mirror of
https://github.com/Z3Prover/z3
synced 2026-06-22 16:40:29 +00:00
d99d5a736f
* Add dual-row shared-factor sandwich for NLA bound propagation
When enabled via arith.nl.monomial_sandwich (default off), monomial_bounds
finds LP term columns whose term has shape a_m * m + a_v * v with exactly
two variables — both factors of a binary monomial m = u*v. The term column's
bound bounds (a_m * m + a_v * v); substituting m = u*v gives v * (a_m*u + a_v),
and sign-aware interval division by v plus an affine shift yields a numeric
bound on u. The derived interval is fed to the existing propagate_value path
so the lemma channel and integer rounding logic are shared with the rest of
NLA's forward/backward propagation; no new emit code.
Catches conflicts of the form
α_v1 * v + α_m * m ≥ k1
α_v2 * v + α_m * m ≤ k2
that today require nlsat (when no single row alone yields infeasibility but
their conjunction tightly bounds u after factoring v).
Scope: binary monomials only (m.size()==2, no squares); cap of 16 term-columns
scanned per call; one lemma per (u,v) attempt to keep the lemma channel quiet.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* Add arith.nl.order.binomial_sign flag (default true)
Granular gate for order_lemma_on_binomial_sign — the only order family that
embeds a model-snapshot literal (x ≷ val(x)) in the lemma body. Disabling it
keeps the always-good structural mon-ol family running while removing the
SAT-splitter shape that cascades under model perturbations (e.g., from
arith.nl.monomial_sandwich tightening factor bounds).
Default true preserves master behaviour; the flag is intended as an
experimental knob to measure how much of an observed cascade is specifically
attributable to the binomial-sign splitter vs. the structural cancellation
lemmas in the same module.
See ord-binom-opportunities.md for the full gap analysis and the
deterministic-replacement directions (sandwich, McCormick) that would let
this flag eventually default to false without regressing leaves where
ord-binom currently carries the proof.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* Add sign-pinned binomial bound for NLA (Opportunity 1 from ord-binom doc)
When enabled via arith.nl.monomial_binomial_sign (default off), monomial_bounds
adds a third pass alongside propagate_down (existing) and propagate_shared_factor
(sandwich). For a binary monomial m = u*v in m_to_refine whose model value mv
disagrees with val(u)*val(v), and where v has a determined sign:
1. synthesize a one-sided interval for m.var() at mv (no deps; the snapshot
enters as a literal in the lemma body, not as an antecedent)
2. divide by v's interval (sign-aware via dep.div<with_deps>) to get a
deterministic interval for u
3. emit a propagate_value-style lemma whose body is
m.var() < mv (or > mv) ∨ u-bound
conditioned on v's bound witness
Targets the case ord-binom currently handles: factors have determined signs,
m.var() may have no LP bound. The clause is sound modulo the monomial
definition (same condition propagate_down, propagate_shared_factor, and
ord-binom already rely on).
A new throttle kind MONOMIAL_BINOMIAL_SIGN keyed on (m.var, u, v, direction)
prevents cascading: without it, each new val(m.var()) snapshot would re-emit
across model changes the same way ord-binom does.
Validated via smt.arith.validate=true: 0 soundness errors across the
32-leaf test corpus.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* Add McCormick box-corner tangent points (Opportunity 2 from ord-binom doc)
When enabled via arith.nl.tangents.box_corners (default off, sub-flag of
arith.nl.tangents), tangent_imp::get_points selects m_a, m_b at the corners
of the bound box [x_lo, x_hi] × [y_lo, y_hi] instead of the model-centered
points val(x) ± delta. The selection follows the classical McCormick
under/over envelope:
- m_below=true (under-approximation):
m_a = (x_lo, y_lo), m_b = (x_hi, y_hi)
- m_below=false (over-approximation):
m_a = (x_lo, y_hi), m_b = (x_hi, y_lo)
The existing generate_plane already produces the McCormick linear form
xy ≷ pl.y·x + pl.x·y − pl.x·pl.y at any chosen point pl. push_point is
skipped in box-corner mode: corners are extremes, so doubling the offset
moves out of the box and would invalidate the McCormick property.
Falls back to the existing model-driven point selection when either factor
has an unbounded side or the box is degenerate (single-point in a
dimension).
Soundness — non-strict inequality at corners. The classical model-driven
flow uses pl strictly in the interior of the box, so generate_plane emits
xy > T (strict). At the box corners the tangent meets the surface along
the box's edges (xy = T when x = pl.x or y = pl.y), so the strict
inequality is violated by any model with x at the box boundary. A new
m_pl_strict_interior member, set false on a successful set_box_corners(),
switches generate_plane's emission to ≥/≤ (non-strict). The model-driven
path keeps strict — its push_point + plane_is_correct_cut chain already
guarantees pl is interior.
Validated via smt.arith.validate=true: 0 validate_conflict() failures
across the 32-leaf test corpus with box_corners=true.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
60 lines
2.4 KiB
C++
60 lines
2.4 KiB
C++
/*++
|
|
Copyright (c) 2020 Microsoft Corporation
|
|
|
|
Author:
|
|
Nikolaj Bjorner (nbjorner)
|
|
Lev Nachmanson (levnach)
|
|
|
|
--*/
|
|
#pragma once
|
|
|
|
#include "math/lp/nla_common.h"
|
|
#include "math/lp/nla_intervals.h"
|
|
#include "util/uint_set.h"
|
|
|
|
namespace nla {
|
|
class core;
|
|
class monomial_bounds : common {
|
|
dep_intervals& dep;
|
|
|
|
bool should_propagate_lower(dep_interval const& range, lpvar v, unsigned p);
|
|
bool should_propagate_upper(dep_interval const& range, lpvar v, unsigned p);
|
|
void propagate_bound(lpvar v, lp::lconstraint_kind cmp, rational const& q, u_dependency* d);
|
|
void var2interval(lpvar v, scoped_dep_interval& i);
|
|
bool is_too_big(mpq const& q) const;
|
|
bool propagate_down(monic const& m, lpvar u);
|
|
bool propagate_value(dep_interval& range, lpvar v);
|
|
bool propagate_value(dep_interval& range, lpvar v, unsigned power);
|
|
void compute_product(unsigned start, monic const& m, scoped_dep_interval& i);
|
|
bool propagate(monic const& m);
|
|
void propagate_fixed_to_zero(monic const& m, lpvar fixed_to_zero);
|
|
void propagate_fixed(monic const& m, rational const& k);
|
|
void propagate_nonfixed(monic const& m, rational const& k, lpvar w);
|
|
u_dependency* explain_fixed(monic const& m, rational const& k);
|
|
lp::explanation get_explanation(u_dependency* dep);
|
|
bool propagate_down(monic const& m, dep_interval& mi, lpvar v, unsigned power, dep_interval& product);
|
|
bool propagate_shared_factor(monic const& m);
|
|
bool propagate_binomial_sign(monic const& m);
|
|
void analyze_monomial(monic const& m, unsigned& num_free, lpvar& free_v, unsigned& power) const;
|
|
bool is_free(lpvar v) const;
|
|
bool is_zero(lpvar v) const;
|
|
bool add_lemma();
|
|
|
|
// monomial propagation
|
|
void unit_propagate(monic & m);
|
|
bool is_linear(monic const& m, lpvar& w, lpvar & fixed_to_zero);
|
|
rational fixed_var_product(monic const& m, lpvar w);
|
|
lpvar non_fixed_var(monic const& m);
|
|
|
|
// fixed variable propagation
|
|
unsigned m_fixed_var_qhead = 0;
|
|
unsigned_vector m_fixed_var_trail;
|
|
void propagate_fixed_vars();
|
|
void propagate_fixed_var(lpvar v);
|
|
void propagate_fixed_var(monic const& m, lpvar v);
|
|
public:
|
|
monomial_bounds(core* core);
|
|
void propagate();
|
|
void unit_propagate();
|
|
};
|
|
}
|