Decision Procedures for Presburger Arithmetic Presented by Constantinos Bartzis

Presburger formulas numeral ::= 0 | 1 | 2… var ::= x | y | z … relop ::= term ::= numeral | var | term + term | -term | numeral  term formula ::= term relop term | formula  formula | formula  formula |  formula |  var. formula |  var. formula numeral  term isn't really multiplication; it's short-hand for term + term + … + term

Decision Procedures Will discuss algorithms for determining truth of formulas of Presburger arithmetic:  Fourier-Motzkin variable elimination (FMVE)  Omega Test  Cooper's algorithm  Automata based

Quantifier Elimination All the methods we'll look at are quantifier elimination procedures. If a formula with no free variables has no quantifiers, then it is easy to determine its truth value, e.g., 10 > 11  3+4 < 5  Quantifier elimination works by taking input P with n quantifiers and turning it into equivalent formula P’ with m quantifiers, and where m < n. So, eventually P  P’  …  Q and Q has no quantifiers. Q will be trivially true or false

Normalization Methods require input formulas to be normalized  e.g., collect coefficients, use only < and ≤ Methods eliminate innermost existential quantifiers. Universal quantifiers are normalized with (  x. P(x))   (  x.  P(x)) In FMVE, the sub-formula under the innermost existential quantifier must be a conjunction of relations. This means the inner formula must be converted to disjunctive normal form (DNF): (c 11  c 12  …  c 1n 1 ) ...  (c m1  c m2  …  c mn m )

Normalization (cont.) The formula under  is in DNF. Next, the  must be moved inwards First over disjuncts, using (  x. P  Q)  (  x. P)  (  x. Q) Must then ensure every conjunct under the quantifier mentions the bound variable. Use (  x. P(x)  Q)  (  x. P(x))  Q For example: (  x. 3 < x  x +2y ≤ 6  y < 0)  (  x. 3 < x  x +2y ≤ 6)  y < 0

Fourier-Motzkin theorems The following simple facts are the basis for a very simple quantifier elimination procedure. Over R (or Q), with a,b > 0: (  x. c ≤ ax  bx ≤ d)  bc ≤ ad (  x. c < ax  bx ≤ d)  bc < ad (  x. c ≤ ax  bx < d)  bc < ad (  x. c < ax  bx < d)  bc < ad In all four, the right hand side is implied by the left because of transitivity  e.g., (x < y  y ≤ z)  x < z

Fourier-Motzkin theorems (cont.) For the other direction: (bc < ad)  (  x. c < ax  bx ≤ d) take x to be d/b : c < a( d/b ), and b( d/b ) ≤ d. For (bc < ad)  (  x. c < ax  bx < d) take x to be (bc+ad)/2ab : c < a(bc+ad)/2ab  2bc < bc+ad  bc < ad Similarly for the other bound

Extending to a full procedure So far: a quantifier elimination procedure for formulas where the scope of each quantifier is 1 upper bound and 1 lower bound. We need to extend the method to cover cases with multiple constraints. No lower bound, many upper bounds: (  x: b 1 x < d 1  b 2 x < d 2 …  b n x < d n ) True! (take min(d i /b i ) as a witness for x) No upper bound, many lower bounds: obviously analogous.

Combining many constraints Example: (  x. c ≤ ax  b 1 x ≤ d 1  b 2 x ≤ d 2 )  b 1 c ≤ ad 1  b 2 c ≤ ad 2 From left to right, the result just depends on transitivity. From right to left, take x to be min(d 1 /b 1, d 2 /b 2 ). In general, with many constraints, combine all possible lower-upper bound pairs. Proof that this is possible is by induction on the number of constraints.

Combining many constraints The core elimination formula is With n constraints initially, evenly divided between upper and lower bounds, this formula generates n 2 /4 new constraints.

FMVE example  x. 20+x ≤ 0   y. 3y +x ≤ 10  20 ≤ y - x (re-arrange)   x. 20+x ≤ 0   y. 20+x ≤ y  3y ≤ 10 - x (eliminate y)   x. 20+x ≤ 0  60+3x ≤ 10 - x (re-arrange)   x. 20+x ≤ 0  4x +50 ≤ 0 (normalize universal)   x. 20+x ≤ 0  0 < 4x +50 (re-arrange)   x. -50 < 4x  x ≤ -20 (eliminate x)   (-50 < -80)  T

Complexity As before, when eliminating an existential over n constraints we may introduce n 2 /4 new constraints. With k quantifiers to eliminate, we might introduce as many as n 2 k /4 k constraints. If dealing with alternating quantifiers, repeated conversions to DNF may become very costly.

Integers

Expressiveness over Integers Can do divisibility by specific numerals: 2|e   x. 2x = e for example:  x. 0 < x < 30   (2|x  3|x  5|x) Can do integer division and modulus, as long as divisor is constant. Use one of the following results (similar for division) P(x mod d)   q,r. (x = qd +r )  (0 ≤ r < d  d < r ≤ 0)  P(r ) P(x mod d)   q,r. (x = qd +r )  (0 ≤ r < d  d < r ≤ 0)  P(r ) Any formula involving modulus or integer division by a constant can be translated to one without.

Expressivity over Integers Any procedure for Z trivially can be extended to one for N (or any mixture of N and Z) too: Add extra constraints stating that variables are  0 Relations < and ≤ can be converted into one another: x ≤ y  x < y +1 x < y  x +1 ≤ y Decision procedures normalize to one of these relations.

Fourier-Motzkin for Integers? Central theorem is false. E.g., (  x  Z. 3 ≤ 2x  2x ≤ 3)  6 ≤ 6 But one direction still works (thanks to transitivity): (  x. c ≤ ax  bx ≤ d)  bc ≤ ad We can compute consequences of existentially quantified formulas /

Fourier-Motzkin for Integers? We know (  x. c ≤ ax  bx ≤ d)  bc ≤ ad Thus an incomplete procedure for universal formulas over Z: Compute negation: (  x. P(x))   (  x.  P(x)) Compute consequences: if (  x.  P(x))   then (  x.  P(x))   and (  x. P(x))  T Repeat for all quantified variables. This is Phase 1 of the Omega Test

Omega Phase 1 - Example  x,y  Z. 0 < x  y < x  y +1 < 2x (normalize)   x,y. 1 ≤ x  y +1 ≤ x  2x ≤ y +1  x,y. 1 ≤ x  y +1 ≤ x  2x ≤ y +1 (eliminate y)   x. 1 ≤ x  2x ≤ x (normalize)   x. 1 ≤ x  x ≤ 0 (eliminate x)  1 ≤ 0  

Omega Phase 1 and the Interactive Theorem Provers The Omega Test's Phase 1 is used by systems like Coq, HOL4, HOL Light and Isabelle to decide arithmetic problems. Cons:  Incomplete  Conversion to DNF  Quadratic increase in numbers of constraints Pros:  Easy to implement  Easy to adapt the procedures to create proofs that can be checked by other tools

Some Shadows Given  x. (  i c i ≤ a i x)  (  j b j x ≤ d j ) The formula  i,j b j c i ≤ a i d j is known as the real shadow. If all of the a i or all of the b j are equal to 1, then the real shadow is exact. If the shadow is exact, then the two formulas are equivalent.

Exact Shadows When a = 1 or b = 1, the core theorem (  x. c ≤ ax  bx ≤ d)  bc ≤ ad is valid because   transitivity still holds   take x = d if b = 1 or x = c if a = 1 Omega Test's inventor, Bill Pugh, claims many problems in his domain (compiler optimization) have exact shadows. Experience suggests the same is true in other domains too, such as hardware model checking. When shadows are exact, we can pretend the problem is over R rather than Z and proceed as before.

Dark Shadows The formula  i,j (a i -1)(b j -1) ≤ a i d j - b j c i is known as the dark shadow. If all a i or all b j are one, then this is the same as the real shadow (or exact). The real shadow provides a test for unsatisfiability. The dark shadow tests for satisfiability, because (a-1)(b-1) ≤ ad - bc  (  x. c ≤ ax  bx ≤ d) This is the Phase 2 of the Omega Test

Omega Test Phases 1 & 2 Problem is  x. P(x) If input is exact for one or more elements of x, then eliminate them  x. P(x)   x’. P’(x’) Otherwise, calculate real shadow R:  x. P(x)  R so, if R  , then input formula is . Otherwise, calculate dark shadow D: D   x. P(x) so, if D = T, then input formula is T.

Omega Phase 2 - Example (a-1)(b-1) ≤ ad - bc  (  x. c ≤ ax  bx ≤ d)  x,y. 3x +2y ≤ 18  3y ≤ 4x  3x ≤ 2y +1 3y ≤ 4x  3x ≤ 2y +1 3y ≤ 4x  3x ≤ y 6 ≤ 8y y 6 ≤ y - 9y y ≤ -2 17y ≤ 66 y ≤ 3 This gives a suitable value for y, and by back- substitution, finds x = -1, y = -2 as a possible solution.

Splinters Purely existential formulas are often proved false by their real shadow; or proved true by their dark shadow But in “rare” cases, the main theorem is needed. Let m be the maximum of all the d j s. Then dark shadow splinter

Splinters A splinter does represent a smaller problem than the original because the extra equality allows x to be eliminated immediately. When quantifiers alternate, and there is no exact shadow, the main theorem is used as an equivalence, and splinters can't be avoided. Splinters must also be checked if neither real nor dark shadows decide an input formula.

Eliminating Equalities In an expression  x. …  cx = e  … the existential can be eliminated. First, multiply all terms involving x so that they have a common coefficient. Formula becomes  x. …c’x …  c’x = e’  …c’x… This is equivalent to …e’…  c’|e’  …e’…

Eliminating Divisibilities  x. …  c | dx + e  … Note: d < c (otherwise, replace d with d mod c). Introduce temporary new existential variable:  x,y. …  cy = dx + e  … Rearrange:  x,y. …  dx = cy -e  … Use equality elimination to derive  y. …  d | cy -e  … Because d < c, this process must terminate with elimination of divisibility term.

Implementation - Normalization Omega Test's main disadvantage is that it requires the matrix of the formula to be in DNF Consider  x. (x  10  x  11  9 < x ≤ 12)  x = 12 Negate, remove , <:  x. (x ≤ 9  11 ≤ x)  (x ≤ 10  12 ≤ x)  10 ≤ x  x ≤ 12  (x ≤ 11  13 ≤ x) Evaluate 8 (= 2 3 ) DNF terms. Clever preparation of input formulas can make orders of magnitude difference

Implementation - Normalization The propositional tautology (p  (q  q’))  (p  q  p  q’) justifies the following procedure: If P is an atomic formula, then when processing P  Q, assume P is true while processing Q:  If a sub-formula Q 0 of Q is such that P  Q 0, then replace Q 0 in Q by T.  If a sub-formula Q 0 of Q is such that P   Q 0, then replace Q 0 in Q by . Similarly, (  p  (q  q’))  (p  q  p  q’) for disjunctions.

Example Over  : 0 ≤ x + y + 4  (0 ≤ x + y + 6  0 ≤ 2x + 3y + 6) is equivalent to 0 ≤ x + y + 4 Whereas 0 ≤ x + y + 4  0 ≤ -x -y -6  0 ≤ 2x + 3y + 6 is equivalent to  Over  : 0 ≤ x + y + 4  0 ≤ x + y + 1  0 ≤ 2x + 3y + 6 is equivalent to 0 ≤ x + y + 4  0 ≤ 2x + 3y + 6

Cooper’s Algorithm

Cooper's Algorithm Cooper's algorithm is a decision procedure for Presburger arithmetic. A non-Fourier-Motzkin alternative It is also a quantifier elimination procedure, which also works from the inside out, eliminating existentials. Its advantage is that it doesn't need to normalize input formulas to DNF. Description is of simplest possible implementation; many tweaks are possible.

Preprocessing To eliminate the quantifier in  x. P(x): 1. Normalize so that only operators are <, and divisibility (c|e), and negations only occur around divisibility leaves. 2. Compute least common multiple c of all coefficients of x, and multiply all terms by appropriate numbers so that in every term the coefficient of x is c. 3. Now apply (  x. P(cx))  (  x. P(x)  c|x).

Preprocessing Example  x,y  Z. 0 < y  x < y  x +1 < 2y (normalize)   x,y. 0 < y  x < y  2y < x +2 (transform y to 2y everywhere)   x,y. 0 < 2y  2x < 2y  2y < x +2 (give y unit coefficient)   x,y. 0 < y  2x < y  y < x +2  2|y

Two ways How might  x. P(x) be true? Either:  there is a least x making P true; or  there is no least x: however small you go, there will be a smaller x that still makes P true Construct two formulas corresponding to both cases.

Case 1: Infinitely many small solutions Look at the atomic formulas in P, and think about their values when x has been made arbitrarily small:  x < e: if x becomes small enough, this will be T  e < x: if x becomes small enough, this will be   c | x+e: unchanged This constructs P - , a formula where x only occurs in divisibility terms. Let  be the l.c.m. of the constants involved in divisibility terms. We just need to test P -  on 1,…, .

P -  example For  y. 0 < y  2x < y  y < x +2  2|y  0 < y will become  as y gets small  2x < y also becomes  as y gets small  y < x +2 will be T as y gets small  2|y doesn't change (it tests if y is even or not) So in this case, P -  (y)  (     T  2|y)  

Case 2: Least solution exists The case when there is a least x satisfying P(x). For there to be a least x satisfying P(x), it must be the case that one of the terms e < x is T, and that if x was any smaller the formula would become . Let B = {b | b < x is a term of P(x)} Need just consider P(b+j), where b  B and 1 ≤ j ≤ . Final elimination formula is:

Example continued For  y. 0 < y  2x < y  y < x +2  2|y least solutions, if they exist, will be at y = 1, y = 2, y = 2x +1, or y = 2x +2 The divisibility constraint eliminates two of these. Original formula is equivalent to: (2x < 2  0 < x)  (0 < 2x +2  x < 0) Which is unsatisfiable.

Automata based

Symbolic Representation We use finite automata to represent the integer solutions (in binary) of atomic linear constraints ,0, ,1, Example: The constraint x 1  x 2  0 has solutions: (0,0), (1,0), (1,1), (2,0), (2,1), (2,2), (3,0), … The corresponding automaton

FA Construction Consider a finite state transducer that computes linear integer expressions / /101/1 1 / / / 1 11/011/0 00/100/1 Example x + 2y  /001/0 1 0 / 0

Equality with 0 Remove transitions that write 1. Make state 0 accepting states / /101/1 1 / / / 1 11/011/0 00/000/0 01/001/0 0 / 0

Inequality (<0) All transitions remain States with negative carries become accepting Same size / /101/1 1 / / / 1 11/011/0 00/000/0 01/001/0 0 / 0

Non-zero Constant Term c Same as before but now -c is the initial state If there is no such state, create one (and possibly some intermediate states) Size / /101/1 1 / / / 1 11/011/0 00/000/0 01/001/0 0 / 0

Boolean Connectives For  compute the intersection For  compute the union For  compute the complement

Conjunction example ,1, , ,1,1 Automaton for x-y< , , ,0,1 1 0, Automaton for 2x-y> , , , , , Automaton for x-y 0 -1,-1 0,-1 -2,-1 -1,0 -2,0-2,1

Existential Quantifier Elimination To eliminate  x, remove the track of x The resulting FA is in general non-deterministic Determinization may cause exponential blowup Rarely occurs in practice ,1, , ,1,1  x. x-y<1 0 1

Extensions / Improvements Using 2’s complement representation the method can be extended to all integers by just doubling the size of the automaton. Can also be extended to rationals, using automata on infinite words. Quantifier elimination can be done in linear time if the existential variable appears in an equation with odd coefficient. Currently outperforms Omega, but loses to DPLL(T) methods for quantifier-free inputs.

Questions?