lpsolver 0.1.0 → 0.2.0

data/ext/lpsolver-highs/highs/ipm/ipx/iterate.h

@@ -0,0 +1,331 @@
+#ifndef IPX_ITERATE_H_
+#define IPX_ITERATE_H_
+
+#include <vector>
+#include "ipm/ipx/model.h"
+
+namespace ipx {
+
+// Iterate manages the IPM iterate consisting of primal variables x[n+m],
+// xl[n+m], xu[n+m] and dual variables y[m], zl[n+m], zu[n+m], where m, n are
+// the # rows and structural columns of the model.
+//
+// Each x[j] is in one of three "states":
+//
+// "fixed":     The variable is not moved by the IPM. The IPM works as if b was
+//              reduced by x[j]*AI[:,j] and column j of AI would not be present.
+// "free":      The variable is treated by the IPM as if lb[j] and ub[j] would
+//              be infinite, regardless of their actual values.
+// "barrier":   The variable has at least one finite bound and a barrier term is
+//              added by the IPM.
+//
+// Notice that the state refers to the way the variable is treated by the IPM,
+// not to its bounds in the model. For example, a variable of fixed state might
+// have zero, one or both finite bounds in the model.
+
+class Iterate {
+public:
+    // Constructs an Iterate object associated with a model. @model must be
+    // valid as long as the object is used; no data is copied. The constructor
+    // initializes variables and states to be consistent with finite/infinite
+    // bounds in the model.
+    explicit Iterate(const Model& model);
+
+    // Copies the given values into the Iterate object and sets the state of
+    // each variable to
+    // - free     if the variable has no finite bound in the model,
+    // - barrier  otherwise.
+    // The variables must have values that are consistent with the bounds of the
+    // model; that is:
+    // - If lb[j]==-inf, then   xl[j]==inf and zl[j]==0.0.
+    // - If lb[j]> -inf, then 0<xl[j]< inf and zl[j]> 0.0.
+    // - If ub[j]==+inf, then   xu[j]==inf and zu[j]==0.0.
+    // - If ub[j]< +inf, then 0<xu[j]< inf and zu[j]> 0.0.
+    // Otherwise an assertion will fail.
+    void Initialize(const Vector& x, const Vector& xl, const Vector& xu,
+                    const Vector& y, const Vector& zl, const Vector& zu);
+
+    // Adds sp*[dx;dxl;dxu] to the primal iterate and sd*[dy;dzl;dzu] to the
+    // dual iterate (restrictions see below). Each of the pointer arguments can
+    // be NULL, in which case its entries are assumed to be 0.0.
+    // - x[j] is updated only if StateOf(j) != State::fixed.
+    // - xl[j] and zl[j] are updated only if has_barrier_lb(j). The update to
+    //                   each of xl[j] and zl[j] is truncated such that
+    //                   xl[j] >= kBarrierMin and zl[j] >= kBarrierMin.
+    // - xu[j] and zu[j] are updated only if has_barrier_ub(j). The update to
+    //                   each of xu[j] and zu[j] is truncated such that
+    //                   xu[j] >= kBarrierMin and zu[j] >= kBarrierMin.
+    // If a variable is not updated, then the entry in the step direction is not
+    // accessed.
+    void Update(double sp, const double* dx, const double* dxl,
+                const double* dxu, double sd, const double* dy,
+                const double* dzl, const double* dzu);
+
+    const Model& model() const { return model_; }
+
+    const Vector& x() const  { return x_; }
+    const Vector& xl() const { return xl_; }
+    const Vector& xu() const { return xu_; }
+    const Vector& y() const  { return y_; }
+    const Vector& zl() const { return zl_; }
+    const Vector& zu() const { return zu_; }
+
+    double x(Int j) const  { return x_[j]; }
+    double xl(Int j) const { return xl_[j]; }
+    double xu(Int j) const { return xu_[j]; }
+    double y(Int j) const  { return y_[j]; }
+    double zl(Int j) const { return zl_[j]; }
+    double zu(Int j) const { return zu_[j]; }
+
+    // Returns const references to the residual vectors
+    // rb = b-AI*x,
+    // rl = lb-x+xl,
+    // ru = ub-x-xu,
+    // rc = c-AI'y-zl+zu.
+    // Warning: The residuals are not evaluated immediately after a change to
+    // an Iterate object (such as by Update() or make_fixed()). Instead, the
+    // four residual vectors are evaluated when any of them is required, e.g.
+    // in calls to rb(), presidual(), etc. Hence if a reference to a residual
+    // vector is stored and the Iterate object is changed, you must call e.g.
+    // presidual() to ensure that the residuals are updated.
+    const Vector& rb() const;
+    const Vector& rl() const;
+    const Vector& ru() const;
+    const Vector& rc() const;
+
+    enum class State { fixed, free, barrier };
+
+    // Returns the state of variable j.
+    State StateOf(Int j) const;
+
+    // Returns true if a barrier term exists for the lower or upper bound.
+    // The following expression is true for each variable:
+    // (StateOf(j)==State::barrier) == (has_barrier_lb(j)||has_barrier_ub(j))
+    bool has_barrier_lb(Int j) const;
+    bool has_barrier_ub(Int j) const;
+
+    // Returns true if variable j is "implied", i.e. StateOf(j)==State::free
+    // but the variable has at least one finite bound in the model.
+    bool is_implied(Int j) const;
+
+    // Changes state of variable j to fixed and sets xl[j]=xu[j]=zl[j]=zu[j]=0.
+    // If @value is given, sets x[j]=@value.
+    void make_fixed(Int j);
+    void make_fixed(Int j, double value);
+
+    // Changes the state of variable j to free.
+    // make_implied_lb(j) lb[j] must be finite.
+    //                    The method leaves zl[j] and zu[j] unchanged. The cost
+    //                    coefficient of variable j in the remaining LP becomes
+    //                    c[j]-zl[j]+zu[j].
+    //                    Postprocess() sets x[j]=lb[j], zu[j]=0 and computes
+    //                    zl[j] to satisfy dual feasibility.
+    // make_implied_ub(j) ub[j] must be finite.
+    //                    The method leaves zl[j] and zu[j] unchanged. The cost
+    //                    coefficient of variable j in the remaining LP becomes
+    //                    c[j]-zl[j]+zu[j].
+    //                    Postprocess() sets x[j]=ub[j], zl[j]=0 and computes
+    //                    zu[j] to satisfy dual feasibility.
+    // make_implied_eq(j) must have lb[j]==ub[j].
+    //                    The method sets zl[j]=0 and zu[j]=0.
+    //                    Postprocess sets x[j]=lb[j] and chooses either zl[j]
+    //                    or zu[j] to be nonzero depending on the sign of the
+    //                    reduced cost.
+    void make_implied_lb(Int j);
+    void make_implied_ub(Int j);
+    void make_implied_eq(Int j);
+
+    // Returns the IPM scaling factor for variable j, which is
+    // 0                                 if StateOf(j)==State::fixed,
+    // inf                               if StateOf(j)==State::free,
+    // 1.0/sqrt(zl[j]/xl[j]+zu[j]/xu[j]) if StateOf(j)==State::barrier.
+    // In the latter case the scaling factor is finite and positive.
+    double ScalingFactor(Int j) const;
+
+    // Returns the primal/dual objective value in the LP that is currently
+    // being solved (after fixing and implying variables).
+    double pobjective() const;
+    double dobjective() const;
+
+    // Returns the primal/dual objective value obtained after postprocessing.
+    double pobjective_after_postproc() const;
+    double dobjective_after_postproc() const;
+
+    // Returns the infinity norm of [rb;rl;ru] and rc.
+    double presidual() const;
+    double dresidual() const;
+
+    // complementarity() returns the sum of the pairwise complementarity
+    // products xl[j]*zl[j] and xu[j]*zu[j] from all barrier terms. mu()
+    // returns the average, mu_min() the minimum and mu_max() the maximum.
+    double complementarity() const;
+    double mu() const;
+    double mu_min() const;
+    double mu_max() const;
+
+    // Returns true if the relative primal and dual residuals are <=
+    // feasibility_tol().
+    bool feasible() const;
+
+    // Returns true if the relative objective gap is <= optimality_tol().
+    bool optimal() const;
+
+    // Returns true if the iterate satisfies the IPM termination
+    // criterion.  If start_crossover_tol() <= 0, the termination
+    // criterion is feasible() && optimal(). If start_crossover_tol()
+    // > 0, it is additionally required that the relative residuals
+    // which result from dropping the iterate to complementarity are
+    // <= start_crossover_tol().
+    bool term_crit_reached() const;
+
+    double feasibility_tol() const { return feasibility_tol_; }
+    double optimality_tol() const { return optimality_tol_; }
+    double start_crossover_tol() const { return start_crossover_tol_; }
+
+    void feasibility_tol(double new_tol) { feasibility_tol_ = new_tol; }
+    void optimality_tol(double new_tol) { optimality_tol_ = new_tol; }
+    void start_crossover_tol(double new_tol) { start_crossover_tol_ = new_tol; }
+
+    // Substitutes x[j], xl[j], xu[j], zl[j] and zu[j] for fixed and implied
+    // variables as defined by their state. After Postprocess() the object is
+    // invalid for use as an IPM iterate.
+    void Postprocess();
+
+    // Calls Model::EvaluateInteriorSolution().
+    void EvaluatePostsolved(Info* info) const;
+
+    // Copies y and constructs x and z from the iterate such that
+    // x[j]==lb[j] || x[j]==ub[j] || z[j]==0.0 is true for each j.
+    // The method can only be called after Postprocess().
+    void DropToComplementarity(Vector& x, Vector& y, Vector& z) const;
+
+    double bounds_measure_;
+    double costs_measure_;
+
+private:
+    // A (primal or dual) variable that is required to be positive in the IPM is
+    // not moved closer to zero than kBarrierMin.
+    static constexpr double kBarrierMin = 1e-30;
+
+    void assert_consistency();
+    void Evaluate() const;
+    void ComputeResiduals() const;
+    void ComputeObjectives() const;
+    void ComputeComplementarity() const;
+
+    // Computes the maximum primal and dual residual that results from dropping
+    // a barrier variable. Dropping means to set either x[j] to a bound or zl[j]
+    // and zu[j] to zero, where the decision is made by the ratios or primal to
+    // dual variables.
+    void ResidualsFromDropping(double* pres, double* dres) const;
+
+    // Internally, the state of each variable is subdivided as follows:
+    //
+    // State::barrier is represented by StateDetail::
+    //
+    //     BARRIER_LB   0 < xl < inf, xu = inf, zl > 0, zu = 0.
+    //                  The variable is moved by the barrier solver. It has a
+    //                  finite lower bound and an infinite upper bound.
+    //    
+    //     BARRIER_UB   xl = inf, 0 < xu < inf, zl = 0, zu > 0.
+    //                  The variable is moved by the barrier solver. It has an
+    //                  infinite lower bound and a finite upper bound.
+    //    
+    //     BARRIER_BOXED 0 < xl < inf, 0 < xu < inf, zl > 0, zu > 0.
+    //                   The variable is moved by the barrier solver. It has a
+    //                   finite lower bound and a finite upper bound.
+    //
+    // State::fixed is represented by StateDetail::
+    //
+    //     FIXED        xl = 0, xu = 0, zl = 0, zu = 0.
+    //                  The variable is fixed at its current x value.
+    //
+    // State::free is represented by StateDetail::
+    //
+    //     FREE         xl = inf, xu = inf, zl = 0, zu = 0.
+    //                  The variable is moved by the barrier solver. It has no
+    //                  finite bound.
+    //    
+    //     IMPLIED_LB   xl = inf, xu = inf, zl >= 0, zu >= 0
+    //                  The variable has a finite lb but is treated as free by
+    //                  the barrier solver. (The ub may or may not be finite.)
+    //                  After termination we set x = lb, zu = 0 and compute zl
+    //                  from dual feasibility.
+    //                  Evaluate() subtracts (zl-zu) from the cost coefficient
+    //                  when computing the primal objective value.
+    //    
+    //     IMPLIED_UB   xl = inf, xu = inf, zl >= 0, zu >= 0
+    //                  The variable has a finite ub but is treated as free by
+    //                  the barrier solver. (The lb may or may not be finite.)
+    //                  After termination we set x = ub, zl = 0 and compute zu
+    //                  from dual feasibility.
+    //                  Evaluate() subtracts (zl-zu) from the cost coefficient
+    //                  when computing the primal objective value.
+    //    
+    //     IMPLIED_EQ   xl = inf, xu = inf, zl = 0, zu = 0
+    //                  The variable has lb == ub but is treated as free by
+    //                  the barrier solver. After termination we set x = lb
+    //                  and compute either zl or zu from dual feasibility and
+    //                  set the other one to zero.
+    //
+    enum class StateDetail { BARRIER_LB, BARRIER_UB, BARRIER_BOXED, FREE, FIXED,
+                             IMPLIED_LB, IMPLIED_UB, IMPLIED_EQ };
+
+    const Model& model_;
+    Vector x_, xl_, xu_, y_, zl_, zu_;
+    std::vector<StateDetail> variable_state_;
+
+    // The mutable quantities have a defined value if evaluated_ is true.
+    // They are computed by Evaluate(), which is called by any method that
+    // accesses any of these quantities.
+    mutable Vector rb_, rl_, ru_, rc_;
+    mutable double pobjective_{0.0};
+    mutable double dobjective_{0.0};
+    mutable double presidual_{0.0};
+    mutable double dresidual_{0.0};
+    mutable double offset_{0.0};
+    mutable double complementarity_{0.0};
+    mutable double mu_{0.0};
+    mutable double mu_min_{0.0};
+    mutable double mu_max_{0.0};
+    mutable bool evaluated_{false};
+    bool postprocessed_{false};
+
+    double feasibility_tol_{1e-6};
+    double optimality_tol_{1e-8};
+    double start_crossover_tol_{-1.0};
+};
+
+inline Iterate::State Iterate::StateOf(Int j) const {
+    switch (variable_state_[j]) {
+    case StateDetail::FIXED:
+        return State::fixed;
+    case StateDetail::FREE:
+    case StateDetail::IMPLIED_LB:
+    case StateDetail::IMPLIED_UB:
+    case StateDetail::IMPLIED_EQ:
+        return State::free;
+    default:
+        return State::barrier;
+    }
+}
+
+inline bool Iterate::has_barrier_lb(Int j) const {
+    return variable_state_[j] == StateDetail::BARRIER_LB
+        || variable_state_[j] == StateDetail::BARRIER_BOXED;
+}
+
+inline bool Iterate::has_barrier_ub(Int j) const{
+    return variable_state_[j] == StateDetail::BARRIER_UB
+        || variable_state_[j] == StateDetail::BARRIER_BOXED;
+}
+
+inline bool Iterate::is_implied(Int j) const {
+    return variable_state_[j] == StateDetail::IMPLIED_LB
+        || variable_state_[j] == StateDetail::IMPLIED_UB
+        || variable_state_[j] == StateDetail::IMPLIED_EQ;
+}
+
+}  // namespace ipx
+
+#endif  // IPX_ITERATE_H_

data/ext/lpsolver-highs/highs/ipm/ipx/kkt_solver.cc

@@ -0,0 +1,23 @@
+#include "ipm/ipx/kkt_solver.h"
+#include "ipm/ipx/timer.h"
+
+namespace ipx {
+
+void KKTSolver::Factorize(Iterate* pt, Info* info) {
+    Timer timer;
+    _Factorize(pt, info);
+    info->time_kkt_factorize += timer.Elapsed();
+}
+
+void KKTSolver::Solve(const Vector& a, const Vector& b, double tol,
+                      Vector& x, Vector& y, Info* info) {
+    Timer timer;
+    _Solve(a, b, tol, x, y, info);
+    info->time_kkt_solve += timer.Elapsed();
+}
+
+Int KKTSolver::iter() const { return _iter(); }
+Int KKTSolver::basis_changes() const { return _basis_changes(); }
+const Basis* KKTSolver::basis() const { return _basis(); }
+
+}  // namespace ipx

data/ext/lpsolver-highs/highs/ipm/ipx/kkt_solver.h

@@ -0,0 +1,70 @@
+#ifndef IPX_KKT_SOLVER_H_
+#define IPX_KKT_SOLVER_H_
+
+#include "ipm/ipx/basis.h"
+#include "ipm/ipx/iterate.h"
+
+namespace ipx {
+
+// Interface to KKT solver implementations. A KKT solver implements a direct or
+// iterative method for solving
+//
+//   [ G   AI' ] (x) = (a) ,                        (1)
+//   [ AI   0  ] (y)   (b)
+//
+// where AI is the m-by-(n+m) matrix defined by a model and G is a positive
+// semidefinite diagonal matrix. The solver may add regularization to G and/or
+// the zero block.
+//
+// An iterative solver must compute an approximate solution of the form
+//
+//   [ G   AI' ] (x) = (a) + (res)                  (2)
+//   [ AI   0  ] (y)   (b)   ( 0 )
+//
+// that satisfies Infnorm(D*res) <= tol, where D is the diagonal matrix with
+// entries D[i,i] = sqrt(1/G[i,i]) if G[i,i] != 0 and D[i,i] = 1 otherwise.
+
+class KKTSolver {
+public:
+    KKTSolver& operator=(const KKTSolver&) = delete;
+    KKTSolver& operator=(KKTSolver&&) = delete;
+    virtual ~KKTSolver() {}
+
+    // Factorizes the KKT matrix (direct solver) or prepares preconditioner
+    // (iterative solver). The diagonal matrix G is built from @iterate.
+    // The implementation is allowed to change variable statuses to eliminate
+    // close-to-converged variables from the IPM solve. Some implementations
+    // allow @iterate to be NULL, in which case G is assumed to be the identity
+    // matrix.
+    void Factorize(Iterate* iterate, Info* info);
+
+    // Solves KKT system. If an iterative method is used, @tol is the required
+    // tolerance for the residual in (2) as specified above.
+    void Solve(const Vector& a, const Vector& b, double tol,
+               Vector& x, Vector& y, Info* info);
+
+    // If an iterative method is used, returns the # iterations in all Solve()
+    // calls since the last call to Factorize(). A direct solver returns the #
+    // iterative refinement steps.
+    Int iter() const;
+
+    // If a basis matrix is maintained, returns the # basis changes in the last
+    // call to Factorize(). Otherwise returns 0.
+    Int basis_changes() const;
+
+    // If a basis matrix is maintained, returns a pointer to it.
+    // Otherwise returns NULL.
+    const Basis* basis() const;
+
+private:
+    virtual void _Factorize(Iterate* iterate, Info* info) = 0;
+    virtual void _Solve(const Vector& a, const Vector& b, double tol,
+                         Vector& x, Vector& y, Info* info) = 0;
+    virtual Int _iter() const = 0;
+    virtual Int _basis_changes() const { return 0; }
+    virtual const Basis* _basis() const { return nullptr; }
+};
+
+}  // namespace ipx
+
+#endif  // IPX_KKT_SOLVER_H_