Вы находитесь на странице: 1из 13

An Upper Bound for

the Path Number


of a Graph -
Alan Donald’
QUEEN’S UNlVERSlN

ABSTRACT
The path number of a graph G, denoted p(G), is the minimum number of
edge-disjoint paths covering the edges of G. Lovasz has proved that if G
has u odd vertices and g even vertices, then p ( G ) s $ u + g - 1 s n - 1 ,
where n is the total number of vertices of G. This paper clears up an
error in Lovasz’s proof of the above result and uses an extension of his
construction to show that p ( G ) s $ u+I$gl=s&l.

INTRODUCTION
In this paper a graph is a finite graph (connected or disconnected) without
loops or multiple edges. A path is a simple path: i.e., one in which no
vertex appears more than once. Similarly, a circuit is a simple circuit: the
only vertex appearing more than once is the initial (=terminal) vertex. A
path (respectively, path-circuit) decomposition of a graph G is a partition
of the edges of G into paths (respectively, paths and circuits). Such a
decomposition is said to be minimal if its cardinality (i.e., the number of
paths or paths and circuits it contains) is no greater than the cardinality of
any other such partition. The path number of a graph G, denoted p(G), is
the cardinality of a minimal path decomposition. This concept appears to
be due to Ore [4, p.401 and Erdos [3]; the term “path number” was
coined by Harary [2, 51.
Gallai has conjectured that for any connected graph G on n vertices
p( G) S [( n + 1)/2]. (The square brackets indicate the greatest integer
function.) In 1966, Lovhsz [3] proved the following theorem and its two
corollaries. (See also Berge [l,pp. 234-2361.)


Present address: Department of Epidemiology and Biostatistics, University of
Western Ontario, London, Ontario.

Journal of Graph Theory, Vol. 4 (1980) 189-201


01980 by John Wiley & Sons, Inc. 0364-9024/80/0004-0189$01 .OO
190 JOURNAL OF GRAPH THEORY

1. Theorem (Lovhsz). If G is a graph (not necessarily connected) on n


vertices, then there exists a partition P of the edges of G into paths and
circuits with [PI G4n.
2. Corollary. If G is a graph with vertices of odd degree only, then
p ( G) = 4n.
3. Corollary. If G is a graph containing exactly u odd vertices and g 2 1
even vertices, then
p ( G) S 4~ + g - 1.

The purpose of this paper is to correct an error in the proof of


Theorem 1 and to present a sharpening of the bound given in Corollary 3.
The new bound is
p ( G ) G $ u+[$g]S[$n].

LOVASZS THEOREM

My reading of the original proof of Theorem 1 led me to think that the


following was known to Lovhsz; however, subsequent correspondence
indicated that it was not.
4. Conjecture. For any graph, there is a minimal path-circuit decompos-
ition in which each odd vertex is the endpoint of exactly one path*
I could not prove Conjecture 4. The presentation of Theorem 1 in [l]
contains a stronger statement: In any minimal path-circuit decomposition,
exactly one path begins at each odd vertex. The graph G in Figure 1 is a
counterexample to this statement.
2

FIGURE 1

* Added in proof: A counterexample to Conjecture 4 has recently been found


by D. de Caen.
PATH NUMBER OF A GRAPH 191

Inspection of G shows that the path-circuit decomposition indicated is


minimal. Vertex 1 is the endpoint of 3 paths in this partition. The graph
G is not a counterexample to Conjecture 4, however, for there is a
minimal path-circuit partition in which each odd vertex is the endpoint of
exactly one path: simply delete the edge (12) from the path (1245) and
add it to either of the other paths beginning at 1.
Conjecture 4 suggests the following:

5. Conjecture. Let G be a graph with u odd vertices and g even


+
vertices. There exists an integer k S [ ( g 1)/2] and a minimal path de-
composition P of the edges of G such that:

(1) Each vertex of odd degree is the endpoint of exactly one path of
P.
(2) There are k vertices of even degree, each of which is the
endpoint of exactly two paths of P.
(3) None of the remaining g- k even vertices is the endpoint of a
path.

Conjecture 5 is stronger than Conjecture 4 and implies Gallai’s conjec-


ture.
Lovlsz has also pointed out (in private correspondence) that Conjec-
ture 4 implies Hajos’ conjecture that a simple graph with n even vertices
and no odd vertices can be decomposed into <[$I] disjoint circuits [3].
We now present a proof of Theorem 1 that is almost the same as the
original, but which does not assume Conjecture 4. Theorem 1 and the
proof of the new bound follow from the following construction, which we
shall call LovCisz’s construction about the vertex x. We use his notation.

6. Lovasz’s Construction. Let G be a graph with at least one noniso-


lated even vertex and let x be a vertex adjacent to an even vertex.
Denote by a,, a,, . . . , ak the even vertices of G adjacent to x. Let G’ be
the graph formed from G by deleting the edges xa,,i = 1 , . . . , k. If H is
the set of vertices adjacent to x in G, then every vertex of H is odd in G’.
Let C be a minimal path or path-circuit decomposition of the edges of G’.
(In proving LovBsz’s Theorem, we let 2 be a path-circuit decomposition;
in proving the new bound we let C be strictly a path decomposition.)
For each i = 1 , . . . , k, we define a sequence of vertices of H : a,,O,
. . . , a,,,,by recursion as follows. Let ai,,,= ai. Suppose ai+ is defined;
since q Pis in H, it is of odd degree in G’ so there is a path Ui,wof 2 that
has ai+ as an endpoint. If U,,&does not pass through x, end the sequence;
192 JOURNAL OF GRAPH THEORY

if x lies on Ui,,, trace the path from a,,, to x and let be the last
vertex on the path before it reaches x. We call each sequence Si = (u,.~,
. . . , a,,,,)a Lovdsz sequence.
Lovisz at this point makes two observations and proves the second
one:
6.1. ai,, is adjacent to x in G’ iff p a 1.

6.2. If at,, = a,,,, then i = j and p = v. For suppose a,,, = q Yand p S v.


If p = 0, then a,,, = a,, so v = 0 and i = j. If p a 1, then a,,, lies on U,,,-I,
which has an endpoint at u ~ , , - ~and contains the edge xu,,. Similarly,
uc,,= 4,. lies on UJ,Y-l which contains xa,,,= xa,,,.Since the paths of C
are edge disjoint, Uc,,-l = UJ,Y-land al,,-, = u , , ~ - ~If. we continue this
retracing process, we arrive at a,, = al,,-,. By the same argument used
for the case p=O, we have v - p = O so p = v and i=j.
Assertion 6.2 also implies that each Lovisz sequence is finite since H is
finite and there is a one-to-one correspondence between the labeled
vertices of H and their labeling. Note that the proof of 6.2 uses no
assumption that each a,,, is the end point of a unique path of C; however,
in constructing the Lovisz sequences, we have chosen for each a,, a path
U,,, and to this extent there is a uniqueness property that enables us to
retrace the sequence of vertices. For the sake of clarity, we make two
more observations:

6.3. The labeling of paths associated with the Lovisz sequences is not
unique: we can have Ui,, = q.+ with ( i , p ) # ( j , v).

6.4. The edge xu,,, lies on Ui,,-l if p21.


We now define a function f that maps elements of C into paths and
circuits that partition the edges of G. Let H’ be the set of indexed vertices
of H. Let U be a path or circuit of X. If U is not a Q,, let f ( U i = U.
Otherwise we distinguish two sets of pairs of mutually exclusive cases.
The first pair is as in Lovisz’s proof:

(a) U = U,,, contains x.


(a‘) U = U,,, does not contain x.

The second pair of cases replaces those used by LovBsz. Some com-
ments on this will follow the construction.

(b) U = U,,, = t J Ywhere (i, p ) # (j, v).


(b’) U = U,,, and the labeling is unique.
PATH NUMBER OF A GRAPH 193

FIGURE 2

We combine these cases into four possible situations and define f(U)
for each.
Case (ab). (a) and (b) hold (see Fig. 2). Let f(Ui,, = Ui,,),be U,,, = Ui.v
with the edges X U ~ , + . + and
~ xj,v+l deleted and the edges xui,, and xi,,
added.
Case (ab'). (a) and (b') hold (see Fig. 3). Let f(Ui,,) be U,,, with X U ~ , , + ~
deleted and xui,, added. (Note that in Fig. 3 vertices y and z might or
might not be in H'.)
Case (a'b). (a') and (b) hold (see Fig. 4).Let f(U,,, = q.,,)
be the circuit
obtained by adding xu,,, and xui,, to Ui,, = U1.,".
Case (a'b'). (a') and (b') hold (see Fig. 5). Let f(q,,)be the path
obtained by 'adding xu,,, to Ui,,.
This completes the definition of f. We now show that the paths and
circuits formed by f are a partition of the edges of G ; i.e., if yz is an edge
of G, then yz appears in exactly one f( U ) .Qearly, if yz is not of the form
xui,,, then yz appears in some U of 2 and also in, and only in, f(U). It is

FIGURE 3
194 JOURNAL OF GRAPH THEORY

f
o x

FIGURE 4

obvious from an examination of the cases that xa,,,, appears in f ( U,,,),so


it remains to prove only that if m k , , is in f ( U l , , ) ,then Uk,p=U,,,.
Let Xakp be an edge of f(UI,,)and assume that a,,,# a,+. U,,, falls
under one of the four cases mentioned above.
If Ul,, is of the form (ab) or (a’b), then U,,, = q,, where (i, p ) # (j, v).
The only edges of f ( U,,,) incident to x are xu,, and xa,,,. Since ak,p# a,,,,
by assumption, it follows that ak,p= a,,,, SO ( k , p) = (j, v) by 6.2.
If Ul,, is of the form (ab’), let U,,, =(aI,,,. . . , a,,,+lr x , y , . . . , z ) . (See
Fig. 3.) If Xak,, is in f(u,,,) and ak,p# a,,,, then ak,p= y. Thus Z = U k , , - 1
and Ul,,= U k , p - l But . U,,, is of type (b’) so it is uniquely labeled. So
( i , p ) = ( k , p - 1) and a,,, = ak,p-l= z. Thus Ul,, must have been a circuit,
which is not the case. So xak,,$f(U,,,)unless ( k , p) = (i, p ) .
If Ul,, is of the form (a’b’),the only edge of f( Ul,,)adjacent to x is a,,,.
Thus f(2)is a partition of the edges of G into paths and circuits and
lf(Z)l= 1x1. This completes Lovbsz’s construction about x .

FIGURE 5
PATH NUMBER OF A GRAPH 195

v FIGURE 6

The above con: ruction differs from LovPsz’s original cons ruction only
in the way in which the paths Ui,& are classified. Instead of cases (b) and
(b’), LOV~SZ distinguished between paths that had one endpoint in H’ and
those that had two. If C is a partition with the characteristic that each odd
vertex is the endpoint of exactly one path, LovPsz’s classification implies
(b) or (b’) and f(2) turns out to be a partition. Since we make no such
assumption about 2 however, classifying paths according to the number
of endpoints in H’ might cause the following situation. Suppose Ui,&has
two endpoints a,+ and q Vin H’. Then a path Uj,”begins at aj,”;it might
be that # Uj,, and under the original instructions for forming f , the
edge xuj,+ is an element of both f(&) and f(Uj,,). (See Fig. 6.)
With this problem somewhat laboriously clarified we are ready to give
the proof of Lovasz’s Theorem. The proof is as LovPsz originally pre-
sented it, with a few details added.
Proof of Theorem 1. Let G be a graph with n nonisolated vertices. If
G contains no even vertices, let G, be the graph obtained by inserting a
vertex y (of degree 2) on some edge. G, contains n1= n + 1 nonisolated
vertices. We claim that if the edges of GI can be partitioned into S $ n ]
disjoint paths and circuits, then so can the edges of G. Note first that
since each vertex of G is odd, n is even, so [inl] = [$TI]. If we partition the
edges of G into a minimal number of paths and circuits, then there are at
least [in] paths in the decomposition. By the hypothesis, there are < [ i n ]
paths and circuits, so in fact the partition contains exactly [in]paths and y
is the endpoint of no path. Thus p(G) = [in].We can now assume that G
contains at least one nonisolated even vertex.
The proof is by induction on h ( G )= 2m - n, where m is the number of
edges of G. If h(G)<O, then n S 2 m =xu d ( u ) , where the summation is
over the n nonisolated vertices. Thus d ( u ) 1 for each such vertex and
196 JOURNAL OF GRAPH THEORY

since none of the vertices are isolated, d(v) = 1 for each one. Thus G is a
set of disjoint edges, possibly with some isolated vertices.
Let A(G) = A and suppose the theorem holds for all graphs G’ with
A(G’)<A(G). Let x be a vertex of G adjacent to even vertices
a,, u 2 , . . . , uk. Delete the edges xu,, i = 1 , 2 , . . . , k to obtain G’, a graph
with n’ nonisolated vertices and m’ edges. Since n - 1S n ’ 4 n (with
n ’ = n - 1 if x is isolated in G’) and m ’ = m - k S m - l , we have A(G)S
2 m - n - 1< A(G). Let C. be a partition of the edges of G’ into S[$n]
paths and circuits and complete the Lovhsz construction about x to obtain
a partition of G into If(C)I = 121<[in] paths and circuits. Theorem 1 is
proved. I
Note that the parity of the degrees of the neighbors of x in G was used
only to ensure that each began a path of 2. This enables us to state the
following convenient generalization of the result of the construction:
7. Lemma. Suppose x is a nonisolated vertex of a graph G, 2 is a
nonempty set of neighbors of x, G’ is the graph obtained from G by
supressing the edges joining Z to x, and C is a path (or path-circuit)
decomposition of G’. If every neighbor of x in G begins some path of 2,
then there is a path-circuit decomposition of the edges of G whose
cardinality is 1x1.

A NEW BOUND

To lower the bound on p ( G ) further than that of Corollary 3, we first


extend the Lovasz construction by breaking any cycles formed by f into
paths. Note that f produces a cycle only when vl,,, = q.,v
and ui,+and ui,,
are the final entries in the Lovhsz sequences Si and Si.Thus:

8. REMARK. If q is the number of circuits formed by f, then q s i k , where


k is the number of even vertices adjacent to x in G.

9. Extension of Lovasz’s Construction. Suppose we have performed the


Lovasz construction about x and obtained a path and circuit covering f ( X )
of G. Suppose further that the original partition 2, of edges of G’ was by
paths only. Then in f(2,) we have q circuits, O ~ q s i k where
, k is the
number of even vertices adjacent to x.
We break these circuits into edge disjoint paths in a simple-minded
manner. Delete from each circuit an edge of the form xu,, to obtain q
paths and q extra edges. Since each of these edges is incident to x, we can
PATH NUMBER OF A GRAPH 197

pair them up to obtain i q paths of length 2 if q is even, or $(q- 1) paths


of length 2 and one path of length 1 if q is odd. (See Fig. 7.) We thus
obtain a path partition of the edges of G of cardinality I2(+ [i(q + 1)].
We will refer to this partition as f’(2).[Strictly speaking, f’ is not
well-defined since the way we break up the circuits of f(2)is to some
extent arbitrary; but this will pose no problems in the future since we are
concerned primarily with the cardinality of f(2).]
We summarize these results in the following lemma:
10. Lemma. Suppose we have performed the L o v b z construction about
x to obtain a path-circuit decomposition f(2)of G. Let k be the number
of even vertices adjacent to x in G. If X consists only of paths of G’, then
in f(2) we have exactly q circuits where

The next lemma will be used later:

11. Lemma. Let G be a graph, x E V(G) a vertex adjacent to an even


vertex. If f’(2)is a path decomposition of the edges of G obtained by
performing the L O V ~ Sconstruction
Z and its extension about x, then at
least one path of f’(X) begins at x.
Proof. Let q rbe the last element of S , and let U = Ul,r.Since U
does not pass through x, it is of type (a’). If U is of type (a’b), f(U) is a

FIGURE 7
198 JOURNAL OF GRAPH THEORY

circuit and the extension of the construction produces a path beginning at


x. If U is of type (a’b’), f(U) is a path beginning at x. I
We are now ready to prove the new upper bound for p ( G ) .

12. Theorem. Let G be a graph with u vertices of odd degree and g


nonisolated vertices of even degree. Then

p(G)==$u +[as].
Proof. The proof is, as in LovBsz’s theorem, by induction on h(G)=
2 m - n , where m is the number of edges in G and n is the number of
nonisolated vertices. If g = 0, Corollary 2 gives the result immediately, so
we assume G contains at least one nonisolated even vertex. If h ( G )= 1,
we have

where the summation is over all nonisolated vertices, and d ( u , ) z 1 for


each ui.We therefore have n - 1 vertices of degree 1 and one vertex of
degree 2, and the result follows by Corollary 3 .
Suppose the hypothesis holds for all A‘ < h ( G ) .

Case I. G contains at least one pair of adjacent even vertices.


Let x be one such vertex and let a,,a,,. .. , a k be the even vertices
adjacent to x. 2

Subcase I(a). If k + 2 (mod 4). Delete the edges xu,, i = 1 , 2 , . . . , k to


obtain a graph G’. If u’ and g’ are the numbers of odd and even vertices
in G’, then p(G’)<$u’+[$g’]. If q circuits are produced by the Lovasz
construction about x, then, by Lemma 10,

+ l)].
p(G)~qu’+[$g’]+[i(q (3)

(i) If k = 0 (mod 4), the vertex x is even in G’, so u’ = u + k and


g ’ s g - k (if x is isolated then g’ = h - k - 1). Also, 2q 6 k so
[;(q + l ) ] s $ k . Thus Eq. (3) gives:

~(G)~~(~+k)+[$(g-kk)l+[~(q+1)1
= +U + [$g] + [$(q+ l)]- $ k
6;u +[&I.
(ii) If k is odd, x is odd in G‘ so u’ = u + k + 1 and g ‘ = g - k - 1.
PATH NUMBER OF A GRAPH 199

Now, 29 S k - 1, since k is odd, so inequality ( 3 ) gives:

p( G) ~ $ + [u&I.
Subcase I@). Suppose k = 2 (mod 4). Here we adjoin to G one vertex
y and the edge a,y to obtain the graph G,. We claim p(G)Sp(G,), for if
2 is a path decomposition of G,, then the edge yak is the initial edge of a
path of C. We can remove this edge, reducing the cardinality of the
partition by 1 if yak is itself a path and by 0 if yak is not an entire path.
We now show that p(GJ s + u +[&I.
As before, we delete xui, i = 1,2, . . . , k - 1 to obtain G:. Since k = 2
(mod 4), k - 151, so we delete at least one edge and do not isolate the
even vertex x. A ( G;) < h(G) = A, and since k - 1 is odd, u: = u + k + 2 and
g; = g - k - 1 where g{ and u ; are the numbers of even and odd vertices
in G:. We do the Lovisz construction to obtain paths and circuits as
before. We perform the extension of the Lovisz construction on the q
cycles.
We have 2q S k - 1= 1 (mod 4). So

[$(q+ I)] (:[ k + l)]= $(k - 2).


Using this inequality and inequality (3) gives:
p(G)Sp(Gi)s$u+[$gI-
Case 11. G contains no pairs of adjacent even vertices and also there
exists a nonisolated even vertex y such that for any other even vertex z,
d(y, z ) z 3 . (This case also includes, trivially, the case in which g = 1.)Let
x be an odd vertex adjacent to y. Delete xy and complete the Lovisz
construction about x. Since x is adjacent to no other even vertex, u' = u,
g' = g, and, since k = 1, q = 0. So Lemma 11 implies:

Case 111. If neither Cases I nor I1 hold, there is a pair of even vertices x
and z such that d(x, z ) = 2 and neither x nor z is adjacent to any even
vertex. So there is an odd vertex y such that xy and yz are edges of G.
Delete the edge xy to obtain the graph GI with u1 odd vertices and g,
even vertices. Now y is even and x is odd in GI. Moreover, x is adjacent
to odd vertices only in G,; and G1 contains a pair of adjacent even
vertices y and z. We now apply the Lovisz construction and its extension
to G1 and y. Since h (G:) < A ( G,) < A (G) = A, this gives a path decompos-
ition 2 of the edges of G1 with 121S $ u l +[$gl]. At least, one path of C has
an endpoint in y, by Lemma 12.
200 JOURNAL OF GRAPH THEORY

We now reintroduce the edge x y without disturbing the cardinality of


the path partition. Since each vertex adjacent to x in G, is odd and since
y is the endpoint of at least one path of 2, Lemma 8 guarantees the
existence of a pathcircuit decomposition C, of G with cardinality 1x1. This
is in fact a path partition for the following reason: the Lovhsz construc-
tion produces circuits only when there is a path Ui,,= U,,, and a,, and
a,,, are the final entries in the sequences Si and Sj with i # j . Since we
obtained G, by deleting only one edge of G, we are dealing with only one
sequence, so no circuits are formed. Thus

where C, is a path decomposition of the edges of G. But u = u , and


g = g,, therefore p ( G ) ~ $+[&I.
u
This completes the proof of Theorem 12. I
It is a consequence of Theorem 12 that for any graph G on n vertices,
p ( G ) s [ $ n ] Although
. this is an improvement over the previous upper
bound of n - 1 (cf. Corollary 3), it appears to be sharp only for some
small graphs. Let G be two disjoint triangles, for example.
It is worth noting, however, that if Gallai’s conjecture is true, then the
best bound possible for the path number of an unconnected graph is [gn].
To see this, suppose G has k nontrivial components G,, G Z ., . . , Gk each
with an odd number of vertices. If G, has ni vertices for i = 1 , . . . , k, then
n i 3 3 so

n= c ni23k.
k

i=l

Now, Gallai’s conjecture implies that

Although the bound proved in this paper exceeds this bound by only
A n , it still appears that Lovhsz’s construction will not yield p(G)SGn],
since in breaking up circuits to form paths we run the risk of creating as
PATH NUMBER OF A GRAPH 201

many as [$(g - l)]new paths. So far, I can see no way of restricting the
number of circuits to avoid this possibility.

ACKNOWLEDGMENTS

The author wishes to thank Professor L. Lovhsz for his helpful reply to
questions about this subject and Professor N. J. Pullman for suggesting
Lemma 7 and for many encouraging discussions. Work on this paper was
supported in part by grant No. A4041 from the National Research
Council of Canada.

References

[l] C. Berge, Graphs and Hypergraphs. Translated by E. Minieka. North


Holland, London (1973).
[2] F. Harary and A. J. Schwenk, Evolution of the path number of a
graph; covering and packing in graphs, 11. In Graph Theory and
Computing. Edited by R. C. Read. Academic, New York (1972)
39-46.
[3] L. Loviisz, On covering of graphs. In Theory of Graphs (Proceedings
of the Colloquium held at Tihany, Hungary, September, 1966). Edited
by P. Erdos and G. Katona. Academic, New York (1968) 231-
236.
[4] 0. Ore, The Theory of Graphs. American Mathematical Society,
Providence, R.I. (1962).
[ 5 ] R. G. Stanton, D. D. Cowan, and L. 0. James, Some results on path
numbers. Proceedings of the Louisiana Conference on Combinatories,
Graph Theory and Computing. Edited by R. C . Mullin, K. B. Reid,
and D. P. Roselle. Louisiana State University, Baton Rouge (1970)
112-135.

Вам также может понравиться