IEEE

238

TRANSACTIONS

matrix (over V) of $ related to a tree %h :I $, and let

o)* be the (only) path between terminals contained in So.
Let Cm be the cyclomatic matrix of Uh8,,. Then it results

c=

c,

(13)

THEORY

June

3) Let $&(&) be the subgraph (eventually nonconnetted) formed by the zero impedance branches of S(G).
It is clear that 6 = $. We shall prove that if 3 5 5 9
such that 2 C 5, then 3 $2 6 : G C 5. Note that such
a tree & can be constructed by the union of the following
subgraphs:
1) S
2) the trees 5h C $, : ai e 5 - 9

C@

with /.i = c

ON CIRCUIT

p,, + p.

3) the subgraphs $ - a:, where s, is a tree of

8 W. Mayeda, Synthesis of switching functions by linear graph

theory, IBM J. of Res. and Dev., vol. 4, pp. 321-329; July, 1960.

Minimum

Feedback Arc Sets for a Directed Graph*

D. H. YOUNGER?

Summary-A minimum feedback arc set is, for a directed graph, a

minimum set of arcs which if removed leaves the resultant graph
free of directed loops. This paper establishes a relationship between
these feedback arcs and order; in particular, such a minimum set of
arcs is shown to be determined by a sequential ordering of the
nodes which minimizes the number of arcs, each of which enters a
node that precedes in the ordering the node it leaves. From this
relationship are developed some simple characteristics of such sets,
as well as properties of the sequential orderings by which these
minimum sets are determined. These properties form the basis of
an algorithm for finding minimum feedback arc sets.

PROBLEM
in the topology of directed graphs
that has attracted some interest in recent years
is the following: to determine, for an arbitrary
directed graph, a minimum set of arcs which, if removed,
leaves the resultant graph free of directed loops. The
problem was originally
suggested by Runyon,
who
observed that the analysis of sequential switching circuits
with feedback paths would be simplified by the knowledge
of such a minimum set. Increased interest in this problem
is in large measure due to Moore, who has encouraged
attempts to find a solution. It was at first h.oped that a
simple and efficient algorithm might be found, perhaps
even an algorithm such that the number of operations

* Received November 26, 1962; revised manuscript

received
February 4, 1963. This work was supported by the National Science
Foundation
under Grant G-15078.
Publication
of this paper is
supported by the Marcellus Hartley Fund.
t Dept. of Electrical Engineering,
Columbia University,
New
York, N. Y.
1 Seshu and Reed [3] include this among a list of research problems given in the appendix.

required would increase linearly with the number of

nodes in the graph. However, the problem has turned
out to be difficult and suggested algorithms have fallen
far short of that goal. One of the more noteworthy approaches is that of Tucker [6] who has formulated the
problem as an Integer Program. More recently, some
necessary conditions on such a minimum set of arcs have
been found [2].
Two names that have been suggested for such a set of
arcs are minimal feedback cutset and minimal chord
set. Much can be said for each, and yet each may convey
an erroneous connotation to some readers. Since the
terminology has not yet become standard, we suggest the
term minimum feedback arc set.
Since directed graphs are finding wide use in many
varied disciplines, a method of finding such a fundamental quantity for a directed graph is certain to be useful.
For example, consider a mathematical system of n variables a,, ... , a,, upon which some set of order conditions
have been imposed, e.g., a, < a,, u3 < a,, a, < u3, a, < a,,
etc. Now assuming that these conditions are transitive,
it may be that some are inconsistent with the rest. Suppose
one wishes to determine a minimum set of conditions
that must be removed from the original set in order that
the subset remaining be consistent. The original set of
conditions may be represented as a directed graph of n
nodes by representing the variables as nodes and the
order conditions as arcs between pairs of nodes; any set of
conditions which corresponds to a minimum feedback
arc set for that graph should be removed.
In engineering, a good deal of success has been achieved
in analyzing complicated systems without feedback. This

1963

Younger: Minimum

has led to methods for the analysis of a system containing

feedback which reduce the system to one without feedback by breaking an appropriate set of feedback arcs [7].
The complexity of this analysis increases drastically with
the number of arcs which must be broken; hence a knowledge of a minimum feedback arc set would be useful.
As suggested above, a linear directed graph is a description of any set of order relations involving the variables represented as nodes of the graph. The object of this
paper is to establish the connection between minimum
feedback arc sets and possible sequential orderings of the
nodes. Consequently, some basic properties of such sets
as well as properties of the sequential orderings related to
them are developed.
DEFINITIONS

239

Feedback Arc Sets for a Directed Graph

AND NOTATIONS

A linear directed graph is shown in Fig. 1. The nodes

have been labeled a through e so that they may be referred
to conveniently. Given a labeling of the nodes, the arcs
(directed edges) may be expressed in terms of these labels;
an arc comiecting node i to j is denoted by the ordered
pair (i, j). If more than one arc connects i to j, these may be
distinguished by subscript: (i, j),, (i, j),, . . . , (i, j),.
Not only may the graphs to be considered have multiple
connections, they may have arcs connecting a node
to itself; the graph of Fig. 1 has such a self-loop at node
a as well as a multiple connection from node a to c.
Except for such basic terms as node, arc and connected
graph [l], [3] other concepts needed are defined below.
Dejkition 1: A (directed) loop is a connected subgraph
such that each node has one arc leaving and one arc
entering.
DeJnition W: A self-loop is a loop which consists of a
single node and an arc from and to itself.
De$nition 3: A (simple) path is a connected subgraph
including a pair of nodes called the initial and final nodes,
such that the subgraph, augmented by an arc from the
final node to the initial node if these nodes are distinct,
forms a loop.
The notation xPy is used to indicate that a simple path
exists from node x to y.
Dejkition 4: For a directed graph, a feedback arc set
is a set of arcs which, if removed, leaves the resultant graph
free of directed loops.
Desnition 6: A feedback arc set is minimum if no other
feedback arc set for that graph consists of a smaller
number of arcs.
A directed graph may be presented in any of three
forms: as a geometrical figure, as a set of ordered pairs,
or as a matrix; if as a matrix, it is assumed to be the
comlection matrix. In any matrix form, the nodes must
be listed in some order. Although the order generally
chosen is that which agrees with the alphabetic order of
the labels, this is certainly not necessary. The following
definitions make explicit the dependence of the connection
matrix upon order.
Dejkition 6: A sequential ordering of a graph of n nodes

c
Q

Fig. 1-A

directed graph.

is a l-l
function from the nodes of the graph to the
integers 1, * . . , n.
DeJinition 7: For a directed graph of n nodes with sequential ordering R, the connection matrix

c = b-ki~R~i~l.
has one row and one column for each node of the graph,
and the entry at row R(i), column R(j), denoted
= number of arcs from node i to node j.
CR(i)R(i)
An example of a sequential ordering of the nodes of the
graph of Fig. 1 is given by
R(a) = 1,

R(c) = 2,

R(e) = 3,

R(b) = 4,

R(d) = 5

(1, 2, 3, 4, 5) = R(a, c, e, b, d).

(1)

which may also be written

The corresponding connection matrix is
a
al

e b d

2 0

10

c00110
C=el

1.

b0

d-0

1 0

(2).

O-

The subgraph which consists of one of the arcs from a to c,

the arc from c to e, and that from e to a, form a loop which
is written as (a, c),(c, e) (e, a). When a sequential order is
given, a loop or path may be alternatively expressed in
terms of the order symbols; the loop just given becomes
(1, 2),(2, 3)(3, 1) under the ordering given by (1). The
self-loop at node a is denoted by (a, a). One simple path
from node a to d is given by (a, b)(b, d), another by
(a, c),(c, e)(e, d). Arcs (a, a), (e, a), (d, c) and (d, e) form
a feedback arc set (which may be shown to be minimum).
Often a graph contains some special characteristics that
simplify the determination of minimum feedback arc sets.
If a graph contains a source (or sink), i.e., a node for
which the arcs incident upon it are all directed away from
(toward) that node, none of these arcs can be contained
in any minimum feedback arc set. If a node has but one
arc entering and one leaving, any minimum feedback arc
set which contains one of these arcs must not contain the
other; however, another minimum set may always be

IEEE

240

ON CIRCUIT

TRANSACTIONS

obtained by replacing the one arc by the other. If a

graph is not connected, a minimum feedback arc set for
the over-all graph is obtained by combining the minimum
feedback arc sets for each connected component. These
statements are almost self-evident and may be easily
proved. The discussion to follow does not depend on such
special characteristics; hence, though the discussion is valid
for all graphs, it is most pertinent to graphs that are connected, have no sources or sinks, and have at least three
arcs touching each node.
THE RELATION

BETWEEN FEEDBACK ARC SETS AND

SEQUENTIAL OBDERINGS

Lemma 1: A graph G has no directed loops if and only

if the nodes can be given a sequential ordering R such
that for each arc (i, j) of the graph, R(i) < R(j).
Proof: Let xPy be the relation that there exists a simple
directed path from node x to y of G. If G is free of directed
loops, then xPy is a strict partial ordering of the nodes,
i.e., xPy is irreflexive, asymmetric and transitive [5].
xPy is irreflexive since xPx asserts the existence of a directed loop involving node x, a contradiction. If xPy and
yPx, then some loop must exist involving node x, a contradiction; hence xPy is asymmetric. If z:Py and yPx,
then either XPZ or the paths from x to y and from y to z
involve some common node other than y, i.t:., there exists
a node w such that wPy and yPw, a contradiction; hence
xPy is transitive.
A .partial ordering may always be extended to a total
ordering; thus the nodes can be sequentially ordered so
that for each pair of nodes x, y such that xPy, R(x) < R(y).
For each arc (i, j), iPj and hence R(i) < R(j).
Conversely, any sequential ordering R such that for
each arc (i, j), R(i) < R(j) holds, is a strict total ordering
containing the ordering given by xPy. If G contains a
loop, involving node x say, then xPx, contradicting
Q.E.D.
irreflexivity. Hence G is free of directed loops,
The condition that R(i) < R(j) for each arc (i, j) of a
graph is equivalent to the condition that the connection
matrix has only zero entries below and on the major
diagonal. That the nodes of a graph can be ordered so that
this condition for the connection matrix holds, is equivalent by Lemma 1 to the condition that the graph be free of
loops.
It follows from Lemma 1 that for any sequential ordering
R of the nodes of a graph G, the removal of those arcs
(i, j) for which R(i) >_ R(j) must eliminate all directed
loops. That is, such a set of arcs must constitute a feedback arc set. Let F, designate the set
{(i, j), R(i)

2 R(j);

i, j E G},

(3)

which may be expressed also as

IW), R(j)),R(i)2 R(j),i, je GJR.

2 See Sierpihski [4], p. 189.

(4)

THEORY

June

For the graph of Fig. 1 with

given by (l),
FR = {(a, a), (e, 4, (d,c),

the sequential

and

ordering

(d,e!)

= ((1, 11, (3, 11, (5,2)

and

(5, 3>}R

constitutes a feedback arc set. Note that the arcs of F,

correspond to those nonzero entries of the connection
matrix on and below the major diagonal.
It is also true that, for any feedback arc set F, some
sequential ordering R must exist such that F, is a subset
of F. To see this, consider removing all arcs of F from
graph G. The resultant graph contains no directed loops
and hence th! nodes may be ordered R so that R(i) < R(j)
for each arc (i, j). If the arcs of F are put back into the
graph, those arcs (i, j) of F which satisfy R(i) > R(j)
constitute a feedback arc set FE which is a subset of F.
Hence, in looking for minimum feedback arc sets, only
those sets given by 1(i, j), R(i) >_ R(j)} for some R,
need be considered. These remarks suffice to establish
the following theorem:
Theorem 1: A minimum feedback arc set for an arbitrary directed graph is given by a minimum set of
arcs ( (i, j), R(i) 2 R(j) } , where the number of elements
of this set has been minimized over all possible sequential
orderings of the nodes of the graph.
Rephrased in terms of connection matrix, the theorem
states: a minimum feedback arc set is given by a minimum
set of arcs corresponding to nonzero entries on and below
the major diagonal of the connection matrix, the minimum taken over all possible orderings of the rows and
corresponding columns.
Corollary to Theorem 1: Given that a set of arcs constitutes a minimum feedback arc set, a sequential ordering
can be found such that {(i, j), R(i) 2 R(j)] is that set.
If, as suggested by the above theorem, attention is
confined to feedback arc sets that correspond to sequential
orderings, a new interpretation may be given to the term
feedback arc. For a sequential ordering R, a feedback
arc is any arc (i, j) for which R(i) 2 R(j). The remaining
arcs are called forward arcs.
PROPERTIES OF MINIMUM

FEEDBACK

ARC SETS

Theorem W: Any subgraph of a directed graph consisting of the arcs of a min@um feedback arc set has no
directed loops except self-loops.
Proof: Let R be a sequential ordering such that
F, = {(i, j), R(i) > R(j)] is a minimum feedback arc
set. The arcs in FB for which R(i) = R(j) are self-loops;
the remaining arcs (i, j) have the property R(i) > R(j) and
hence, as in Lemma 1, form no directed loops.
It follows from this theorem that an arbitrary directed
graph without self-loops can always be partitioned into
two subgraphs such that each is itself free of directed
loops. This remark can be elaborated to prove that the
edges of any nondirected graph without connection from

1963

Younger: Minimum

241

Feedback Arc Sets for a Directed Graph

a node to itself can be oriented so that the resultant

directed graph is free of directed loops.
Theorem 3: The removal of the arcs of any minimum
feedback arc set of a connected n-node graph leaves a
subgraph of n nodes that is still connected.
Proof: Let F, be a minimum feedback arc set for an
n-node graph G. Suppose the removal of the arcs of F,
from G leaves a subgraph G which contains a part (or
node) isolated from the remaining portion of G. Since the
over-all graph G is taken to be connected, one or more of
its arcs that are in F, must connect the part isolated in G
to the remaining portion. The addition of any one of these
arcs to G certainly adds no loop to that subgraph, contradicting the minimality of F,. Hence, the assumption of
an isolated part or node in G is untenable.
The subgraph G obtained by removing the arcs of F,
can thus be described as a maximum connected subgraph
of graph G which contains all the nodes of G and is free
of directed loops. Omitting the word directed,
this
description is suitable for a tree of a nondirected graph.
The closeness of this analogy has prompted the use of the
term minimal chord set for the set of arcs to be deleted [Z].
IROPERTIES OF THOSE SEQUENTIAL ORDERINGS THAT
CORRESPOND TO MINIMUM FEEDBACK ARC SETS

Dejkition 8: An optimum ordering R is a sequential

ordering of the nodes of a directed graph for which
{ (i, j), R(i) >_ R(j) } is a minimum feedback arc set.
DeJnition 9: For a sequential ordering R of a directed
graph, a consecutive subgraph is any (nonempty) subgraph composed of nodes consecutively ordered by R
and the arcs connecting these in the over-all graph.
For the graph of Fig. 1, the feedback arc set F, corresponding to the sequential ordering given by (1) has
already been asserted to be minimum; hence, R is an
optimum ordering for that graph. Examples of consecutive subgraphs under R are the single node c (indeed,
any single node), the subgraph framed on nodes c, e, and
b (which consists of the edges (c, e) and (c, b)), and the
over-all graph.
Theorem ~$1The set of optimum orderings for a given
graph is invariant under the removal of self-loops and
loops involving two arcs.
In the case of multiple connections between nodes,
e.g., say these are p arcs from node i to j, and Q arcs from
node j to i, with p > Q, the theorem should be interpreted
as permitting the removal of up to 4 arcs from i to j
and the same number of arcs from j to i.
Proof: An arc corresponding to a self-loop is of the
form (i, i) and so will be included in the set FE = { (i, j),
R(i) > R(j)] for any R. A loop involving two arcs is one
of the form (i, j)(j, i), i # j. If R(i) > R(j), the arc (i, j)
is in F, but (j, i) is not; if R(i) < R(j) the arc (j, i) is in
F, but (i, j) is not. In either case one and only one of these

is a feedback arc for any R. Hence the choice of sequential

orderings which minimize F, cannot depend on loops
consisting of only one or two arcs.
In accordance with Theorem 4, two directed graphs
are said to be order equivalent if the removal of all selfloops and loops of two arcs from each graph results in
graphs with the same connection matrix. A graph without
such loops is termed reduced and is the representative of
all those to which it is order equivalent. The reduced
graph and hence the class it represents may be expressed
in a simpler form than the connection matrix, as follows:
Definition 10: For a reduced graph of n nodes with
sequential ordering R, the connection triangle
1 I R(i) < R(j) I n}

C = Niww,

has one row and one column for each node of the graph,
and GcijRtil = number of arcs from node i to j minus
number of arcs from node j to i.
For the (set of graphs order equivalent to the) graph
of Fig. 1 with the sequential ordering given by (I), the
connection triangle is

C=

a 2

-11

0
1

-1
(5)

d
From this array, the reduced graph may immediately
be obtained as a figure: in this case it differs from Fig. 1
in that it lacks the self-loop at node a and the loop
(e, 4 (4 e).
Theorem 5: Given an optimum ordering R of a directed
graph G, let Gl be any consecutive subgraph:
a) G, must have as minimum feedback arc set those
arcs of G, that are feedback arcs according to R;
b) similarly, the subgraph H obtained from G by
deleting all arcs and coalescing all nodes of G, must have as
minimum feedback arc set those arcs of H that are feedback arcs by R.3
Proof: Partition the feedback arc set F, into two
subsets:

Fi = ((i, j), R(i) 2 R(j); i, j EG},

(6)

Fi = {(i, j), R(i) 2 R(j); i or j# G,].

(7)

F$ is the set of feedback arcs by R that are in G,, and

F, is the set of feedback arcs that are in H. Suppose,
first, that Fi is not minimum: that there exists a set F
containing fewer members whose deletion also leaves G,
free of directed loops. Reorder G, by R so that
{(i, j), R(i) 2 R(j); i, j E G,)

(8)

3 Part (b) of this theorem WBSsllggested by one of the reviewers;

this part helps to simplify the proof of Theorem 6.

242

IEEE

TRANSACTIONS

ON CIRCUIT

is equal to F. This reordering of G, does not affect the

membership of F,2 since the nodes of G, are consecutively
numbered by R. It follows that F U F,2 is a feedback arc
set for G with fewer members than F,, a contradiction.
Hence F$ is minimum for G,. Assuming F, not minimum
for H leads to a similar contradiction.
As an illustration, consider once again the R given by
(l), asserted to be optimum for the graph of Fig. 1.
Using order symbols rather than node labels, the feedback arcs are
FR = ((1, I>, (3, 11, (5, 2) and (5, 3) lR.

(9)

If G, is the consecutive subgraph framed on nodes 2

through 5, Fi is {5, 2) and (5, 3) } R = { (d, c) and (d, e) }
and F, is ((1, 1) and (3, l))R = ((a, a) and (e, a)}. These
are, by the theorem, the minimum feedback arc sets for
G, and H, respectively.
It follows from Theorem 5 a) that for an optimum
ordering R for a graph G, the number of arcs from node i
to node j, where R(j) = R(i) + 1, is greater than or
equals the number of arcs from j to i. That is, in the connection triangle, ck(i),R(i)+l 2 0. In fact, a stronger
statement that includes this one is made by the theorem
which follows.
Notation: For subgraphs G, and G, of a dnected graph
G which. contain no common node, let ccXc, denote the
number of arcs each of which leaves a node in G, and
enters a node in G,.
Theorem 6: Given an optimum ordering R for a directed
graph G, let G, and G, be consecutive subgraphs of nl
and n, nodes such that the highest numbered node in G,
is one less than the lowest numbered node in G,; then:
a) CG,G~ 2 CC,C~,
b) if cG,,& = CG,G,, then the ordering R is also opti-

June

THEORY

5 b), the feedback arc set corresponding to cGSG,is minimum for H and hence, cc,,& 2 cGZGI,
the feedback arc set under R
b) For CG>G, = cG,G1,
contains the same number of members as the feedback
arc set under R, and since R is optimum, so is R, Q.E.D.
An examination of the graph of Fig. 1 reveals that the
ordering given by (1) does indeed satisfy this theorem.
It is simple enough to choose for this graph an ordering
which does not, but instead consider the graph of Fig. 3
with the sequential ordering
(11)

(1, 2, 3, 4, 5, 6, 7) = R(f, d, b, a, c, e, g).

This ordering cannot be optimum; for taking GI the single

node d (with R(d) = 2) and G, the single node b (with
R(b) = 3) gives
CO,G,(=CZ = 0) < CG,G~(=C.Q= 1);
see Fig. 4(a). Alternatively, consider G, the single node a
(with R(u) = 4) and G, the consecutive subgraph c, e, g
[with (5, 6, 7) = R(c, e, g)] for which
cG,G,(=c45 + c46+ c47 = 1)
<c

+ 64 + ci4 = 2);

G,G,(=h

see Fig. 4(b). The condition CG,G~2 cG,G, is not satisfied

for all G,, G,.
This condition can also be tested directly in terms of
the connection triangle. Consider the special case where
G, is a single node lc and G, any succeeding consecutive
subgraph. Theorem 6 for this case becomes the following:
the summation to the right of elements in row R(k) of
the connection triangle can at no stage fall below zero
for an optimum ordering R. The connection triangle for
the graph of Fig. 3 (and its order equivalents) is, for R
given by (II),

mum, where for each node i of G

R(i) = R(i),
R(i) = R(i) -

i eG,
n,,

R(i) = R(i) + n,,

i E G,

or

G,
(10)

i e G,.

A diagram illustrating the theorem is shown in Fig. 2.

The number of arcs from nodes of G, to nodes of G, is
= 4, whereas CG~G~= 3. The condition cG,G, >_ CG,G,
CG,G.
is necessary for R to be optimum; of course, this condition
must also hold for all other possible G,, Gz as well.4
Proof: a) Let G be the consecutive subgraph which
contains all nodes of G, and G,. The subgraph H obtained from G by first deleting the arcs and coalescing
the nodes of G, and then deleting the arcs and coalescing
the nodes of G, is a subgraph of two nodes whose forward
arcs are those contributing to cc,& and whose feedback
arcs correspond to CG,G,. By Theorem 5 a) followed by
4 An interesting result by Hakimi [2] states that for any cutset
of a directed graph, the number of arcs cut which are elements of a
minimum feedback arc set is not greater than the minimum of the
number of arcs cut which agree with the cutset orientation and the
number of arcs which oppose the cutset orientation.

70
C =

O-l

IU

-1

1
1

-1

03

9
Note that c:,( = - 1) < 0 and c:, + cl, + c:,( = - 1) < 0;
these are the illustrations already chosen but are more
easily recognized in this form.
When for a graph G with sequential ordering R there
exists consecutive subgraphs G,, G, such that cG,G,<&,G,,
then there are fewer feedback arcs under R given by (lo),
i.e., replacing R by R gives a better ordering. By iterating
this procedure a sequential ordering that does satisfy
CG,G,
> cG,G, for all G,, G, can be found. In the example
just considered, if the order of d and b are first reversed
and then the order of a shifted beyond that of c, e, g,

Younger: Minimum

1963

243

Feedback Arc Sets for a Directed Graph

the sequential ordering obtained is
(1, 2, 3, 4, 5, 6, 7) = R(f,

b, 4 c, e, g, 4

(13)

which does satisfy CG,G, 2 cG*G, for all G,, Gz. That an
ordering satisfies this condition is not sufficient for it to
be optimum, but any such ordering appears, intuitively at
least, to be a fairly good one.
DEFINITION

Fig. 2-Illustration

of a property

of an optimum

ordering.

Fig. 3-A

directed graph used to illustrate

Thcorcm

6.

(a)

(b)
Fig. 4-Application

of Theorem 6 to the graph of Fig. 3.

ANU A PROPERTY OF AIIMISSIBLE

ORDERINGS

The viewpoint in this section will be broadened to include feedback arc sets and sequential orderings which,
while possessing the properties discussed in the previous
two sections, may be less than optimum.
DeJkition 11: A feedback arc set for a directed graph
is minimal if it contains no proper subset that is also a
feedback arc set for this graph.
To test a feedback arc set F, of a directed graph G for
minimality, the arcs of F, are removed from G leaving a
subgraph G free of directed loops. The arcs of FE are
then added individually to G; if any of these arcs with
G form no directed loops then F, is not minimal; however, if all such arcs are deleted from F,, the resultant
feedback arc set is minimal. A sequential ordering of the
nodes which determines this resultant feedback arc set
can then be found. In case it is not clear whether an arc
with G forms any loop, a simple conclusive test for the
nonexistence of dir&ted loops in a graph may be used;
this test is implied by the lemma which follows, a restatement of a result of Hakimi [2].
Lemma 2: A graph is free of directed loops if, and
only if, successively removing source nodes and the arcs
leaving each, exhausts the graph.
In this statement, which follows from Lemma 1, the
order in which sources are removed when a choice exists
does not matter. An equally valid test for the lack of loops
in a graph is the successive removal of sink nodes.
DeJinition lb: An ordering R for a directed graph is
admissible if
a) the graph ordered by R satisfies cG,G. > CG,G, for
all appropriate consecutive subgraphs G,, Gz;
b) the feedback arc set determined by R is minimal.
The first criterion should generally be the more critical,
but does not imply the second. The test of each criterion,
if not satisfied, indicates a simple means to develop an
ordering which does meet the test. Hence, the securing of
an admissible ordering is really an easily accomplished
task and is the starting point from which to search for
optimum orderings. It is convenient to call this admissible
ordering first obtained the admissible reference ordering
and to relabel the graph a, b, c, . . . in accordance with
this reference ordering.
As a further bit of terminology, call an n-tuple of nodes
(i, i . . . , Ic) for which (1, 2, . . . , n) = R(i, j, . . . , k)
the n-tuple representation of an ordering R. It is often
convenient to order a set of n-tuples lexicographically,
with the left-most symbol of greatest significance. Lexicographic order depends directly on the labeling used; the

IEEE

244

TRANSACTIONS

most suitable labels in this work are those which correspond to the admissible reference ordering.
DeJinition 1s: Two sequential orderings of a graph are
identical with respect to feedbaclc, or F identical if they
determine the same feedback arc set. An F- identical class
of orderings is a set of sequential orderings identical with
respect to feedback. The ordering in an F-identical class
whose representation as an n-tuple is lexicographically
the smallest with respect to the admissible reference
labeling is the F representative.
Definition 1.4: Given a sequential ordering R, a sequent
derived from R is an ordered pair of nodes [i, j] for which
R(j) = R(i) + 1. If node i precedes node j according to
the reference labeling, then [i, j] is an up sequent; if i
succeeds j then [i, j] is a down sequent.
An F-identical class for the graph of Fig. 3 is given by
(f, 4 b, a, c, e, d
(f, d, a, b, c, e, d
I

(f, a, 4 b, c, e, d .

(f, 4 a, c, b, e, d
I (f, a, d, 6, b, e, cd,
The third of these tuples is the smallest lexicographically
and hence is the F representative. Using the labeling
given, this third tuple has down sequents [f, a] and [d, b]
and up sequents [a, d], [b, c], [c, e] and [e, g].
Theorem 7: To every down sequent [i, 91 derived from
an admissible F representative R,, there corresponds one
or more arcs (i, j) in the reduced graph G.
Note: These arcs are feedback arcs according to the
admissible reference ordering.
Proof: By definition of reduced graph, there cannot
be arcs both from i to j and from j to i in G; suppose that
G contains one or more arcs from j to i. Modifying R,
by reversing the order of nodes j and i alters the corresponding feedback arc set only by removing these arcs,
and thereby contradicts the admissibility
of R,. The
second alternative, that no arcs connect nodes i and j,
contradicts (since [i, j] is a down sequent) the designation
of R, as F representative. Hence one or more arcs leave
node i and enter node j.
AN

ALGORITHM

FOR FINDING

AX

MINIMUM

FEEDBACK

SETS

For a directed graph G with ordering R, let Q(R) represent the number of arcs in F,, the feedback arc set
determined by R. It has been shown that determining a
minimum feedback arc set for G is equivalent to finding
an R for which Q(R) is minimum. In this section, a more
general version of this basic problem is considered: that
of finding all minimum feedback arc sets, or equivalently,
all F representatives R for which Q(R) attains its absolute
minimum.
The search for optimum orderings begins with any
ordering which is admissible; with this as reference, an

-..- ClKCUlI
-----I-UN

-------LHlCUEY

J me

ordering R with smaller Q(R) is sought. If such a better

ordering is obtained it becomes the new reference; this
process is iterated until a reference is obtained which
cannot be improved; such an ordering is optimum. During
this process, those orderings R for which Q(R) = Q(R)
for reference R are noted since if the reference is established as optimum, they become with the reference the
set of optimum orderings.
In this search, the orderings that are examined are
modifications of the reference ordering. The scope of
modifications which may lead to an improvement are
restricted by Theorem 7 to those orderings whose down
sequents are feedback arcs under the reference ordering.
This restriction leaves, potentially at least, a wide region
of search. But any such modification can be reduced to a
series of perturbations in the ordering, each of which
involves the shift of a single node, other nodes remaining
fixed except to accommodate such a shift. These single
node perturbations are the only direct modifications of the
reference ordering that are considered. A further restriction is in fact made that the perturbation
must
establish an allowable down sequent. Under this single
node perturbation, some feedback arc under the reference
ordering becomes an arc which enters the node immediately
succeeding the node it leaves.
Since each perturbation of the reference ordering must
establish some feedback arc as a down sequent, there are
at most Q(R) such perturbations for the reference ordering
R. Not all Q(R) possible perturbations need always be
considered since some down sequents corresponding to
feedback arcs under the reference ordering can sometimes
be rejected as inferior to other sequents. In fact, if less
than a complete search is demanded, down sequents can
be rejected as unlikely if they cause a large net change of
forward arcs to feedback arcs.
If one of these initial perturbations does not establish
a new reference, then each of these is further perturbed
in the same way. The search thus branches out and, were
it not for an effective limiting mechanism, would become
unwieldy. This limitation arises from the nature of the
single node perturbations permitted.
A single node perturbation of the type permitted can be
describe as a cyclic shift by one order position of the nodes
of a consecutive subgraph, where the subgraph has,
before shift, a feedback arc connecting its two extreme
nodes. This feedback arc becomes by this cyclic shift a
forward arc between nodes adjacent in order. Of the two
possible directions for this cyclic shift, it is convenient to
choose the one changing fewer forward arcs to feedback
arcs.
Because of the branching nature of the search, once a
down sequent is established, the pair of nodes comprising
it are permanently linked in this relationship: the pair
defines a consecutive subgraph. By Theorem 5, a consecutive subgraph can be considered a single node with
respect to the remaining portion of the graph. The pair
may be in fact united to produce a single node; the arc or
arcs between them (forward arcs under the perturbed

Younger: Minimum

Feedback Arc Sets for a Directed Graph

ordering) may then be deleted. Hence each perturbation

reduces by one the number of nodes. This reduction,
favorable in itself, has a significant auxiliary effect.
Uniting two nodes of a graph generally creates loops
comprised of two arcs. By Theorem 4, arcs forming such
loops may be deleted. Each loop deleted inevitably contributes exactly one of its two arcs to any minimum
feedback arc set; hence the number of loops deleted should
be subtracted from Q(R) (where R is the reference ordering) to thus make more stringent the requirement for a
perturbed ordering to be perferable to the reference. In
fact, if the number of loops deleted exceeds Q(R), then
the down sequent leading to this reduction must be rejected. This provides the mechanism for limiting this
branching search.
Perturbing an ordering generally upsets its admissibility.
Hence after uniting the nodes of the down sequent and
canceling loops comprised of two arcs, rearrangement
sufficient to reestablish admissibility
should be made.
Then, unless the rearranged ordering is better than the
reference, further perturbations are considered.
The organization of the algorithm is made explicit
in the outline which follows. The operations on the graph
should be interpreted in terms of the connection triangle
representation.
1) Preliminary: Obtain an admissible ordering R,
relabeling the graph according to this ordering. Find the
feedback arc set determined by R and note the number
of elements in this set Q(R).
2) Branching operation: For each feedback arc, perform
a single node perturbation on R to establish the down
sequent corresponding to this arc. Unite the nodes of this
down sequent into a single node, eliminating loops of
two arcs created by this union. Add the number of such
loops to an index number I, where I indicates for a given
perturbed ordering the total number of such cancellations.

Theorem 6 b). For each optimum ordering,

the corresponding feedback arc set.
CoNCLUsIoN

A minimum feedback arc set has been shown to be

determined by a sequential ordering of the nodes of a
directed graph which minimizes the number of arcs
each of which enters a node that precedes in the ordering
the node it leaves. If feedback arc sets for all n! possible
sequential orderings of an n-node graph are scanned, all
optimum orderings (and hence, all minimum feedback arc
sets) can be identified. For large n (say 30) the number
of possible orderings exceeds reasonable bounds; to analyze
these larger graphs, the domain to be scanned must be
made smaller. This has led to a study of the properties of
sequential orderings which correspond to minimum feedback arc sets. These properties have, in turn, led to a
search algorithm. While the efficiency of this algorithm is
not easy to estimate, it appears to be suitable for graphs
of fairly large size; the author found in a few hours by
hand calculation the minimum feedback arc sets for a
12-node 41-arc graph, for which the arcs were chosen by a
random process.
For a graph too large for a complete search, a partial
search can be made by being more selective about the
down sequents to be investigated. For still larger graphs,
one may be content with an admissible ordering. But for
any graph, an ordering as good as time limitations and
other resources permit can be developed.
Although the primary purpose of the paper is to lay a
foundation for an algorithm, it is hoped that the properties
of minimum feedback arc sets and the sequential orderings
which determine them will prove to be useful in themselves, and serve to emphasize the fundamental role of
order in the analysis of directed graphs.

3) Evaluation and transfer:

a) If Q(R) - I < 0, reject the down sequent under
consideration. Return to the next feedback arc in step 2,
resetting I to the number it was before consideration of
this rejected down sequent. If all feedback arcs at a given
branching point have been examined, then proceed back
to the next arc of the branching point of the next higher
level; if at the first level, proceed to step 4.
b) If Q(R) - I 2 0, rearrange the perturbed ordering
until it is admissible. If the number of elements in the
feedback arc set determined by the perturbed ordering
Q(R,,,,) < Q(R) - I, consider it the new reference
ordering, and begin the program again at step 1. If
Q(R,,,,) = Q(R) - I, list R,,,, on a list of potentially
optimum F representatives. If Q(R,,,,) 2 Q(R) - I,
note the feedback arcs determined by this ordering and
for this new level of branching, proceed back to step 2.
4) Final processing: From the reference ordering and all
others on the potentially optimum list, find all optimum
F representatives
by the equivalences indicated in

enumerate

ACKNOTVLEDGMENT

A discussion with S. L. Hakimi, J. P. Runyon and

E. F. Moore of the Bell Telephone Laboratories helped
to deepen the authors understanding of the problem.
The author is particularly thankful for the guidance and
support of his doctoral advisor Professor 0. Wing.
BIBLIOGRAPHY

Dl C. Berge, The Theory of Graphs and its Applications, Methucn

and Co. Ltd., London; 1962. (English translation).
121S. L. Hakimi, On the Degrees of the Vertices of a Directed
Graph, to be published.
131 S. Seshu and M. B. Reed, Linear Graphs and Electrical Networks, Addison-Wesley
Publishing Co., Inc., Reading, Mass.;
lcG1
&I.

and Ordinal Numbers,

Panstwowe
[41 W. Sierpinski, Cardinal
Wydawnictwo
Naukowe, Warsaw, Poland; 1958. (In English.)
Set Theory,
D. Van Nostrand, Inc.,
[51 P. Suppes, Axiomatic
Princeton, N. J.; 1960.
161A. W. Tucker, On Directed Graphs and Integer Programs,
presented at 1960 Symp. on Combinatorial
Problems, Princeton
University, Princeton, N. J.
[71 S. H. Unger, A Study of Asynchronous Logical Feedback Networks, Research Lab. of Electronics, M.I.T., Cambridge, Mass.,
Tech. Rept. 320; April 26, 1957.