From: Andrew Reynolds <andrew.j.reynolds@gmail.com>
Date: Wed, 8 Jul 2020 00:50:01 +0000 (-0500)
Subject: Improve and fix secant and tangent lemmas in the transcendental solver (#4689)
X-Git-Tag: cvc5-1.0.0~3144
X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=de1018fc9d12f722a5c88c0a862a7e94e3de37ac;p=cvc5.git

Improve and fix secant and tangent lemmas in the transcendental solver (#4689)

Fixes #4686.

This benchmark failed an assertion that corresponded to the fact that a refinement lemma did not evaluate to false in the current model (and hence does not rule out the current model).

This was caused by applying the rewriter in a way that led to an incorrect approximation. This meant that some tangent and secant lemmas would be ineffective.

The benchmark in that issue now times out, but makes progress in the refinement lemmas it generates.

This also simplifies and improves the use of approximated values (instead of model values) when constructing tangent and secant lemmas.
---

diff --git a/src/theory/arith/nl/transcendental_solver.cpp b/src/theory/arith/nl/transcendental_solver.cpp
index 1377aa54a..1a80342bc 100644
--- a/src/theory/arith/nl/transcendental_solver.cpp
+++ b/src/theory/arith/nl/transcendental_solver.cpp
@@ -776,6 +776,13 @@ bool TranscendentalSolver::checkTfTangentPlanesFun(Node tf,
   bool is_tangent = false;
   bool is_secant = false;
   std::pair<Node, Node> mvb = getTfModelBounds(tf, d);
+  // this is the approximated value of tf(c), which is a value such that:
+  //    M_A(tf(c)) >= poly_appox_c >= tf(c) or
+  //    M_A(tf(c)) <= poly_appox_c <= tf(c)
+  // In other words, it is a better approximation of the true value of tf(c)
+  // in the case that we add a refinement lemma. We use this value in the
+  // refinement schemas below.
+  Node poly_approx_c;
   for (unsigned r = 0; r < 2; r++)
   {
     Node pab = poly_approx_bounds[r][csign];
@@ -792,6 +799,7 @@ bool TranscendentalSolver::checkTfTangentPlanesFun(Node tf,
       Trace("nl-ext-tftp-debug2") << "...got : " << compr << std::endl;
       if (compr == d_true)
       {
+        poly_approx_c = v_pab;
         // beyond the bounds
         if (r == 0)
         {
@@ -830,17 +838,8 @@ bool TranscendentalSolver::checkTfTangentPlanesFun(Node tf,
   }
 
   // Figure 3: P( c )
-  Node poly_approx_c;
   if (is_tangent || is_secant)
   {
-    Assert(!poly_approx.isNull());
-    std::vector<Node> taylor_subs;
-    taylor_subs.push_back(c);
-    Assert(taylor_vars.size() == taylor_subs.size());
-    poly_approx_c = poly_approx.substitute(taylor_vars.begin(),
-                                           taylor_vars.end(),
-                                           taylor_subs.begin(),
-                                           taylor_subs.end());
     Trace("nl-ext-tftp-debug2")
         << "...poly approximation at c is " << poly_approx_c << std::endl;
   }
@@ -1450,11 +1449,18 @@ std::pair<Node, Node> TranscendentalSolver::getTfModelBounds(Node tf,
     Node pab = pbounds[index];
     if (!pab.isNull())
     {
-      // { x -> tf[0] }
-      pab = pab.substitute(tfv, tfs);
+      // { x -> M_A(tf[0]) }
+      // Notice that we compute the model value of tfs first, so that
+      // the call to rewrite below does not modify the term, where notice that
+      // rewrite( x*x { x -> M_A(t) } ) = M_A(t)*M_A(t)
+      // is not equal to
+      // M_A( x*x { x -> t } ) = M_A( t*t )
+      // where M_A denotes the abstract model.
+      Node mtfs = d_model.computeAbstractModelValue(tfs);
+      pab = pab.substitute(tfv, mtfs);
       pab = Rewriter::rewrite(pab);
-      Node v_pab = d_model.computeAbstractModelValue(pab);
-      bounds.push_back(v_pab);
+      Assert (pab.isConst());
+      bounds.push_back(pab);
     }
     else
     {