On the threshold for k-regular subgraphs of random graphs
Pawel Pralat, Jacques Verstraete, Nicholas Wormald

TL;DR
This paper proves that for large enough k, the (k+2)-core of a random graph almost surely contains a spanning k-regular subgraph, pinpointing the threshold for the emergence of such subgraphs.
Contribution
It establishes that the threshold for the appearance of a k-regular subgraph coincides with the threshold for the (k+2)-core in random graphs for large k.
Findings
The (k+2)-core of a random graph contains a spanning k-regular subgraph asymptotically almost surely.
The threshold for the emergence of a k-regular subgraph is at most the threshold for the (k+2)-core.
The threshold window for the appearance of a k-regular subgraph is roughly 2/n for large n and k.
Abstract
The -core of a graph is the largest subgraph of minimum degree at least . We show that for sufficiently large, the -core of a random graph asymptotically almost surely has a spanning -regular subgraph. Thus the threshold for the appearance of a -regular subgraph of a random graph is at most the threshold for the -core. In particular, this pins down the point of appearance of a -regular subgraph in to a window for of width roughly for large and moderately large .
Click any figure to enlarge with its caption.
Figure 1Peer Reviews
No public reviews on file for this paper yet. If you reviewed it on a platform where reviews are public (OpenReview, ICLR, NeurIPS, ICML), you can paste yours below so the community can read it here.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsLimits and Structures in Graph Theory · Graph theory and applications · Advanced Graph Theory Research
On the threshold for -regular subgraphs
of
random graphs.
Paweł Prałat
Department of Mathematics and Statistics
Dalhousie University
Halifax NS, Canada
Jacques Verstraëte
Department of Mathematics and Statistics
McGill University
Montreal QC, Canada
Nicholas Wormald
Department of Combinatorics and Optimization
University of Waterloo
Waterloo ON, Canada Supported by the Canada Research Chairs Program and NSERC
Abstract
The -core of a graph is the largest subgraph of minimum degree at least . We show that for sufficiently large, the -core of a random graph asymptotically almost surely has a spanning -regular subgraph. Thus the threshold for the appearance of a -regular subgraph of a random graph is at most the threshold for the -core. In particular, this pins down the point of appearance of a -regular subgraph in to a window for of width roughly for large and moderately large .
1 Introduction
In this paper, we study the appearance of -regular subgraphs of random graphs. The -core of a graph is the unique largest subgraph of of minimum degree at least (note that the -core may be empty). Evidently, the -core of a graph can be found be repeatedly deleting vertices of degree less than from the graph. In the case , this corresponds to the appearance of cycles in , which is well-researched, and precise results concerning the distribution of cycles may be found in Janson [6] and Flajolet, Knuth and Pittel [5]. By analysing the vertex deletion algorithm for the Erdős-Rényi model of random graphs, Pittel, Spencer and Wormald [12] proved that for fixed , there exists a constant such that is a sharp threshold for a -core in . (When discussing thresholds of -cores and -regular subgraphs, we mean thresholds for nonempty -cores and nonempty -regular subgraphs.) Here
[TABLE]
where is defined by
[TABLE]
and is the positive number minimising the right hand side of (1). Recently, a number of simpler proofs establishing the threshold for the -core have been published (see Kim [9], Cain and Wormald [4], and Janson and Luczak [7]).
In what follows, we write a.a.s. to denote an event which occurs with probability tending to one as . In comparison to studying the -core in random graphs, it appears to be substantially more difficult to analyse the appearance of -regular subgraphs when . One reason is that it is NP-hard to determine whether a graph contains such a subgraph, and there is no analogue of the simple vertex deletion algorithm which produces the -core. As every -regular subgraph is contained in the -core, we deduce that a.a.s. does not contain a -regular subgraph whenever is below the threshold for the -core described in (1) and (2). Bollobás, Kim and Verstraëte [2] showed that a.a.s. contains a -regular subgraph when is, roughly, larger than , and conjectured a sharp threshold for the appearance of -regular subgraphs in . In the same paper it was shown that for some , the -core of has no 3-regular subgraph a.a.s., whereas for , the -core of contains a 3-regular subgraph a.a.s. In support of the conjecture of a sharp threshold, Pretti and Weigt [13] numerically analysed equations arising from the cavity method of statistical physics to conclude empirically that indeed, there is a sharp threshold for the appearance of a -regular subgraph of a random graph. For they concluded that it is the same as the threshold for the -core, which is at odds with [2, Conjecture 1.3]. For , these thresholds differ, as shown using the first moment method in [2].
In this paper, we improve the window of the threshold for -regular subgraphs in by proving Theorem 1 below. A -factor of a graph is a spanning -regular subgraph, and a graph is -factor critical if, whenever we delete a vertex from the graph, we obtain a graph which has a -factor.
Theorem 1**.**
There exists an absolute constant such that for , the -core of a random graph is nonempty and contains a -factor or is -factor-critical a.a.s.
Theorem 1 will be proved in Section 4. We remark that the first nonempty -core of the random graph process a.a.s. contains many vertices of degree adjacent to vertices of degree , so the -core cannot contain a -factor and cannot be -factor critical a.a.s. Bollobás, Cooper, Fenner and Frieze [1] conjectured that the -core contains edge disjoint hamiltonian cycles a.a.s., so Theorem 1 supports this conjecture.
The value of can be determined approximately for large as follows. This corrects, and sharpens, the error term of the formula given in [12]. All logarithms in this paper are natural, and is the set of positive integers.
Lemma 1**.**
For any , let . Then
[TABLE]
Lemma 1 is proved in Section 5. It follows immediately from this lemma that
[TABLE]
Hence, combining the lemma and Theorem 1, we have pinned down the threshold for the appearance of -regular subgraphs in to a window for of width . The following questions remain: (1) to determine whether there is a sharp threshold for the appearance of a -regular subgraph, and (2) whether the -core of a random graph, when it is a.a.s. nonempty, contains a -factor or is -factor critical a.a.s.
Throughout the paper, we denote by the Erdős-Rényi model of random graphs. If is a graph with vertex set , then denotes the number of edges of with one endpoint in and one endpoint in , where . If , we write instead. The number of components of a graph is denoted by .
2 Factors of Graphs
In this section, we allow graphs to contain multiple edges. Let be a graph and let . Recall that a -factor of is a spanning subgraph of all of whose vertices have degree . A graph is -factor-critical if the deletion of any vertex of results in a graph with a -factor. In particular, a 1-factor of is a perfect matching of . Tutte’s -Factor Theorem gives the following necessary and sufficient condition for a graph to have a 1-factor.
Theorem 2**.**
Let be a graph, and let denote the number of components of with an odd number of vertices. Then has a 1-factor if and only if
[TABLE]
for every set .
Using Tutte’s 1-Factor Theorem applied to a new graph, , a necessary and sufficient condition can be found for a graph to have a -factor (this is a special case of Tutte’s -factor theorem; see Lovász and Plummer [10] for details). To construct , let , and let , where and are disjoint sets and is partitioned into independent sets with and is partitioned into sets such that . Then consists of all edges between and for , together with a matching on such that when we contract all the independent sets in to single vertices, we obtain . An example is shown below in Figure 1, where is a quadrilateral, , and a -factor in corresponds to a -factor in .
Figure 1 : 2-factor in and corresponding 1-factor in .
It is not hard to see that has a -factor if and only if has a -factor. We will show that the following condition on is enough to guarantee a -factor in . It is convenient to define
[TABLE]
Lemma 2**.**
Let , and let be a connected graph such that for every pair of disjoint sets for which ,
[TABLE]
Then has a -factor or is -factor critical according as or .
Proof.
We first consider the case . Let . To show that has a -factor, it is sufficient to show that for all , , by (3). If , then this follows from the fact that is connected and
[TABLE]
In what remains, we verify (3) for . Suppose that for some , . Since and form a complete bipartite graph, we may delete one of the vertices of from , and the number of components of does not change. Then the right side of (3) decreases, and the left side decreases by at most 1. Hence we assume or for all .
Define the following sets of vertices of :
[TABLE]
and the following sets of vertices of :
[TABLE]
Suppose that for some , (that is, and ). But then we may delete all the vertices of from , and both the left side and the right side of (3) decreases by . Hence we assume .
For convenience, given a subgraph of , we write for the subgraph of obtained from by deleting all vertices of and contracting all the sets . For each component of containing a vertex of , either is an isolated vertex in , or is a component of . So at most k|T|+\omega\big{(}G-(S\cup T)\big{)} components of contain a vertex of . Now let be a component of containing no vertices of . Then , so the only components of containing no vertices of and contributing to are isolated vertices – and these are vertices of which are adjacent to a vertex of . So the number of these isolated vertices is:
[TABLE]
where is the number of edges between and in . It now follows using (4) that
[TABLE]
So we have shown for every set , as required. This completes the proof for .
Finally, suppose and let . We show that has a -factor. Let be the graph obtained by adding a vertex to and joining to with parallel edges. Then is connected, and . Furthermore, it is straightforward to check that (4) is satisfied in . By the first part of the proof, has a -factor. Deleting both and from this -factor of , we get a -factor in , as required.
3 Structure of the -core
In this section we describe the structure of the -core in ; this material will be used throughout the proof of Theorem 1. We will assume throughout that where , so that the -core of is a.a.s. nonempty. We let denote this nonempty -core.
In the first lemma, denotes the set of edges of with exactly one endpoint in a set . The lemma seems to be well known, and follows, for example, from Benjamini, Kozma and Wormald [3, Lemma 5.3]. (That lemma concerns graphs with a given degree sequence, all degrees between 3 and . See the proof of [3, Theorem 4.2] to find the connection with the following.)
Lemma 3**.**
There is a positive constant such that the following holds. Fix . Then a.a.s. every set of at most vertices satisfies
[TABLE]
Throughout the rest of the paper, denotes the constant appearing in Lemma 3.
Lemma 4**.**
Let . Then a.a.s. for every set of size at most , contains a component with more than vertices.
Proof.
We first show that a.a.s. there are no sets of , vertices in inducing at least edges: the expected number of subgraphs of vertices with at least edges, for some , is at most
[TABLE]
So the claim follows from Markov’s inequality. Thus, we may assume that all sets of vertices in induce at most edges. We may also assume that a.a.s. the property in Lemma 3 holds.
It suffices to show that if is the vertex set of a union of components of and , then . Suppose . From the property in Lemma 3,
[TABLE]
Suppose . By averaging, some of size satisfies
[TABLE]
However , which is a contradiction.
In fact Łuczak [11] showed that the -core is -connected a.a.s., as stated in the next lemma.
Lemma 5**.**
For and , the -core of is -connected a.a.s.
Our final lemma is a large deviation result for the degrees of the vertices of the -core. Essentially, the degree of a vertex in has (asymptotically) a truncated Poisson distribution, which gives a precise bound on the number of vertices which deviate from degree in .
Lemma 6**.**
For all there exists such that for and , it is a.a.s. true that
[TABLE]
for at most vertices of .
Proof.
Let , and fix , and . From [4, Corollary 3 and Erratum], if denotes the number of vertices of degree in , then a.a.s.
[TABLE]
where is the larger of the two positive solutions of the equation
[TABLE]
(The fact that there are two such solutions is known to be guaranteed by the fact that .) Let , and suppose that . Then, since the Poisson distribution is asymptotically normal with variance equal to its mean, we have for sufficiently large
[TABLE]
Also, by Lemma 1, we may assume that . Suppose that is substituted for in (6). It is then elementary to obtain that the right hand side of (6) is greater than . Recalling also that , this is greater than the left hand side of (6). On the other hand, if anthing larger than is substituted for in (6) then the left hand side is greater than the right, since the right is equal to a probability strictly less then 1. So by continuity, . Taking slightly smaller than , the lemma follows from (7) and (5).
4 Proof of Theorem 1
In this section, we denote by the -core of where and , for . To prove Theorem 1, we show that there exists such that for , (4) holds in a.a.s. The value of will not be optimized in the proof to follow. To prove the theorem, we consider a number of cases according to the sizes of the sets and in the lemma, where . It is convenient throughout to let .
**Case 1 ** .
Let . By Lemma 4, contains a component with more than vertices a.a.s. Let denote the number of components of of size less than , and let be the set of vertices in these components. As in the proof of Lemma 4, and a.a.s. However, every vertex of has degree at least , by definition of , so a.a.s. since holds a.a.s. It follows that
[TABLE]
Therefore, since ,
[TABLE]
Since , and \omega\big{(}G-(S\cup T)\big{)}\geq 1,
[TABLE]
If or , then this is less than , as required. If and , then a.a.s. (see Lemma 5), which implies
[TABLE]
Therefore (4) holds a.a.s. in , so has a -factor or is -factor critical, by Lemma 2.
For the rest of the proof, is an absolute constant; we will take for definiteness.
**Case 2 ** , , and .
Let . In this case we estimate and separately. First we show that a.a.s. provided that is large enough to ensure that (so this tells us is required in our proof). It suffices to show that if is the vertex set of any union of components of , then a.a.s. or . Suppose that . Then Lemma 3 shows . Let
[TABLE]
The expected number of pairs of sets in satisfying the above requirements is
[TABLE]
where the sums are over . So in fact the expected number of sets and as described above is . By Markov’s Inequality, we conclude that a.a.s.
To finish verifying inequality (4) in case 2, it remains to show that a.a.s.
[TABLE]
Let and and let (here and are allowed to depend on ). Then the number of ways of choosing the sets and is bounded above by
[TABLE]
The probability that there are at least edges between and is at most
[TABLE]
Multiplying by the bound on the number of choices of and , the bound (8) is true a.a.s. if
[TABLE]
for which it suffices that
[TABLE]
Since and for large enough ,
[TABLE]
To prove that the expression on the left in (9) is less than , we can assume
[TABLE]
It follows that and therefore, since ,
[TABLE]
**Case 3 ** and and .
To prove that the requirements of Lemma 2 are satisfied a.a.s. for sufficiently large, it is enough to show that a.a.s. for every pair of sets under consideration, . This is because the sum of vertex degrees in is at least , and also \omega\big{(}G-(S\cup T)\big{)}\leq n<\frac{1}{4}k|S| for large enough (as is fixed). We may assume that , and so if is sufficiently large, Lemma 1 shows that . Thus it suffices to show , since this is at most . This follows immediately once we show that a.a.s. all sets of at most vertices (in particular, ) have total degree at most .
It is well known that in the random graph , the vertex degrees are have asymptotically Poisson distribution with mean : the number of vertices of degree is a.a.s. asymptotic to . It follows that the sum of the degrees of those vertices of degree less than is a.a.s. asymptotic to
[TABLE]
Since the Poisson distribution is asymptotically normal with mean , for large enough (i.e. large enough ) we have
[TABLE]
Since the sum of vertex degrees is a.a.s. asymptotic to , the ones of degree at least a.a.s. have total degree less than . Assuming this is true, there are at most such vertices, and any set of at most vertices thus has total degree at most
[TABLE]
as required.
**Case 4 ** .
By Lemmas 1 and 6, for all , if is sufficiently large, then a.a.s.
[TABLE]
So by Lemma 2 and the fact that , it is enough to show, for some , that a.a.s.
[TABLE]
We will prove this by considering the cases and separately, where . For , we will use
[TABLE]
For this, we may assume that if then contains the vertices of largest degree in (and that they have the same degrees in ). Using the argument about the degrees of as in Case 3, it is straightforward to show that a.a.s. these vertices have total degree at most
[TABLE]
for sufficiently large, which is at most
[TABLE]
since . This gives (10).
It only remains to treat those sets for which . Then using the same argument as with in Case 3, the sum of degrees of vertices in is a.a.s. at most \big{(}k+(1+\eta)\sqrt{k\log k}\big{)}|S|.
For a set of this size in , the expected value of is
[TABLE]
since . Moreover, is binomially distributed. So by Chernoff’s inequality (see for example [8, Theorem 2.1]),
[TABLE]
for sufficiently large (recall that is an absolute constant). Hence, a.a.s. every set that is this large induces a subgraph of at least edges. Provided , it contains exactly the same number of edges in as in . Hence we have that a.a.s. for all such and ,
[TABLE]
for large enough . This gives (10), as required.
5 Proof of Lemma 1
A weakened version of the main result in Pittel, Spencer and Wormald [12] is that if is fixed, a.a.s. has no -core if , a.a.s. has one if , where is defined in (1). A little calculation shows that and ( is also defined in (1)) satisfy
[TABLE]
with defined in (2). Substituting (12) and (2) into (11) gives
[TABLE]
(where square brackets denote falling factorials) and so, multiplying by , we obtain
[TABLE]
Since the right hand side is an increasing function of , the value of is uniquely determined. Moreover, since is exactly the ratio of the th to the th term in the summation, the largest term in the summation occurs for and from elementary considerations it is easy to see that . Thus, putting
[TABLE]
we know that . In addition, rewriting (2) as
[TABLE]
we now see that , and hence also .
To get a slightly better bound on straight away, substitute (12) into (11), use Stirling’s formula with its correction term due to Robbins: , and take logarithms to give
[TABLE]
Recalling from above that and , we may expand to show that
[TABLE]
where .
Taking out a factor of from the terms in the summation in (14), using (12) we obtain
[TABLE]
The terms in the summation are monotonically decreasing. Since , we see that, for any , the terms for sum to . For , we see after expanding that the product over is
[TABLE]
Putting and recalling , we now have
[TABLE]
To estimate the first error term we approximate the summation by an integral, so that term becomes
[TABLE]
using (16). The other error term is similarly . The main term in the summation is a truncated geometric series with the truncated terms negligible, so we have
[TABLE]
Using this with (11) shows that
[TABLE]
So we may continue with
[TABLE]
Hence
[TABLE]
Next, substitute this in the left side of (15), and into the right side. We know that from (16), and we may expand as . The upshot is that
[TABLE]
This determines , and since , we have from (13) that . Now using (17) and the formula for immediately above (which in particular gives ), we obtain Lemma 1.
The reference list from the paper itself. Each links out to its DOI / PubMed record.
- 1[1] B. Bollobás, C. Cooper, T. Fenner, and A. Frieze, Edge disjoint Hamilton cycles in sparse random graphs of minimum degree at least k 𝑘 k . J. Graph Theory 34 (2000), no. 1, 42–59.
- 2[2] B. Bollobás, J. H. Kim, and J. Verstraëte, Regular subgraphs of random graphs, Random Structures & Algorithms , 29 (2006), 1-13.
- 3[3] I. Benjamini, G. Kozma, and N. Wormald, The mixing time of the giant component of a random graph, Preprint .
- 4[4] J. Cain and N. Wormald, Encores on cores, Electronic Journal of Combinatorics 13 (2006), RP 81.
- 5[5] P. Flajolet, D. Knuth, and B. Pittel, The first cycles in an evolving graph. Graph theory and combinatorics (Cambridge, 1988). Discrete Math. 75 (1989), no. 1-3, 167–215.
- 6[6] S. Janson, Cycles and unicyclic components in random graphs. Combin. Probab. Comput. 12 (2003), no. 1, 27–52.
- 7[7] S. Janson and M. Luczak, A simple solution to the k 𝑘 k -core problem. Tech. Report 2005:31, Uppsala.
- 8[8] S. Janson, T. Łuczak, and A. Ruciński, Random Graphs , Wiley, New York, 2000.
