parameters:

fun - objective function f : R^n -> R

grad - analytic gradient

hess - analytic Hessian

x0 - initial point (n x 1)

constraints - array of LinearConstraint / NonlinearConstraint (may be null)

maxIter - maximum outer iterations

xtol - stop when trustRadius < xtol

gtol - stop when optimality < gtol and constrViolation < gtol

Returns:

populated OptimizeResult

minimizeTrustConstr

public static OptimizeResult minimizeTrustConstr(ToDoubleBiFunction<Matrix,Object> fun,
 Matrix x0,
 Object args,
 BiFunction<Matrix,Object,Matrix> grad,
 BiFunction<Matrix,Object,Matrix> hess,
 HessianProduct hessp,
 Bounds bounds,
 Object constraints,
 double xTol,
 double gTol,
 double barrierTol,
 Optional<Boolean> sparseJacobian,
 IterationCallback callback,
 int maxIter,
 int verbose,
 Matrix finiteDifferenceRelStep,
 double initialConstraintPenalty,
 double initialTrustRadius,
 double initialBarrierParameter,
 double initialBarrierTolerance,
 ProjectionMethod factorizationMethod,
 boolean disp)

Minimize a scalar function subject to constraints. Full scipy-shape entry point: see Conn, Gould & Toint, Trust Region Methods (SIAM, 2000), p.19 for the algorithm and parameter recommendations.

Parameters:

fun - objective (x, args) -> f(x)

x0 - starting point (n x 1)

args - extra arguments forwarded to fun/grad/hess; may be null

grad - gradient (x, args) -> gradf(x); if null, computed by 2-point finite differences

hess - Hessian (x, args) -> grad^2f(x); if null, approximated by BFGS

hessp - Hessian-vector product; honored when hess is null (materialised via HessianLinearOperator)

bounds - box bounds on x; folded into the constraint set as a LinearConstraint; may be null

constraints - either a single LinearConstraint / NonlinearConstraint, an Object[] of such, or null

xTol - tolerance for termination by change in x: stop when tr_radius < xTol

gTol - tolerance for termination by Lagrangian gradient norm and constraint violation

barrierTol - IP-only: termination requires barrier parameter below this threshold

sparseJacobian - true/false forces all constraint Jacobians to that representation via SparsityForcedConstraint; Optional.empty() uses the constraint's native auto-detect

callback - per-iteration callback; return true to terminate. May be null

maxIter - maximum outer iterations

verbose - verbosity level (0 silent, 1 termination report, 2-3 per-iteration progress)

finiteDifferenceRelStep - relative step size for FD gradient/Hessian; may be null

initialConstraintPenalty - initial constraint penalty for the merit function f(x) + rho ||c(x)||_2

initialTrustRadius - initial trust-region radius

initialBarrierParameter - IP-only: initial barrier parameter for the log-barrier subproblem

initialBarrierTolerance - IP-only: initial inner-loop tolerance for the barrier subproblem

factorizationMethod - how to factor the equality-Jacobian for the projection: one of AUGMENTED_SYSTEM, QR_FACTORIZATION, SVD_FACTORIZATION, or null for auto-select (QR for dense, AUGMENTED_SYSTEM for sparse)

disp - if true, bumps verbose to 1 when it was 0

Returns:

populated OptimizeResult

minimizeTrustConstr

public static OptimizeResult minimizeTrustConstr(ToDoubleBiFunction<Matrix,Object> fun,
 Matrix x0,
 Object args,
 BiFunction<Matrix,Object,Matrix> grad,
 BiFunction<Matrix,Object,Matrix> hess,
 HessianProduct hessp,
 Bounds bounds,
 Object constraints,
 double xTol,
 double gTol,
 double barrierTol,
 Optional<Boolean> sparseJacobian,
 IterationCallback callback,
 int maxIter,
 int verbose,
 Matrix finiteDifferenceRelStep,
 double initialConstraintPenalty,
 double initialTrustRadius,
 double initialBarrierParameter,
 double initialBarrierTolerance,
 ProjectionMethod factorizationMethod,
 boolean disp,
 ExecutorService workers)

Strict scipy _minimize_trustregion_constr(..., workers=N) counterpart. Same arguments as minimizeTrustConstr(ToDoubleBiFunction, Matrix, Object, BiFunction, BiFunction, HessianProduct, Bounds, Object, double, double, double, Optional, IterationCallback, int, int, Matrix, double, double, double, double, ProjectionMethod, boolean), plus a final ExecutorService for parallel FD evaluation.

workers only affects the FD-gradient path (when grad is null); analytic gradients are unaffected.

Parameters:

fun - objective (x, args) -> f

x0 - starting iterate (n x 1)

args - extra arguments forwarded to fun/grad/hess

grad - analytic gradient (x, args) -> grad f(x), or null for FD

hess - analytic Hessian (x, args) -> H(x), or null for BFGS

hessp - matrix-free Hessian-vector product (x, p, args) -> H(x) p, or null

bounds - box bounds on the decision variables, or null

constraints - LinearConstraint, NonlinearConstraint, Object[] mix, or null

xTol - step-size convergence tolerance

gTol - KKT-optimality convergence tolerance

barrierTol - IP-path barrier-parameter convergence tolerance

sparseJacobian - forces sparse / dense Jacobian; Optional.empty() auto-detects

callback - per-iteration callback, returns true to terminate; null disables

maxIter - iteration limit

verbose - verbosity level (0 silent, 1 progress table)

finiteDifferenceRelStep - FD relative step (per-element vector); null uses the eps-derived default

initialConstraintPenalty - initial merit-function penalty

initialTrustRadius - initial trust-region radius

initialBarrierParameter - IP-path: initial barrier coefficient

initialBarrierTolerance - IP-path: initial barrier-subproblem tolerance

factorizationMethod - projection factorisation; null picks QR for dense / AugmentedSystem for sparse

disp - bumps verbose to 1 when it was 0

workers - executor for parallel FD evaluation; null = serial

Returns:

populated OptimizeResult