def PerfectMatchingPrinciple(G):
"""Generates the clauses for the graph perfect matching principle.
The principle claims that there is a way to select edges to such
that all vertices have exactly one incident edge set to 1.
Parameters
----------
G : undirected graph
"""
cnf=CNF()
# Describe the formula
name="Perfect Matching Principle"
if hasattr(G,'name'):
cnf.header=name+" of graph:\n"+G.name+"\n"+cnf.header
else:
cnf.header=name+".\n"+cnf.header
def var_name(u,v):
if u<=v:
return 'x_{{{0},{1}}}'.format(u,v)
else:
return 'x_{{{0},{1}}}'.format(v,u)
# Each vertex has exactly one edge set to one.
for v in enumerate_vertices(G):
edge_vars = [var_name(u,v) for u in neighbors(G,v)]
cnf.add_equal_to(edge_vars,1)
return cnf
def CountingPrinciple(M, p):
"""Generates the clauses for the counting matching principle.
The principle claims that there is a way to partition M in sets of
size p each.
Arguments:
- `M` : size of the domain
- `p` : size of each class
"""
cnf = CNF()
# Describe the formula
name = "Counting Principle: {0} divided in parts of size {1}.".format(M, p)
cnf.header = name + "\n" + cnf.header
def var_name(tpl):
return "Y_{{" + ",".join("{0}".format(v) for v in tpl) + "}}"
# Incidence lists
incidence = [[] for _ in range(M)]
for tpl in combinations(list(range(M)), p):
for i in tpl:
incidence[i].append(tpl)
# Each element of the domain is in exactly one part.
for el in range(M):
edge_vars = [var_name(tpl) for tpl in incidence[el]]
cnf.add_equal_to(edge_vars, 1)
return cnf
def CountingPrinciple(M,p):
"""Generates the clauses for the counting matching principle.
The principle claims that there is a way to partition M in sets of
size p each.
Arguments:
- `M` : size of the domain
- `p` : size of each class
"""
cnf=CNF()
# Describe the formula
name="Counting Principle: {0} divided in parts of size {1}.".format(M,p)
cnf.header=name+"\n"+cnf.header
def var_name(tpl):
return "Y_{{"+",".join("{0}".format(v) for v in tpl)+"}}"
# Incidence lists
incidence=[[] for _ in range(M)]
for tpl in combinations(range(M),p):
for i in tpl:
incidence[i].append(tpl)
# Each element of the domain is in exactly one part.
for el in range(M):
edge_vars = [var_name(tpl) for tpl in incidence[el]]
cnf.add_equal_to(edge_vars,1)
return cnf
def TseitinFormula(graph,charges=None):
"""Build a Tseitin formula based on the input graph.
Odd charge is put on the first vertex by default, unless other
vertices are is specified in input.
Arguments:
- `graph`: input graph
- `charges': odd or even charge for each vertex
"""
V=enumerate_vertices(graph)
if charges==None:
charges=[1]+[0]*(len(V)-1) # odd charge on first vertex
else:
charges = [bool(c) for c in charges] # map to boolean
if len(charges)<len(V):
charges=charges+[0]*(len(V)-len(charges)) # pad with even charges
# init formula
tse=CNF()
for e in sorted(graph.edges(),key=sorted):
tse.add_variable(edgename(e))
# add constraints
for v,c in zip(V,charges):
# produce all clauses and save half of them
names = [ edgename((u,v)) for u in neighbors(graph,v) ]
for cls in parity_constraint(names,c):
tse.add_clause(list(cls),strict=True)
return tse
def PerfectMatchingPrinciple(graph):
"""Generates the clauses for the graph perfect matching principle.
The principle claims that there is a way to select edges to such
that all vertices have exactly one incident edge set to 1.
Arguments:
- `graph` : undirected graph
"""
cnf=CNF()
# Describe the formula
name="Perfect Matching Principle"
if hasattr(graph,'name'):
cnf.header=name+" of graph:\n"+graph.name+"\n"+cnf.header
else:
cnf.header=name+".\n"+cnf.header
def var_name(u,v):
if u<=v:
return 'x_{{{0},{1}}}'.format(u,v)
else:
return 'x_{{{0},{1}}}'.format(v,u)
# Each vertex has exactly one edge set to one.
for v in graph.nodes():
edge_vars = [var_name(u,v) for u in graph.adj[v]]
for cls in equal_to_constraint(edge_vars,1):
cnf.add_clause(cls)
return cnf
def test_one_eq(self) :
opb="""\
* #variable= 5 #constraint= 1
*
+1 x1 +1 x2 +1 x3 +1 x4 +1 x5 = 2;
"""
F=CNF()
F.add_equal_to(["a","b","c","d","e"],2)
self.assertCnfEqualsOPB(F,opb)
def test_weighted_gt(self) :
opb="""\
* #variable= 5 #constraint= 1
*
+1 x1 +2 x2 +3 x3 -1 x4 -2 x5 >= 3;
"""
F=CNF()
F.add_linear(1,"a",2,"b",3,"c",-1,"d",-2,"e",">",2)
self.assertCnfEqualsOPB(F,opb)
def test_one_eq(self):
opb = """\
* #variable= 5 #constraint= 1
*
+1 x1 +1 x2 +1 x3 +1 x4 +1 x5 = 2;
"""
F = CNF()
F.add_equal_to(["a", "b", "c", "d", "e"], 2)
self.assertCnfEqualsOPB(F, opb)
def test_one_leq(self):
opb = """\
* #variable= 5 #constraint= 1
*
-1 x1 -1 x2 -1 x3 -1 x4 -1 x5 >= -2;
"""
F = CNF()
F.add_less_or_equal(["a", "b", "c", "d", "e"], 2)
self.assertCnfEqualsOPB(F, opb)
def test_one_gt(self):
opb = """\
* #variable= 5 #constraint= 1
*
+1 x1 +1 x2 +1 x3 +1 x4 +1 x5 >= 3;
"""
F = CNF()
F.add_strictly_greater_than(["a", "b", "c", "d", "e"], 2)
self.assertCnfEqualsOPB(F, opb)
def test_one_clause(self):
opb = """\
* #variable= 4 #constraint= 1
*
+1 x1 +1 x2 -1 x3 -1 x4 >= -1;
"""
F = CNF()
F.add_clause([(True, "a"), (True, "b"), (False, "c"), (False, "d")])
self.assertCnfEqualsOPB(F, opb)
def test_weighted_gt(self):
opb = """\
* #variable= 5 #constraint= 1
*
+1 x1 +2 x2 +3 x3 -1 x4 -2 x5 >= 3;
"""
F = CNF()
F.add_linear(1, "a", 2, "b", 3, "c", -1, "d", -2, "e", ">", 2)
self.assertCnfEqualsOPB(F, opb)
def test_weighted_leq(self):
opb = """\
* #variable= 5 #constraint= 1
*
-1 x1 -2 x2 -3 x3 +1 x4 +2 x5 >= -2;
"""
F = CNF()
F.add_linear(1, "a", 2, "b", 3, "c", -1, "d", -2, "e", "<=", 2)
self.assertCnfEqualsOPB(F, opb)
def test_one_lt(self):
opb = """\
* #variable= 5 #constraint= 1
*
-1 x1 -1 x2 -1 x3 -1 x4 -1 x5 >= -1;
"""
F = CNF()
F.add_strictly_less_than(["a", "b", "c", "d", "e"], 2)
self.assertCnfEqualsOPB(F, opb)
def test_one_lt(self) :
opb="""\
* #variable= 5 #constraint= 1
*
-1 x1 -1 x2 -1 x3 -1 x4 -1 x5 >= -1;
"""
F=CNF()
F.add_strictly_less_than(["a","b","c","d","e"],2)
self.assertCnfEqualsOPB(F,opb)
def test_one_gt(self) :
opb="""\
* #variable= 5 #constraint= 1
*
+1 x1 +1 x2 +1 x3 +1 x4 +1 x5 >= 3;
"""
F=CNF()
F.add_strictly_greater_than(["a","b","c","d","e"],2)
self.assertCnfEqualsOPB(F,opb)
def test_one_leq(self) :
opb="""\
* #variable= 5 #constraint= 1
*
-1 x1 -1 x2 -1 x3 -1 x4 -1 x5 >= -2;
"""
F=CNF()
F.add_less_or_equal(["a","b","c","d","e"],2)
self.assertCnfEqualsOPB(F,opb)
def test_one_clause(self) :
opb="""\
* #variable= 4 #constraint= 1
*
+1 x1 +1 x2 -1 x3 -1 x4 >= -1;
"""
F=CNF()
F.add_clause([(True,"a"),(True,"b"),(False,"c"),(False,"d")])
self.assertCnfEqualsOPB(F,opb)
def test_weighted_leq(self) :
opb="""\
* #variable= 5 #constraint= 1
*
-1 x1 -2 x2 -3 x3 +1 x4 +2 x5 >= -2;
"""
F=CNF()
F.add_linear(1,"a",2,"b",3,"c",-1,"d",-2,"e","<=",2)
self.assertCnfEqualsOPB(F,opb)
def test_strict_clause_insertion(self):
F=CNF()
F.mode_strict()
F.add_variable("S")
F.add_variable("T")
F.add_variable("U")
self.assertTrue(len(list(F.variables()))==3)
F.add_clause([(True,"S"),(False,"T")])
F.add_clause([(True,"T"),(False,"U")])
self.assertRaises(ValueError, F.add_clause,
[(True,"T"),(False,"V")])
def test_parity(self) :
opb="""\
* #variable= 3 #constraint= 4
*
+1 x1 +1 x2 +1 x3 >= 1;
+1 x1 -1 x2 -1 x3 >= -1;
-1 x1 +1 x2 -1 x3 >= -1;
-1 x1 -1 x2 +1 x3 >= -1;
"""
F=CNF()
F.add_parity(["a","b","c"],1)
self.assertCnfEqualsOPB(F,opb)
def EvenColoringFormula(G):
"""Even coloring formula
The formula is defined on a graph :math:`G` and claims that it is
possible to split the edges of the graph in two parts, so that
each vertex has an equal number of incident edges in each part.
The formula is defined on graphs where all vertices have even
degree. The formula is satisfiable only on those graphs with an
even number of vertices in each connected component [1]_.
Arguments
---------
G : networkx.Graph
a simple undirected graph where all vertices have even degree
Raises
------
ValueError
if the graph in input has a vertex with odd degree
Returns
-------
CNF object
References
----------
.. [1] Locality and Hard SAT-instances, Klas Markstrom
Journal on Satisfiability, Boolean Modeling and Computation 2 (2006) 221-228
"""
F = CNF()
F.mode_strict()
F.header = "Even coloring formula on graph " + G.name + "\n" + F.header
def var_name(u, v):
if u <= v:
return 'x_{{{0},{1}}}'.format(u, v)
else:
return 'x_{{{0},{1}}}'.format(v, u)
for (u, v) in enumerate_edges(G):
F.add_variable(var_name(u, v))
# Defined on both side
for v in enumerate_vertices(G):
if G.degree(v) % 2 == 1:
raise ValueError(
"Markstrom formulas requires all vertices to have even degree."
)
edge_vars = [var_name(u, v) for u in neighbors(G, v)]
# F.add_exactly_half_floor would work the same
F.add_exactly_half_ceil(edge_vars)
return F
def test_parity(self):
opb = """\
* #variable= 3 #constraint= 4
*
+1 x1 +1 x2 +1 x3 >= 1;
+1 x1 -1 x2 -1 x3 >= -1;
-1 x1 +1 x2 -1 x3 >= -1;
-1 x1 -1 x2 +1 x3 >= -1;
"""
F = CNF()
F.add_parity(["a", "b", "c"], 1)
self.assertCnfEqualsOPB(F, opb)
def build_cnf(args):
"""Build an disjunction
Arguments:
- `args`: command line options
"""
clause = [ (True,"x_{}".format(i)) for i in range(args.P) ] + \
[ (False,"y_{}".format(i)) for i in range(args.N) ]
orcnf = CNF([clause])
orcnf.header = "Clause with {} positive and {} negative literals\n\n".format(args.P,args.N) + \
orcnf.header
return orcnf
def build_cnf(args):
"""Build a conjunction
Arguments:
- `args`: command line options
"""
clauses = [ [(True,"x_{}".format(i))] for i in range(args.P) ] + \
[ [(False,"y_{}".format(i))] for i in range(args.N) ]
andcnf = CNF(clauses)
andcnf.header = "Singleton clauses: {} positive and {} negative\n\n""".format(args.P,args.N) +\
andcnf.header
return andcnf
def build_cnf(args):
"""Build a conjunction
Arguments:
- `args`: command line options
"""
clauses = [ [(True,"x_{}".format(i))] for i in range(args.P) ] + \
[ [(False,"y_{}".format(i))] for i in range(args.N) ]
andcnf = CNF(clauses)
andcnf.header = "Singleton clauses: {} positive and {} negative\n\n""".format(args.P,args.N) +\
andcnf.header
return andcnf
def build_cnf(args):
"""Build an disjunction
Arguments:
- `args`: command line options
"""
clause = [ (True,"x_{}".format(i)) for i in range(args.P) ] + \
[ (False,"y_{}".format(i)) for i in range(args.N) ]
orcnf = CNF([clause])
orcnf.header = "Clause with {} positive and {} negative literals\n\n".format(args.P,args.N) + \
orcnf.header
return orcnf
def getFormula(self):
self.f = CNF()
self.f._header += headerInfo
self.addAllVariables()
result = fe.And()
for i in range(self.numRows):
result.append(self.addRowConstraint(i))
for j in range(self.numCols):
result.append(self.addColConstraint(j))
fe.toCNFgen(self.f, result)
return self.f
def BinaryPigeonholePrinciple(pigeons, holes):
"""Binary Pigeonhole Principle CNF formula
The pigeonhole principle claims that no M pigeons can sit in
N pigeonholes without collision if M>N. This formula encodes the
principle using binary strings to identify the holes.
Parameters
----------
pigeon : int
number of pigeons
holes : int
number of holes
"""
bphp = CNF()
bphp.header="Binary Pigeonhole Principle for {0} pigeons and {1} holes\n".format(pigeons,holes)\
+ bphp.header
bphpgen = bphp.binary_mapping(xrange(1, pigeons + 1),
xrange(1, holes + 1),
injective=True)
for v in bphpgen.variables():
bphp.add_variable(v)
for c in bphpgen.clauses():
bphp.add_clause_unsafe(c)
return bphp
def PythagoreanTriples(N):
"""There is a Pythagorean triples free coloring on N
The formula claims that it is possible to bicolor the numbers from
1 to :math:`N` so that there is no monochromatic triplet
:math:`(x,y,z)` so that :math:`x^2+y^2=z^2`.
Parameters
----------
N : int
size of the interval
Return
------
A CNF object
Raises
------
ValueError
Parameters are not positive integers
References
----------
.. [1] M. J. Heule, O. Kullmann, and V. W. Marek.
Solving and verifying the boolean pythagorean triples problem via cube-and-conquer.
arXiv preprint arXiv:1605.00723, 2016.
"""
ptn = CNF()
ptn.header = dedent("""
It is possible to bicolor the numbers from
1 to {} so that there is no monochromatic triplets
(x,y,z) such that x^2+y^2=z^2
""".format(N)) + ptn.header
def V(i):
return "x_{{{}}}".format(i)
# Variables represent the coloring of the number
for i in range(1, N + 1):
ptn.add_variable(V(i))
for x, y in combinations(list(range(1, N + 1)), 2):
z = int(sqrt(x**2 + y**2))
if z <= N and z**2 == x**2 + y**2:
ptn.add_clause([(True, V(x)), (True, V(y)), (True, V(z))],
strict=True)
ptn.add_clause([(False, V(x)), (False, V(y)), (False, V(z))],
strict=True)
return ptn
def test_empty(self):
opb = """\
* #variable= 0 #constraint= 0
*
"""
F = CNF()
self.assertCnfEqualsOPB(F, opb)
def GraphAutomorphism(G):
"""Graph Automorphism formula
The formula is the CNF encoding of the statement that a graph G
has a nontrivial automorphism, i.e. an automorphism different from
the idential one.
Parameter
---------
G : a simple graph
Returns
-------
A CNF formula which is satiafiable if and only if graph G has a
nontrivial automorphism.
"""
tmp = CNF()
header = "Graph automorphism formula for graph " + G.name + "\n" + tmp.header
F = GraphIsomorphism(G, G)
F.header = header
var = _graph_isomorphism_var
F.add_clause([(False, var(u, u)) for u in enumerate_vertices(G)],
strict=True)
return F
def EvenColoringFormula(G):
"""Even coloring formula
The formula is defined on a graph :math:`G` and claims that it is
possible to split the edges of the graph in two parts, so that
each vertex has an equal number of incident edges in each part.
The formula is defined on graphs where all vertices have even
degree. The formula is satisfiable only on those graphs with an
even number of vertices in each connected component [1]_.
Arguments
---------
G : networkx.Graph
a simple undirected graph where all vertices have even degree
Raises
------
ValueError
if the graph in input has a vertex with odd degree
Returns
-------
CNF object
References
----------
.. [1] Locality and Hard SAT-instances, Klas Markstrom
Journal on Satisfiability, Boolean Modeling and Computation 2 (2006) 221-228
"""
F = CNF()
F.mode_strict()
F.header = "Even coloring formula on graph " + G.name + "\n" + F.header
def var_name(u,v):
if u<=v:
return 'x_{{{0},{1}}}'.format(u,v)
else:
return 'x_{{{0},{1}}}'.format(v,u)
for (u, v) in enumerate_edges(G):
F.add_variable(var_name(u, v))
# Defined on both side
for v in enumerate_vertices(G):
if G.degree(v) % 2 == 1:
raise ValueError("Markstrom formulas requires all vertices to have even degree.")
edge_vars = [ var_name(u, v) for u in neighbors(G, v) ]
F.add_exactly_half_ceil(edge_vars) # F.add_exactly_half_floor would work the same
return F
def BinaryPigeonholePrinciple(pigeons,holes):
"""Binary Pigeonhole Principle CNF formula
The pigeonhole principle claims that no M pigeons can sit in
N pigeonholes without collision if M>N. This formula encodes the
principle using binary strings to identify the holes.
Parameters
----------
pigeon : int
number of pigeons
holes : int
number of holes
"""
bphp=CNF()
bphp.header="Binary Pigeonhole Principle for {0} pigeons and {1} holes\n".format(pigeons,holes)\
+ bphp.header
bphpgen=bphp.binary_mapping(xrange(1,pigeons+1), xrange(1,holes+1), injective = True)
for v in bphpgen.variables():
bphp.add_variable(v)
for c in bphpgen.clauses():
bphp.add_clause(c,strict=True)
return bphp
def HiddenPigeonholePrinciple(pigeons, items, holes):
"""Hidden Pigeonhole Principle CNF formula
Each pigeon has multiple items that should be fit into a hole, however each
hole can only contain items of at most one pigeon.
It is called hidden pigeon hole principle as it is only adding more
constraints to the origianal pigeon hole principle, which makes it difficult
to detect "the right" at-most-one constraints.
Arguments:
- `pigeons`: number of pigeons
- `items`: number of items per pigeon
- `holes`: number of holes
"""
def var(p, i, h):
return 'p_{{{0},{1},{2}}}'.format(p, h, i)
php = CNF()
php.header = "hidden pigeon hole principle formula for {0} pigeons, each having {2} items ,and {1} holes\n".format(pigeons, holes, items)\
+ php.header
for p in range(pigeons):
for i in range(items):
php.add_clause([(True, var(p, i, h)) for h in range(holes)])
for h in range(holes):
for p1 in range(pigeons):
for p2 in range(p1):
for i1 in range(items):
for i2 in range(items):
php.add_clause([(False, var(p1, i1, h)),
(False, var(p2, i2, h))])
return php
def BinaryPigeonholePrinciple(pigeons, holes):
"""Binary Pigeonhole Principle CNF formula
The pigeonhole principle claims that no M pigeons can sit in
N pigeonholes without collision if M>N. This formula encodes the
principle using binary strings to identify the holes.
Parameters
----------
pigeon : int
number of pigeons
holes : int
number of holes
"""
bphp = CNF()
bphp.header = "Binary Pigeonhole Principle for {0} pigeons and {1} holes\n".format(pigeons, holes)\
+ bphp.header
mapping = binary_mapping(range(1, pigeons + 1),
range(1, holes + 1),
injective=True)
bphp.mode_unchecked()
mapping.load_variables_to_formula(bphp)
mapping.load_clauses_to_formula(bphp)
bphp.mode_default()
return bphp
def build_cnf(args):
"""Build an empty CNF formula
Parameters
----------
args : ignored
command line options
"""
return CNF()
def build_cnf(args):
"""Build a CNF formula with an empty clause
Parameters
----------
args : ignored
command line options
"""
return CNF([[]])
def test_one(self):
ass = [{'x_1': True, 'x_2': False}]
F = RandomKCNF(2, 2, 3, planted_assignments=ass)
G = CNF([
[(True, 'x_1'), (False, 'x_2')],
[(False, 'x_1'), (False, 'x_2')],
[(True, 'x_1'), (True, 'x_2')],
])
self.assertCnfEqual(F, G)
def test_most(self):
ass = [
{'x_1': True, 'x_2': False},
{'x_1': False, 'x_2': False},
{'x_1': True, 'x_2': True},
]
F = RandomKCNF(2, 2, 1, planted_assignments=ass)
G = CNF([[(True, 'x_1'), (False, 'x_2')]])
self.assertCnfEqual(F, G)
def test_opposite_literals(self):
F = CNF()
F.auto_add_variables = True
F.allow_opposite_literals = True
F.add_clause([(True, "S"), (False, "T"), (False, "S")])
F.allow_opposite_literals = False
self.assertRaises(ValueError, F.add_clause, [(True, "T"), (True, "V"),
(False, "T")])
def BinaryCliqueFormula(G, k):
"""Test whether a graph has a k-clique.
Given a graph :math:`G` and a non negative value :math:`k`, the
CNF formula claims that :math:`G` contains a :math:`k`-clique.
This formula uses the binary encoding, in the sense that the
clique elements are indexed by strings of bits.
Parameters
----------
G : networkx.Graph
a simple graph
k : a non negative integer
clique size
Returns
-------
a CNF object
"""
F = CNF()
F.header = "Binary {0}-clique formula\n".format(k) + F.header
clauses_gen = F.binary_mapping(xrange(1, k + 1),
G.nodes(),
injective=True,
nondecreasing=True)
for v in clauses_gen.variables():
F.add_variable(v)
for c in clauses_gen.clauses():
F.add_clause(c, strict=True)
for (i1, i2), (v1, v2) in product(combinations(xrange(1, k + 1), 2),
combinations(G.nodes(), 2)):
if not G.has_edge(v1, v2):
F.add_clause(clauses_gen.forbid_image(i1, v1) +
clauses_gen.forbid_image(i2, v2),
strict=True)
return F
def RamseyLowerBoundFormula(s,k,N):
"""Formula claiming that Ramsey number r(s,k) > N
Arguments:
- `s`: independent set size
- `k`: clique size
- `N`: vertices
"""
ram=CNF()
ram.mode_strict()
ram.header=dedent("""\
CNF encoding of the claim that there is a graph of %d vertices
with no independent set of size %d and no clique of size %d
""" % (N,s,k)) + ram.header
#
# One variable per edge (indices are ordered)
#
for edge in combinations(xrange(1,N+1),2):
ram.add_variable('e_{{{0},{1}}}'.format(*edge))
#
# No independent set of size s
#
for vertex_set in combinations(xrange(1,N+1),s):
clause=[]
for edge in combinations(vertex_set,2):
clause += [(True,'e_{{{0},{1}}}'.format(*edge))]
ram.add_clause(clause)
#
# No clique of size k
#
for vertex_set in combinations(xrange(1,N+1),k):
clause=[]
for edge in combinations(vertex_set,2):
clause+=[(False,'e_{{{0},{1}}}'.format(*edge))]
ram.add_clause(clause)
return ram
def RandomKCNF(k, n, m, seed=None, planted_assignments=[]):
"""Build a random k-CNF
Sample :math:`m` clauses over :math:`n` variables, each of width
:math:`k`, uniformly at random. The sampling is done without
repetition, meaning that whenever a randomly picked clause is
already in the CNF, it is sampled again.
Parameters
----------
k : int
width of each clause
n : int
number of variables to choose from. The resulting CNF object
will contain n variables even if some are not mentioned in the clauses.
m : int
number of clauses to generate
seed : hashable object
seed of the random generator
planted_assignments : iterable(dict), optional
a set of total/partial assigments such that all clauses in the formula
will be satisfied by all of them.
Returns
-------
a CNF object
Raises
------
ValueError
when some paramenter is negative, or when k>n.
"""
if seed:
random.seed(seed)
if n<0 or m<0 or k<0:
raise ValueError("Parameters must be non-negatives.")
if k>n:
raise ValueError("Clauses cannot have more {} literals.".format(n))
F = CNF()
F.header = "Random {}-CNF over {} variables and {} clauses\n".format(k,n,m) + F.header
indices = xrange(1,n+1)
for i in indices:
F.add_variable('x_{0}'.format(i))
try:
for clause in sample_clauses(k, indices, m, planted_assignments):
F.add_clause(list(clause), strict=True)
except ValueError:
raise ValueError("There are fewer clauses available than the number requested")
return F
def PebblingFormula(digraph):
"""Pebbling formula
Build a pebbling formula from the directed graph. If the graph has
an `ordered_vertices` attribute, then it is used to enumerate the
vertices (and the corresponding variables).
Arguments:
- `digraph`: directed acyclic graph.
"""
if not is_dag(digraph):
raise ValueError(
"Pebbling formula is defined only for directed acyclic graphs")
peb = CNF()
if hasattr(digraph, 'name'):
peb.header = "Pebbling formula of: " + digraph.name + "\n\n" + peb.header
else:
peb.header = "Pebbling formula\n\n" + peb.header
# add variables in the appropriate order
vertices = enumerate_vertices(digraph)
position = dict((v, i) for (i, v) in enumerate(vertices))
for v in vertices:
peb.add_variable(
v, description="There is a pebble on vertex ${}$".format(v))
# add the clauses
for v in vertices:
# If predecessors are pebbled the vertex must be pebbled
pred = sorted(digraph.predecessors(v), key=lambda x: position[x])
peb.add_clause_unsafe([(False, p) for p in pred] + [(True, v)])
if digraph.out_degree(v) == 0: #the sink
peb.add_clause_unsafe([(False, v)])
return peb
def PerfectMatchingPrinciple(G):
"""Generates the clauses for the graph perfect matching principle.
The principle claims that there is a way to select edges to such
that all vertices have exactly one incident edge set to 1.
Parameters
----------
G : undirected graph
"""
cnf = CNF()
# Describe the formula
name = "Perfect Matching Principle"
if hasattr(G, 'name'):
cnf.header = name + " of graph:\n" + G.name + "\n" + cnf.header
else:
cnf.header = name + ".\n" + cnf.header
def var_name(u, v):
if u <= v:
return 'x_{{{0},{1}}}'.format(u, v)
else:
return 'x_{{{0},{1}}}'.format(v, u)
# Each vertex has exactly one edge set to one.
for v in enumerate_vertices(G):
edge_vars = [var_name(u, v) for u in neighbors(G, v)]
cnf.add_equal_to(edge_vars, 1)
return cnf
def BinaryPigeonholePrinciple(pigeons,holes):
"""Binary Pigeonhole Principle CNF formula
The pigeonhole principle claims that no M pigeons can sit in
N pigeonholes without collision if M>N. This formula encodes the
principle using binary strings to identify the holes.
Parameters
----------
pigeon : int
number of pigeons
holes : int
number of holes
"""
bphp=CNF()
bphp.header="Binary Pigeonhole Principle for {0} pigeons and {1} holes\n".format(pigeons,holes)\
+ bphp.header
mapping=binary_mapping(xrange(1,pigeons+1),
xrange(1,holes+1), injective = True)
bphp.mode_unchecked()
mapping.load_variables_to_formula(bphp)
mapping.load_clauses_to_formula(bphp)
bphp.mode_default()
return bphp
def __call__(self):
mn = self.m * self.n
A = [0, 1] * (mn // 2) + [0] * (mn % 2)
random.shuffle(A)
A = [[A[i * self.n + j] for j in range(self.n)] for i in range(self.m)]
b = [0, 1] * (self.m // 2) + [0] * (self.m % 2)
self.f = CNF()
terms = [(-1, self.x(i)) for i in range(self.n)]
if self.eq:
self.f.add_linear(*toArgs(terms, "==", -self.k))
else:
self.f.add_linear(*toArgs(terms, ">=", -self.k))
for i in range(self.m):
terms = [(1, self.x(j)) for j in range(self.n) if A[i][j] == 1]
if len(terms) > 0:
self.addXor(terms, b[i])
return self.f
def test_opposite_literals(self):
F=CNF()
F.auto_add_variables = True
F.allow_opposite_literals = True
F.add_clause([(True,"S"),(False,"T"),(False,"S")])
F.allow_opposite_literals = False
self.assertRaises(ValueError, F.add_clause,
[(True,"T"),(True,"V"),(False,"T")])
def BinaryCliqueFormula(G,k):
"""Test whether a graph has a k-clique.
Given a graph :math:`G` and a non negative value :math:`k`, the
CNF formula claims that :math:`G` contains a :math:`k`-clique.
This formula uses the binary encoding, in the sense that the
clique elements are indexed by strings of bits.
Parameters
----------
G : networkx.Graph
a simple graph
k : a non negative integer
clique size
Returns
-------
a CNF object
"""
F=CNF()
F.header="Binary {0}-clique formula\n".format(k) + F.header
mapping=CNF.binary_mapping(xrange(1,k+1), G.nodes(),
injective = True,
nondecreasing = True)
mapping.load_variables_to_formula(F)
mapping.load_clauses_to_formula(F)
for (i1,i2),(v1,v2) in product(combinations(xrange(1,k+1),2),
combinations(G.nodes(),2)):
if not G.has_edge(v1,v2):
F.add_clause( clauses_gen.forbid_image(i1,v1) +
clauses_gen.forbid_image(i2,v2))
return F
def RamseyLowerBoundFormula(s, k, N):
"""Formula claiming that Ramsey number r(s,k) > N
Arguments:
- `s`: independent set size
- `k`: clique size
- `N`: vertices
"""
ram = CNF()
ram.header = dedent("""\
CNF encoding of the claim that there is a graph of %d vertices
with no independent set of size %d and no clique of size %d
""" % (N, s, k)) + ram.header
#
# One variable per edge (indices are ordered)
#
for edge in combinations(range(1, N + 1), 2):
ram.add_variable('e_{{{0},{1}}}'.format(*edge))
#
# No independent set of size s
#
for vertex_set in combinations(range(1, N + 1), s):
clause = []
for edge in combinations(vertex_set, 2):
clause += [(True, 'e_{{{0},{1}}}'.format(*edge))]
ram.add_clause(clause, strict=True)
#
# No clique of size k
#
for vertex_set in combinations(range(1, N + 1), k):
clause = []
for edge in combinations(vertex_set, 2):
clause += [(False, 'e_{{{0},{1}}}'.format(*edge))]
ram.add_clause(clause, strict=True)
return ram
def PebblingFormula(digraph):
"""Pebbling formula
Build a pebbling formula from the directed graph. If the graph has
an `ordered_vertices` attribute, then it is used to enumerate the
vertices (and the corresponding variables).
Arguments:
- `digraph`: directed acyclic graph.
"""
if not is_dag(digraph):
raise ValueError("Pebbling formula is defined only for directed acyclic graphs")
peb=CNF()
if hasattr(digraph,'name'):
peb.header="Pebbling formula of: "+digraph.name+"\n\n"+peb.header
else:
peb.header="Pebbling formula\n\n"+peb.header
# add variables in the appropriate order
vertices=enumerate_vertices(digraph)
position=dict((v,i) for (i,v) in enumerate(vertices))
for v in vertices:
peb.add_variable(v,description="There is a pebble on vertex ${}$".format(v))
# add the clauses
for v in vertices:
# If predecessors are pebbled the vertex must be pebbled
pred=sorted(digraph.predecessors(v),key=lambda x:position[x])
peb.add_clause_unsafe([(False,p) for p in pred]+[(True,v)])
if digraph.out_degree(v)==0: #the sink
peb.add_clause_unsafe([(False,v)])
return peb
def PythagoreanTriples(N):
"""There is a Pythagorean triples free coloring on N
The formula claims that it is possible to bicolor the numbers from
1 to :math:`N` so that there is no monochromatic triplet
:math:`(x,y,z)` so that :math:`x^2+y^2=z^2`.
Parameters
----------
N : int
size of the interval
Return
------
A CNF object
Raises
------
ValueError
Parameters are not positive integers
References
----------
.. [1] M. J. Heule, O. Kullmann, and V. W. Marek.
Solving and verifying the boolean pythagorean triples problem via cube-and-conquer.
arXiv preprint arXiv:1605.00723, 2016.
"""
ptn=CNF()
ptn.header=dedent("""
It is possible to bicolor the numbers from
1 to {} so that there is no monochromatic triplets
(x,y,z) such that x^2+y^2=z^2
""".format(N)) + ptn.header
def V(i):
return "x_{{{}}}".format(i)
# Variables represent the coloring of the number
for i in xrange(1,N+1):
ptn.add_variable(V(i))
for x,y in combinations(range(1,N+1),2):
z = int(sqrt(x**2 + y**2))
if z <=N and z**2 == x**2 + y**2:
ptn.add_clause([ (True, V(x)), (True, V(y)), (True, V(z))])
ptn.add_clause([ (False,V(x)), (False,V(y)), (False,V(z))])
return ptn
def VertexCover(G, d):
F = CNF()
def D(v):
return "x_{{{0}}}".format(v)
def N(v):
return tuple(sorted([e for e in G.edges(v)]))
# Fix the vertex order
V = enumerate_vertices(G)
# Create variables
for v in V:
F.add_variable(D(v))
# Not too many true variables
F.add_less_or_equal([D(v) for v in V], d)
# Every edge must have a true D variable
for e in G.edges():
F.add_clause([(True, D(v)) for v in e])
return F
def VertexCover(G,d):
F=CNF()
def D(v):
return "x_{{{0}}}".format(v)
def N(v):
return tuple(sorted([ e for e in G.edges(v) ]))
# Fix the vertex order
V=enumerate_vertices(G)
# Create variables
for v in V:
F.add_variable(D(v))
# Not too many true variables
F.add_less_or_equal([D(v) for v in V],d)
# Every edge must have a true D variable
for e in G.edges():
F.add_clause([ (True,D(v)) for v in e])
return F
def DominatingSet(G,d, alternative = False):
r"""Generates the clauses for a dominating set for G of size <= d
The formula encodes the fact that the graph :math:`G` has
a dominating set of size :math:`d`. This means that it is possible
to pick at most :math:`d` vertices in :math:`V(G)` so that all remaining
vertices have distance at most one from the selected ones.
Parameters
----------
G : networkx.Graph
a simple undirected graph
d : a positive int
the size limit for the dominating set
alternative : bool
use an alternative construction that
is provably hard from resolution proofs.
Returns
-------
CNF
the CNF encoding for dominating of size :math:`\leq d` for graph :math:`G`
"""
F=CNF()
if not isinstance(d,int) or d<1:
ValueError("Parameter \"d\" is expected to be a positive integer")
# Describe the formula
name="{}-dominating set".format(d)
if hasattr(G,'name'):
F.header=name+" of graph:\n"+G.name+".\n\n"+F.header
else:
F.header=name+".\n\n"+F.header
# Fix the vertex order
V=enumerate_vertices(G)
def D(v):
return "x_{{{0}}}".format(v)
def M(v,i):
return "g_{{{0},{1}}}".format(v,i)
def N(v):
return tuple(sorted([ v ] + [ u for u in G.neighbors(v) ]))
# Create variables
for v in V:
F.add_variable(D(v))
for i,v in product(range(1,d+1),V):
F.add_variable(M(v,i))
# No two (active) vertices map to the same index
if alternative:
for u,v in combinations(V,2):
for i in range(1,d+1):
F.add_clause( [ (False,D(u)),(False,D(v)), (False,M(u,i)), (False,M(v,i)) ])
else:
for i in range(1,d+1):
for c in CNF.less_or_equal_constraint([M(v,i) for v in V],1):
F.add_clause(c)
# (Active) Vertices in the sequence are not repeated
if alternative:
for v in V:
for i,j in combinations(range(1,d+1),2):
F.add_clause([(False,D(v)),(False,M(v,i)),(False,M(v,j))])
else:
for i,j in combinations_with_replacement(range(1,d+1),2):
i,j = min(i,j),max(i,j)
for u,v in combinations(V,2):
u,v = max(u,v),min(u,v)
F.add_clause([(False,M(u,i)),(False,M(v,j))])
# D(v) = M(v,1) or M(v,2) or ... or M(v,d)
if not alternative:
for i,v in product(range(1,d+1),V):
F.add_clause([(False,M(v,i)),(True,D(v))])
for v in V:
F.add_clause([(False,D(v))] + [(True,M(v,i)) for i in range(1,d+1)])
# Every neighborhood must have a true D variable
neighborhoods = sorted( set(N(v) for v in V) )
for N in neighborhoods:
F.add_clause([ (True,D(v)) for v in N])
return F
def GraphColoringFormula(G,colors,functional=True):
"""Generates the clauses for colorability formula
The formula encodes the fact that the graph :math:`G` has a coloring
with color set ``colors``. This means that it is possible to
assign one among the elements in ``colors``to that each vertex of
the graph such that no two adjacent vertices get the same color.
Parameters
----------
G : networkx.Graph
a simple undirected graph
colors : list or positive int
a list of colors or a number of colors
Returns
-------
CNF
the CNF encoding of the coloring problem on graph ``G``
"""
col=CNF()
col.mode_strict()
if isinstance(colors,int) and colors>=0:
colors = range(1,colors+1)
if not isinstance(list, collections.Iterable):
ValueError("Parameter \"colors\" is expected to be a iterable")
# Describe the formula
name="graph colorability"
if hasattr(G,'name'):
col.header=name+" of graph:\n"+G.name+".\n\n"+col.header
else:
col.header=name+".\n\n"+col.header
# Fix the vertex order
V=enumerate_vertices(G)
# Create the variables
for vertex in V:
for color in colors:
col.add_variable('x_{{{0},{1}}}'.format(vertex,color))
# Each vertex has a color
for vertex in V:
clause = []
for color in colors:
clause += [(True,'x_{{{0},{1}}}'.format(vertex,color))]
col.add_clause(clause)
# unique color per vertex
if functional:
for (c1,c2) in combinations(colors,2):
col.add_clause([
(False,'x_{{{0},{1}}}'.format(vertex,c1)),
(False,'x_{{{0},{1}}}'.format(vertex,c2))])
# This is a legal coloring
for (v1,v2) in enumerate_edges(G):
for c in colors:
col.add_clause([
(False,'x_{{{0},{1}}}'.format(v1,c)),
(False,'x_{{{0},{1}}}'.format(v2,c))])
return col
def StoneFormula(D,nstones):
"""Stone formulas
The stone formulas have been introduced in [2]_ and generalized in
[1]_. They are one of the classic examples that separate regular
resolutions from general resolution [1]_.
A \"Stones formula\" from a directed acyclic graph :math:`D`
claims that each vertex of the graph is associated with one on
:math:`s` stones (not necessarily in an injective way).
In particular for each vertex :math:`v` in :math:`V(D)` and each
stone :math:`j` we have a variable :math:`P_{v,j}` that claims
that stone :math:`j` is associated to vertex :math:`v`.
Each stone can be either red or blue, and not both.
The propositional variable :math:`R_j` if true when the stone
:math:`j` is red and false otherwise.
The clauses of the formula encode the following constraints.
If a stone is on a source vertex (i.e. a vertex with no incoming
edges), then it must be red. If all stones on the predecessors of
a vertex are red, then the stone of the vertex itself must be red.
The formula furthermore enforces that the stones on the sinks
(i.e. vertices with no outgoing edges) are blue.
.. note:: The exact formula structure depends by the graph and on
its topological order, which is determined by the
``enumerate_vertices(D)``.
Parameters
----------
D : a directed acyclic graph
it should be a directed acyclic graph.
nstones : int
the number of stones.
Raises
------
ValueError
if :math:`D` is not a directed acyclic graph
ValueError
if the number of stones is negative
References
----------
.. [1] M. Alekhnovich, J. Johannsen, T. Pitassi and A. Urquhart
An Exponential Separation between Regular and General Resolution.
Theory of Computing (2007)
.. [2] R. Raz and P. McKenzie
Separation of the monotone NC hierarchy.
Combinatorica (1999)
"""
if not is_dag(D):
raise ValueError("Stone formulas are defined only for directed acyclic graphs.")
if nstones<0:
raise ValueError("There must be at least one stone.")
cnf = CNF()
if hasattr(D, 'name'):
cnf.header = "Stone formula of: " + D.name + "\nwith " + str(nstones) + " stones\n" + cnf.header
else:
cnf.header = "Stone formula with " + str(nstones) + " stones\n" + cnf.header
# Add variables in the appropriate order
vertices=enumerate_vertices(D)
position=dict((v,i) for (i,v) in enumerate(vertices))
stones=range(1,nstones+1)
# Caching variable names
color_vn = {}
stone_vn = {}
# Stones->Vertices variables
for v in vertices:
for j in stones:
stone_vn[(v,j)] = "P_{{{0},{1}}}".format(v,j)
cnf.add_variable(stone_vn[(v,j)],
description="Stone ${1}$ on vertex ${0}$".format(v,j))
# Color variables
for j in stones:
color_vn[j] = "R_{{{0}}}".format(j)
cnf.add_variable(color_vn[j],
description="Stone ${}$ is red".format(j))
# Each vertex has some stone
for v in vertices:
cnf.add_clause_unsafe([(True,stone_vn[(v,j)]) for j in stones])
# If predecessors have red stones, the sink must have a red stone
for v in vertices:
for j in stones:
pred=sorted(D.predecessors(v),key=lambda x:position[x])
for stones_tuple in product([s for s in stones if s!=j],repeat=len(pred)):
cnf.add_clause_unsafe([(False, stone_vn[(p,s)]) for (p,s) in zip(pred,stones_tuple)] +
[(False, stone_vn[(v,j)])] +
[(False, color_vn[s]) for s in _uniqify_list(stones_tuple)] +
[(True, color_vn[j])])
if D.out_degree(v)==0: #the sink
for j in stones:
cnf.add_clause_unsafe([ (False,stone_vn[(v,j)]), (False,color_vn[j])])
return cnf
def GraphIsomorphism(G1, G2):
"""Graph Isomorphism formula
The formula is the CNF encoding of the statement that two simple
graphs G1 and G2 are isomorphic.
Parameters
----------
G1 : networkx.Graph
an undirected graph object
G2 : networkx.Graph
an undirected graph object
Returns
-------
A CNF formula which is satiafiable if and only if graphs G1 and G2
are isomorphic.
"""
F = CNF()
F.header = "Graph Isomorphism problem between graphs " +\
G1.name + " and " + G2.name + "\n" + F.header
U=enumerate_vertices(G1)
V=enumerate_vertices(G2)
var = _graph_isomorphism_var
for (u, v) in product(U,V):
F.add_variable(var(u, v))
# Defined on both side
for u in U:
F.add_clause([(True, var(u, v)) for v in V], strict=True)
for v in V:
F.add_clause([(True, var(u, v)) for u in U], strict=True)
# Injective on both sides
for u in U:
for v1, v2 in combinations(V, 2):
F.add_clause([(False, var(u, v1)),
(False, var(u, v2))], strict=True)
for v in V:
for u1, u2 in combinations(U, 2):
F.add_clause([(False, var(u1, v)),
(False, var(u2, v))], strict=True)
# Edge consistency
for u1, u2 in combinations(U, 2):
for v1, v2 in combinations(V, 2):
if G1.has_edge(u1, u2) != G2.has_edge(v1, v2):
F.add_clause([(False, var(u1, v1)),
(False, var(u2, v2))], strict=True)
F.add_clause([(False, var(u1, v2)),
(False, var(u2, v1))], strict=True)
return F
def SparseStoneFormula(D,B):
"""Sparse Stone formulas
This is a variant of the :py:func:`StoneFormula`. See that for
a description of the formula. This variant is such that each
vertex has only a small selection of which stone can go to that
vertex. In particular which stones are allowed on each vertex is
specified by a bipartite graph :math:`B` on which the left
vertices represent the vertices of DAG :math:`D` and the right
vertices are the stones.
If a vertex of :math:`D` correspond to the left vertex :math:`v`
in :math:`B`, then its neighbors describe which stones are allowed
for it.
The vertices in :math:`D` do not need to have the same name as the
one on the left side of :math:`B`. It is only important that the
number of vertices in :math:`D` is the same as the vertices in the
left side of :math:`B`.
In that case the element at position :math:`i` in the ordered
sequence ``enumerate_vertices(D)`` corresponds to the element of
rank :math:`i` in the sequence of left side vertices of
:math:`B` according to the output of ``Left, Right =
bipartite_sets(B)``.
Standard :py:func:`StoneFormula` is essentially equivalent to
a sparse stone formula where :math:`B` is the complete graph.
Parameters
----------
D : a directed acyclic graph
it should be a directed acyclic graph.
B : bipartite graph
Raises
------
ValueError
if :math:`D` is not a directed acyclic graph
ValueError
if :math:`B` is not a bipartite graph
ValueError
when size differs between :math:`D` and the left side of
:math:`B`
See Also
--------
StoneFormula
"""
if not is_dag(D):
raise ValueError("Stone formulas are defined only for directed acyclic graphs.")
if not has_bipartition(B):
raise ValueError("Vertices to stones mapping must be specified with a bipartite graph")
Left, Right = bipartite_sets(B)
nstones = len(Right)
if len(Left) != D.order():
raise ValueError("Formula requires the bipartite left side to match #vertices of the DAG.")
cnf = CNF()
if hasattr(D, 'name'):
cnf.header = "Sparse Stone formula of: " + D.name + "\nwith " + str(nstones) + " stones\n" + cnf.header
else:
cnf.header = "Sparse Stone formula with " + str(nstones) + " stones\n" + cnf.header
# add variables in the appropriate order
vertices=enumerate_vertices(D)
stones=range(1,nstones+1)
mapping = cnf.unary_mapping(vertices,stones,sparsity_pattern=B)
stone_formula_helper(cnf,D,mapping)
return cnf