0% found this document useful (0 votes)
25 views

Factors in Random Graphs

This document summarizes a research paper on factors in random graphs. It contains the following key points: 1) It proves that in a random graph G with n vertices and edge probability p, if p is such that the graph has minimum degree at least r with high probability, then any factor f less than or equal to r will exist with high probability as n approaches infinity. 2) It outlines an approach to proving this using alternating paths to augment subfactors until a full factor is obtained. 3) It describes an algorithm to construct disjoint trees Fx and Fy from two initially unsaturated vertices x and y, with the goal of connecting them and showing the graph can be fully factored.

Uploaded by

Evan Brackett
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
25 views

Factors in Random Graphs

This document summarizes a research paper on factors in random graphs. It contains the following key points: 1) It proves that in a random graph G with n vertices and edge probability p, if p is such that the graph has minimum degree at least r with high probability, then any factor f less than or equal to r will exist with high probability as n approaches infinity. 2) It outlines an approach to proving this using alternating paths to augment subfactors until a full factor is obtained. 3) It describes an algorithm to construct disjoint trees Fx and Fy from two initially unsaturated vertices x and y, with the goal of connecting them and showing the graph can be fully factored.

Uploaded by

Evan Brackett
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 7

ISRAEL JOURNAL OF MATHEMATICS, Vol, 39, No.

4, 1981

ON FACTORS IN RANDOM GRAPHS

BY
E. SHAMIR AND E. UPFAL

ABSTRACT
The following result is proved: Let G,.p be a random graph with n vertices and
probability p for an edge. If p is such that the random graph has min-degree at
least r with probability l, then any f-factor I _-<f _- r exists with probability 1, as
yl ----) r

1. Definitions and results

Let G be a graph with vertex set V(G), edge set E ( G ) . Given a map f of
V ( G ) into the set of non-negative integers, define an f-factor of G as a
spanning subgraph of G in which the degree (valency) of x is f ( x ) . When f - 1, a
one-factor is obtained, which is also called a perfect matching, i.e. a set of
non-intersecting edges covering V ( G ) . One may want to study a maximal
matching when a perfect matching does not exist.
We shall study the existence of f-factors in the context of random graphs,
following the works of ErdSs-Renyi [2, 3, 4]. Let 1,..., n be a fixed labelling of
the vertices. Let {el,}, l<=i<j<=n, be an array of independent random
variables, each e,, assuming the value 1 with probability p, 0 with probability
1 - p. This array determines a random graph on {1,.. -, n}, where (ij) is an edge if
and only if e,, = 1. This probability space is denoted by G,.p. We allow p to be a
function of n, and study the asymptotic behavior of probabilities of events in G..p
as n--~ oo. In particular, here we study the property of having an f-factor.
The principal results about properties of G..p were obtained by Erd6s and
R6nyi [2, 3, 4]. They use a somewhat different space G,,N~,~ of graphs with N
random edges. The passage from G,.N~,~to G,., with p ( n ) ~ 2 N ( n ) / n 2 is usually
quite simple [1].

THEOREM 1 [4]. Let n be even, p = (1/n)(log n + w(n)), with lim,~| w ( n ) =


oo. Consider the event E : G E G,.p, G has a 1-factor. Then

Received September 1, 1980

296
Vol. 39, 1981 RANDOM GRAPHS 297

(1.1) lim Prob E = 1 (or E holds a.a.s.),


n ~

where a.a.s, is an abbreviation for "asymptotically almost surely"

THEOREM 2 [3]. Let p = ( 1 / n ) ( l o g n + ( r - 1 ) l o g l o g n + w ( n ) ) , r>=l,


l i m , ~ w ( n ) = ~ . Then in G~.p, Mindeg G _-r, a.s.s.

Our main result is the following

THEOREM 3. Let p be as in Theorem 2. Let 1 <=f(x~) <-_r, ET=~f(xi) even. Then


G has an f-factor, a.a.s.

REMARK. Our proof will hold also for the case p = (1/n)(log n + c), where
a.a.s. G consists of a huge component and some isolated points. A.a.s. there is a
perfect matching [1-factor] in the huge component [if it is even].

Notice that by Theorem 2 we may assume that deg(x) >- r, x E V(G). Also it is
easy to see that the graphic condition, assuring the existence of a graph with
degree sequence f(x,), is satisfied for large n.

2. Alternating paths, trees and augmentation

Existence of factors can be approached by Tutte's characterization theorem


[4, 8]. For one-factor, this approach was followed in [4, 7]. Here we shall follow
another method, using augmentation of sub-factors by alternating paths, used
extensively in the algorithmic studies of matching and flow problems [1, 5]. In
this respect it is closer to Posa's proof for a Hamiltonian path in random graphs
[6].
A sub-f-factor of G is a subgraph M such that E ( M ) C E ( G ) and
degM (x) <=f(x). In case of strict inequality, the vertex x is unsaturated. Edges in
E ( G ) \ E ( M ) are free edges.
Consider a path x0, x t , ' - ' , xm such that
x,_,x, is a free edge for i odd, an edge of M for i even, 0 < i =< m.

This is an alternating path with respect to M. A vertex with an odd [even] index
is an/-vertex [T-vertex]. If m is odd, x0 and x,, are unsaturated, such a path can
be used to augment M:Dropxi_tx,, i even, add the free edge x~_~xi, i odd,
0 < i <= m. The resulting M' is still a sub-f-factor, with a degree increased by 1 at
x,,, x,, but unchanged otherwise.
We outline the proof of Theorem 3. Consider the event
NM : There exists in G a sub-f-factor M, with an
(2.1) unsaturated vertex, which admits no augmentation.
298 E. SHAMIR AND E. UPFAL Israel J. Math.

./r by a construction which is the core of the proof, the existence of two
(probably) large disjoint trees F~, F, with edges in E(G). Being disjoint implies
disconnection between two large sets of vertices. Hence the probability of
finding such trees in G is small, in fact o(1) as n--,co.
Assume the event .N~ holds for G. Without loss of generality, there are two
unsaturated vertices x, y (if say d e g ( x ) < f ( x ) - 2 , we add some xu and drop
some uy). The trees F~, Fy are obtained by a greedy algorithm, trying to catch
many alternating paths from x and from y, while keeping a balance between I x
and Fy.

3. The parallel construction of the trees Fx, Vy and the auxiliary set N

The construction proceeds in steps. Vertices added at even [odd] steps are T
vertices [1 vertices], respectively.

Step O. x is the root ofUx, y is the root of Vy.

N ={w [zw E E ( M ) for z = x or z = y}.


Step I. A = {w [ w ~ N, wx or wy i s a f r e e e d g e } = A x U A , U B (disjoint
union) where for z = x, y, Az C_A is the set of w's which are connected to Fz
only, B is the set connected to both. Let Az C_At be a maximal set such that two
vertices in ,4~ are not connected by an edge of M. Extend F~ by connecting a set
A~ C , ~ U B such that

(3.1) Ax h A , = O , lAx I -- IA, I = [ IBIJ +Mintl, I,l 4, I).


An edge of M connecting a vertex u in B with a vertex w in B, Ax or A , or an
edge connecting a vertex u in Ax to a vertex w in A, closes an alternating path
between x and y, hence an augmentation for M. Thus the construction of A, and
A, gives the maximum possible additions which are disjoint, equal and no edge
of M connects two vertices in Ax U Ay.

Step A J even. Connect to F~ any w such that

(:lu) (u an/-vertex of F~, uw E E(M)),

under the provision that the set of w's connected to both trees is split evenly
between them (if a single one remains, it is connected to the smaller tree; if they
are equal, it is connected to Fx).

Note: the vertex w is new (outside Fx U Fy). Indeed, w was not added in the
( j - 1) step since vw is an edge of M.
Vol. 39, 1981 RANDOM GRAPHS 299

If w had been added in a previous odd [even] step in F, U Fy as an/-vertex


[T-vertex], then u would be a T-vertex [/-vertex with u E N], which is
impossible.
Add to N all vertices connected to the (newly introduced) T-vertices by edges
of M.

Step ], j odd. A = {w I w ~ N, w is new and connected to a T-vertex in


Fx U F, by a free edge} = A, U A, U B (disjoint union).
The decomposition of A and how to extend F~ by new /-vertices is done
precisely as described in Step 1 above.

The construction terminates when it is impossible to extend the trees and


preserve equality in the number of /-vertices in F~, Fr Upon termination all
T-vertices of one of the trees are connected only to vertices in F, U Fy U N (and
perhaps one more vertex in case (3.1) was 0 at the terminating odd step, since
[5,z [= 0 if and only if [Az [= 0).

4. The probable size of the trees

CLAIM 1. J~..~ implies that the trees F,, z = x, y, each have at least two
T- t,ertices, a.a.s.

PROOF. The root z itself is one T-vertex. Being unsaturated, z has one free
edge which is not xy (else M can be augmented). Each Fz has an/-vertex uz.
Otherwise x and y each has a single free edge, which is connected to the same
vertex u (and each has < r edges of M). The probability of this event is
estimated by

< n(log n + w(n)2"e -2~'§ O(n-').

Now if Fz does not contain another T-vertex (connected to u,), then all the
/-vertices of F~ U Fy are connected by M to one and the same vertex w. Only up
to r vertices may be connected to w by M, hence lAx U A I =<r in the first step
and 2 _-<IAx U A, U B I =<r~ (since r 1,4z I =>IA, I). Thus x and y are connected to
at most r2+2r vertices and at least two of these vertices have a common
neighbour. We can express the existence of such a configuration in a graph
G E G,.p and estimate its probability by
300 E. SHAMIR AND E. UPFAL Israel J. Math.

,2+2, n n log n
~: (2)(k)(1)( +w(n))k§176 z'"-k-',

-_<~ n(log n + w(n))k§ -2'~


k

Let K be the number o f / - v e r t i c e s in Ft.

CLAIM 2. Each tree has at least K / 2 r and at most rK + 1 T-vertices.

PROOF. Each /-vertex is connected in M to at least one T-vertex. Each


T-vertex is connected in M to at most r vertices. Thus after the splitting between
trees, each tree has at least K / 2 r T-vertices (including its root).
Each T-vertex (except for the root) is connected in M to at least one/-vertex.
E a c h / - v e r t e x is connected in M to at most r vertices. Hence there are at most
rK + 1 T-vertices.

CLAIM 3. Each tree has at most 3rK vertices.

PROOF. There are K / - v e r t i c e s , at most rK + 1 T-vertices.

CLAIM 4. I N[<-_4r2K.

PROOF. Multiply the estimate for T-vertices in Fx U Fr by r.

CLAIM 5. Each tree has at least n/80r 5 T-vertices, a.a.s.

PROOF. Let Fx be the tree which caused the construction to terminate. All its
T-vertices are connected to Fx t_JFy U N and perhaps one more vertex s.
Consider FC(Fx t_JFy U N U { s } ) which is the tree obtained from Fx by (i)
reconnecting all vertices which are connected to a T-vertex of Fx, but went to F,
upon the splitting of B, (ii) connecting a w E N which is connected to a T-vertex
by an edge of M, (iii) connecting s. Note that F is indeed a tree, we added vertices
to the leaves of Fx with one connection each (as we did throughout).

]F]=<IFx U F , U N U{s}[---_ 10r~K.

Let t be the number of T-vertices in Fx. Since by Claim 1, t -> 2 and by Claim
2, t >=k/2r,
t =>Max(2, Flrl/20rq).
Consider the event

E: There is a tree F with 2 <=! <=n/2 vertices in which t vertices,


t = Max(2, [l/20r ~]), are not connected in G to a vertex outside
F.
Vol. 39, 1981 RANDOM GRAPHS 301

2 0 r 3 . P r o b E _<-- ~ i z log n + w 2~ + w ( n ) '~"-~)


2~t ~nl4~r 3 n r[

<=~ n(2e l o g n
f
+ w(n))'exp (-t(n n l)(logn + w(n)))
---
2_~u<log n
+ E
logn~t<=n/40r 3
--E'+Y;';

~'<-- n(2e log n + w(n))4~ e -2'~ ~"--< n2(log n)'Og"e -t~

Thus Fx has at least n/40r 3 T-vertices a.a.s. Hence, by Claim 2, K >-_n/40r 4, and
so Fr also has at least n/80r 5 T-vertices.

5. Conclusion of the proof

CLAIM. 6. limo_~ Prob (NM) = 0.

PROOF. Let q = n/N/log n. Take a set A of T-vertices in F , I A I = q. It is


connected by M to at most rq vertices. Thus Fy contains another set D of q
vertices which (since NM holds) has no connection in G to A. Indeed a free edge
connecting a T-vertex b in Fx to a T-vertex c in Fr closes an alternating path
between Fx and F , hence an augmentation for M. The probability that such A
and D exist is bounded by

n n-q 1-1~ q~<__ qexp - (logn+w(n))


q q n

<_-(N/log n e -V,og,),/V~og,,

which has a sub-exponential decrease, as n--~ oo.

REFERENCES

1. D. Angluin and L. Valiant, Fast probabilistic algorithm for Hamiltonian circuits and
matchings, J. Comput. Syst. Sci. 18 (1979), 155-193.
2. P. Erd6s and A. R6nyi, On the evolution of random graphs, Publ. Math. Inst. Hungar. Acad.
Sci. 5A (1960), 17-61.
3. P. Erd6s and A. R6nyi, On the strength of connectedness of a random graph, Acta Math.
Acad. Sci. Hungar. 12 (1961), 261-267.
4. P. Erd6s and A. R~nyi, On the existence of a factor of degree one of a connected random graph,
Acta Math. Acad. Sci. Hungar. 17 (1966), 359-368.
302 E. SHAMIR AND E. UPFAL Israel J. Math.

5. E.L. Lawler, Combinatorial Optimization: Networks and Matroids, Holt, Rinehart and
Winston, 1976.
6. L. P6sa, Hamiltonian circuits in random graphs, Discrete Math. 14 (1976), 359-364.
7. E. Shamir and E. Upfal, One factor in random graphs based on vertex choice, submitted, 1980.
8. W. T. Tutte, The subgraph problem, Ann. Discrete Math. 3 (1978), 289-295.

INS1TrUTE OF MATHEMATICS
THE HEBREW UNIVERSITYOF JERUSALEM
JERUSALEM, ISRAEL

DEPARTMENT OF APPLIED MATHEMATICS


THE WEIZMANN INSTITUTE OF SCIENCE
REHOVO'r, ISRAEL

You might also like