3
0
Fork 0
mirror of https://github.com/Z3Prover/z3 synced 2025-04-11 03:33:35 +00:00
z3/src/util/lp/lar_solver.h
Nikolaj Bjorner 4069e76ab0 remove unused column function field, #1021
Signed-off-by: Nikolaj Bjorner <nbjorner@microsoft.com>
2017-05-16 21:27:43 -07:00

1550 lines
59 KiB
C++

/*
Copyright (c) 2017 Microsoft Corporation
Author: Lev Nachmanson
*/
#pragma once
#include "util/vector.h"
#include <utility>
#include "util/debug.h"
#include "util/buffer.h"
#include <unordered_map>
#include <unordered_set>
#include <string>
#include "util/lp/lar_constraints.h"
#include <functional>
#include "util/lp/lar_core_solver.h"
#include <algorithm>
#include "util/lp/numeric_pair.h"
#include "util/lp/scaler.h"
#include "util/lp/lp_primal_core_solver.h"
#include "util/lp/random_updater.h"
#include <stack>
#include "util/lp/stacked_map.h"
#include "util/lp/stacked_value.h"
#include "util/lp/stacked_vector.h"
#include "util/lp/stacked_unordered_set.h"
#include "util/lp/iterator_on_pivot_row.h"
#include "util/lp/implied_bound.h"
#include "util/lp/bound_analyzer_on_row.h"
#include "util/lp/iterator_on_term_with_basis_var.h"
#include "util/lp/iterator_on_row.h"
#include "util/lp/quick_xplain.h"
#include "util/lp/conversion_helper.h"
namespace lean {
class lar_solver : public column_namer {
//////////////////// fields //////////////////////////
lp_settings m_settings;
stacked_value<lp_status> m_status;
stacked_value<simplex_strategy_enum> m_simplex_strategy;
std::unordered_map<unsigned, var_index> m_ext_vars_to_columns;
vector<unsigned> m_columns_to_ext_vars_or_term_indices;
stacked_vector<ul_pair> m_vars_to_ul_pairs;
vector<lar_base_constraint*> m_constraints;
stacked_value<unsigned> m_constraint_count;
// the set of column indices j such that bounds have changed for j
int_set m_columns_with_changed_bound;
int_set m_rows_with_changed_bounds;
int_set m_basic_columns_with_changed_cost;
stacked_value<int> m_infeasible_column_index; // such can be found at the initialization step
stacked_value<unsigned> m_term_count;
vector<lar_term*> m_terms;
vector<lar_term*> m_orig_terms;
const var_index m_terms_start_index;
indexed_vector<mpq> m_column_buffer;
public:
lar_core_solver m_mpq_lar_core_solver;
unsigned constraint_count() const {
return m_constraints.size();
}
const lar_base_constraint& get_constraint(unsigned ci) const {
return *(m_constraints[ci]);
}
////////////////// methods ////////////////////////////////
static_matrix<mpq, numeric_pair<mpq>> & A_r() { return m_mpq_lar_core_solver.m_r_A;}
static_matrix<mpq, numeric_pair<mpq>> const & A_r() const { return m_mpq_lar_core_solver.m_r_A;}
static_matrix<double, double> & A_d() { return m_mpq_lar_core_solver.m_d_A;}
static_matrix<double, double > const & A_d() const { return m_mpq_lar_core_solver.m_d_A;}
static bool valid_index(unsigned j){ return static_cast<int>(j) >= 0;}
public:
lp_settings & settings() { return m_settings;}
lp_settings const & settings() const { return m_settings;}
void clear() {lean_assert(false); // not implemented
}
lar_solver() : m_status(OPTIMAL),
m_infeasible_column_index(-1),
m_terms_start_index(1000000),
m_mpq_lar_core_solver(m_settings, *this)
{}
void set_propagate_bounds_on_pivoted_rows_mode(bool v) {
m_mpq_lar_core_solver.m_r_solver.m_pivoted_rows = v? (& m_rows_with_changed_bounds) : nullptr;
}
virtual ~lar_solver(){
for (auto c : m_constraints)
delete c;
for (auto t : m_terms)
delete t;
for (auto t : m_orig_terms)
delete t;
}
#include "util/lp/init_lar_solver.h"
numeric_pair<mpq> const& get_value(var_index vi) const { return m_mpq_lar_core_solver.m_r_x[vi]; }
bool is_term(var_index j) const {
return j >= m_terms_start_index && j - m_terms_start_index < m_terms.size();
}
unsigned adjust_term_index(unsigned j) const {
lean_assert(is_term(j));
return j - m_terms_start_index;
}
bool use_lu() const { return m_settings.simplex_strategy() == simplex_strategy_enum::lu; }
bool sizes_are_correct() const {
lean_assert(strategy_is_undecided() || !m_mpq_lar_core_solver.need_to_presolve_with_double_solver() || A_r().column_count() == A_d().column_count());
lean_assert(A_r().column_count() == m_mpq_lar_core_solver.m_r_solver.m_column_types.size());
lean_assert(A_r().column_count() == m_mpq_lar_core_solver.m_r_solver.m_costs.size());
lean_assert(A_r().column_count() == m_mpq_lar_core_solver.m_r_x.size());
return true;
}
void print_implied_bound(const implied_bound& be, std::ostream & out) const {
out << "implied bound\n";
unsigned v = be.m_j;
if (is_term(v)) {
out << "it is a term number " << be.m_j << std::endl;
print_term(*m_orig_terms[be.m_j - m_terms_start_index], out);
}
else {
out << get_column_name(v);
}
out << " " << lconstraint_kind_string(be.kind()) << " " << be.m_bound << std::endl;
// for (auto & p : be.m_explanation) {
// out << p.first << " : ";
// print_constraint(p.second, out);
// }
// m_mpq_lar_core_solver.m_r_solver.print_column_info(be.m_j< m_terms_start_index? be.m_j : adjust_term_index(be.m_j), out);
out << "end of implied bound" << std::endl;
}
bool implied_bound_is_correctly_explained(implied_bound const & be, const vector<std::pair<mpq, unsigned>> & explanation) const {
std::unordered_map<unsigned, mpq> coeff_map;
auto rs_of_evidence = zero_of_type<mpq>();
unsigned n_of_G = 0, n_of_L = 0;
bool strict = false;
for (auto & it : explanation) {
mpq coeff = it.first;
constraint_index con_ind = it.second;
const auto & constr = *m_constraints[con_ind];
lconstraint_kind kind = coeff.is_pos() ? constr.m_kind : flip_kind(constr.m_kind);
register_in_map(coeff_map, constr, coeff);
if (kind == GT || kind == LT)
strict = true;
if (kind == GE || kind == GT) n_of_G++;
else if (kind == LE || kind == LT) n_of_L++;
rs_of_evidence += coeff*constr.m_right_side;
}
lean_assert(n_of_G == 0 || n_of_L == 0);
lconstraint_kind kind = n_of_G ? GE : (n_of_L ? LE : EQ);
if (strict)
kind = static_cast<lconstraint_kind>((static_cast<int>(kind) / 2));
if (!is_term(be.m_j)) {
if (coeff_map.size() != 1)
return false;
auto it = coeff_map.find(be.m_j);
if (it == coeff_map.end()) return false;
mpq ratio = it->second;
if (ratio < zero_of_type<mpq>()) {
kind = static_cast<lconstraint_kind>(-kind);
}
rs_of_evidence /= ratio;
} else {
const lar_term * t = m_orig_terms[adjust_term_index(be.m_j)];
const auto first_coeff = *t->m_coeffs.begin();
unsigned j = first_coeff.first;
auto it = coeff_map.find(j);
if (it == coeff_map.end())
return false;
mpq ratio = it->second / first_coeff.second;
for (auto & p : t->m_coeffs) {
it = coeff_map.find(p.first);
if (it == coeff_map.end())
return false;
if (p.second * ratio != it->second)
return false;
}
if (ratio < zero_of_type<mpq>()) {
kind = static_cast<lconstraint_kind>(-kind);
}
rs_of_evidence /= ratio;
rs_of_evidence += t->m_v * ratio;
}
return kind == be.kind() && rs_of_evidence == be.m_bound;
}
void analyze_new_bounds_on_row(
unsigned row_index,
bound_propagator & bp) {
lean_assert(!use_tableau());
iterator_on_pivot_row<mpq> it(m_mpq_lar_core_solver.get_pivot_row(), m_mpq_lar_core_solver.m_r_basis[row_index]);
bound_analyzer_on_row ra_pos(it,
zero_of_type<numeric_pair<mpq>>(),
row_index,
bp
);
ra_pos.analyze();
}
void analyze_new_bounds_on_row_tableau(
unsigned row_index,
bound_propagator & bp
) {
if (A_r().m_rows[row_index].size() > settings().max_row_length_for_bound_propagation)
return;
iterator_on_row<mpq> it(A_r().m_rows[row_index]);
lean_assert(use_tableau());
bound_analyzer_on_row::analyze_row(it,
zero_of_type<numeric_pair<mpq>>(),
row_index,
bp
);
}
void substitute_basis_var_in_terms_for_row(unsigned i) {
// todo : create a map from term basic vars to the rows where they are used
unsigned basis_j = m_mpq_lar_core_solver.m_r_solver.m_basis[i];
for (unsigned k = 0; k < m_terms.size(); k++) {
if (term_is_used_as_row(k))
continue;
if (!m_terms[k]->contains(basis_j))
continue;
m_terms[k]->subst(basis_j, m_mpq_lar_core_solver.m_r_solver.m_pivot_row);
}
}
void calculate_implied_bounds_for_row(unsigned i, bound_propagator & bp) {
if(use_tableau()) {
analyze_new_bounds_on_row_tableau(i, bp);
} else {
m_mpq_lar_core_solver.calculate_pivot_row(i);
substitute_basis_var_in_terms_for_row(i);
analyze_new_bounds_on_row(i, bp);
}
}
/*
void process_new_implied_evidence_for_low_bound(
implied_bound_explanation& implied_evidence, // not pushed yet
vector<bound_evidence> & bound_evidences,
std::unordered_map<unsigned, unsigned> & improved_low_bounds) {
unsigned existing_index;
if (try_get_val(improved_low_bounds, implied_evidence.m_j, existing_index)) {
// we are improving the existent bound
bound_evidences[existing_index] = fill_bound_evidence(implied_evidence);
} else {
improved_low_bounds[implied_evidence.m_j] = bound_evidences.size();
bound_evidences.push_back(fill_bound_evidence(implied_evidence));
}
}
void fill_bound_evidence_on_term(implied_bound & ie, implied_bound& be) {
lean_assert(false);
}
void fill_implied_bound_on_row(implied_bound & ie, implied_bound& be) {
iterator_on_row<mpq> it(A_r().m_rows[ie.m_row_or_term_index]);
mpq a; unsigned j;
bool toggle = ie.m_coeff_before_j_is_pos;
if (!ie.m_is_low_bound)
toggle = !toggle;
while (it.next(a, j)) {
if (j == ie.m_j) continue;
const ul_pair & ul = m_vars_to_ul_pairs[j];
if (is_neg(a)) { // so the monoid has a positive coeff on the right side
constraint_index witness = toggle ? ul.m_low_bound_witness : ul.m_upper_bound_witness;
lean_assert(is_valid(witness));
be.m_explanation.emplace_back(a, witness);
}
}
}
*/
/*
implied_bound fill_implied_bound_for_low_bound(implied_bound& ie) {
implied_bound be(ie.m_j, ie.m_bound.y.is_zero() ? GE : GT, ie.m_bound.x);
if (is_term(ie.m_row_or_term_index)) {
fill_implied_bound_for_low_bound_on_term(ie, be);
}
else {
fill_implied_bound_for_low_bound_on_row(ie, be);
}
return be;
}
implied_bound fill_implied_bound_for_upper_bound(implied_bound& implied_evidence) {
lean_assert(false);
be.m_j = implied_evidence.m_j;
be.m_bound = implied_evidence.m_bound.x;
be.m_kind = implied_evidence.m_bound.y.is_zero() ? LE : LT;
for (auto t : implied_evidence.m_vector_of_bound_signatures) {
const ul_pair & ul = m_vars_to_ul_pairs[t.m_column_index];
constraint_index witness = t.m_low_bound ? ul.m_low_bound_witness : ul.m_upper_bound_witness;
lean_assert(is_valid(witness));
be.m_explanation.emplace_back(t.m_coeff, witness);
}
}
*/
/*
void process_new_implied_evidence_for_upper_bound(
implied_bound& implied_evidence,
vector<implied_bound> & implied_bounds,
std::unordered_map<unsigned, unsigned> & improved_upper_bounds) {
unsigned existing_index;
if (try_get_val(improved_upper_bounds, implied_evidence.m_j, existing_index)) {
implied_bound & be = implied_bounds[existing_index];
be.m_explanation.clear();
// we are improving the existent bound improve the existing bound
be = fill_implied_bound_for_upper_bound(implied_evidence);
} else {
improved_upper_bounds[implied_evidence.m_j] = implied_bounds.size();
implied_bounds.push_back(fill_implied_bound_for_upper_bound(implied_evidence));
}
}
*/
// implied_bound * get_existing_
linear_combination_iterator<mpq> * create_new_iter_from_term(unsigned term_index) const {
lean_assert(false); // not implemented
return nullptr;
// new linear_combination_iterator_on_vector<mpq>(m_terms[adjust_term_index(term_index)]->coeffs_as_vector());
}
unsigned adjust_column_index_to_term_index(unsigned j) const {
unsigned ext_var_or_term = m_columns_to_ext_vars_or_term_indices[j];
return ext_var_or_term < m_terms_start_index ? j : ext_var_or_term;
}
void propagate_bounds_on_a_term(const lar_term& t, bound_propagator & bp, unsigned term_offset) {
lean_assert(false); // not implemented
}
void explain_implied_bound(implied_bound & ib, bound_propagator & bp) {
unsigned i = ib.m_row_or_term_index;
int bound_sign = ib.m_is_low_bound? 1: -1;
int j_sign = (ib.m_coeff_before_j_is_pos ? 1 :-1) * bound_sign;
unsigned m_j = ib.m_j;
if (is_term(m_j)) {
m_j = m_ext_vars_to_columns[m_j];
}
for (auto const& r : A_r().m_rows[i]) {
unsigned j = r.m_j;
mpq const& a = r.get_val();
if (j == m_j) continue;
if (is_term(j)) {
j = m_ext_vars_to_columns[j];
}
int a_sign = is_pos(a)? 1: -1;
int sign = j_sign * a_sign;
const ul_pair & ul = m_vars_to_ul_pairs[j];
auto witness = sign > 0? ul.upper_bound_witness(): ul.low_bound_witness();
lean_assert(is_valid(witness));
bp.consume(a, witness);
}
// lean_assert(implied_bound_is_correctly_explained(ib, explanation));
}
bool term_is_used_as_row(unsigned term) const {
lean_assert(is_term(term));
return contains(m_ext_vars_to_columns, term);
}
void propagate_bounds_on_terms(bound_propagator & bp) {
for (unsigned i = 0; i < m_terms.size(); i++) {
if (term_is_used_as_row(i + m_terms_start_index))
continue; // this term is used a left side of a constraint,
// it was processed as a touched row if needed
propagate_bounds_on_a_term(*m_terms[i], bp, i);
}
}
// goes over touched rows and tries to induce bounds
void propagate_bounds_for_touched_rows(bound_propagator & bp) {
if (!use_tableau())
return; // ! todo : enable bound propagaion here. The current bug is that after the pop
// the changed terms become incorrect!
for (unsigned i : m_rows_with_changed_bounds.m_index) {
calculate_implied_bounds_for_row(i, bp);
}
m_rows_with_changed_bounds.clear();
if (!use_tableau()) {
propagate_bounds_on_terms(bp);
}
}
lp_status get_status() const { return m_status;}
void set_status(lp_status s) {m_status = s;}
lp_status find_feasible_solution() {
if (strategy_is_undecided())
decide_on_strategy_and_adjust_initial_state();
m_mpq_lar_core_solver.m_r_solver.m_look_for_feasible_solution_only = true;
return solve();
}
lp_status solve() {
if (m_status == INFEASIBLE) {
return m_status;
}
solve_with_core_solver();
if (m_status != INFEASIBLE) {
if (m_settings.bound_propagation())
detect_rows_with_changed_bounds();
}
m_columns_with_changed_bound.clear();
return m_status;
}
void fill_explanation_from_infeasible_column(vector<std::pair<mpq, constraint_index>> & evidence) const{
// this is the case when the lower bound is in conflict with the upper one
const ul_pair & ul = m_vars_to_ul_pairs[m_infeasible_column_index];
evidence.push_back(std::make_pair(numeric_traits<mpq>::one(), ul.upper_bound_witness()));
evidence.push_back(std::make_pair(-numeric_traits<mpq>::one(), ul.low_bound_witness()));
}
unsigned get_total_iterations() const { return m_mpq_lar_core_solver.m_r_solver.total_iterations(); }
// see http://research.microsoft.com/projects/z3/smt07.pdf
// This method searches for a feasible solution with as many different values of variables, reverenced in vars, as it can find
// Attention, after a call to this method the non-basic variables don't necesserarly stick to their bounds anymore
vector<unsigned> get_list_of_all_var_indices() const {
vector<unsigned> ret;
for (unsigned j = 0; j < m_mpq_lar_core_solver.m_r_heading.size(); j++)
ret.push_back(j);
return ret;
}
void push() {
m_simplex_strategy = m_settings.simplex_strategy();
m_simplex_strategy.push();
m_status.push();
m_vars_to_ul_pairs.push();
m_infeasible_column_index.push();
m_mpq_lar_core_solver.push();
m_term_count = m_terms.size();
m_term_count.push();
m_constraint_count = m_constraints.size();
m_constraint_count.push();
}
static void clean_large_elements_after_pop(unsigned n, int_set& set) {
vector<int> to_remove;
for (unsigned j: set.m_index)
if (j >= n)
to_remove.push_back(j);
for (unsigned j : to_remove)
set.erase(j);
}
static void shrink_inf_set_after_pop(unsigned n, int_set & set) {
clean_large_elements_after_pop(n, set);
set.resize(n);
}
void pop(unsigned k) {
int n_was = static_cast<int>(m_ext_vars_to_columns.size());
m_status.pop(k);
m_infeasible_column_index.pop(k);
unsigned n = m_vars_to_ul_pairs.peek_size(k);
for (unsigned j = n_was; j-- > n;)
m_ext_vars_to_columns.erase(m_columns_to_ext_vars_or_term_indices[j]);
m_columns_to_ext_vars_or_term_indices.resize(n);
if (m_settings.use_tableau()) {
pop_tableau();
}
m_vars_to_ul_pairs.pop(k);
m_mpq_lar_core_solver.pop(k);
clean_large_elements_after_pop(n, m_columns_with_changed_bound);
unsigned m = A_r().row_count();
clean_large_elements_after_pop(m, m_rows_with_changed_bounds);
clean_inf_set_of_r_solver_after_pop();
lean_assert(m_settings.simplex_strategy() == simplex_strategy_enum::undecided ||
(!use_tableau()) || m_mpq_lar_core_solver.m_r_solver.reduced_costs_are_correct_tableau());
lean_assert(ax_is_correct());
lean_assert(m_mpq_lar_core_solver.m_r_solver.inf_set_is_correct());
m_constraint_count.pop(k);
for (unsigned i = m_constraint_count; i < m_constraints.size(); i++)
delete m_constraints[i];
m_constraints.resize(m_constraint_count);
m_term_count.pop(k);
for (unsigned i = m_term_count; i < m_terms.size(); i++) {
delete m_terms[i];
delete m_orig_terms[i];
}
m_terms.resize(m_term_count);
m_orig_terms.resize(m_term_count);
m_simplex_strategy.pop(k);
m_settings.simplex_strategy() = m_simplex_strategy;
lean_assert(sizes_are_correct());
lean_assert((!m_settings.use_tableau()) || m_mpq_lar_core_solver.m_r_solver.reduced_costs_are_correct_tableau());
}
vector<constraint_index> get_all_constraint_indices() const {
vector<constraint_index> ret;
constraint_index i = 0;
while ( i < m_constraints.size())
ret.push_back(i++);
return ret;
}
bool maximize_term_on_tableau(const vector<std::pair<mpq, var_index>> & term,
impq &term_max) {
if (settings().simplex_strategy() == simplex_strategy_enum::undecided)
decide_on_strategy_and_adjust_initial_state();
m_mpq_lar_core_solver.solve();
if (m_mpq_lar_core_solver.m_r_solver.get_status() == UNBOUNDED)
return false;
term_max = 0;
for (auto & p : term)
term_max += p.first * m_mpq_lar_core_solver.m_r_x[p.second];
return true;
}
bool costs_are_zeros_for_r_solver() const {
for (unsigned j = 0; j < m_mpq_lar_core_solver.m_r_solver.m_costs.size(); j++) {
lean_assert(is_zero(m_mpq_lar_core_solver.m_r_solver.m_costs[j]));
}
return true;
}
bool reduced_costs_are_zeroes_for_r_solver() const {
for (unsigned j = 0; j < m_mpq_lar_core_solver.m_r_solver.m_d.size(); j++) {
lean_assert(is_zero(m_mpq_lar_core_solver.m_r_solver.m_d[j]));
}
return true;
}
void set_costs_to_zero(const vector<std::pair<mpq, var_index>> & term) {
auto & rslv = m_mpq_lar_core_solver.m_r_solver;
auto & jset = m_mpq_lar_core_solver.m_r_solver.m_inf_set; // hijack this set that should be empty right now
lean_assert(jset.m_index.size()==0);
for (auto & p : term) {
unsigned j = p.second;
rslv.m_costs[j] = zero_of_type<mpq>();
int i = rslv.m_basis_heading[j];
if (i < 0)
jset.insert(j);
else {
for (auto & rc : A_r().m_rows[i])
jset.insert(rc.m_j);
}
}
for (unsigned j : jset.m_index)
rslv.m_d[j] = zero_of_type<mpq>();
jset.clear();
lean_assert(reduced_costs_are_zeroes_for_r_solver());
lean_assert(costs_are_zeros_for_r_solver());
}
void prepare_costs_for_r_solver(const vector<std::pair<mpq, var_index>> & term) {
auto & rslv = m_mpq_lar_core_solver.m_r_solver;
rslv.m_using_infeas_costs = false;
lean_assert(costs_are_zeros_for_r_solver());
lean_assert(reduced_costs_are_zeroes_for_r_solver());
rslv.m_costs.resize(A_r().column_count(), zero_of_type<mpq>());
for (auto & p : term) {
unsigned j = p.second;
rslv.m_costs[j] = p.first;
if (rslv.m_basis_heading[j] < 0)
rslv.m_d[j] += p.first;
else
rslv.update_reduced_cost_for_basic_column_cost_change(- p.first, j);
}
lean_assert(rslv.reduced_costs_are_correct_tableau());
}
bool maximize_term_on_corrected_r_solver(const vector<std::pair<mpq, var_index>> & term,
impq &term_max) {
settings().backup_costs = false;
switch (settings().simplex_strategy()) {
case simplex_strategy_enum::tableau_rows:
prepare_costs_for_r_solver(term);
settings().simplex_strategy() = simplex_strategy_enum::tableau_costs;
{
bool ret = maximize_term_on_tableau(term, term_max);
settings().simplex_strategy() = simplex_strategy_enum::tableau_rows;
set_costs_to_zero(term);
m_mpq_lar_core_solver.m_r_solver.set_status(OPTIMAL);
return ret;
}
case simplex_strategy_enum::tableau_costs:
prepare_costs_for_r_solver(term);
{
bool ret = maximize_term_on_tableau(term, term_max);
set_costs_to_zero(term);
m_mpq_lar_core_solver.m_r_solver.set_status(OPTIMAL);
return ret;
}
case simplex_strategy_enum::lu:
lean_assert(false); // not implemented
return false;
default:
lean_unreachable(); // wrong mode
}
return false;
}
// starting from a given feasible state look for the maximum of the term
// return true if found and false if unbounded
bool maximize_term(const vector<std::pair<mpq, var_index>> & term,
impq &term_max) {
lean_assert(m_mpq_lar_core_solver.m_r_solver.current_x_is_feasible());
m_mpq_lar_core_solver.m_r_solver.m_look_for_feasible_solution_only = false;
return maximize_term_on_corrected_r_solver(term, term_max);
}
const lar_term & get_term(unsigned j) const {
lean_assert(j >= m_terms_start_index);
return *m_terms[j - m_terms_start_index];
}
void pop_core_solver_params() {
pop_core_solver_params(1);
}
void pop_core_solver_params(unsigned k) {
A_r().pop(k);
A_d().pop(k);
}
void set_upper_bound_witness(var_index j, constraint_index ci) {
ul_pair ul = m_vars_to_ul_pairs[j];
ul.upper_bound_witness() = ci;
m_vars_to_ul_pairs[j] = ul;
}
void set_low_bound_witness(var_index j, constraint_index ci) {
ul_pair ul = m_vars_to_ul_pairs[j];
ul.low_bound_witness() = ci;
m_vars_to_ul_pairs[j] = ul;
}
void substitute_terms(const mpq & mult,
const vector<std::pair<mpq, var_index>>& left_side_with_terms,
vector<std::pair<mpq, var_index>> &left_side, mpq & right_side) const {
for (auto & t : left_side_with_terms) {
if (t.second < m_terms_start_index) {
lean_assert(t.second < A_r().column_count());
left_side.push_back(std::pair<mpq, var_index>(mult * t.first, t.second));
} else {
const lar_term & term = * m_terms[adjust_term_index(t.second)];
substitute_terms(mult * t.first, left_side_with_terms, left_side, right_side);
right_side -= mult * term.m_v;
}
}
}
void detect_rows_of_bound_change_column_for_nbasic_column(unsigned j) {
if (A_r().row_count() != m_column_buffer.data_size())
m_column_buffer.resize(A_r().row_count());
else
m_column_buffer.clear();
lean_assert(m_column_buffer.size() == 0 && m_column_buffer.is_OK());
m_mpq_lar_core_solver.m_r_solver.solve_Bd(j, m_column_buffer);
for (unsigned i : m_column_buffer.m_index)
m_rows_with_changed_bounds.insert(i);
}
void detect_rows_of_bound_change_column_for_nbasic_column_tableau(unsigned j) {
for (auto & rc : m_mpq_lar_core_solver.m_r_A.m_columns[j])
m_rows_with_changed_bounds.insert(rc.m_i);
}
bool use_tableau() const { return m_settings.use_tableau(); }
bool use_tableau_costs() const {
return m_settings.simplex_strategy() == simplex_strategy_enum::tableau_costs;
}
void detect_rows_of_column_with_bound_change(unsigned j) {
if (m_mpq_lar_core_solver.m_r_heading[j] >= 0) { // it is a basic column
// just mark the row at touched and exit
m_rows_with_changed_bounds.insert(m_mpq_lar_core_solver.m_r_heading[j]);
return;
}
if (use_tableau())
detect_rows_of_bound_change_column_for_nbasic_column_tableau(j);
else
detect_rows_of_bound_change_column_for_nbasic_column(j);
}
void adjust_x_of_column(unsigned j) {
lean_assert(false);
}
bool row_is_correct(unsigned i) const {
numeric_pair<mpq> r = zero_of_type<numeric_pair<mpq>>();
for (const auto & c : A_r().m_rows[i])
r += c.m_value * m_mpq_lar_core_solver.m_r_x[c.m_j];
return is_zero(r);
}
bool ax_is_correct() const {
for (unsigned i = 0; i < A_r().row_count(); i++) {
if (!row_is_correct(i))
return false;
}
return true;
}
bool tableau_with_costs() const {
return m_settings.simplex_strategy() == simplex_strategy_enum::tableau_costs;
}
bool costs_are_used() const {
return m_settings.simplex_strategy() != simplex_strategy_enum::tableau_rows;
}
void change_basic_x_by_delta_on_column(unsigned j, const numeric_pair<mpq> & delta) {
if (use_tableau()) {
for (const auto & c : A_r().m_columns[j]) {
unsigned bj = m_mpq_lar_core_solver.m_r_basis[c.m_i];
m_mpq_lar_core_solver.m_r_x[bj] -= A_r().get_val(c) * delta;
if (tableau_with_costs()) {
m_basic_columns_with_changed_cost.insert(bj);
}
m_mpq_lar_core_solver.m_r_solver.update_column_in_inf_set(bj);
}
} else {
m_column_buffer.clear();
m_column_buffer.resize(A_r().row_count());
m_mpq_lar_core_solver.m_r_solver.solve_Bd(j, m_column_buffer);
for (unsigned i : m_column_buffer.m_index) {
unsigned bj = m_mpq_lar_core_solver.m_r_basis[i];
m_mpq_lar_core_solver.m_r_x[bj] -= m_column_buffer[i] * delta;
m_mpq_lar_core_solver.m_r_solver.update_column_in_inf_set(bj);
}
}
}
void update_x_and_inf_costs_for_column_with_changed_bounds(unsigned j) {
if (m_mpq_lar_core_solver.m_r_heading[j] >= 0) {
if (costs_are_used()) {
bool was_infeas = m_mpq_lar_core_solver.m_r_solver.m_inf_set.contains(j);
m_mpq_lar_core_solver.m_r_solver.update_column_in_inf_set(j);
if (was_infeas != m_mpq_lar_core_solver.m_r_solver.m_inf_set.contains(j))
m_basic_columns_with_changed_cost.insert(j);
} else {
m_mpq_lar_core_solver.m_r_solver.update_column_in_inf_set(j);
}
} else {
numeric_pair<mpq> delta;
if (m_mpq_lar_core_solver.m_r_solver.make_column_feasible(j, delta))
change_basic_x_by_delta_on_column(j, delta);
}
}
void detect_rows_with_changed_bounds_for_column(unsigned j) {
if (m_mpq_lar_core_solver.m_r_heading[j] >= 0) {
m_rows_with_changed_bounds.insert(m_mpq_lar_core_solver.m_r_heading[j]);
return;
}
if (use_tableau())
detect_rows_of_bound_change_column_for_nbasic_column_tableau(j);
else
detect_rows_of_bound_change_column_for_nbasic_column(j);
}
void detect_rows_with_changed_bounds() {
for (auto j : m_columns_with_changed_bound.m_index)
detect_rows_with_changed_bounds_for_column(j);
}
void update_x_and_inf_costs_for_columns_with_changed_bounds() {
for (auto j : m_columns_with_changed_bound.m_index)
update_x_and_inf_costs_for_column_with_changed_bounds(j);
}
void update_x_and_inf_costs_for_columns_with_changed_bounds_tableau() {
lean_assert(ax_is_correct());
for (auto j : m_columns_with_changed_bound.m_index)
update_x_and_inf_costs_for_column_with_changed_bounds(j);
if (tableau_with_costs()) {
for (unsigned j : m_basic_columns_with_changed_cost.m_index)
m_mpq_lar_core_solver.m_r_solver.update_inf_cost_for_column_tableau(j);
lean_assert(m_mpq_lar_core_solver.m_r_solver.reduced_costs_are_correct_tableau());
}
}
void solve_with_core_solver() {
if (!use_tableau())
add_last_rows_to_lu(m_mpq_lar_core_solver.m_r_solver);
if (m_mpq_lar_core_solver.need_to_presolve_with_double_solver()) {
add_last_rows_to_lu(m_mpq_lar_core_solver.m_d_solver);
}
m_mpq_lar_core_solver.prefix_r();
if (costs_are_used()) {
m_basic_columns_with_changed_cost.clear();
m_basic_columns_with_changed_cost.resize(m_mpq_lar_core_solver.m_r_x.size());
}
if (use_tableau())
update_x_and_inf_costs_for_columns_with_changed_bounds_tableau();
else
update_x_and_inf_costs_for_columns_with_changed_bounds();
m_mpq_lar_core_solver.solve();
set_status(m_mpq_lar_core_solver.m_r_solver.get_status());
lean_assert(m_status != OPTIMAL || all_constraints_hold());
}
numeric_pair<mpq> get_basic_var_value_from_row_directly(unsigned i) {
numeric_pair<mpq> r = zero_of_type<numeric_pair<mpq>>();
unsigned bj = m_mpq_lar_core_solver.m_r_solver.m_basis[i];
for (const auto & c: A_r().m_rows[i]) {
if (c.m_j == bj) continue;
const auto & x = m_mpq_lar_core_solver.m_r_x[c.m_j];
if (!is_zero(x))
r -= c.m_value * x;
}
return r;
}
numeric_pair<mpq> get_basic_var_value_from_row(unsigned i) {
if (settings().use_tableau()) {
return get_basic_var_value_from_row_directly(i);
}
numeric_pair<mpq> r = zero_of_type<numeric_pair<mpq>>();
m_mpq_lar_core_solver.calculate_pivot_row(i);
for (unsigned j : m_mpq_lar_core_solver.m_r_solver.m_pivot_row.m_index) {
lean_assert(m_mpq_lar_core_solver.m_r_solver.m_basis_heading[j] < 0);
r -= m_mpq_lar_core_solver.m_r_solver.m_pivot_row.m_data[j] * m_mpq_lar_core_solver.m_r_x[j];
}
return r;
}
template <typename K, typename L>
void add_last_rows_to_lu(lp_primal_core_solver<K,L> & s) {
auto & f = s.m_factorization;
if (f != nullptr) {
auto columns_to_replace = f->get_set_of_columns_to_replace_for_add_last_rows(s.m_basis_heading);
if (f->m_refactor_counter + columns_to_replace.size() >= 200 || f->has_dense_submatrix()) {
delete f;
f = nullptr;
} else {
f->add_last_rows_to_B(s.m_basis_heading, columns_to_replace);
}
}
if (f == nullptr) {
init_factorization(f, s.m_A, s.m_basis, m_settings);
if (f->get_status() != LU_status::OK) {
delete f;
f = nullptr;
}
}
}
bool x_is_correct() const {
if (m_mpq_lar_core_solver.m_r_x.size() != A_r().column_count()) {
// std::cout << "the size is off " << m_r_solver.m_x.size() << ", " << A().column_count() << std::endl;
return false;
}
for (unsigned i = 0; i < A_r().row_count(); i++) {
numeric_pair<mpq> delta = A_r().dot_product_with_row(i, m_mpq_lar_core_solver.m_r_x);
if (!delta.is_zero()) {
// std::cout << "x is off (";
// std::cout << "m_b[" << i << "] = " << m_b[i] << " ";
// std::cout << "left side = " << A().dot_product_with_row(i, m_r_solver.m_x) << ' ';
// std::cout << "delta = " << delta << ' ';
// std::cout << "iters = " << total_iterations() << ")" << std::endl;
// std::cout << "row " << i << " is off" << std::endl;
return false;
}
}
return true;;
}
bool var_is_registered(var_index vj) const {
if (vj >= m_terms_start_index) {
if (vj - m_terms_start_index >= m_terms.size())
return false;
} else if ( vj >= A_r().column_count()) {
return false;
}
return true;
}
unsigned constraint_stack_size() const {
return m_constraint_count.stack_size();
}
void fill_last_row_of_A_r(static_matrix<mpq, numeric_pair<mpq>> & A, const lar_term * ls) {
lean_assert(A.row_count() > 0);
lean_assert(A.column_count() > 0);
unsigned last_row = A.row_count() - 1;
lean_assert(A.m_rows[last_row].size() == 0);
for (auto & t : ls->m_coeffs) {
lean_assert(!is_zero(t.second));
var_index j = t.first;
A.set(last_row, j, - t.second);
}
unsigned basis_j = A.column_count() - 1;
A.set(last_row, basis_j, mpq(1));
}
template <typename U, typename V>
void create_matrix_A(static_matrix<U, V> & matr) {
lean_assert(false); // not implemented
/*
unsigned m = number_or_nontrivial_left_sides();
unsigned n = m_vec_of_canonic_left_sides.size();
if (matr.row_count() == m && matr.column_count() == n)
return;
matr.init_empty_matrix(m, n);
copy_from_mpq_matrix(matr);
*/
}
template <typename U, typename V>
void copy_from_mpq_matrix(static_matrix<U, V> & matr) {
matr.m_rows.resize(A_r().row_count());
matr.m_columns.resize(A_r().column_count());
for (unsigned i = 0; i < matr.row_count(); i++) {
for (auto & it : A_r().m_rows[i]) {
matr.set(i, it.m_j, convert_struct<U, mpq>::convert(it.get_val()));
}
}
}
bool try_to_set_fixed(column_info<mpq> & ci) {
if (ci.upper_bound_is_set() && ci.low_bound_is_set() && ci.get_upper_bound() == ci.get_low_bound() && !ci.is_fixed()) {
ci.set_fixed_value(ci.get_upper_bound());
return true;
}
return false;
}
column_type get_column_type(const column_info<mpq> & ci) {
auto ret = ci.get_column_type_no_flipping();
if (ret == column_type::boxed) { // changing boxed to fixed because of the no span
if (ci.get_low_bound() == ci.get_upper_bound())
ret = column_type::fixed;
}
return ret;
}
std::string get_column_name(unsigned j) const {
if (j >= m_terms_start_index)
return std::string("_t") + T_to_string(j);
if (j >= m_columns_to_ext_vars_or_term_indices.size())
return std::string("_s") + T_to_string(j);
return std::string("v") + T_to_string(m_columns_to_ext_vars_or_term_indices[j]);
}
bool all_constrained_variables_are_registered(const vector<std::pair<mpq, var_index>>& left_side) {
for (auto it : left_side) {
if (! var_is_registered(it.second))
return false;
}
return true;
}
constraint_index add_constraint(const vector<std::pair<mpq, var_index>>& left_side_with_terms, lconstraint_kind kind_par, const mpq& right_side_parm) {
/*
mpq rs = right_side_parm;
vector<std::pair<mpq, var_index>> left_side;
substitute_terms(one_of_type<mpq>(), left_side_with_terms, left_side, rs);
lean_assert(left_side.size() > 0);
lean_assert(all_constrained_variables_are_registered(left_side));
lar_constraint original_constr(left_side, kind_par, rs);
unsigned j; // j is the index of the basic variables corresponding to the left side
canonic_left_side ls = create_or_fetch_canonic_left_side(left_side, j);
mpq ratio = find_ratio_of_original_constraint_to_normalized(ls, original_constr);
auto kind = ratio.is_neg() ? flip_kind(kind_par) : kind_par;
rs/= ratio;
lar_normalized_constraint normalized_constraint(ls, ratio, kind, rs, original_constr);
m_constraints.push_back(normalized_constraint);
constraint_index constr_ind = m_constraints.size() - 1;
update_column_type_and_bound(j, kind, rs, constr_ind);
return constr_ind;
*/
lean_assert(false); // not implemented
return 0;
}
bool all_constraints_hold() const {
if (m_settings.get_cancel_flag())
return true;
std::unordered_map<var_index, mpq> var_map;
get_model(var_map);
for (unsigned i = 0; i < m_constraints.size(); i++) {
if (!constraint_holds(*m_constraints[i], var_map)) {
print_constraint(i, std::cout);
return false;
}
}
return true;
}
bool constraint_holds(const lar_base_constraint & constr, std::unordered_map<var_index, mpq> & var_map) const {
mpq left_side_val = get_left_side_val(constr, var_map);
switch (constr.m_kind) {
case LE: return left_side_val <= constr.m_right_side;
case LT: return left_side_val < constr.m_right_side;
case GE: return left_side_val >= constr.m_right_side;
case GT: return left_side_val > constr.m_right_side;
case EQ: return left_side_val == constr.m_right_side;
default:
lean_unreachable();
}
return false; // it is unreachable
}
bool the_relations_are_of_same_type(const vector<std::pair<mpq, unsigned>> & evidence, lconstraint_kind & the_kind_of_sum) const {
unsigned n_of_G = 0, n_of_L = 0;
bool strict = false;
for (auto & it : evidence) {
mpq coeff = it.first;
constraint_index con_ind = it.second;
lconstraint_kind kind = coeff.is_pos() ?
m_constraints[con_ind]->m_kind :
flip_kind(m_constraints[con_ind]->m_kind);
if (kind == GT || kind == LT)
strict = true;
if (kind == GE || kind == GT) n_of_G++;
else if (kind == LE || kind == LT) n_of_L++;
}
the_kind_of_sum = n_of_G ? GE : (n_of_L ? LE : EQ);
if (strict)
the_kind_of_sum = static_cast<lconstraint_kind>((static_cast<int>(the_kind_of_sum) / 2));
return n_of_G == 0 || n_of_L == 0;
}
static void register_in_map(std::unordered_map<var_index, mpq> & coeffs, const lar_base_constraint & cn, const mpq & a) {
for (auto & it : cn.get_left_side_coefficients()) {
unsigned j = it.second;
auto p = coeffs.find(j);
if (p == coeffs.end())
coeffs[j] = it.first * a;
else {
p->second += it.first * a;
if (p->second.is_zero())
coeffs.erase(p);
}
}
}
bool the_left_sides_sum_to_zero(const vector<std::pair<mpq, unsigned>> & evidence) const {
std::unordered_map<var_index, mpq> coeff_map;
for (auto & it : evidence) {
mpq coeff = it.first;
constraint_index con_ind = it.second;
lean_assert(con_ind < m_constraints.size());
register_in_map(coeff_map, *m_constraints[con_ind], coeff);
}
if (!coeff_map.empty()) {
std::cout << "left side = ";
vector<std::pair<mpq, var_index>> t;
for (auto & it : coeff_map) {
t.push_back(std::make_pair(it.second, it.first));
}
print_linear_combination_of_column_indices(t, std::cout);
std::cout << std::endl;
return false;
}
return true;
}
bool the_right_sides_do_not_sum_to_zero(const vector<std::pair<mpq, unsigned>> & evidence) {
mpq ret = numeric_traits<mpq>::zero();
for (auto & it : evidence) {
mpq coeff = it.first;
constraint_index con_ind = it.second;
lean_assert(con_ind < m_constraints.size());
const lar_constraint & constr = *m_constraints[con_ind];
ret += constr.m_right_side * coeff;
}
return !numeric_traits<mpq>::is_zero(ret);
}
bool explanation_is_correct(const vector<std::pair<mpq, unsigned>>& explanation) const {
#ifdef LEAN_DEBUG
lconstraint_kind kind;
lean_assert(the_relations_are_of_same_type(explanation, kind));
lean_assert(the_left_sides_sum_to_zero(explanation));
mpq rs = sum_of_right_sides_of_explanation(explanation);
switch (kind) {
case LE: lean_assert(rs < zero_of_type<mpq>());
break;
case LT: lean_assert(rs <= zero_of_type<mpq>());
break;
case GE: lean_assert(rs > zero_of_type<mpq>());
break;
case GT: lean_assert(rs >= zero_of_type<mpq>());
break;
case EQ: lean_assert(rs != zero_of_type<mpq>());
break;
default:
lean_assert(false);
return false;
}
#endif
return true;
}
bool inf_explanation_is_correct() const {
#ifdef LEAN_DEBUG
vector<std::pair<mpq, unsigned>> explanation;
get_infeasibility_explanation(explanation);
return explanation_is_correct(explanation);
#endif
return true;
}
mpq sum_of_right_sides_of_explanation(const vector<std::pair<mpq, unsigned>> & explanation) const {
mpq ret = numeric_traits<mpq>::zero();
for (auto & it : explanation) {
mpq coeff = it.first;
constraint_index con_ind = it.second;
lean_assert(con_ind < m_constraints.size());
ret += (m_constraints[con_ind]->m_right_side - m_constraints[con_ind]->get_free_coeff_of_left_side()) * coeff;
}
return ret;
}
bool has_lower_bound(var_index var, constraint_index& ci, mpq& value, bool& is_strict) {
if (var >= m_vars_to_ul_pairs.size()) {
// TBD: bounds on terms could also be used, caller may have to track these.
return false;
}
const ul_pair & ul = m_vars_to_ul_pairs[var];
ci = ul.low_bound_witness();
if (ci != static_cast<constraint_index>(-1)) {
auto& p = m_mpq_lar_core_solver.m_r_low_bounds()[var];
value = p.x;
is_strict = p.y.is_pos();
return true;
}
else {
return false;
}
}
bool has_upper_bound(var_index var, constraint_index& ci, mpq& value, bool& is_strict) {
if (var >= m_vars_to_ul_pairs.size()) {
// TBD: bounds on terms could also be used, caller may have to track these.
return false;
}
const ul_pair & ul = m_vars_to_ul_pairs[var];
ci = ul.upper_bound_witness();
if (ci != static_cast<constraint_index>(-1)) {
auto& p = m_mpq_lar_core_solver.m_r_upper_bounds()[var];
value = p.x;
is_strict = p.y.is_neg();
return true;
}
else {
return false;
}
}
void get_infeasibility_explanation(vector<std::pair<mpq, constraint_index>> & explanation) const {
explanation.clear();
if (m_infeasible_column_index != -1) {
fill_explanation_from_infeasible_column(explanation);
return;
}
if (m_mpq_lar_core_solver.get_infeasible_sum_sign() == 0) {
return;
}
// the infeasibility sign
int inf_sign;
auto inf_row = m_mpq_lar_core_solver.get_infeasibility_info(inf_sign);
get_infeasibility_explanation_for_inf_sign(explanation, inf_row, inf_sign);
lean_assert(explanation_is_correct(explanation));
}
void get_infeasibility_explanation_for_inf_sign(
vector<std::pair<mpq, constraint_index>> & explanation,
const vector<std::pair<mpq, unsigned>> & inf_row,
int inf_sign) const {
for (auto & it : inf_row) {
mpq coeff = it.first;
unsigned j = it.second;
int adj_sign = coeff.is_pos() ? inf_sign : -inf_sign;
const ul_pair & ul = m_vars_to_ul_pairs[j];
constraint_index bound_constr_i = adj_sign < 0 ? ul.upper_bound_witness() : ul.low_bound_witness();
lean_assert(bound_constr_i < m_constraints.size());
explanation.push_back(std::make_pair(coeff, bound_constr_i));
}
}
void get_model(std::unordered_map<var_index, mpq> & variable_values) const {
mpq delta = mpq(1, 2); // start from 0.5 to have less clashes
lean_assert(m_status == OPTIMAL);
unsigned i;
do {
// different pairs have to produce different singleton values
std::unordered_set<impq> set_of_different_pairs;
std::unordered_set<mpq> set_of_different_singles;
delta = m_mpq_lar_core_solver.find_delta_for_strict_bounds(delta);
for (i = 0; i < m_mpq_lar_core_solver.m_r_x.size(); i++ ) {
const numeric_pair<mpq> & rp = m_mpq_lar_core_solver.m_r_x[i];
set_of_different_pairs.insert(rp);
mpq x = rp.x + delta * rp.y;
set_of_different_singles.insert(x);
if (set_of_different_pairs.size()
!= set_of_different_singles.size()) {
delta /= mpq(2);
break;
}
variable_values[i] = x;
}
} while (i != m_mpq_lar_core_solver.m_r_x.size());
}
std::string get_variable_name(var_index vi) const {
return get_column_name(vi);
}
// ********** print region start
void print_constraint(constraint_index ci, std::ostream & out) const {
if (ci >= m_constraints.size()) {
out << "constraint " << T_to_string(ci) << " is not found";
out << std::endl;
return;
}
print_constraint(m_constraints[ci], out);
}
void print_constraints(std::ostream& out) const {
for (auto c : m_constraints) {
print_constraint(c, out);
}
}
void print_terms(std::ostream& out) const {
for (auto it : m_terms) {
print_term(*it, out);
out << "\n";
}
}
void print_left_side_of_constraint(const lar_base_constraint * c, std::ostream & out) const {
print_linear_combination_of_column_indices(c->get_left_side_coefficients(), out);
mpq free_coeff = c->get_free_coeff_of_left_side();
if (!is_zero(free_coeff))
out << " + " << free_coeff;
}
void print_term(lar_term const& term, std::ostream & out) const {
if (!numeric_traits<mpq>::is_zero(term.m_v)) {
out << term.m_v << " + ";
}
print_linear_combination_of_column_indices(term.coeffs_as_vector(), out);
}
mpq get_left_side_val(const lar_base_constraint & cns, const std::unordered_map<var_index, mpq> & var_map) const {
mpq ret = cns.get_free_coeff_of_left_side();
for (auto & it : cns.get_left_side_coefficients()) {
var_index j = it.second;
auto vi = var_map.find(j);
lean_assert(vi != var_map.end());
ret += it.first * vi->second;
}
return ret;
}
void print_constraint(const lar_base_constraint * c, std::ostream & out) const {
print_left_side_of_constraint(c, out);
out << " " << lconstraint_kind_string(c->m_kind) << " " << c->m_right_side << std::endl;
}
void fill_var_set_for_random_update(unsigned sz, var_index const * vars, vector<unsigned>& column_list) {
for (unsigned i = 0; i < sz; i++) {
var_index var = vars[i];
if (var >= m_terms_start_index) { // handle the term
for (auto & it : m_terms[var - m_terms_start_index]->m_coeffs) {
column_list.push_back(it.first);
}
} else {
column_list.push_back(var);
}
}
}
void random_update(unsigned sz, var_index const * vars) {
vector<unsigned> column_list;
fill_var_set_for_random_update(sz, vars, column_list);
random_updater ru(m_mpq_lar_core_solver, column_list);
ru.update();
}
void try_pivot_fixed_vars_from_basis() {
m_mpq_lar_core_solver.m_r_solver.pivot_fixed_vars_from_basis();
}
void pop() {
pop(1);
}
bool column_represents_row_in_tableau(unsigned j) {
return m_vars_to_ul_pairs()[j].m_i != static_cast<row_index>(-1);
}
void make_sure_that_the_bottom_right_elem_not_zero_in_tableau(unsigned i, unsigned j) {
// i, j - is the indices of the bottom-right element of the tableau
lean_assert(A_r().row_count() == i + 1 && A_r().column_count() == j + 1);
auto & last_column = A_r().m_columns[j];
int non_zero_column_cell_index = -1;
for (unsigned k = last_column.size(); k-- > 0;){
auto & cc = last_column[k];
if (cc.m_i == i)
return;
non_zero_column_cell_index = k;
}
lean_assert(non_zero_column_cell_index != -1);
lean_assert(static_cast<unsigned>(non_zero_column_cell_index) != i);
m_mpq_lar_core_solver.m_r_solver.transpose_rows_tableau(last_column[non_zero_column_cell_index].m_i, i);
}
void remove_last_row_and_column_from_tableau(unsigned j) {
lean_assert(A_r().column_count() == m_mpq_lar_core_solver.m_r_solver.m_costs.size());
auto & slv = m_mpq_lar_core_solver.m_r_solver;
unsigned i = A_r().row_count() - 1; //last row index
make_sure_that_the_bottom_right_elem_not_zero_in_tableau(i, j);
if (slv.m_basis_heading[j] < 0) {
slv.pivot_column_tableau(j, i);
}
auto & last_row = A_r().m_rows[i];
mpq &cost_j = m_mpq_lar_core_solver.m_r_solver.m_costs[j];
bool cost_is_nz = !is_zero(cost_j);
for (unsigned k = last_row.size(); k-- > 0;) {
auto &rc = last_row[k];
if (cost_is_nz) {
m_mpq_lar_core_solver.m_r_solver.m_d[rc.m_j] += cost_j*rc.get_val();
}
A_r().remove_element(last_row, rc);
}
lean_assert(last_row.size() == 0);
lean_assert(A_r().m_columns[j].size() == 0);
A_r().m_rows.pop_back();
A_r().m_columns.pop_back();
slv.m_b.pop_back();
}
void remove_last_column_from_tableau(unsigned j) {
lean_assert(j == A_r().column_count() - 1);
// the last column has to be empty
lean_assert(A_r().m_columns[j].size() == 0);
A_r().m_columns.pop_back();
}
void remove_last_column_from_basis_tableau(unsigned j) {
auto& rslv = m_mpq_lar_core_solver.m_r_solver;
int i = rslv.m_basis_heading[j];
if (i >= 0) { // j is a basic var
int last_pos = static_cast<int>(rslv.m_basis.size()) - 1;
lean_assert(last_pos >= 0);
if (i != last_pos) {
unsigned j_at_last_pos = rslv.m_basis[last_pos];
rslv.m_basis[i] = j_at_last_pos;
rslv.m_basis_heading[j_at_last_pos] = i;
}
rslv.m_basis.pop_back(); // remove j from the basis
} else {
int last_pos = static_cast<int>(rslv.m_nbasis.size()) - 1;
lean_assert(last_pos >= 0);
i = - 1 - i;
if (i != last_pos) {
unsigned j_at_last_pos = rslv.m_nbasis[last_pos];
rslv.m_nbasis[i] = j_at_last_pos;
rslv.m_basis_heading[j_at_last_pos] = - i - 1;
}
rslv.m_nbasis.pop_back(); // remove j from the basis
}
rslv.m_basis_heading.pop_back();
lean_assert(rslv.m_basis.size() == A_r().row_count());
lean_assert(rslv.basis_heading_is_correct());
}
void remove_column_from_tableau(unsigned j) {
auto& rslv = m_mpq_lar_core_solver.m_r_solver;
lean_assert(j == A_r().column_count() - 1);
lean_assert(A_r().column_count() == m_mpq_lar_core_solver.m_r_solver.m_costs.size());
if (column_represents_row_in_tableau(j)) {
remove_last_row_and_column_from_tableau(j);
if (rslv.m_basis_heading[j] < 0)
rslv.change_basis_unconditionally(j, rslv.m_basis[A_r().row_count()]); // A_r().row_count() is the index of the last row in the basis still
}
else {
remove_last_column_from_tableau(j);
}
rslv.m_x.pop_back();
rslv.m_d.pop_back();
rslv.m_costs.pop_back();
remove_last_column_from_basis_tableau(j);
lean_assert(m_mpq_lar_core_solver.r_basis_is_OK());
lean_assert(A_r().column_count() == m_mpq_lar_core_solver.m_r_solver.m_costs.size());
}
void pop_tableau() {
lean_assert(m_mpq_lar_core_solver.m_r_solver.m_costs.size() == A_r().column_count());
lean_assert(m_mpq_lar_core_solver.m_r_solver.m_basis.size() == A_r().row_count());
lean_assert(m_mpq_lar_core_solver.m_r_solver.basis_heading_is_correct());
// We remove last variables starting from m_column_names.size() to m_vec_of_canonic_left_sides.size().
// At this moment m_column_names is already popped
for (unsigned j = A_r().column_count(); j-- > m_columns_to_ext_vars_or_term_indices.size();)
remove_column_from_tableau(j);
lean_assert(m_mpq_lar_core_solver.m_r_solver.m_costs.size() == A_r().column_count());
lean_assert(m_mpq_lar_core_solver.m_r_solver.m_basis.size() == A_r().row_count());
lean_assert(m_mpq_lar_core_solver.m_r_solver.basis_heading_is_correct());
}
void clean_inf_set_of_r_solver_after_pop() {
vector<unsigned> became_feas;
clean_large_elements_after_pop(A_r().column_count(), m_mpq_lar_core_solver.m_r_solver.m_inf_set);
std::unordered_set<unsigned> basic_columns_with_changed_cost;
auto inf_index_copy = m_mpq_lar_core_solver.m_r_solver.m_inf_set.m_index;
for (auto j: inf_index_copy) {
if (m_mpq_lar_core_solver.m_r_heading[j] >= 0) {
continue;
}
// some basic columns might become non-basic - these columns need to be made feasible
numeric_pair<mpq> delta;
if (m_mpq_lar_core_solver.m_r_solver.make_column_feasible(j, delta))
change_basic_x_by_delta_on_column(j, delta);
became_feas.push_back(j);
}
for (unsigned j : became_feas) {
lean_assert(m_mpq_lar_core_solver.m_r_solver.m_basis_heading[j] < 0);
m_mpq_lar_core_solver.m_r_solver.m_d[j] -= m_mpq_lar_core_solver.m_r_solver.m_costs[j];
m_mpq_lar_core_solver.m_r_solver.m_costs[j] = zero_of_type<mpq>();
m_mpq_lar_core_solver.m_r_solver.m_inf_set.erase(j);
}
became_feas.clear();
for (unsigned j : m_mpq_lar_core_solver.m_r_solver.m_inf_set.m_index) {
lean_assert(m_mpq_lar_core_solver.m_r_heading[j] >= 0);
if (m_mpq_lar_core_solver.m_r_solver.column_is_feasible(j))
became_feas.push_back(j);
}
for (unsigned j : became_feas)
m_mpq_lar_core_solver.m_r_solver.m_inf_set.erase(j);
if (use_tableau_costs()) {
for (unsigned j : became_feas)
m_mpq_lar_core_solver.m_r_solver.update_inf_cost_for_column_tableau(j);
for (unsigned j : basic_columns_with_changed_cost)
m_mpq_lar_core_solver.m_r_solver.update_inf_cost_for_column_tableau(j);
lean_assert(m_mpq_lar_core_solver.m_r_solver.reduced_costs_are_correct_tableau());
}
}
void shrink_explanation_to_minimum(vector<std::pair<mpq, constraint_index>> & explanation) const {
// implementing quickXplain
quick_xplain::run(explanation, *this);
lean_assert(this->explanation_is_correct(explanation));
}
};
}