Z3 Performance with Non-Linear Arithmetic - z3

We are running into performance problems with what I believe is the part of Z3 that treats non-linear arithmetic. Here is a simple concrete Boogie example, that when verified with Z3 (version 4.1) takes a long time (on the order of 3 minutes) to complete.
const D: int;
function f(n: int) returns (int) { n * D }
procedure test() returns ()
{
var a, b, c: int;
var M: [int]int;
var t: int;
assume 0 < a && 1000 * a < f(1);
assume 0 < c && 1000 * c < f(1);
assume f(100) * b == a * c;
assert M[t] > 0;
}
It seems that the problem is caused by an interaction of functions, range assumption on integer variables as well as multiplications of (unknown) integer values. The assertion in the end should not be provable. Instead of failing quickly, it seems that Z3 has ways to instantiate lots of terms somehow, as its memory consumptions grows fairly quickly to about 300 MB, at which point it gives up.
I'm wondering if this is a bug, or whether it is possible to improve the heuristics on when Z3 should stop the search in the particular direction it is currently trying to solve the problem.
One interesting thing is that inlining the function by using
function {:inline} f(n: int) returns (int) { n * D }
makes the verification terminate very quickly.
Background: This is a minimal test case for a problem that we see in our verifier Chalice. There, the Boogie programs get much longer, potentially with multiple assumptions of a similar kind. Often, the verification does not appear to be terminating at all.
Any ideas?

Yes, the non-termination is due to nonlinear integer arithmetic. Z3 has a new nonlinear solver, but it is for "nonlinear real arithmetic", and can only be used in quantifier free problems that only use arithmetic (i.e., no uninterpreted functions like in your example).
Thus, the old arithmetic solver is used in your example. This solver has very limited support for integer arithmetic. Your analysis of the problem is correct. Z3 has trouble finding a solution for the block of nonlinear integer constraints. Note that if we replace f(100) * b == a * c with f(100) * b <= a * c, then Z3 returns immediately with an "unknown" answer.
We can avoid the non-termination by limiting the number of nonlinear arithmetic reasoning in Z3. The option NL_ARITH_ROUNDS=100 will use the nonlinear module at most 100 times for each Z3 query. This is not an ideal solution, but at least, we avoid the non-termination.

Related

What alternatives exist for representing sets in Z3?

Per this answer the Z3 set sort is implemented using arrays, which makes sense given the SetAdd and SetDel methods available in the API. It is also claimed here that if the array modification functions are never used, it's wasteful overhead to use arrays instead of uninterpreted functions. Given that, if my only uses of a set are to apply constraints with IsMember (either on individual values or as part of a quantification), is it a better idea to use an uninterpreted function mapping from the underlying element sort to booleans? So:
from z3 import *
s = Solver()
string_set = SetSort(StringSort())
x = String('x')
s.add(IsMember(x, string_set))
becomes
from z3 import *
s = Solver()
string_set = Function('string_set', StringSort(), BoolSort())
x = String('x')
s.add(string_set(x))
Are there any drawbacks to this approach? Alternative representations with even less overhead?
Those are really your only options, as long as you want to restrict yourself to the standard interface. In the past, I also had luck with representing sets (and in general relations) outside of the solver, keeping the processing completely outside. Here's what I mean:
from z3 import *
def FSet_Empty():
return lambda x: False
def FSet_Insert(val, s):
return lambda x: If(x == val, True, s(val))
def FSet_Delete(val, s):
return lambda x: If(x == val, False, s(val))
def FSet_Member(val, s):
return s(val)
x, y, z = Ints('x y z')
myset = FSet_Insert(x, FSet_Insert(y, FSet_Insert(z, FSet_Empty())))
s = Solver()
s.add(FSet_Member(2, myset))
print(s.check())
print(s.model())
Note how we model sets by unary relations, i.e., functions from values to booleans. You can generalize this to arbitrary relations and the ideas carry over. This prints:
sat
[x = 2, z = 4, y = 3]
You can easily add union (essentially Or), intersection (essentially And), and complement (essentially Not) operations. Doing cardinality is harder, especially in the presence of complement, but that's true for all the other approaches too.
As is usual with these sorts of modeling questions, there's no single approach that will work best across all problems. They'll all have their strengths and weaknesses. I'd recommend creating a single API, and implementing it using all three of these ideas, and benchmarking your problem domain to see what works the best; keeping in mind if you start working on a different problem the answer might be different. Please report your findings!

Dafny as a SAT-QBF solver is not giving right results

I am trying to get the habit to use Dafny as a friendly SAT-QBF solver for some simple formulae, since doing it in, for instance, Z3 is too uncomfortable.
The context for this is that I have implemented Cooper's algorithm for quantifier elimination and, when all the variables are bounded, it can be used as a decision procedure: therefore, I want to know which is the result that I should get before executing it.
However, I encountered a problem in Dafny.
Let us raise, for instance, this formula (written in Dafny):
assert forall x_1: int :: exists y_1: int :: forall x_2: int :: exists y_2 : int
:: (y_2<y_1) && (x_2<y_2) && (x_1<x_2);
In my Cooper, it returns False, while Dafny returns assertion violation (along with the typical triggerswarnings), which I interpret as False too. Okay, so no problem with this.
But if I raise:
assert exists x_1: int :: exists y_1: int :: exists x_2: int :: exists y_2 : int
:: (y_2<y_1) && (x_2<y_2) && (x_1<x_2);
In my Cooper, it returns True, while Dafny also returns assertion violation. I have done a manual Cooper execution (pencil and paper) and I think the True is right one.
Any idea of what is going on?
PS: I have not tried it in Z3, yet, because I am doing first other attempts with other theories.
EDIT
Trigger warnings can be avoided using a simple trick to instantiate the quantified variables: creating an uninterpreted function.
method Main() {
assert exists x_1 : int {:trigger P(x_1)} :: exists y_1: int {:trigger P(y_1)}
:: exists x_2: int {:trigger P(x_2)} :: exists y_2 : int {:trigger P(y_2)}
:: (y_2<y_1) && (x_2<y_2) && (x_1<x_2);
}
predicate P(a: int)
{
true
}
You cannot do this with Dafny. While Dafny supports quantifiers, booleans, arithmetic, and many other things (recursive functions, sets, sequences, objects and references, multi-dimensional arrays, induction, inductive and coinductive datatypes, bitvectors, greatest and least fixpoints of monotonic functions, etc.), it is not suitable for SAT-QBF (or QBF + artihmetic) benchmarks.
Dafny's errors, including the assertion violation, tell you that the verifier was not able to do the proof. It may be that the property still holds, but you'll need to supply more of the proof yourself. In other words, you should interpret the assertion violation as a "don't know" answer. Stated differently, you cannot decide (only semi-decide) formulas with Dafny.
Dafny uses quantifiers in the SMT solvers via matching patterns, aka triggers. When a quantifier has no good triggers, which is what Dafny's "no trigger" warning is telling you, you may see bad performance, unstable verifications, and so-called butterfly effects (where a small and seemingly unrelated part of the program causes a change in the automatic construction of other proofs). Triggers are driven by uninterpreted function symbols, which your example doesn't have at all.
If you want a readable syntax, you may be able to do what you're trying through Boogie. I have not tried that, but you could try putting Boogie in its monomorphic mode and then supplying prover options to ask for the SAT-QBF or something similar (see Boogie's /help). Otherwise, if you're interested in deciding these problems, then going directly to SMT solver is the way to go.

Floor and Ceiling Function implementation in Z3

I have tried to implement Floor and Ceiling Function as defined in the following link
https://math.stackexchange.com/questions/3619044/floor-or-ceiling-function-encoding-in-first-order-logic/3619320#3619320
But Z3 query returning counterexample.
Floor Function
_X=Real('_X')
_Y=Int('_Y')
_W=Int('_W')
_n=Int('_n')
_Floor=Function('_Floor',RealSort(),IntSort())
..
_s.add(_X>=0)
_s.add(_Y>=0)
_s.add(Implies(_Floor(_X)==_Y,And(Or(_Y==_X,_Y<_X),ForAll(_W,Implies(And(_W>=0,_W<_X),And(_W ==_Y,_W<_Y))))))
_s.add(Implies(And(Or(_Y==_X,_Y<_X),ForAll(_W,Implies(And(_W>=0,_W<_X),And(_W==_Y,_W<_Y))),_Floor(_X)==_Y))
_s.add(Not(_Floor(0.5)==0))
Expected Result - Unsat
Actual Result - Sat
Ceiling Function
_X=Real('_X')
_Y=Int('_Y')
_W=Int('_W')
_Ceiling=Function('_Ceiling',RealSort(),IntSort())
..
..
_s.add(_X>=0)
_s.add(_Y>=0)
_s.add(Implies(_Ceiling(_X)==_Y,And(Or(_Y==_X,_Y<_X),ForAll(_W,Implies(And(_W>=0,_W<_X),And(_W ==_Y,_Y<_W))))))
_s.add(Implies(And(Or(_Y==_X,_Y<_X),ForAll(_W,Implies(And(_W>=0,_W<_X),And(_W==_Y,_Y<_W)))),_Ceiling(_X)==_Y))
_s.add(Not(_Ceilng(0.5)==1))
Expected Result - Unsat
Actual Result - Sat
[Your encoding doesn't load to z3, it gives a syntax error even after eliminating the '..', as your call to Implies needs an extra argument. But I'll ignore all that.]
The short answer is, you can't really do this sort of thing in an SMT-Solver. If you could, then you can solve arbitrary Diophantine equations. Simply cast it in terms of Reals, solve it (there is a decision procedure for Reals), and then add the extra constraint that the result is an integer by saying Floor(solution) = solution. So, by this argument, you can see that modeling such functions will be beyond the capabilities of an SMT solver.
See this answer for details: Get fractional part of real in QF_UFNRA
Having said that, this does not mean you cannot code this up in Z3. It just means that it will be more or less useless. Here's how I would go about it:
from z3 import *
s = Solver()
Floor = Function('Floor',RealSort(),IntSort())
r = Real('R')
f = Int('f')
s.add(ForAll([r, f], Implies(And(f <= r, r < f+1), Floor(r) == f)))
Now, if I do this:
s.add(Not(Floor(0.5) == 0))
print(s.check())
you'll get unsat, which is correct. If you do this instead:
s.add(Not(Floor(0.5) == 1))
print(s.check())
you'll see that z3 simply loops forever. To make this usefull, you'd want the following to work as well:
test = Real('test')
s.add(test == 2.4)
result = Int('result')
s.add(Floor(test) == result)
print(s.check())
but again, you'll see that z3 simply loops forever.
So, bottom line: Yes, you can model such constructs, and z3 will correctly answer the simplest of queries. But with anything interesting, it'll simply loop forever. (Essentially whenever you'd expect sat and most of the unsat scenarios unless they can be constant-folded away, I'd expect z3 to simply loop.) And there's a very good reason for that, as I mentioned: Such theories are just not decidable and fall well out of the range of what an SMT solver can do.
If you are interested in modeling such functions, your best bet is to use a more traditional theorem prover, like Isabelle, Coq, ACL2, HOL, HOL-Light, amongst others. They are much more suited for working on these sorts of problems. And also, give a read to Get fractional part of real in QF_UFNRA as it goes into some of the other details of how you can go about modeling such functions using non-linear real arithmetic.

Modulo Power function in J

If I want to compute a^b mod c then there is an efficient way of doing this, without computing a^b in full.
However, when programming, if I write f g x then g(x) is computed regardless of f.
J provides for composing f and g in special cases, and the modulo power function is one of them. For instance, the following runs extremely quickly.
1000&| # (2&^) 10000000x
This is because the 'atop' conjunction # tells the language to compose the functions if possible. If I remove it, it goes unbearably slowly.
If however I want to work with x^x , then ^~ no longer works and I get limit errors for large values. Binding those large values however does work.
So
999&| # (100333454&^) 100333454x
executes nice and fast but
999&| # ^~ 100333454x
gives me a limit error - the RHS is too big.
Am I right in thinking that in this case, J is not using an efficient power-modulo algorithm?
Per the special code page:
m&|#^ dyad avoids exponentiation for integer arguments
m&|#(n&^) monad avoids exponentiation for integer arguments
The case for ^~ is not supported by special code.
https://github.com/Pascal-J/BN-openssl-bindings-for-J
includes bindings for openssl (distributed with J on windows, and pre-installed on all other compatible systems)'s BN library including modexp. The x ^ x case will be especially efficient due to fewer parameters converted from J to "C" (BN)

Determine upper/lower bound for variables in an arbitrary propositional formula [closed]

Closed. This question does not meet Stack Overflow guidelines. It is not currently accepting answers.
This question does not appear to be about programming within the scope defined in the help center.
Closed 8 years ago.
Improve this question
Given an arbitrary propositional formula PHI (linear constraints on some variables), what is the best way to determine the (approximate) upper and lower bound for each variable?
Some variables may be unbounded. In this case, an algorithm should conclude that there's no upper/lower bound for those variables.
e.g., PHI = (x=3 AND y>=1). The upper and lower bound for x are both 3. The lower bound for y is 1, and y does not have an upper bound.
This is a very simple example, but is there a solution in general (perhaps using SMT solver)?
This assumes the SAT/UNSAT domain of each variable is continuous.
Use an SMT solver to check for a solution to the formula. If there's no solution then that means no upper/lower bounds, so stop.
For each variable in the formula, conduct two binary searches over the domain of the variable, one searching for the lower bound, the other for the upper bound. The starting value in the search for each variable is the value for the variable in the solution found in step 1. Use the SMT solver to probe each search value for satisfiability and methodically bracket the bounds for each variable.
Pseudo code for the search functions, assuming integer domain variables:
lower_bound(variable, start, formula)
{
lo = -inf;
hi = start;
last_sat = start;
incr = 1;
do {
variable = (lo + hi) / 2;
if (SMT(formula) == UNSAT) {
lo = variable + incr;
} else {
last_sat = variable;
hi = variable - incr;
}
} while (lo <= hi);
return last_sat;
}
and
upper_bound(variable, start, formula)
{
lo = start;
hi = +inf;
last_sat = start;
do {
variable = (lo + hi) / 2;
if (SMT(formula) == SAT) {
last_sat = variable;
lo = variable + incr;
} else {
hi = variable - incr;
}
} while (lo <= hi);
return last_sat;
}
-inf/+inf are the smallest/largest values representable in the domain of each variable. If lower_bound returns -inf then the variable has no lower bound. If upper_bound returns +inf then the variable has no upper bound.
In practice most of such optimization problems require some sort of iterate-until-maximum/minimum kind of external driver on top of the SMT solver. Quantified approaches are also possible that can leverage the particular capabilities of the SMT solvers, but in practice such constraints end up being too hard for the underlying solver. See this discussion in particular: How to optimize a piece of code in Z3? (PI_NON_NESTED_ARITH_WEIGHT related)
Having said that, most high-level language bindings include the necessary machinery to simplify your life. For instance, if you use the Haskell SBV library to script z3, you can have:
Prelude> import Data.SBV
Prelude Data.SBV> maximize Quantified head 2 (\[x, y] -> x.==3 &&& y.>=1)
Just [3,1]
Prelude Data.SBV> maximize Quantified (head . tail) 2 (\[x, y] -> x.==3 &&& y.>=1)
Nothing
Prelude Data.SBV> minimize Quantified head 2 (\[x, y] -> x.==3 &&& y.>=1)
Just [3,1]
Prelude Data.SBV> minimize Quantified (head . tail) 2 (\[x, y] -> x.==3 &&& y.>=1)
Just [3,1]
The first result states x=3, y=1 maximizes x with respect to the predicate x==3 && y>=1.
The second result says there is no value that maximizes y with respect to the same predicate.
Third call says x=3,y=1 minimizes the predicate with respect to x.
Fourth call says x=3,y=1 minimizes the predicate with respect to y.
(See http://hackage.haskell.org/packages/archive/sbv/0.9.24/doc/html/Data-SBV.html#g:34 for details.)
You can also use the option "Iterative" (instead of Quantified) to have the library do the optimization iteratively instead of using quantifiers. These two techniques are not equivalent as the latter can get stuck in a local minima/maxima, but iterative approaches might solve problems where the quantified version is way too much to handle for the SMT solver.

Resources