/*++ Copyright (c) 2012 Microsoft Corporation Module Name: nlsat_explain.cpp Abstract: Functor that implements the "explain" procedure defined in Dejan and Leo's paper. Author: Leonardo de Moura (leonardo) 2012-01-13. Revision History: --*/ #include"nlsat_explain.h" #include"nlsat_assignment.h" #include"nlsat_evaluator.h" #include"algebraic_numbers.h" #include"ref_buffer.h" namespace nlsat { typedef polynomial::polynomial_ref_vector polynomial_ref_vector; typedef ref_buffer polynomial_ref_buffer; struct explain::imp { solver & m_solver; assignment const & m_assignment; atom_vector const & m_atoms; atom_vector const & m_x2eq; anum_manager & m_am; polynomial::cache & m_cache; pmanager & m_pm; polynomial_ref_vector m_ps; polynomial_ref_vector m_psc_tmp; polynomial_ref_vector m_factors; scoped_anum_vector m_roots_tmp; bool m_simplify_cores; bool m_full_dimensional; bool m_minimize_cores; struct todo_set { polynomial::cache & m_cache; polynomial_ref_vector m_set; svector m_in_set; todo_set(polynomial::cache & u):m_cache(u), m_set(u.pm()) {} void reset() { pmanager & pm = m_set.m(); unsigned sz = m_set.size(); for (unsigned i = 0; i < sz; i++) { m_in_set[pm.id(m_set.get(i))] = false; } m_set.reset(); } void insert(poly * p) { pmanager & pm = m_set.m(); p = m_cache.mk_unique(p); unsigned pid = pm.id(p); if (m_in_set.get(pid, false)) return; m_in_set.setx(pid, true, false); m_set.push_back(p); } bool empty() const { return m_set.empty(); } // Return max variable in todo_set var max_var() const { pmanager & pm = m_set.m(); var max = null_var; unsigned sz = m_set.size(); for (unsigned i = 0; i < sz; i++) { var x = pm.max_var(m_set.get(i)); SASSERT(x != null_var); if (max == null_var || x > max) max = x; } return max; } /** \brief Remove the maximal polynomials from the set and store them in max_polys. Return the maximal variable */ var remove_max_polys(polynomial_ref_vector & max_polys) { max_polys.reset(); var x = max_var(); pmanager & pm = m_set.m(); unsigned sz = m_set.size(); unsigned j = 0; for (unsigned i = 0; i < sz; i++) { poly * p = m_set.get(i); var y = pm.max_var(p); SASSERT(y <= x); if (y == x) { max_polys.push_back(p); m_in_set[pm.id(p)] = false; } else { m_set.set(j, p); j++; } } m_set.shrink(j); return x; } }; // temporary field for store todo set of polynomials todo_set m_todo; // temporary fields for preprocessing core scoped_literal_vector m_core1; scoped_literal_vector m_core2; // temporary fields for storing the result scoped_literal_vector * m_result; svector m_already_added_literal; evaluator & m_evaluator; imp(solver & s, assignment const & x2v, polynomial::cache & u, atom_vector const & atoms, atom_vector const & x2eq, evaluator & ev): m_solver(s), m_assignment(x2v), m_atoms(atoms), m_x2eq(x2eq), m_am(x2v.am()), m_cache(u), m_pm(u.pm()), m_ps(m_pm), m_psc_tmp(m_pm), m_factors(m_pm), m_roots_tmp(m_am), m_todo(u), m_core1(s), m_core2(s), m_result(0), m_evaluator(ev) { m_simplify_cores = false; m_full_dimensional = false; m_minimize_cores = false; } ~imp() { } void display(std::ostream & out, polynomial_ref const & p) const { m_pm.display(out, p, m_solver.display_proc()); } void display(std::ostream & out, polynomial_ref_vector const & ps, char const * delim = "\n") const { for (unsigned i = 0; i < ps.size(); i++) { if (i > 0) out << delim; m_pm.display(out, ps.get(i), m_solver.display_proc()); } } void display(std::ostream & out, literal l) const { m_solver.display(out, l); } void display_var(std::ostream & out, var x) const { m_solver.display(out, x); } void display(std::ostream & out, unsigned sz, literal const * ls) { for (unsigned i = 0; i < sz; i++) { display(out, ls[i]); out << "\n"; } } void display(std::ostream & out, literal_vector const & ls) { display(out, ls.size(), ls.c_ptr()); } void display(std::ostream & out, scoped_literal_vector const & ls) { display(out, ls.size(), ls.c_ptr()); } /** \brief Add literal to the result vector. */ void add_literal(literal l) { SASSERT(m_result != 0); SASSERT(l != true_literal); if (l == false_literal) return; unsigned lidx = l.index(); if (m_already_added_literal.get(lidx, false)) return; TRACE("nlsat_explain", tout << "adding literal: " << lidx << "\n"; m_solver.display(tout, l); tout << "\n";); m_already_added_literal.setx(lidx, true, false); m_result->push_back(l); } /** \brief Reset m_already_added vector using m_result */ void reset_already_added() { SASSERT(m_result != 0); unsigned sz = m_result->size(); for (unsigned i = 0; i < sz; i++) m_already_added_literal[(*m_result)[i].index()] = false; } /** \brief evaluate the given polynomial in the current interpretation. max_var(p) must be assigned in the current interpretation. */ int sign(polynomial_ref const & p) { TRACE("nlsat_explain", tout << "p: " << p << "\n";); SASSERT(max_var(p) == null_var || m_assignment.is_assigned(max_var(p))); return m_am.eval_sign_at(p, m_assignment); } /** \brief Wrapper for factorization */ void factor(polynomial_ref & p, polynomial_ref_vector & fs) { // TODO: add params, caching TRACE("nlsat_factor", tout << "factor\n" << p << "\n";); fs.reset(); m_cache.factor(p.get(), fs); } /** \brief Wrapper for psc chain computation */ void psc_chain(polynomial_ref & p, polynomial_ref & q, unsigned x, polynomial_ref_vector & result) { // TODO caching SASSERT(max_var(p) == max_var(q)); SASSERT(max_var(p) == x); m_cache.psc_chain(p, q, x, result); } /** \breif Store in ps the polynomials occurring in the given literals. */ void collect_polys(unsigned num, literal const * ls, polynomial_ref_vector & ps) { ps.reset(); for (unsigned i = 0; i < num; i++) { atom * a = m_atoms[ls[i].var()]; SASSERT(a != 0); if (a->is_ineq_atom()) { unsigned sz = to_ineq_atom(a)->size(); for (unsigned j = 0; j < sz; j++) ps.push_back(to_ineq_atom(a)->p(j)); } else { ps.push_back(to_root_atom(a)->p()); } } } /** \brief Add literal p != 0 into m_result. */ ptr_vector m_zero_fs; svector m_is_even; void add_zero_assumption(polynomial_ref & p) { // If p is of the form p1^n1 * ... * pk^nk, // then only the factors that are zero in the current interpretation needed to be considered. // I don't want to create a nested conjunction in the clause. // Then, I assert p_i1 * ... * p_im != 0 factor(p, m_factors); unsigned num_factors = m_factors.size(); m_zero_fs.reset(); m_is_even.reset(); polynomial_ref f(m_pm); for (unsigned i = 0; i < num_factors; i++) { f = m_factors.get(i); if (sign(f) == 0) { m_zero_fs.push_back(m_factors.get(i)); m_is_even.push_back(false); } } SASSERT(!m_zero_fs.empty()); // one of the factors must be zero in the current interpretation, since p is zero in it. literal l = m_solver.mk_ineq_literal(atom::EQ, m_zero_fs.size(), m_zero_fs.c_ptr(), m_is_even.c_ptr()); l.neg(); TRACE("nlsat_explain", tout << "adding (zero assumption) literal:\n"; display(tout, l); tout << "\n";); add_literal(l); } void add_simple_assumption(atom::kind k, poly * p, bool sign = false) { SASSERT(k == atom::EQ || k == atom::LT || k == atom::GT); bool is_even = false; bool_var b = m_solver.mk_ineq_atom(k, 1, &p, &is_even); literal l(b, !sign); add_literal(l); } void add_assumption(atom::kind k, poly * p, bool sign = false) { // TODO: factor add_simple_assumption(k, p, sign); } /** \brief Eliminate "vanishing leading coefficients" of p. That is, coefficients that vanish in the current interpretation. The resultant p is a reduct of p s.t. its leading coefficient does not vanish in the current interpretation. If all coefficients of p vanish, then the resultant p is the zero polynomial. */ void elim_vanishing(polynomial_ref & p) { SASSERT(!is_const(p)); var x = max_var(p); unsigned k = degree(p, x); SASSERT(k > 0); polynomial_ref lc(m_pm); polynomial_ref reduct(m_pm); while (true) { if (is_const(p)) return; if (k == 0) { // x vanished from p, peek next maximal variable x = max_var(p); SASSERT(x != null_var); k = degree(p, x); } if (m_pm.nonzero_const_coeff(p, x, k)) return; // lc is a nonzero constant lc = m_pm.coeff(p, x, k, reduct); if (!is_zero(lc)) { if (sign(lc) != 0) return; // lc is not the zero polynomial, but it vanished in the current interpretaion. // so we keep searching... add_zero_assumption(lc); } if (k == 0) { // all coefficients of p vanished in the current interpretation, // and were added as assumptions. p = m_pm.mk_zero(); return; } k--; p = reduct; } } /** Eliminate vanishing coefficients of polynomials in ps. The coefficients that are zero (i.e., vanished) are added as assumptions into m_result. */ void elim_vanishing(polynomial_ref_vector & ps) { unsigned j = 0; unsigned sz = ps.size(); polynomial_ref p(m_pm); for (unsigned i = 0; i < sz; i++) { p = ps.get(i); elim_vanishing(p); if (!is_const(p)) { ps.set(j, p); j++; } } ps.shrink(j); } /** Normalize literal with respect to given maximal variable. The basic idea is to eliminate vanishing (leading) coefficients from a (arithmetic) literal, and factors from lower stages. The vanishing coefficients and factors from lower stages are added as assumptions to the lemma being generated. Example 1) Assume - l is of the form (y^2 - 2)*x^3 + y*x + 1 > 0 - x is the maximal variable - y is assigned to sqrt(2) Thus, (y^2 - 2) the coefficient of x^3 vanished. This method returns y*x + 1 > 0 and adds the assumption (y^2 - 2) = 0 to the lemma Example 2) Assume - l is of the form (x + 2)*(y - 1) > 0 - x is the maximal variable - y is assigned to 0 (x + 2) < 0 is returned and assumption (y - 1) < 0 is added as an assumption. Remark: root atoms are not normalized */ literal normalize(literal l, var max) { bool_var b = l.var(); if (b == true_bool_var) return l; SASSERT(m_atoms[b] != 0); if (m_atoms[b]->is_ineq_atom()) { polynomial_ref_buffer ps(m_pm); sbuffer is_even; polynomial_ref p(m_pm); ineq_atom * a = to_ineq_atom(m_atoms[b]); int atom_sign = 1; unsigned sz = a->size(); bool normalized = false; // true if the literal needs to be normalized for (unsigned i = 0; i < sz; i++) { p = a->p(i); if (max_var(p) == max) elim_vanishing(p); // eliminate vanishing coefficients of max if (is_const(p) || max_var(p) < max) { int s = sign(p); if (!is_const(p)) { SASSERT(max_var(p) != null_var); SASSERT(max_var(p) < max); // factor p is a lower stage polynomial, so we should add assumption to justify p being eliminated if (s == 0) add_simple_assumption(atom::EQ, p); // add assumption p = 0 else if (a->is_even(i)) add_simple_assumption(atom::EQ, p, true); // add assumption p != 0 else if (s < 0) add_simple_assumption(atom::LT, p); // add assumption p < 0 else add_simple_assumption(atom::GT, p); // add assumption p > 0 } if (s == 0) { bool atom_val = a->get_kind() == atom::EQ; bool lit_val = l.sign() ? !atom_val : atom_val; return lit_val ? true_literal : false_literal; } else if (s == -1 && a->is_odd(i)) { atom_sign = -atom_sign; } normalized = true; } else { if (p != a->p(i)) { SASSERT(!m_pm.eq(p, a->p(i))); normalized = true; } is_even.push_back(a->is_even(i)); ps.push_back(p); } } if (ps.empty()) { SASSERT(atom_sign != 0); // LHS is positive or negative. It is positive if atom_sign > 0 and negative if atom_sign < 0 bool atom_val; if (a->get_kind() == atom::EQ) atom_val = false; else if (a->get_kind() == atom::LT) atom_val = atom_sign < 0; else atom_val = atom_sign > 0; bool lit_val = l.sign() ? !atom_val : atom_val; return lit_val ? true_literal : false_literal; } else if (normalized) { atom::kind new_k = a->get_kind(); if (atom_sign < 0) new_k = atom::flip(new_k); literal new_l = m_solver.mk_ineq_literal(new_k, ps.size(), ps.c_ptr(), is_even.c_ptr()); if (l.sign()) new_l.neg(); return new_l; } else { SASSERT(atom_sign > 0); return l; } } else { return l; } } /** Normalize literals (in the conflicting core) with respect to given maximal variable. The basic idea is to eliminate vanishing (leading) coefficients (and factors from lower stages) from (arithmetic) literals, */ void normalize(scoped_literal_vector & C, var max) { unsigned sz = C.size(); unsigned j = 0; for (unsigned i = 0; i < sz; i++) { literal new_l = normalize(C[i], max); if (new_l == true_literal) continue; if (new_l == false_literal) { // false literal was created. The assumptions added are sufficient for implying the conflict. C.reset(); return; } C.set(j, new_l); j++; } C.shrink(j); } var max_var(poly const * p) { return m_pm.max_var(p); } /** \brief Return the maximal variable in a set of nonconstant polynomials. */ var max_var(polynomial_ref_vector const & ps) { if (ps.empty()) return null_var; var max = max_var(ps.get(0)); SASSERT(max != null_var); // there are no constant polynomials in ps unsigned sz = ps.size(); for (unsigned i = 1; i < sz; i++) { var curr = m_pm.max_var(ps.get(i)); SASSERT(curr != null_var); if (curr > max) max = curr; } return max; } polynomial::var max_var(literal l) { atom * a = m_atoms[l.var()]; if (a != 0) return a->max_var(); else return null_var; } /** \brief Return the maximal variable in the given set of literals */ var max_var(unsigned sz, literal const * ls) { var max = null_var; for (unsigned i = 0; i < sz; i++) { literal l = ls[i]; atom * a = m_atoms[l.var()]; if (a != 0) { var x = a->max_var(); SASSERT(x != null_var); if (max == null_var || x > max) max = x; } } return max; } /** \brief Move the polynomials in q in ps that do not contain x to qs. */ void keep_p_x(polynomial_ref_vector & ps, var x, polynomial_ref_vector & qs) { unsigned sz = ps.size(); unsigned j = 0; for (unsigned i = 0; i < sz; i++) { poly * q = ps.get(i); if (max_var(q) != x) { qs.push_back(q); } else { ps.set(j, q); j++; } } ps.shrink(j); } /** \brief Add factors of p to todo */ void add_factors(polynomial_ref & p) { if (is_const(p)) return; elim_vanishing(p); if (is_const(p)) return; #if 1 TRACE("nlsat_explain", tout << "adding factors of\n"; display(tout, p); tout << "\n";); factor(p, m_factors); polynomial_ref f(m_pm); for (unsigned i = 0; i < m_factors.size(); i++) { f = m_factors.get(i); elim_vanishing(f); if (!is_const(f)) { TRACE("nlsat_explain", tout << "adding factor:\n"; display(tout, f); tout << "\n";); m_todo.insert(f); } } #else m_todo.insert(normalize(p)); #endif } /** \brief Add leading coefficients of the polynomials in ps. \pre all polynomials in ps contain x Remark: the leading coefficients do not vanish in the current model, since all polynomials in ps were pre-processed using elim_vanishing. */ void add_lc(polynomial_ref_vector & ps, var x) { polynomial_ref p(m_pm); polynomial_ref lc(m_pm); unsigned sz = ps.size(); for (unsigned i = 0; i < sz; i++) { p = ps.get(i); unsigned k = degree(p, x); SASSERT(k > 0); TRACE("nlsat_explain", tout << "add_lc, x: "; display_var(tout, x); tout << "\nk: " << k << "\n"; display(tout, p); tout << "\n";); if (m_pm.nonzero_const_coeff(p, x, k)) { TRACE("nlsat_explain", tout << "constant coefficient, skipping...\n";); continue; } lc = m_pm.coeff(p, x, k); SASSERT(sign(lc) != 0); SASSERT(!is_const(lc)); add_factors(lc); } } /** \brief Add v-psc(p, q, x) into m_todo */ void psc(polynomial_ref & p, polynomial_ref & q, var x) { polynomial_ref_vector & S = m_psc_tmp; polynomial_ref s(m_pm); TRACE("nlsat_explain", tout << "computing psc of\n"; display(tout, p); tout << "\n"; display(tout, q); tout << "\n";); psc_chain(p, q, x, S); unsigned sz = S.size(); for (unsigned i = 0; i < sz; i++) { s = S.get(i); TRACE("nlsat_explain", tout << "processing psc(" << i << ")\n"; display(tout, s); tout << "\n";); if (is_zero(s)) { TRACE("nlsat_explain", tout << "skipping psc is the zero polynomial\n";); continue; } if (is_const(s)) { TRACE("nlsat_explain", tout << "done, psc is a constant\n";); return; } if (sign(s) == 0) { TRACE("nlsat_explain", tout << "psc vanished, adding zero assumption\n";); add_zero_assumption(s); continue; } TRACE("nlsat_explain", tout << "adding v-psc of\n"; display(tout, p); tout << "\n"; display(tout, q); tout << "\n---->\n"; display(tout, s); tout << "\n";); // s did not vanish completely, but its leading coefficient may have vanished add_factors(s); return; } } /** \brief For each p in ps, add v-psc(x, p, p') into m_todo \pre all polynomials in ps contain x Remark: the leading coefficients do not vanish in the current model, since all polynomials in ps were pre-processed using elim_vanishing. */ void psc_discriminant(polynomial_ref_vector & ps, var x) { polynomial_ref p(m_pm); polynomial_ref p_prime(m_pm); unsigned sz = ps.size(); for (unsigned i = 0; i < sz; i++) { p = ps.get(i); if (degree(p, x) < 2) continue; p_prime = derivative(p, x); psc(p, p_prime, x); } } /** \brief For each p and q in ps, p != q, add v-psc(x, p, q) into m_todo \pre all polynomials in ps contain x Remark: the leading coefficients do not vanish in the current model, since all polynomials in ps were pre-processed using elim_vanishing. */ void psc_resultant(polynomial_ref_vector & ps, var x) { polynomial_ref p(m_pm); polynomial_ref q(m_pm); unsigned sz = ps.size(); for (unsigned i = 0; i < sz - 1; i++) { p = ps.get(i); for (unsigned j = i + 1; j < sz; j++) { q = ps.get(j); psc(p, q, x); } } } void add_root_literal(atom::kind k, var y, unsigned i, poly * p) { scoped_mpz c(m_pm.m()); bool_var b; bool lsign; if (m_pm.degree(p, y) == 1 && m_pm.const_coeff(p, y, 1, c)) { SASSERT(!m_pm.m().is_zero(c)); // literal can be expressed using a linear ineq_atom polynomial_ref p_prime(m_pm); p_prime = p; if (m_pm.m().is_neg(c)) p_prime = neg(p_prime); p = p_prime.get(); switch (k) { case atom::ROOT_EQ: k = atom::EQ; lsign = false; break; case atom::ROOT_LT: k = atom::LT; lsign = false; break; case atom::ROOT_GT: k = atom::GT; lsign = false; break; case atom::ROOT_LE: k = atom::GT; lsign = true; break; case atom::ROOT_GE: k = atom::LT; lsign = true; break; default: UNREACHABLE(); break; } bool is_even = false; b = m_solver.mk_ineq_atom(k, 1, &p, &is_even); } else { b = m_solver.mk_root_atom(k, y, i, p); lsign = false; } lsign = !lsign; // adding as an assumption literal l(b, lsign); TRACE("nlsat_explain", tout << "adding literal\n"; display(tout, l); tout << "\n";); add_literal(l); } /** Add one or two literals that specify in which cell of variable y the current interpretation is. One literal is added for the cases: - y in (-oo, min) where min is the minimal root of the polynomials p2 in ps We add literal ! (y < root_1(p2)) - y in (max, oo) where max is the maximal root of the polynomials p1 in ps We add literal ! (y > root_k(p1)) where k is the number of real roots of p - y = r where r is the k-th root of a polynomial p in ps We add literal ! (y = root_k(p)) Two literals are added when - y in (l, u) where (l, u) does not contain any root of polynomials p in ps, and l is the i-th root of a polynomial p1 in ps, and u is the j-th root of a polynomial p2 in ps. We add literals ! (y > root_i(p1)) or !(y < root_j(p2)) */ void add_cell_lits(polynomial_ref_vector & ps, var y) { SASSERT(m_assignment.is_assigned(y)); bool lower_inf = true; bool upper_inf = true; scoped_anum_vector & roots = m_roots_tmp; scoped_anum lower(m_am); scoped_anum upper(m_am); anum const & y_val = m_assignment.value(y); TRACE("nlsat_explain", tout << "adding literals for "; display_var(tout, y); tout << " -> "; m_am.display_decimal(tout, y_val); tout << "\n";); polynomial_ref p_lower(m_pm); unsigned i_lower; polynomial_ref p_upper(m_pm); unsigned i_upper; polynomial_ref p(m_pm); unsigned sz = ps.size(); for (unsigned k = 0; k < sz; k++) { p = ps.get(k); if (max_var(p) != y) continue; roots.reset(); // Variable y is assigned in m_assignment. We must temporarily unassign it. // Otherwise, the isolate_roots procedure will assume p is a constant polynomial. m_am.isolate_roots(p, undef_var_assignment(m_assignment, y), roots); unsigned num_roots = roots.size(); for (unsigned i = 0; i < num_roots; i++) { TRACE("nlsat_explain", tout << "comparing root: "; m_am.display_decimal(tout, roots[i]); tout << "\n";); int s = m_am.compare(y_val, roots[i]); if (s == 0) { // y_val == roots[i] // add literal // ! (y = root_i(p)) add_root_literal(atom::ROOT_EQ, y, i+1, p); return; } else if (s < 0) { // y_val < roots[i] // check if roots[i] is a better upper bound if (upper_inf || m_am.lt(roots[i], upper)) { upper_inf = false; m_am.set(upper, roots[i]); p_upper = p; i_upper = i+1; } } else if (s > 0) { // roots[i] < y_val // check if roots[i] is a better lower bound if (lower_inf || m_am.lt(lower, roots[i])) { lower_inf = false; m_am.set(lower, roots[i]); p_lower = p; i_lower = i+1; } } } } if (!lower_inf) add_root_literal(m_full_dimensional ? atom::ROOT_GE : atom::ROOT_GT, y, i_lower, p_lower); if (!upper_inf) add_root_literal(m_full_dimensional ? atom::ROOT_LE : atom::ROOT_LT, y, i_upper, p_upper); } /** \brief Return true if all polynomials in ps are univariate in x. */ bool all_univ(polynomial_ref_vector const & ps, var x) { unsigned sz = ps.size(); for (unsigned i = 0; i < sz; i++) { poly * p = ps.get(i); if (max_var(p) != x) return false; if (!m_pm.is_univariate(p)) return false; } return true; } /** \brief Apply model-based projection operation defined in our paper. */ void project(polynomial_ref_vector & ps, var max_x) { if (ps.empty()) return; m_todo.reset(); for (unsigned i = 0; i < ps.size(); i++) m_todo.insert(ps.get(i)); var x = m_todo.remove_max_polys(ps); // Remark: after vanishing coefficients are eliminated, ps may not contain max_x anymore if (x < max_x) add_cell_lits(ps, x); while (true) { if (all_univ(ps, x) && m_todo.empty()) { m_todo.reset(); break; } TRACE("nlsat_explain", tout << "project loop, processing var "; display_var(tout, x); tout << "\npolynomials\n"; display(tout, ps); tout << "\n";); add_lc(ps, x); psc_discriminant(ps, x); psc_resultant(ps, x); if (m_todo.empty()) break; x = m_todo.remove_max_polys(ps); add_cell_lits(ps, x); } } bool check_already_added() const { for (unsigned i = 0; i < m_already_added_literal.size(); i++) { SASSERT(m_already_added_literal[i] == false); } return true; } /* Conflicting core simplification using equations. The idea is to use equations to reduce the complexity of the conflicting core. Basic idea: Let l be of the form h > 0 and eq of the form p = 0 Using pseudo-division we have that: lc(p)^d h = q p + r where q and r are the pseudo-quotient and pseudo-remainder d is the integer returned by the pseudo-division algorithm. lc(p) is the leading coefficient of p If d is even or sign(lc(p)) > 0, we have that sign(h) = sign(r) Otherwise sign(h) = -sign(r) flipped the sign We have the following rules If (C and h > 0) implies false Then 1. (C and p = 0 and lc(p) != 0 and r > 0) implies false if d is even 2. (C and p = 0 and lc(p) > 0 and r > 0) implies false if lc(p) > 0 and d is odd 3. (C and p = 0 and lc(p) < 0 and r < 0) implies false if lc(p) < 0 and d is odd If (C and h = 0) implies false Then (C and p = 0 and lc(p) != 0 and r = 0) implies false If (C and h < 0) implies false Then 1. (C and p = 0 and lc(p) != 0 and r < 0) implies false if d is even 2. (C and p = 0 and lc(p) > 0 and r < 0) implies false if lc(p) > 0 and d is odd 3. (C and p = 0 and lc(p) < 0 and r > 0) implies false if lc(p) < 0 and d is odd Good cases: - lc(p) is a constant - p = 0 is already in the conflicting core - p = 0 is linear We only use equations from the conflicting core and lower stages. Equations from lower stages are automatically added to the lemma. */ struct eq_info { poly const * m_eq; polynomial::var m_x; unsigned m_k; poly * m_lc; int m_lc_sign; bool m_lc_const; bool m_lc_add; bool m_lc_add_ineq; void add_lc_ineq() { m_lc_add = true; m_lc_add_ineq = true; } void add_lc_diseq() { if (!m_lc_add) { m_lc_add = true; m_lc_add_ineq = false; } } }; void simplify(literal l, eq_info & info, var max, scoped_literal & new_lit) { bool_var b = l.var(); atom * a = m_atoms[b]; SASSERT(a != 0); if (a->is_root_atom()) { new_lit = l; return; } ineq_atom * _a = to_ineq_atom(a); unsigned num_factors = _a->size(); if (num_factors == 1 && _a->p(0) == info.m_eq) { new_lit = l; return; } TRACE("nlsat_simplify_core", tout << "trying to simplify literal\n"; display(tout, l); tout << "\nusing equation\n"; m_pm.display(tout, info.m_eq, m_solver.display_proc()); tout << "\n";); int atom_sign = 1; bool modified_lit = false; polynomial_ref_buffer new_factors(m_pm); sbuffer new_factors_even; polynomial_ref new_factor(m_pm); for (unsigned s = 0; s < num_factors; s++) { poly * f = _a->p(s); bool is_even = _a->is_even(s); if (m_pm.degree(f, info.m_x) < info.m_k) { new_factors.push_back(f); new_factors_even.push_back(is_even); continue; } modified_lit = true; unsigned d; m_pm.pseudo_remainder(f, info.m_eq, info.m_x, d, new_factor); // adjust sign based on sign of lc of eq if (d % 2 == 1 && // d is odd !is_even && // degree of the factor is odd info.m_lc_sign < 0 // lc of the eq is negative ) { atom_sign = -atom_sign; // flipped the sign of the current literal } if (is_const(new_factor)) { int s = sign(new_factor); if (s == 0) { bool atom_val = a->get_kind() == atom::EQ; bool lit_val = l.sign() ? !atom_val : atom_val; new_lit = lit_val ? true_literal : false_literal; return; } else if (s == -1 && !is_even) { atom_sign = -atom_sign; } } else { new_factors.push_back(new_factor); new_factors_even.push_back(is_even); if (!info.m_lc_const) { if (d % 2 == 0 || // d is even is_even || // rewriting a factor of even degree, sign flip doesn't matter _a->get_kind() == atom::EQ) { // rewriting an equation, sign flip doesn't matter info.add_lc_diseq(); } else { info.add_lc_ineq(); } } } } if (modified_lit) { atom::kind new_k = _a->get_kind(); if (atom_sign < 0) new_k = atom::flip(new_k); new_lit = m_solver.mk_ineq_literal(new_k, new_factors.size(), new_factors.c_ptr(), new_factors_even.c_ptr()); if (l.sign()) new_lit.neg(); TRACE("nlsat_simplify_core", tout << "simplified literal:\n"; display(tout, new_lit); tout << "\n";); if (max_var(new_lit) < max) { // The conflicting core may have redundant literals. // We should check whether new_lit is true in the current model, and discard it if that is the case lbool val = m_solver.value(new_lit); SASSERT(val != l_undef); if (m_solver.value(new_lit) == l_false) add_literal(new_lit); new_lit = true_literal; return; } new_lit = normalize(new_lit, max); TRACE("nlsat_simplify_core", tout << "simplified literal after normalization:\n"; display(tout, new_lit); tout << "\n";); } else { new_lit = l; } } bool simplify(scoped_literal_vector & C, poly const * eq, var max) { bool modified_core = false; eq_info info; info.m_eq = eq; info.m_x = m_pm.max_var(info.m_eq); info.m_k = m_pm.degree(eq, info.m_x); polynomial_ref lc_eq(m_pm); lc_eq = m_pm.coeff(eq, info.m_x, info.m_k); info.m_lc = lc_eq.get(); info.m_lc_sign = sign(lc_eq); info.m_lc_add = false; info.m_lc_const = m_pm.is_const(lc_eq); SASSERT(info.m_lc != 0); scoped_literal new_lit(m_solver); unsigned sz = C.size(); unsigned j = 0; for (unsigned i = 0; i < sz; i++) { literal l = C[i]; new_lit = null_literal; simplify(l, info, max, new_lit); SASSERT(new_lit != null_literal); if (l == new_lit) { C.set(j, l); j++; continue; } modified_core = true; if (new_lit == true_literal) continue; if (new_lit == false_literal) { // false literal was created. The assumptions added are sufficient for implying the conflict. j = 0; // force core to be reset break; } C.set(j, new_lit); j++; } C.shrink(j); if (info.m_lc_add) { if (info.m_lc_add_ineq) add_assumption(info.m_lc_sign < 0 ? atom::LT : atom::GT, info.m_lc); else add_assumption(atom::EQ, info.m_lc, true); } return modified_core; } /** \brief (try to) Select an equation from C. Returns 0 if C does not contain any equality. This method selects the equation of minimal degree in max. */ poly * select_eq(scoped_literal_vector & C, var max) { poly * r = 0; unsigned min_d = UINT_MAX; unsigned sz = C.size(); for (unsigned i = 0; i < sz; i++) { literal l = C[i]; if (l.sign()) continue; bool_var b = l.var(); atom * a = m_atoms[b]; SASSERT(a != 0); if (a->get_kind() != atom::EQ) continue; ineq_atom * _a = to_ineq_atom(a); if (_a->size() > 1) continue; if (_a->is_even(0)) continue; unsigned d = m_pm.degree(_a->p(0), max); SASSERT(d > 0); if (d < min_d) { r = _a->p(0); min_d = d; if (min_d == 1) break; } } return r; } /** \brief Select an equation eq s.t. max_var(eq) < max, and it can be used to rewrite a literal in C. Return 0, if such equation was not found. */ var_vector m_select_tmp; ineq_atom * select_lower_stage_eq(scoped_literal_vector & C, var max) { var_vector & xs = m_select_tmp; unsigned sz = C.size(); for (unsigned i = 0; i < sz; i++) { literal l = C[i]; bool_var b = l.var(); atom * a = m_atoms[b]; if (a->is_root_atom()) continue; // we don't rewrite root atoms ineq_atom * _a = to_ineq_atom(a); unsigned num_factors = _a->size(); for (unsigned j = 0; j < num_factors; j++) { poly * p = _a->p(j); xs.reset(); m_pm.vars(p, xs); unsigned xs_sz = xs.size(); for (unsigned k = 0; k < xs_sz; k++) { var y = xs[k]; if (y >= max) continue; atom * eq = m_x2eq[y]; if (eq == 0) continue; SASSERT(eq->is_ineq_atom()); SASSERT(to_ineq_atom(eq)->size() == 1); SASSERT(!to_ineq_atom(eq)->is_even(0)); poly * eq_p = to_ineq_atom(eq)->p(0); SASSERT(m_pm.degree(eq_p, y) > 0); // TODO: create a parameter // In the current experiments, using equations with non constant coefficients produces a blowup if (!m_pm.nonzero_const_coeff(eq_p, y, m_pm.degree(eq_p, y))) continue; if (m_pm.degree(p, y) >= m_pm.degree(eq_p, y)) return to_ineq_atom(eq); } } } return 0; } /** \brief Simplify the core using equalities. */ void simplify(scoped_literal_vector & C, var max) { // Simplify using equations in the core while (!C.empty()) { poly * eq = select_eq(C, max); if (eq == 0) break; TRACE("nlsat_simplify_core", tout << "using equality for simplifying core\n"; m_pm.display(tout, eq, m_solver.display_proc()); tout << "\n";); if (!simplify(C, eq, max)) break; } // Simplify using equations using variables from lower stages. while (!C.empty()) { ineq_atom * eq = select_lower_stage_eq(C, max); if (eq == 0) break; SASSERT(eq->size() == 1); SASSERT(!eq->is_even(0)); poly * eq_p = eq->p(0); VERIFY(simplify(C, eq_p, max)); // add equation as an assumption add_literal(literal(eq->bvar(), true)); } } /** \brief Main procedure. The explain the given unsat core, and store the result in m_result */ void main(unsigned num, literal const * ls) { if (num == 0) return; collect_polys(num, ls, m_ps); var max_x = max_var(m_ps); TRACE("nlsat_explain", tout << "polynomials in the conflict:\n"; display(tout, m_ps); tout << "\n";); elim_vanishing(m_ps); project(m_ps, max_x); } void process2(unsigned num, literal const * ls) { if (m_simplify_cores) { m_core2.reset(); m_core2.append(num, ls); var max = max_var(num, ls); SASSERT(max != null_var); normalize(m_core2, max); TRACE("nlsat_explain", tout << "core after normalization\n"; display(tout, m_core2);); simplify(m_core2, max); TRACE("nlsat_explain", tout << "core after simplify\n"; display(tout, m_core2);); main(m_core2.size(), m_core2.c_ptr()); m_core2.reset(); } else { main(num, ls); } } // Auxiliary method for core minimization. literal_vector m_min_newtodo; bool minimize_core(literal_vector & todo, literal_vector & core) { SASSERT(!todo.empty()); literal_vector & new_todo = m_min_newtodo; new_todo.reset(); interval_set_manager & ism = m_evaluator.ism(); interval_set_ref r(ism); // Copy the union of the infeasible intervals of core into r. unsigned sz = core.size(); for (unsigned i = 0; i < sz; i++) { literal l = core[i]; atom * a = m_atoms[l.var()]; SASSERT(a != 0); interval_set_ref inf = m_evaluator.infeasible_intervals(a, l.sign()); r = ism.mk_union(inf, r); if (ism.is_full(r)) { // Done return false; } } TRACE("nlsat_mininize", tout << "interval set after adding partial core:\n" << r << "\n";); if (todo.size() == 1) { // Done core.push_back(todo[0]); return false; } // Copy the union of the infeasible intervals of todo into r until r becomes full. sz = todo.size(); for (unsigned i = 0; i < sz; i++) { literal l = todo[i]; atom * a = m_atoms[l.var()]; SASSERT(a != 0); interval_set_ref inf = m_evaluator.infeasible_intervals(a, l.sign()); r = ism.mk_union(inf, r); if (ism.is_full(r)) { // literal l must be in the core core.push_back(l); new_todo.swap(todo); return true; } else { new_todo.push_back(l); } } UNREACHABLE(); return true; } literal_vector m_min_todo; literal_vector m_min_core; void minimize(unsigned num, literal const * ls, scoped_literal_vector & r) { literal_vector & todo = m_min_todo; literal_vector & core = m_min_core; todo.reset(); core.reset(); todo.append(num, ls); while (true) { TRACE("nlsat_mininize", tout << "core minimization:\n"; display(tout, todo); tout << "\nCORE:\n"; display(tout, core);); if (!minimize_core(todo, core)) break; std::reverse(todo.begin(), todo.end()); TRACE("nlsat_mininize", tout << "core minimization:\n"; display(tout, todo); tout << "\nCORE:\n"; display(tout, core);); if (!minimize_core(todo, core)) break; } TRACE("nlsat_mininize", tout << "core:\n"; display(tout, core);); r.append(core.size(), core.c_ptr()); } void process(unsigned num, literal const * ls) { if (m_minimize_cores && num > 1) { m_core1.reset(); minimize(num, ls, m_core1); process2(m_core1.size(), m_core1.c_ptr()); m_core1.reset(); } else { process2(num, ls); } } void operator()(unsigned num, literal const * ls, scoped_literal_vector & result) { SASSERT(check_already_added()); SASSERT(num > 0); TRACE("nlsat_explain", tout << "[explain] set of literals is infeasible in the current interpretation\n"; display(tout, num, ls);); // exit(0); m_result = &result; process(num, ls); reset_already_added(); m_result = 0; TRACE("nlsat_explain", tout << "[explain] result\n"; display(tout, result);); CASSERT("nlsat", check_already_added()); } }; explain::explain(solver & s, assignment const & x2v, polynomial::cache & u, atom_vector const & atoms, atom_vector const & x2eq, evaluator & ev) { m_imp = alloc(imp, s, x2v, u, atoms, x2eq, ev); } explain::~explain() { dealloc(m_imp); } void explain::reset() { m_imp->m_core1.reset(); m_imp->m_core2.reset(); } void explain::set_simplify_cores(bool f) { m_imp->m_simplify_cores = f; } void explain::set_full_dimensional(bool f) { m_imp->m_full_dimensional = f; } void explain::set_minimize_cores(bool f) { m_imp->m_minimize_cores = f; } void explain::operator()(unsigned n, literal const * ls, scoped_literal_vector & result) { (*m_imp)(n, ls, result); } };