Expressive Power of Digraph Solvability

Micha l Walicki
1,
, Clemens Grabmayer
2
, and Marc Bezem
1
1
University of Bergen, Department of Informatics,
P.O. Box 7803, 5020 Bergen, Norway
(

on leave at UTL, IST, Lisbon)

michal,[email protected]
2
Universiteit Utrecht, Department of Philosophy,
Heidelberglaan 6, 3584 CS Utrecht, The Netherlands
[email protected]
Abstract. A kernel of a directed graph is a set of vertices without edges
between them, such that every other vertex has a directed edge to a ver-
tex in the kernel. A digraph possessing a kernel is called solvable. We
show that solvability of digraphs is equivalent to satisability of theo-
ries of propositional logic. Based on a new normal form for such theories,
this equivalence relates nitely branching digraphs to propositional logic,
and arbitrary digraphs to innitary propositional logic. Furthermore, we
show that the existence of a kernel for a digraph is equivalent to the
existence of a kernel for its lifting to an innitely-branching dag. While
the computational complexity of solvability diers between nite dags
(trivial, since always solvable) and nite digraphs (NP-complete), this
dierence disappears in the innite case: we prove that solvability for re-
cursive dags and digraphs is
1
1
-complete. This implies that satisability
for recursive theories of innitary propositional logic is also
1
1
-complete.
Finally, using solvability of dags we formulate a new equivalent of the
axiom of choice.
1 Introduction
In a digraph (directed graph), a subset X of its vertices is called independent
if the predecessors of vertices in X are not in X. A kernel of a digraph is an
independent subset K of vertices such that there is an edge from every vertex
outside of K to a vertex in K. Equivalently, a kernel of a digraph is a subset
K of vertices such that a vertex is in K if and only if none of its successors is
in K. This concept corresponds to, and originates from, the concept of solution
of binary relations as introduced by von Neumann and Morgenstern in their
famous book Theory of Games and Economic Behavior [14]. In order to explain
the connection with game theory, consider a two-player game with alternating
moves, let the vertices of a digraph G represent the positions and the edges
the possible moves. Then kernels of G describe a stable situation for one of the
players: a player A in a position outside a kernel K of G can always choose to
move to a position in K, forcing the opponent B to move out of K, and so on.
Thus A can stay outside K for the rest of the game, whereas B is forced to
2 Micha l Walicki, Clemens Grabmayer, and Marc Bezem
stay inside K. Depending on the other rules of the game, this can be a winning
strategy for A.
Today, kernel theory is an active research eld in graph theory that draws
inspirations from, and nds applications in, game theory as well as in logic. Its
main question concerns sucient conditions for the existence of kernels in nite
digraphs, e.g., [1, 79], with a recent overview in [2].
In this paper we aim at more general formulations, asking about the set-
theoretical status of the solvability of the main classes of digraphs. We recall,
from [4, 6], the equivalence between the existence of kernels of digraphs and the
satisability of corresponding theories of propositional logic, and give a new,
simple proof. This correspondence relates innitely branching digraphs to inni-
tary propositional logic, and nitely branching ones to the usual propositional
logic.
Overview. In Section 2 we start with the denitions of kernels and solutions of
digraphs, and introduce correspondence functions between digraphs and propo-
sitional theories that map satisable theories to solvable digraphs and vice versa.
In Section 3 we establish some independent results, useful for a general study of
the digraph solvability, which are then used in Section 4, presenting the main
results of the paper:
1
1
-completeness of the solvability of recursive dags, as well
as of satisability of recursive theories in innitary propositional logic, and a new
equivalent of the axiom of choice, in terms of the solvability of certain digraphs.
2 Basic denitions and facts
A directed graph (a digraph) is a pair G = G, E, where G is a set of vertices
and E G G is a binary relation representing the directed edges of G.
3
A
directed acyclic graph (a dag) is a digraph without cycles.
For a vertex x G, we denote by E(x) := y G [ E(x, y) the set of
successors of x, and by E

(x) := y G [ E(y, x) the set of predecessors of x

with respect to the edge relation of G. This notation is extended to subsets of
vertices, for example, for all X G, we let E

(X) :=

xX
E

(x). A sink in
G is a vertex x G without successors. Then sinks(G) = x G [ E(x) =
denotes the set of sinks of G.
A kernel of a digraph G = G, E is a subset of vertices K G such that:
(i) G K E

(K) (K is an independent set in G), and

(ii) G K E

(K) (from every non-kernel vertex there is at least one edge to

a kernel vertex).
The equivalence between the existence of kernels and the satisability of propo-
sitional theories that we explore in this paper arises from an equivalent denition
3
Directed graphs as dened here correspond to simple ones according to a more
general notion of directed graph in which some vertices u and v may be connected
by more than one edge with source u and target v; simple directed graphs do not
contain such multiple edges.
Expressive Power of Digraph Solvability 3
of kernels, the notion of solution. Let G = G, E be a digraph. An assignment
0, 1
G
(of truth-values to the vertices of G) is a solution of G if for every
x G: (x) = 1 (E(x)) 0 or, equivalently, if for every x G :
_
(x) = 1 (E(x)) 0
_

_
(x) = 0 1 (E(x))
_
(2.1)
The set of solutions of G is denoted by sol(G). G is called solvable i sol(G) ,= .
By
1
we denote the set x G [ (x) = 1.
The simplest example of an unsolvable digraph is

.
. For all digraphs G
and all assignments 0, 1
G
it holds:
sol(G)
1
= G E

(
1
)
1
is a kernel of G. (2.2)
Now, a digraph G induces a (possibly innitary) propositional theory T (G) by
taking, for each x G, the formula x E

(x), where E

(x) =
_
yE(x)
y with
the convention that
_
= 1.
4
Letting mod(T) denote all models of a theory T,
it is easy to see that (2.1) entails:
sol(G) = mod(T (G)) . (2.3)
As a consequence of (2.2) and (2.3), determining kernels of digraphs can be
viewed as a special case of determining the models of theories in propositional
logic (and hence digraph solvability can be reduced to consistency of proposi-
tional theories). These theories are in ordinary, nitary propositional logic (PL),
if G is nitely branching, and in innitary propositional logic (PL

), otherwise.
PL

denotes propositional logic with formulas formed over an arbitrary set

of propositional variables by unary negation, and innite conjunction
5
. PL

de-
notes the restriction of PL

to propositions over a countable set of propositional

variables and with conjunction of arity .
Conversely, also consistency of propositional theories can be reduced to solv-
ability of corresponding digraphs. In fact, every PL

-theory T can be repre-

sented as a digraph ((T) such that mod(T) = sol(((T)). First, assume a theory
T given as a set of equivalences of the form
y

iIy
x
i
, (2.4)
where all y, x
i
are variables, and where every variable occurs at most once on
the left of . The digraph ((T) is obtained by taking variables as vertices
and, for every formula, introducing edges y, x
i
for all i I
y
. In addition, for
every variable z not occurring on the left of any , we add a new vertex z and
two edges z, z and z, z. This last addition ensures that each variable z of
4
Satisability of such a theory is equivalent to the existence of solutions for the corres-
ponding system of boolean equations. This motivates the use of the name solution.
This was also the name used for kernels in the early days of kernel theory, e.g., in
[14], p.588, or [12].
5
By de Morgan, innite disjunction can be used instead of innite conjunction.
4 Micha l Walicki, Clemens Grabmayer, and Marc Bezem
T which would become a sink of ((T), and hence could only be assigned 1 by
any solution of ((T), can actually be also assigned 0 (when the respective z is
assigned 1). Letting V (T) denote all variables of T, and sol(X)[
Y
the restriction
of assignments in sol(X) to the variables in Y , we have that
mod(T) = sol(((T))[
V (T)
. (2.5)
Now, an arbitrary theory T can be transformed to the above form. It can be
done in many ways, and we give only one example, assuming T to be given as
a set of clauses. A clause is a set of literals. A literal is an atom or a negated
atom. Logically, a clause is the disjunction of its literals, with the convention
that
_
= 0.
Given a clause C, we denote the subset of its positive (negative) literals by
C
+
(C

), and dene the set P (the set N) as the index set of the positive
(negative) literals in C. In other words, C
+
= x
p
[ p P and C

= x
n
[
n N. First, let a
C
be a new variable. The formula C

: a
C
a
C
C
is equivalent to a
C
C, and C and C

are equisatisable with mod(C

) =
mod(C) a
C
, 0. Substituting for C, we obtain a more explicit form of
C

: a
C
a
C

_
pP
x
p

_
nN
x
n
. In addition, for every variable in the
intial theory x V (T), we introduce a new variable x. For every such pair we
introduce the two formulae (i), and for every clause C the formula (ii):
(i) x x and x x.
(ii) a
C
a
C

_
pP
x
p

_
nN
x
n
The theory C

containing formulae (i) and (ii) is equisatisable to C, with

mod(C) = mod(C

)[
V (C)
. Dening T

=

CT
C

and letting ((T) = ((T

),
the equality (2.5) remains valid.
When no confusion can arise, we use the denotation of the clause C itself for
the node representing the new propositional variable a
C
.
Example 2.1. For T
1
= x and T
2
= x y, we obtain the digraphs:
C
1

x

x
((T
1
)
C
2

K
K
K
K
K
K
K
x

x
((T
2
)
y

y
The theory T
3
with one innitary clause C =
_
iN
x
i
and the literal x
i
, for all
i N, gives rise to the following digraph ((T
3
):
x
1

x
1
C
1

...
C

v
v
v
v
v
v
v
v
v
v
v
v

x
i

x
i
C
i

...
Expressive Power of Digraph Solvability 5
((T
3
) can be obtained from the nite (sub)graph G
1
induced by C, x
1
, x
1
, C
1

by replicating the subgraph induced by these vertices except C. The inconsis-

tency of T
3
is reected by the lack of solutions for ((T
3
) which, in turn, reduces
to the lack of solutions of the nite subgraph G
1
. This suggests a possibility of
reducing satisability of theories in PL

to solvability of nite graphs, instead

of to satisability of nite subtheories. Such an investigation, however, is not the
topic of the present paper.
The equation (2.3) for digraphs, and equation (2.5), for propositional theories
establish a back-and-forth correspondence between satisable propositional theo-
ries and solvable digraphs that enables the transfer of results from propositional
logic to digraphs and vice versa. For instance, that kernel existence of nite
graphs is NP-complete [4] follows now from the fact that ( denes a polynomial
reduction. (As an aside: also T denes a polynomial reduction.) Conversely, for
every theory T in our normal form, i.e., with all formulae in the form (2.4) and
such that T = T (((T)), we obtain the following equivalent of satisability in
terms of kernels. Each x V (T) occurs on the left of one x E

(x), and
V (E

(x)) denotes the variables in the formula E

(x). We have that

mod(T) ,=
(2.5)(2.2)
K V (T) :
K
_
xK
V (E

(x)) = y , K : V (E

(y)) K ,= .
(2.6)
The rst conjunct reects (i) and the second one (ii) of the denition of a kernel.
Every model of T is uniquely determined by such a subset K, giving the variables
to be assigned 1. Thus, various statements of sucient conditions for the exis-
tence of kernels, e.g., [1, 2, 79], can be now applied for determining satisability
of PL theories.
3 Some general facts about solvability
This section presents some results on solvability that are of independent, general
interest; some of them will play a role in the next section as well. We show in
Section 3.1 that every digraph has a sinkless subgraph with essentially the same
solution set. The proof also yields the well-known fact that every nite dag
has a unique solution, since the appropriate sinkless subgraph of a nite dag is
empty. In Section 3.2 we give a simple characterization of the solutions of weakly
complete digraphs, and in 3.3 we show that solutions for arbitrary digraphs can
be represented as solutions for appropriate, innitely branching dags.
3.1 Induced assignment
In this subsection we will use induction on the set of ordinals with cardinality at
most the cardinality of the graph in question. All quantications etc. are relative
to this set of ordinals and we use to denote such ordinals ( for limits). We
6 Micha l Walicki, Clemens Grabmayer, and Marc Bezem
C0 = G, for the given digraph G = G, E

= sinks(C)

= E

(
1

) C
C+1 = C \ (
1

) and C

<
C for limit
C is the subgraph induced by C for > 0
G

C and G

is the induced subgraph

v
=

, for v {0, 1}
The induced assignment is given by = {x, v | x
v
} .
Fig. 1: Denition by ordinal induction of an induced sinkless subgraph G

, for every
given digraph G.
will inductively remove vertices from the graph until a xed-point, a sinkless
subgraph with essentially the same solution set, is reached.
Assigning 1 to sinks(G) may force values at some other vertices. This was
observed already in [10] for irreexive graphs. Since irreexivity is unnecessary,
we spell out and justify the construction in full generality. It is based on repeat-
edly removing sinks and their predecessors. The induced (partial) assignment
is dened by ordinal induction in Figure 1.
Note that is well-dened since there is no overlap between the sets
v

, when
or v varies. For nitely branching digraphs iterations suce. In general, even
if any path to a sink is nite, one may need transnite ordinals to reach a xed-
point, but one never needs ordinals with cardinality larger than that of the graph.
In the following example the (empty) xed-point is reached in + iterations,
while the innitely branching graph is countable.
Example 3.1. In the digraph below, after iterations only vertices at level 1
obtain induced values. The digraph obtains the induced (unique) solution after
+ iterations. Note that G

is empty here.
2
.

<
<
<
<
<
<
<
<

S
S
S
S
S
S
S
S
S
S
S
S
S
S
S
S
S
S
S
S

X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X

. . . . . .
1
. . . . . . .
The following proposition allows one to consider sinkless digraphs only.
Proposition 3.2. For any G, with , C

and G

as dened in Figure 1:
1. G

= C

= C
+1
for some with cardinality at most [G[
2. sinks(G

) =
3. sol(G) = [ sol(G

), in particular, sol(G) ,= sol(G

) ,= .
Proof.
1. For nite graphs this is obvious, so let G be innite and assume by contradic-
tion that C

C
+1
is non-empty for all with cardinality at most [G[. Then
Expressive Power of Digraph Solvability 7
there would be an injection : [[ [G[ G, which is impossible.
2. This follows directly from the previous point, since C

= C
+1
implies that
there are no sinks in C

= G

.
3. Let sol(G). By induction we show that for all :
1

1
and
0

0
.
This is obvious for
1
0
= sinks(G) and, consequently, also for
0
0
= E

(sinks(G)).
Inductively, if x
1

= sinks(C

), then E(x)

<

(since y E(x)
1

would imply x
0

and hence x ,
1

). By the induction hypothesis we get

E(x)
0
, and so (x) = 1. If x
0

then x E

(
1

) E

(
1
), so (x) = 0.
This proves that any sol(G) extends .
Now let x G

and y E(x). If y , G

, then y
0
, since y
1
would
imply x , G

. In other words, all successors of x outside G

have (x) = 0,
which means that restricted to G

is a solution of G

. By similar arguments,
any solution of G

can be extended to a solution of G by joining . 2

When G

= , sol() = , = and, by point 3, G has only one solution .

This is the case, for instance, for nite dags, which appears to be the rst theorem
in kernel theory from [14]. The above formulation implies also the existence of
a unique solution, for instance, for arbitrary well-founded dags (i.e., having no
innite path of connected edges x
0
, x
1
, x
1
, x
2
, x
2
, x
3
, ....).
3.2 Weakly complete digraphs
An undirected graph (or, shorter, a graph) is a pair G = G, E, where G is a set
of vertices and E is a set of undirected edges between the vertices in G, that is,
a set E u, v : u, v G consisting of unordered pairs of vertices in G. A
graph G = G, E is called complete if E u, v : u, v G, u ,= v, that is, if
every pair of distinct vertices of G is connected by an edge. Note that a complete
undirected graph may, but need not, contain self-loops, that is, undirected edges
that connect vertices with themselves.
Properties of undirected graphs can be used to dene weak counterparts
for directed graphs by using the forgetful projection from digraphs to graphs
in which the direction of the edges is removed. Given a digraph G, let G denote
the undirected graph obtained by turning every directed edge x, y into an
undirected one x, y. For every property P of undirected graphs, we say that
G is weakly P if G is P. For example, G is weakly connected, complete, etc., if G
is connected, complete, etc.
Note that distinct vertices u and v in a weakly complete digraph are con-
nected by at least one of the directed edges u, v and v, u. In addition to weak
completeness, there is also the following stronger notion of completeness: a di-
graph G = G, E is called complete if E = u, v : u, v G, u ,= v, that is,
if the edges of G connect each pair of distinct vertices of G in both directions.
Note that a complete digraph does not contain any self-loop, that is, a directed
edge that connects a vertex with itself.
For a kernel sol(G),
1
must be independent and dominating in G. These
two properties are equivalent to
1
being a maximal independent subset of G.
Conversely, given such a maximal independent subset K G, we determine if
8 Micha l Walicki, Clemens Grabmayer, and Marc Bezem
Fig. 2: Solutions of complete digraphs assign 1 to precisely one of the vertices.
it is a kernel of G by verifying that every vertex x G K has a directed edge
into K (an edge in G might be only to x.)
Consequently, a trivial (and inecient) algorithm for nding a kernel is to
unorient the digraph G, nd Gs maximal independent subsets, and for each such
check if every vertex outside it has a directed edge to the subset. A simple fact
follows from this observation.
Proposition 3.3. For a weakly complete digraph G, the following hold:
1. for any sol(G) : [
1
[ = 1.
2. sol(G) ,= x G : E

(x) = G x
(each such x gives a distinct solution).
Proof. Part 1 follows directly from the preceding observations:
1
must be a
maximal independent subset of a complete G. As to part 2, the implication to the
left holds for any digraph since, if such an x exists, it satises x = G E

(x),
i.e., is a kernel. Conversely, if sol(G) then, by 1, [
1
[ = 1. So assume a solu-
tion with (x) = 1 for some x. Then x , E

(x). Also, if y ,= x, then (y) = 0

and so y E

(x). 2
Example 3.4. The digraph C
3
, a cycle with 3 vertices, which is weakly complete,
is unsolvable due to item 2 of the proposition. But, according to the same item,
adding a single reverse edge makes it solvable.
Every complete digraph G is solvable: every solution of G picks a vertex u
such that (u) = 1 and (v) = 0 for all v ,= u, as illustrated by the examples in
Figure 2. Hence complete digraphs have precisely as many solutions as vertices.
The innite Yablo dag [15], the digraph N, <, is unsolvable: it is weakly
complete, but it does not contain a vertex as required in item 2 of the proposition.
3.3 Lifting digraphs to dags
Every digraph G (with at least one edge) can be transformed into an innitely
branching dag G

preserving and reecting the solutions as follows.

The (dag-)lifting of a digraph G = G, E is the digraph G

= G

, E

with:
G

:= G
E

:= n
i
, m
j
[ n, m E i < j
(3.7)
Expressive Power of Digraph Solvability 9
where, here and below, the vertices of G

, pairs in G, are denoted by index-

ing the vertex in the rst component, that is, a pair n, i is written as n
i
. The
graph G

is indeed a dag: it contains no cycles, since there can be a path of posi-

tive length from y
i
to y
j
only in case that i < j. Also, sinks(G

) = sinks(G)
and G is not well-founded (has a cycle or an -path) i G

has an -path.
For every function f : G X, we dene its lifting f

: G

X by:
f

(n
i
) := f(n) (for all n G and i ) . (3.8)
For a set (of functions) F we denote F

= f

[ f F.
Lemma 3.6. For every G : (sol(G))

sol(G

).
Proof. By denition, for every vertex x G and for all i :
E

(x
i
) =

mE(x),j>i
m
j
(so x sinks(G) : E

(x
i
) = 1).
Let sol(G), then (x) = (E

(x)). By (3.8) we have

(x
i
) = (x) =
(E

(x)) =

(E

(x
i
)) for all x, i. It follows that

sol(G

). 2
We say that a sol(G

) is stable on a vertex n G if i, j : (n
i
) = (n
j
)
and call stable if is stable on every vertex of G.
Lemma 3.7. For every G, every sol(G

) is stable.
Proof. G

has the property that n G i, j > i : E(n

j
) E(n
i
). Now,
if (n
i
) = 1, that is, (E(n
i
)) 0, then also (n
k
) = 1 for all k i. If
(n
i
) = 0, there is an m
j
E(n
i
) : (m
j
) = 1 and, by the previous case,
(m
j
) = 1 for all j

j. Hence (n
k
) = 0 for all k i. 2
The immediate corollary of the two lemmata is the following:
Theorem 3.8. For every G : (sol(G))

= sol(G

).
In particular, G is solvable if and only if G

is. A special case of the above

gives, for a nite cyclic G, its innite, acyclic counterpart. The paradigmatic
example is lifting a single loop to the innite Yablo dag. This special case was
addressed in [5] and we merely generalized its results allowing innite graphs with
sinks. When digraphs are innitely branching, the theorem allows us to equate
the problem of solvability of arbitrary digraphs and the problem of solvability
of dags, as will be done in the following section.
4 Solvability of digraphs and dags
This section contains the main results of the paper concerning the complex-
ity of solvability of digraphs and dags. In Section 4.1, we briey comment on
the universal solvability of nite, well-founded and nitely branching dags and
NP-completeness of solvability of nite digraphs. In Section 4.2, we show that
solvability of recursive digraphs and dags is
1
1
-complete and that, as a conse-
quence, this also holds for satisability of recursive PL

-theories. In Section 4.3,

we give a new equivalent of the axiom of choice in terms of solvability of innitely
branching digraphs and dags.
10 Micha l Walicki, Clemens Grabmayer, and Marc Bezem
4.1 The nite case
As noted after Proposition 3.2, every nite dag has a unique kernel and such
dags are obviously less expressive than arbitrary nite digraphs. E.g., a kernel
for a dag can be constructed in linear time while the problem of kernel existence
for general (nite) digraphs is NP-complete, [4]. The linear reductions T and (
from Section 2 give a new proof of this fact, which has been noticed earlier in [6].
Two kinds of innite dags, always possessing solutions, can be seen under
the current heading of niteness. (1) A well-founded dag may possess innite
branchings but has only nite paths and is solvable by Proposition 3.2.3. (2) A
nitely branching dag, on the other hand, may possess innite paths but no
innitely branching vertices. It, too, is always solvable, as we now explain.
Any nitely branching digraph G represents a propositional theory T = T (G)
and, by (2.3), its solvability is equivalent to the consistency of T. G is innite i
T is, and we consider this case. Every nite subtheory of T corresponds to an
appropriate subgraph of G (with additional 2-cycles added at new sinks s, arising
by removing parts of the graph; these 2-cycles can be viewed as denitional
extensions of T by s

s). By compactness, G is solvable i all such nite

subgraphs of G are solvable. Since every subgraph of a dag is a dag and all nite
dags are solvable, there is no need to add 2-cycles in the case of nitely branching
dags (we get models of nite subsets of T strengthened by some axioms s for
new sinks s). Compactness thus implies consistency of T and hence solvability
of G. This is the gist of the argument for solvability of nitely branching dags,
which can be found in [11].
4.2 Recursive dags
A consistent, recursive, propositional theory may have no recursive models or, in
terms of digraphs, a solvable, recursive digraph may have no recursive solutions.
Since lifting (3.7) of a recursive digraph yields a recursive dag, there are recursive
dags with no recursive solutions. The following gives a direct proof of this fact
even for nitely branching dags, using (a variation of) the Kleene-tree.
Theorem 4.1. There exists a recursive binary tree T without recursive solutions.
Proof. The argument is based on the existence of two recursively enumerable
but inseparable sets A and B. This means that A B = and for no recursive
set C we have A C and B C. Let recursive functions a and b enumerate
these sets A and B, respectively. We dene, uniformly recursive in n, linear trees
T
n
consisting of all sequences 0
0
, 0
1
, 0
2
, ..., 0
k
where 0
0
= and k is such that:
(1) a(i) ,= n b(i) ,= n for all i < k/2,
(2) k = 2i if i is minimal such that a(i) = n, and
(3) k = 2i + 1 if i is minimal such that b(i) = n.
This means that T
n
consists of all sequences 0
k
if n , A B. Otherwise, T
n
is a nite path with an even number of edges if n A and an odd number
Expressive Power of Digraph Solvability 11
if n B. The recursive tree T consists now of all prexes of sequences 0
n
10
k
for all n N and 0
k
T
n
. If there exists a recursive sol(T), then the set
C = n N [ (0
n
1) = 1 is recursive and separates A and B. Contradiction. 2
The theorem above concerns nitely branching trees, and a fortiori, nitely
branching dags. As we will see in Theorem 4.3 below, the complexity of solv-
ability of arbitrary, possibly innitely branching dags, leaves the arithmetical
hierarchy and is
1
1
-complete. So let us observe that going over now to digraphs
with innite branching, the dierence in expressive power between dags and
general digraphs disappears. On the one hand, dags with innite branching may
be unsolvable (e.g., the Yablo dag), while on the other hand, for an arbitrary
digraph G it holds that the solutions of G correspond bijectively to the solutions
of the dag-lifting of G, by Theorem 3.8. Therefore we will be able to show the
results in the sequel both for arbitrary dags, and for arbitrary digraphs.
For the statement of our complexity results, we now specify the digraph
solvability problem GSOL, and the
1
1
-complete problem to which we will link
it, the non-well-foundedness problem NWF for binary relations. In both cases
we employ encodings R of recursive binary relations R N N as natural
number codes of Turing machines that compute the characteristic function of R
(assuming standard constructions for these codes, we do not specify the details).
Digraph Solvability Problem GSOL
Instance: An encoding E of the edge relation E of a rec. digraph G = N, E
Question: Is G solvable?
Recognition problem: E : G = N, E is a recursive dag that is solvable
A binary relation R AA on a set A is called well-founded if there are no
innitely descending R-chains, that is, no innite sequences x
0
, x
1
, x
2
, . . . such
that for all n N : x
n+1
, x
n
R.
Non-Well-Foundedness Problem NWF
Instance: an encoding R of a recursive binary relation R on N
Question: Is R not well-founded?
Recognition problem: R : R is a rec., not well-founded, binary relation on N
Let A, B N. We write A
m
B to denote the fact that there is a many-one
reduction from A to B, that is, a total computable function f : N N such
that, for every n N, n A if and only if f(n) B.
Lemma 4.2. NWF
m
GSOL.
Proof. For every binary relation R N N we call G(R) = N, R

, where R

is the converse of R, the digraph associated with R; and we say, for x, y N,

that y is reachable from x via an R-chain if there are z
0
, z
1
, . . . , z
n
N such
that y = z
n
R z
n1
R . . . R z
1
R z
0
= x.
In the construction below of a computable many-one reduction from NWF
to GSOL we will assume that given recursive relations R NN are such that:
0, 0 / R (i, j R)
_
j (and hence also i)
is reachable from 0 via an R-chain

.
(4.9)
12 Micha l Walicki, Clemens Grabmayer, and Marc Bezem
0
G(R)
0
T(R)
Fig. 3: Left: example of a well-founded relation R N N such that (4.9) holds,
illustrated via the digraph G(R) associated with R. Right: the tree unravelling T(R) :=
T
G(R);0
of G(R) at vertex 0.
M(T(R)) M(T(R))
Fig. 4: Left: the modication M(T(R)) of the tree unravelling T(R) of G(R) as in
Fig. 3, which is obtained by splitting the edges in T(R), thereby creating intermediate
vertices. Right: its unique solution, which maps each of the intermediate vertices to 0.
M(T(R))
t
H(R)
t
M(T(R))
H(R)
Fig. 5: Left: the digraph H(R) constructed for R as chosen for G(R) in Fig. 3 by
extending M(T(R)) (see Fig. 4) with a new vertex t on top, an edge that is a self-loop
on t, and edges from t to all of the intermediate vertices in M(T(R)). Right: the unique
solution of its subgraph M(T(R)) cannot be extended to a solution of H(R).
Expressive Power of Digraph Solvability 13
0
G(R)
0
T(R)
Fig. 6: Left: example of a non-well-founded relation R N N such that (4.9) holds,
illustrated via the digraph G(R) associated with R. Right: the tree unravelling T(R) :=
T
G(R);0
of G(R) at vertex 0. Both digraphs contain an innite path.
M(T(R))
M(T(R))
Fig. 7: Left: the modication M(T(R)) of the tree unravelling T(R) of G(R) as in
Fig. 6 which is obtained by splitting the edges in T(R). Right: solutions are uniquely
determined on the well-founded part of M(T(R)), but not on the set of vertices that
are situated on, or are the starting points of, innite paths.
14 Micha l Walicki, Clemens Grabmayer, and Marc Bezem
M(T(R))
Fig. 8: There is a solution of M(T(R)) as in Fig. 7 that assigns 1 to all those interme-
diate vertices in M(T(R)) that are situated on innite paths.
t
H(R)
M(T(R))
H(R)
t
M(T(R))
Fig. 9: Left: the digraph H(R) constructed for R as chosen for G(R) in Fig. 6 by
extending M(T(R)) (see Fig. 7) with a new vertex t on top, an edge that is a self-
loop on t, and edges from t to all of the intermediate vertices in M(T(R)). Right: The
solution of its subgraph M(T(R)) as illustrated in Fig. 8 can be extended to a solution
of H(R) by assigning 0 to the vertex t on the top.
Expressive Power of Digraph Solvability 15
We are justied in doing so, because every binary recursive relation R on N can
be embedded into a binary recursive relation R

on N such that (4.9) holds for

R := R

, and such that well-foundedness is preserved and reected: given an

arbitrary recursive binary relation R N N, the binary relation:
R

:= i, 0 [ i > 0 i + 1, j + 1 [ i, j R N N
is recursive, satises (4.9) for R := R

, and furthermore, R

is well-founded if
and only if R is well-founded.
Let H = N, P be a digraph, and v N a vertex of H. By the tree unravelling
of H at v we mean the digraph T
H;v
=
H;v
, E with:

H;v
:= v
0
v
1
. . . v
n
N

[ v
i
N, v
0
= v, i 0, 1, . . . , n 1.v
i
, v
i+1
P
E := wv, wvv

H;v

H;v
[ w N

, v, v

N, v, v

P
That is, the set of vertices of T
H;v
is the set of paths in H from v, and the edge
relation of T
H;v
is obtained as a lifting from the edge relation of H. It is easy
to see that T
H;v
is a tree with root v, which has a recursive edge-relation if this
holds for H. Note that if a binary relation R N N has the property (4.9),
then every path in G(R) extends to a path that starts at the vertex 0; it follows
that R is well-founded if and only if T
G(R);0
does not contain an innite path.
Now we dene a computable many-one reduction from NWF to GSOL as
follows. For a given recursive binary relation R N N with (4.9), a recursive
digraph H(R) is constructed in a recursive way such that H(R) consists of:
a single vertex t with a loop on the top;
a modication M(T(R)) of the tree T(R) := T
G(R);0
, the tree unravelling
at 0 of the digraph G(R) associated with R; the modication is obtained
by splitting each edge of T(R) into two consecutive edges, inserting a new,
intermediate vertex between the edge-connected old vertices;
an edge from the top vertex t to every new, intermediate vertex in M(T(R)).
Figure 5 shows, on the left, the example of the digraph H(R) that is constructed
for a well-founded binary relation R on N such that (4.9) holds with associated
digraph G(R) as in Figure 3. The tree unravelling T(R) of G(R) is well-founded
(see Figure 3, on the right). Consequently, this also holds for its modication
M(T(R)), the tree that resulted by splitting the edges of T(R) (see Figure 4, on
the left). The unique solution of M(T(R)) assigns 1 to the old vertices symbolized
by black lled circles, and 0 to the new, intermediate vertices symbolized by
unlled circles (see Figure 4, on the right). The digraph H(R) itself is unsolvable:
All edges from t to vertices in its subgraph M(T(R)) target intermediate vertices,
and there always meet the value 0. Hence unsolvability of H(R) follows from
the unsolvability of its t-loop by Proposition 3.2.3.
Figure 9 shows the example of the digraph H(R) that is constructed for a
non-well-founded relation R such that (4.9) with associated digraph G(R) as
in Figure 6, on the left. In this case both of T(R) and M(T(R)) are non-well-
founded trees (see Figure 6 on the right, and Figure 7 on the left). In particular,
16 Micha l Walicki, Clemens Grabmayer, and Marc Bezem
M(T(R)) contains an innite path. As a consequence there exists a solution of
M(T(R)) that assigns 1 to all intermediate vertices on the innite path, and 0
to intermediate vertices with no innite paths below them (see Figure 8). This
solution of M(T(R)) induces 0 at the top vertex t with the loop, and thereby
extends to a solution for H(R) (see Figure 9). Hence H(R) is solvable.
Reasoning as above for the examples in Figures 35 and Figures 69 justies
the rst of the following four steps, which hold for every binary relation R on N
with (4.9):
H(R) is solvable M(T(R)) has an innite path
T(R) = T
G(R);0
has an innite path
G(R) has an innite path
R is not well-founded
(The last three steps are easy to verify.) Together with the facts that for every
recursive binary relation R on N a recursive relation R

with (4.9) such that

R

well-founded if and only if R is well-founded can be constructed, that the

digraphs T(R

), M(T(R

)), and H(R

) are recursive, and that encodings of the

edge-relation in each of these digraphs can be obtained from the code R of R by
a recursive procedure, it follows that the described transformation from recursive
relations R to recursive digraphs H(R

) constitutes a computable many-one

reduction from NWF to GSOL. .
Theorem 4.3. The solvability problem GSOL for digraphs is
1
1
-complete.
Proof. Since NWF is
1
1
-complete [13, 3], and NWF
m
GSOL by Lemma 4.2, it
remains to show that GSOL
1
1
. For this we describe a
1
1
-formula (m) that
expresses solvability for recursive digraphs, and more precisely, has the property:
(E) G is solvable (for all recursive digraphs G = N, E).
We let (m) be the
1
1
-formula:
K n
_
n K n

( EdgeIn(n, n

, m) n

/ K )

(n, n

vary over natural numbers, and K over sets of natural numbers), where
EdgeIn(n, n

, m) is a
0
0
-formula with the property that:
EdgeIn(n, n

, E) E(n, n

) (for all n, n

N)
holds for all recursive digraphs G = N, E (based on the chosen Turing maching
encoding, it is routine work to construct such a formula EdgeIn). .
The following corollary states
1
1
-completeness also for two problems that are
closely related to GSOL: the solvability problem DSOL for dags, and the sat-
isability problem for recursive PL

-theories, ISAT. Here we call a PL

-theory
recursive if it contains a countable set of propositions, the set of (the codes of)
Expressive Power of Digraph Solvability 17
its propositions is recursive, and if each proposition (typically an innite object)
has itself a recursive code.
6
Corollary 4.4. The following problems are
1
1
-complete:
GSOL: Is a given recursive digraph solvable?
DSOL: Is a given recursive dag solvable?
ISAT: Is a given, recursive, PL

theory satisable?
Proof. GSOL
m
DSOL is a consequence of Theorem 3.8: the dag-lifting G

of a
recursive digraph G is a recursive dag that is solvable if and only if G is solvable.
Since DSOL is a sub-problem of GSOL, this implies
1
1
-completeness of DSOL
in view of the theorem.
Together with ISAT
1
1
, which is not dicult to show for concrete encod-
ings
7
of recursive PL

-theories,
1
1
-completeness of ISAT follows from the fact
that GSOL
m
ISAT holds: this in turn can be proved by applying the reduction
used in Section 2 to show (2.3). .
Note that a direct proof for dags can be obtained showing NWF
m
DSOL
by replacing, in the proof of Lemma 4.2, the loop at the top of H(R) by a Yablo
dag, with edges from all its vertices to all the intermediate vertices in M(T(R)).
4.3 Arbitrary digraphs
We proceed now to arbitrary, innitely branching digraphs, and formulate a new
equivalent to the axiom of choice in terms of digraph solvability.
Let T = X
i
[ i I be an indexed family of sets. The cartesian product of
sets in T is dened by

iI
X
i
:=
_
f : I

iI
[ i I : f(i) X
i
_
and the
disjoint union by

iI
X
i
:=
_
x, i [ i I, x X
i
_
.
Let X be a set. A choice function on X is a function f : X

X such
that f(x) x for all x X. The axiom of choice (AC) is the statement: for
every set X, there exists a choice function on X.
Let ( = G
i
: i I be an indexed family of digraphs with G
i
= G
i
, E
i

for all i I. The disjoint union of the graphs in ( is dened by

iI
G
i
:=

iI
G
i
, E with E := v, i, v

, i [ i I, v G
i
, v

E
i
(v).
6
The complexity result for ISAT in Corollary 4.4 is invariant under reasonable precise
denitions of the concept recursive PL

-theory and under reasonable encodings.

One possibility is to limit considerations to PL

-theories in innitary clausal form,

in which innite disjunction is restricted to sets of literals. A recursive PL

-theory T
in innitary clause form can be encoded by a recursive ternary predicate C on N such
that the set of clauses Ci := {pj | C(i, j, 1)} {pj | C(i, j, 2)} for i N represents
the innitary clause form of T.
7
For every recursive PL

-theory T in innitary clausal form that is encoded through

a ternary predicate C
T
as described in footnote 6, satisability of T can be expressed,
using a
1
1
-formula, as: Mij(j M R(C
T
, i, j, 1) j / M R(C
T
, i, j, 2)),
where M varies over sets of natural numbers, and where R is a recursive predi-
cate such that, for all recursive ternary predicates P, and for all i, j, k, it holds:
R(P, i, j, k) P(i, j, k).
18 Micha l Walicki, Clemens Grabmayer, and Marc Bezem
Theorem 4.5. Over ZF, AC is equivalent with the following statement:
(GS) For every indexed family G
i
[ i I of solvable digraphs, the disjoint
union

iI
G
i
is solvable.
Proof. We use the well-known fact that, over ZF, AC is equivalent to:
(CP) For every indexed family X
i
: i I of non-empty sets, the cartesian
product

iI
X
i
is non-empty.
Hence it suces to show the two implications in (CP) (GS).
(CP) (GS). Let G
i
= G
i
, E
i
[ i I be an indexed family of solvable
digraphs. By (CP) it follows that the product

iI
sol(G
i
) is non-empty. Since
every f

iI
sol(G
i
) denes a solution
f
:

iI
G
i
0, 1 of

iI
G
i
by
letting
f
(v, i) := f(i)(v) for all v, i

iI
G
i
, it follows that

iI
G
i
is
solvable.
(GS) (CP). Let X
i
: i I be a collection of non-empty sets. For each i I,
let G
i
be the complete digraph with X
i
as its set of vertices. By Proposition 3.3
G
i
has [X
i
[ solutions, each picking one element of X
i
. For every solution of the
disjoint union

iI
G
i
it holds that the restriction [
Gi
of to G
i
is a solution
of G
i
. Consequently, every solution of the disjoint union of the G
i
s induces
a function f : I

iI
X
i
in

iI
X
i
by dening, for every i I, f(i) as the
unique x X
i
such that [
Gi
(x) = 1. 2
By employing dag-liftings and Theorem 3.8, this result can be lifted to dags.
Corollary 4.6. Over ZF, AC is equivalent with the following statements:
(DS) For every indexed family D
i
[ i I of solvable dags, the disjoint union

iI
D
i
is solvable.
Proof. Since (GS) (DS) is obvious, in view of Theorem 4.5 it suces to show
(DS) (GS). For this, let us assume (DS), and let G
i
[ i I be an indexed
family of solvable digraphs. By Theorem 3.8, the dag-lifting G

i
of G
i
is solvable
for every i I. Then it follows by (DS) that the dag D :=

iI
G

i
is solvable.
Since, as is easy to prove in ZF, D is isomorphic to G

for G :=

iI
G
i
, it follows
that G

is solvable, and hence, by Theorem 3.8 again, that G is solvable. .

See Figure 10 for a picture that suggests a direct proof of (DS)(AC): Every
solution of the dag-lifting G

of a disjoint union

iI
G
i
of complete digraphs
G
i
induces a choice function on the vertices of the disjoint union. In the picture
this is explained for the dag-lifting H

of the complete two-vertex digraph H.

5 Conclusion
Kernel theory is an active research eld of graph theory; a recent overview can
be found in [2]. Unlike most of the research in kernel theory, we have studied
graph kernels from the point of view of mathematical logic. We have elaborated
constructions for the following:
Expressive Power of Digraph Solvability 19
b
a
b
a
b
a
b
a
b
a
b
a
Fig. 10: Every solution of the dag-lifting H

of the complete two-vertex digraph H =

{a, b}, {a, b, b, a} induces a choice between a and b. (For the rst step, note that
a solution of H

cannot assign 0 to every vertex of H

.)
20 Micha l Walicki, Clemens Grabmayer, and Marc Bezem
1. For every digraph G a (possibly innitary) propositional theory T (G), the
model class of which corresponds to the set of kernels of G.
2. For every propositional theory T a digraph ((T) the set of kernels of which
corresponds to model class of T.
3. For every digraph an innite dag having essentially the same kernels.
4. For every binary relation R a digraph which has a kernel if and only if R is
not well-founded.
All constructions preserve recursiveness. These constructions yield, among other
results, the following insights, of which only the rst has been noticed before:
1. Propositional SAT and the existence of kernels of nitely branching digraphs
are equivalent problems. In the nite case, both are NP-complete.
2. The problem of the existence of a kernel of a recursive digraph is
1
1
-complete.
3. Since SAT of recursive theories in innitary logic is equivalent to the exis-
tence of kernels of recursive, innitely branching digraphs, this version of
SAT is
1
1
-complete, too.
4. The problem of the existence of a kernel is equally dicult for (recursive)
dags and for (recursive) digraphs.
5. The existence of kernels for disjoint unions of digraphs (or respectively, dis-
joint unions of dags) that have kernels is equivalent to the Axiom of Choice.
Acknowledgments. The proof of Theorem 4.1 is due to Dag Normann, who
deserves thanks also for discussion of several other points addressed in the paper.
We thank also Sjur Dyrkolbotn, Ulrich Kohlenbach, Albert Visser, and Vincent
van Oostrom for comments and discussion of various issues related to the pre-
sented results.
References
1. Martine Anciaux-Mendeleer and Pierre Hansen. On kernels in strongly connected
graphs. Networks, 7(3):263266, 1977.
2. Endre Boros and Vladimir Gurvich. Perfect graphs, kernels and cooperative games.
Discrete Mathematics, 306:23362354, 2006.
3. J. Castro and F. Cucker. Non-deterministic -computations and the analytical
hierarchy. Zentralblatt f. Math. Logik u. Grundl. d. Mathematik, 35:333342, 1989.
4. Vasek Chv atal. On the computational complexity of nding a kernel. Technical
Report CRM-300, Centre de Recherches Mathematiques, Univeriste de Montreal,
1973. http://users.encs.concordia.ca/
~
chvatal.
5. Roy Cook. Patterns of paradox. The Journal of Symbolic Logic, 69(3):767774,
2004.
6. Nadia Creignou. The class of problems that are linearly equivalent to Satisability
or a uniform method for proving NP-completeness. Theoretical Computer Science,
145:111145, 1995.
7. Pierre Duchet. Graphes noyau-paraits, II. Annals of Discrete Mathematics, 9:93
101, 1980.
Expressive Power of Digraph Solvability 21
8. Pierre Duchet and Henry Meyniel. Une jeneralization du theoreme de Richard-
son sur lexistence de noyaux dans les graphes orientes. Discrete Mathematics,
43(1):2127, 1983.
9. Hortensia Galeana-S anchez and Victor Neumann-Lara. On kernels and semikernels
of digraphs. Discrete Mathematics, 48(1):6776, 1984.
10. John R. Isbell. On a theorem of Richardson. Proceedings of the AMS, 8(5):928929,
1957.
11. Eric C. Milner and Robert E. Woodrow. On directed graphs with an independent
covering set. Graphs and Combinatorics, 5:363369, 1989.
12. Moses Richardson. Solutions of irreexive relations. The Annals of Mathematics,
Second Series, 58(3):573590, 1953.
13. H. Rogers. Theory of Recursive Functions and Eective Computability. MacGraw
Hill, 1967.
14. John von Neumann and Oskar Morgenstern. Theory of Games and Economic
Behavior. Princeton University Press, 1944 (1947).
15. Stephen Yablo. Paradox without self-reference. Analysis, 53(4):251252, 1993.