Variance of random greedy independent sets in triangle-free graphs

Abstract

Inspect the vertices of a finite simple graph in uniformly random order, accepting each vertex if none of its neighbors has previously been accepted. Let $X_{G}$ be the number of accepted vertices. For every triangle-free graph with $n\geq 2$ vertices and $e$ edges, we prove

$$\operatorname{Var}(X_{G})\leq e\left(\frac{n-2}{n}\right)^{2},$$

with equality precisely for edgeless graphs and connected stars. In particular, among trees of a given order, the star uniquely maximizes the variance, with value $(n-1)(n-2)^{2}/n^{2}$. Known expected vertex-deletion stability already yields the elementary baseline $\operatorname{Var}(X_{G})\leq e$. We obtain the sharp finite-order refinement by combining a stronger centered first-choice estimate with a triangle-free edge-count identity in the law of total variance.

1 Introduction

Random-order greedy acceptance produces a maximal independent set, but generally not a maximum independent set. Its cardinality is random even when the underlying graph is fixed. Equivalently, assign independent uniform arrival times in $[0,1]$ to the vertices and occupy a vertex on arrival if no neighbor is already occupied. This is terminal blocking random sequential adsorption, as studied by Penrose and Sudbury [6].

We consider the variance of the accepted cardinality. Throughout, graphs are finite, simple, and undirected; they need not be connected. Let $X_{G}$ denote the accepted count, and write $e(G)=|E(G)|$.

Theorem 1.

If $G$ is triangle-free and $n=|V(G)|\geq 2$, then

$$\operatorname{Var}(X_{G})\leq e(G)\left(\frac{n-2}{n}\right)^{2}.$$ (1)

Equality holds if and only if $G$ is edgeless or $G\cong K_{1,n-1}$.

Corollary 2.

For every nonempty tree $T$ of order $n$,

$$\operatorname{Var}(X_{T})\leq\frac{(n-1)(n-2)^{2}}{n^{2}},$$

with equality if and only if $T\cong K_{1,n-1}$. The singleton is included as $K_{1,0}$.

Gamarnik and Goldberg [3, Theorem 6, equation (9)] give $\operatorname{Var}(X_{G})\leq C(r)n$ under a maximum-degree bound $r\geq 3$, without a girth restriction. Penrose and Sudbury [6, Section 2.7, discussion following Theorem 7] also note a linear variance upper bound in the uniformly bounded-degree setting, alongside their correlation and normal-approximation results. Krivelevich, Mészáros, Michaeli, and Shikhelman [4, Theorem 1.2] prove concentration under subfactorial path growth and local convergence. Their Theorem 1.4 makes the path mean-minimal among trees, while Section 8 makes the star uniquely mean-maximal. Corollary 2 gives the variance analogue of the latter extremal statement; a bound on the mean alone does not control the variance. In terms of average degree $\bar{d}=2e(G)/n$, (1) reads

$$\operatorname{Var}(X_{G}/n)\leq\frac{\bar{d}}{2n}\left(1-\frac{2}{n}\right)^{2}.$$

Our bound imposes no restriction on the maximum degree.

The degree-free baseline $\operatorname{Var}(X_{G})\leq e(G)$ is already an elementary consequence of the expected deletion stability of Censor-Hillel, Haramaty, and Karnin [1], together with first-choice conditioning and a triangle-free edge count; we give this deduction in Proposition 6. Related dynamic-MIS and sensitivity methods appear in [2, 8]. Thus the point of Theorem 1 is the sharp finite-order coefficient and equality classification, not the existence of a linear-in-edge-count variance bound. The refinement uses a stronger centered first-choice estimate. We include a proof of the needed deletion-stability statement for self-containment.

Sudbury [7, Theorem 1] compares single-vertex occupation probabilities on regular and branching-random trees. For a uniformly random labelled tree, Panholzer [5, Theorem 4] obtains an exact accepted-count distribution, variance $n/16+O(1)$, and a central limit theorem. These are averaged over the random tree as well as the arrival order; they do not by themselves give a variance bound for every fixed tree.

2 Deletion stability and first-choice means

Write $I(G,\pi)$ for the independent set obtained from an order $\pi$, and put

$$\mu(G)=\mathbb{E}X_{G},\qquad v(G)=\operatorname{Var}(X_{G}),\qquad\mu(\varnothing)=v(\varnothing)=0.$$

When comparing induced subgraphs, use the restrictions of the same uniform permutation. Let $N_{G}(x)$ and $N_{G}[x]$ be the open and closed neighborhoods, and let $d_{G}(x)=|N_{G}(x)|$. For a nonempty graph set

$$R_{u}=G-N_{G}[u],\qquad a_{u}(G)=1+\mu(R_{u}),\qquad b_{u}(G)=\mu(G)-a_{u}(G).$$

Conditional on $u$ being first, the remaining accepted count has the law of $X_{R_{u}}$. Therefore, with $n=|V(G)|$,

$$\mu(G)=\frac{1}{n}\sum_{u}a_{u}(G),\qquad\sum_{u}b_{u}(G)=0.$$ (2)

Define $c_{0}=c_{1}=0$ and $c_{n}=(n-2)/n$ for $n\geq 2$. Then $0\leq c_{n}\leq 1$, and $c_{n}$ is strictly increasing for $n\geq 2$.

Lemma 3 (Deletion stability, [1, Theorem 1]).

For any graph $G$ and fixed vertex $x\in V(G)$,

$$\mathbb{E}|I(G,\pi)\triangle I(G-x,\pi)|\leq 1.$$ (3)

Consequently $|\mu(G)-\mu(G-x)|\leq 1$, and for subsets $A,B\subseteq V(G)$,

$$|\mu(G[A])-\mu(G[B])|\leq|A\triangle B|.$$ (4)

Proof.

Induct on $n=|V(G)|$, the singleton case being immediate. Fix $x$, write $d=d_{G}(x)$, and condition on the first vertex $u$. The remaining relative order is uniform.

If $u\in N_{G}(x)$, both runs accept $u$ and thereafter reduce to the same graph $R_{u}$, so their discrepancy is zero. If $u\notin N_{G}[x]$, both runs accept $u$ and the discrepancy equals that for deleting $x$ from $R_{u}$. Its conditional expectation is at most one by induction.

If $u=x$, the full run gives $\{x\}\cup I(R_{x},\pi)$, whereas the deleted run gives $I(G-x,\pi)$. Delete the $d$ neighbors of $x$ successively from $G-x$, in any fixed order. The triangle inequality bounds the symmetric difference by one plus the sum of these $d$ deletion discrepancies. Each intermediate graph is smaller than $G$ and has a uniform restricted order, so the conditional expectation is at most $1+d$. Averaging gives

$$\mathbb{E}|I(G,\pi)\triangle I(G-x,\pi)|\leq\frac{d\cdot 0+(n-d-1)\cdot 1+(d+1)}{n}=1.$$

The mean assertion follows by bounding cardinality difference by symmetric difference. Telescoping deletions from $A$ and $B$ to $A\cap B$ proves (4). ∎

Deleting the open neighborhood $N_{G}(x)$ leaves $x$ isolated. Hence, for every graph $G$ and vertex $x$, (4) gives the elementary centered estimate

$$|b_{x}(G)|=|\mu(G)-\mu(G-N_{G}(x))|\leq d_{G}(x),\qquad\mu(G-N_{G}(x))=1+\mu(R_{x}).$$ (5)

The next lemma sharpens this estimate for triangle-free graphs.

Lemma 4 (Centered first-choice estimate).

For every triangle-free graph $G$ of order $n\geq 2$ and every $x\in V(G)$,

$$|b_{x}(G)|\leq c_{n}d_{G}(x).$$ (6)

For a singleton, $b_{x}(G)=0$. Hence, for a uniform vertex $u$,

$$\operatorname{Var}_{u}(a_{u}(G))=\frac{1}{n}\sum_{u}b_{u}(G)^{2}\leq\frac{c_{n}^{2}}{n}\sum_{u}d_{G}(u)^{2}.$$ (7)

Proof.

The singleton case is immediate; assume $n\geq 2$.

Fix $x$, put $d=d_{G}(x)$, and let $U=V(G)\setminus N_{G}[x]$. We first establish

$$nb_{x}(G)=\sum_{u\in N_{G}(x)}\bigl(\mu(R_{u})-\mu(R_{x})\bigr)+\sum_{u\in U}b_{x}(R_{u}).$$ (8)

For $u\in U$, deleting the two closed neighborhoods commutes:

$$R_{u}-N_{R_{u}}[x]=G-(N_{G}[u]\cup N_{G}[x])=R_{x}-N_{R_{x}}[u].$$

Thus $\mu(R_{u})=1+\mu(R_{x}-N_{R_{x}}[u])+b_{x}(R_{u})$. Sum over $u\in U=V(R_{x})$ and apply (2) to $R_{x}$ to obtain

$$\sum_{u\in U}\mu(R_{u})=|U|\mu(R_{x})+\sum_{u\in U}b_{x}(R_{u}).$$

When $U=\varnothing$ both sides are zero. Substitution into $n\mu(G)=n+\sum_{u}\mu(R_{u})$ gives (8).

If $u$ is adjacent to $x$, triangle-freeness gives $N_{G}[u]\cap N_{G}[x]=\{u,x\}$. By (4),

$$|\mu(R_{u})-\mu(R_{x})|\leq d_{G}(u)+d-2.$$

For $u\in U$, apply (5) to $R_{u}$ to get $|b_{x}(R_{u})|\leq d_{R_{u}}(x)$. A count of ordered pairs $(u,w)$, where $w\in N_{G}(x)$ is removed by deleting $N_{G}[u]$, gives

$$\sum_{u\in U}d_{R_{u}}(x)=(n-d-1)d-\sum_{w\in N_{G}(x)}(d_{G}(w)-1).$$ (9)

Indeed, for each such $w$, every vertex of $N_{G}(w)\setminus\{x\}$ lies in $U$, because there are no triangles. The count allows different common neighbors, so it does not exclude four-cycles.

Taking absolute values in (8) now gives

$$n|b_{x}(G)| \leq\sum_{u\in N_{G}(x)}(d_{G}(u)+d-2)+(n-d-1)d-\sum_{w\in N_{G}(x)}(d_{G}(w)-1)$$

$$=d(n-2).$$

This proves (6), also for $n=2$ and for isolated $x$. Equation (7) follows from (2). ∎

3 Variance and equality

Lemma 5 (Edge loss).

For a nonempty triangle-free graph $G$ of order $n$,

$$\frac{1}{n}\sum_{u}e(R_{u})+\frac{1}{n}\sum_{u}d_{G}(u)^{2}=e(G).$$ (10)

Proof.

Every edge removed by deleting $N_{G}[u]$ is incident to a neighbor of $u$. No edge joins two such neighbors, so

$$e(G)-e(R_{u})=\sum_{w\in N_{G}(u)}d_{G}(w).$$

Summing over $u$ proves (10). ∎

Proposition 6 (Elementary edge-count bound).

For every finite triangle-free graph $G$, $v(G)\leq e(G)$.

Proof.

Induct on the order, with $v(\varnothing)=e(\varnothing)=0$. For a nonempty graph, condition on its first vertex. Using induction on the smaller residual graphs, (5), and Lemma 5,

$$v(G)=\frac{1}{n}\sum_{u}v(R_{u})+\frac{1}{n}\sum_{u}b_{u}(G)^{2}\leq\frac{1}{n}\sum_{u}e(R_{u})+\frac{1}{n}\sum_{u}d_{G}(u)^{2}=e(G).$$

This argument uses the known deletion estimate, not Lemma 4 or Theorem 1. ∎

Proof of Theorem 1.

We first prove the bound by induction on the order over all triangle-free graphs. Counts are deterministic for orders zero, one, and two. For $n\geq 3$, the law of total variance, conditional on the first vertex, gives

$$v(G)=\frac{1}{n}\sum_{u}v(R_{u})+\frac{1}{n}\sum_{u}b_{u}(G)^{2}.$$ (11)

The first term is the expected conditional variance; the second is the variance of the conditional mean. With $s_{u}=|V(R_{u})|<n$, induction and monotonicity give $v(R_{u})\leq c_{s_{u}}^{2}e(R_{u})\leq c_{n}^{2}e(R_{u})$, including empty residual graphs. Applying (7) and (10) to (11) yields

$$v(G)\leq\frac{c_{n}^{2}}{n}\left(\sum_{u}e(R_{u})+\sum_{u}d_{G}(u)^{2}\right)=c_{n}^{2}e(G).$$

For equality, edgeless graphs attain the zero bound. If $e(G)>0$ and $n=2$, then $G=K_{1,1}$. Suppose $n\geq 3$. If some $R_{u}$ has an edge, then $2\leq s_{u}<n$ and $c_{s_{u}}<c_{n}$, making the corresponding induction estimate strict. Thus equality requires every $R_{u}$ to be edgeless.

Choose an edge $xy$. Every other vertex must be adjacent to at least one of $x,y$, since otherwise $xy$ survives in its residual graph; triangle-freeness prevents adjacency to both. Therefore $A=N_{G}(y)$ and $B=N_{G}(x)$ partition $V(G)$ into independent sets, with $x\in A$ and $y\in B$. If $z\in A\setminus\{x\}$ and $w\in B$ were nonadjacent, then the edge $xw$ would survive in $R_{z}$, a contradiction. Hence $G=K_{a,b}$ for positive integers $a=|A|$, $b=|B|$.

The first vertex of $K_{a,b}$ determines the accepted part. Its count is $a$ with probability $a/(a+b)$ and $b$ with probability $b/(a+b)$, so

$$v(K_{a,b})=\frac{ab(a-b)^{2}}{(a+b)^{2}},\qquad e(K_{a,b})=ab.$$

As $a+b=n$, equality in (1) is equivalent to $|a-b|=n-2$, or $\min(a,b)=1$. These are exactly the connected stars, and the same computation proves that they attain equality. ∎

Proof of Corollary 2.

A tree of order $n\geq 2$ is triangle-free and has $n-1$ edges, so apply Theorem 1. For $n=1$, the count is identically one and the bound is zero. ∎

4 Scope of the bound

Theorem 1 is sharp uniformly and gives the exact maximum over trees of each order. It does not give an attained maximum for every prescribed pair $(n,e)$. Nor does it assert a central limit theorem.

Relative to Proposition 6, the finite-order improvement is $4e(G)(n-1)/n^{2}$. For a tree it is

$$(n-1)-\frac{(n-1)(n-2)^{2}}{n^{2}}=4\left(1-\frac{1}{n}\right)^{2},$$

which tends to four. The coefficient one in the baseline cannot be decreased uniformly in $n$: the explicit star law gives $v(K_{1,n-1})/e(K_{1,n-1})=(1-2/n)^{2}\to 1$. The refinement supplies the exact finite-order tree maximum and its equality cases.

Triangle-freeness is used in both (9) and (10). If $\tau(G)$ is the number of triangles in an arbitrary graph, the latter identity becomes

$$\frac{1}{n}\sum_{u}e(R_{u})+\frac{1}{n}\sum_{u}d_{G}(u)^{2}=e(G)+\frac{3\tau(G)}{n}.$$

Indeed, the removed-edge count is $\sum_{w\in N_{G}(u)}d_{G}(w)-e(G[N_{G}(u)])$, and $\sum_{u}e(G[N_{G}(u)])=3\tau(G)$. The argument above does not settle whether (1) continues to hold when triangles are present.