arXiv reader

A 2/3 Bound for Vizing’s Conjecture

Mohsen Aliabadi and Elliot Krop

arXiv
2607.01109v3 · PDF · 9 September 2026
Categories
math.CO
Licence
CC BY 4.0 The site shows the text.

Abstract

Vizing’s conjecture, dating back to 1963, asserts that

\gamma(G\mathbin{\square}H) \geq \gamma(G)\gamma(H)

for all finite graphs G and H, where \gamma denotes the domination number and \square denotes the Cartesian product. In 2000, Clark and Suen proved the universal bound

\gamma(G\mathbin{\square}H) \geq \frac{1}{2}\gamma(G)\gamma(H).

Recently, Steiner obtained the first constant-factor improvement of the Clark–Suen bound, proving that

\gamma(G\mathbin{\square}H) \geq \frac{5+\sqrt{73}}{24}\gamma(G)\gamma(H) \approx 0.5643\,\gamma(G)\gamma(H).

In this paper, we further improve the universal constant by proving that

\gamma(G\mathbin{\square}H) \geq \frac{2}{3}\gamma(G)\gamma(H)

for all finite graphs G and H. Thus, we raise the best known universal constant in the approximate form of Vizing’s conjecture from (5+\sqrt{73})/24 to 2/3.

Items 35

Theorem

Theorem 1.1

QU3H

Let G and H be graphs, let P\subseteqV(H) be a 2-packing of H, and put

r:=|P|,\qquad W:=V(H)\setminusN_H[P],\qquad s:=\gamma_H(W).

Then

\gamma(G\mathbin{\square}H) \geq \frac{r\gamma(G)}{3} +\frac{(r+s)(2\gamma(G)-\rho(G))}{3}.
Proof (Proof of Theorem 1.1) of Theorem 1.1

Adopt Setting 3.1 and write \gamma=\gamma(G) and \rho=\rho(G). By (3.7) and Lemma 3.4,

s\gamma\leq w+\frac{b+w}{2}+\frac{s\rho}{2},

and hence

w\geq\frac23s\left(\gamma-\frac\rho2\right)-\frac b3.

Therefore

|D|=a+b+w \geq a+\frac23b+\frac23s\left(\gamma-\frac\rho2\right).

By Lemma 3.5 and (3.6),

\begin{aligned} a+\frac23b &=\frac13(a+b)+\frac23\left(a+\frac b2\right)\\ &\geq\frac13r\gamma +\frac23r\left(\gamma-\frac\rho2\right). \end{aligned}

Combining (Proof) and (Proof) yields

\gamma(G\mathbin{\square}H)=|D| \geq\frac{r\gamma}{3} +\frac23(r+s)\left(\gamma-\frac\rho2\right) =\frac{r\gamma}{3}+\frac{(r+s)(2\gamma-\rho)}3.
Corollary

Corollary 1.2

9DXE

For all graphs G and H,

\gamma(G\mathbin{\square}H) \geq \frac{2\gamma(G)\gamma(H)+\rho(H)\gamma(G)-\rho(G)\gamma(H)}{3},

and hence, by symmetry,

\gamma(G\mathbin{\square}H) \geq \frac{2\gamma(G)\gamma(H) +|\rho(H)\gamma(G)-\rho(G)\gamma(H)|}{3}.
Used by Proof, Remark 5.1
Proof (Proof of Corollary 1.2) of Corollary 1.2

Apply Theorem 1.1 to a maximum 2-packing P of H. Then r=\rho(H) and, as observed above, r+s\geq\gamma(H). By Lemma 2.1, 2\gamma(G)-\rho(G)\geq\gamma(G)>0, so (1.1) gives

\begin{aligned} \gamma(G\mathbin{\square}H) &\geq\frac{\rho(H)\gamma(G)}3 +\frac{\gamma(H)(2\gamma(G)-\rho(G))}3\\ &=\frac{2\gamma(G)\gamma(H)+\rho(H)\gamma(G)-\rho(G)\gamma(H)}3, \end{aligned}

which is (1.2). Since G\mathbin{\square}H\cong H\mathbin{\square}G, the same inequality holds with the roles of G and H interchanged. Taking the larger of the two bounds gives (1.2).

Theorem

Theorem 1.3

J5DW

For all graphs G and H,

\gamma(G\mathbin{\square}H)\geq \frac23\gamma(G)\gamma(H).
Proof (Proof of Theorem 1.3) of Theorem 1.3

This follows immediately from (1.2), since the absolute-value term is nonnegative.

Theorem

Theorem 1.4

BMH1

In the setting of Theorem 1.1,

\gamma(G\mathbin{\square}H) \geq r\gamma(G)-\frac{r\rho^{\{2\}}(G)}8 +\frac{s(3\gamma(G)-\rho^{\{2\}}(G))}{4}.

Consequently, for all graphs G and H,

\gamma(G\mathbin{\square}H) \geq \frac{2(3\gamma(G)-\rho^{\{2\}}(G))\gamma(H) +\rho(H)(2\gamma(G)+\rho^{\{2\}}(G))}{8}.
Proof (Proof of Theorem 1.4) of Theorem 1.4

Write \gamma=\gamma(G) and \rho^{\{2\}}=\rho^{\{2\}}(G). By (3.7) and Lemma 3.4,

s\gamma\leq w+\frac{b+w}{3}+\frac{s\rho^{\{2\}}}{3},

so

w\geq\frac34s\left(\gamma-\frac{\rho^{\{2\}}}{3}\right)-\frac b4.

Thus

|D|\geq a+\frac34b +\frac34s\left(\gamma-\frac{\rho^{\{2\}}}{3}\right).

By Lemma 3.5 and (3.6),

\begin{aligned} a+\frac34b &=\frac58(a+b)+\frac38\left(a+\frac b3\right)\\ &\geq\frac58r\gamma +\frac38r\left(\gamma-\frac{\rho^{\{2\}}}{3}\right) =r\gamma-\frac{r\rho^{\{2\}}}{8}. \end{aligned}

This proves (1.4).

For (1.4), let P be a maximum 2-packing of H, so r=\rho(H). Since P dominates N_H[P] and Q dominates W, the set P\cup Q dominates H. Hence r+s\geq\gamma(H) and s\geq\gamma(H)-\rho(H). By Lemma 2.1, 3\gamma-\rho^{\{2\}}\geq\gamma>0, so substituting the lower bound on s into (1.4) gives

\begin{aligned} \gamma(G\mathbin{\square}H) &\geq \rho(H)\gamma-\frac{\rho(H)\rho^{\{2\}}}{8} +\frac{(\gamma(H)-\rho(H))(3\gamma-\rho^{\{2\}})}4\\ &=\frac{2(3\gamma-\rho^{\{2\}})\gamma(H) +\rho(H)(2\gamma+\rho^{\{2\}})}8. \end{aligned}
Lemma

Lemma 2.1

FHIG

For every graph G,

\rho(G)\leq\gamma(G) \qquad\text{and}\qquad \rho^{\{2\}}(G)\leq2\gamma(G).
Used by Proof, Proof
Proof of Lemma 2.1

Let D be a minimum dominating set of G. If P is a 2-packing, each d\in D dominates at most one vertex of P, whereas D dominates all of P; hence |P|\leq|D|. If f is a 2-packing function, every vertex lies in N_G[d] for some d\in D, so, as f\geq0,

f(V(G))\leq\sum_{d\in D}f(N_G[d])\leq2|D|.
Lemma

Lemma 2.2 (Brešar; Steiner)

7HDF

For every graph G and every S\subseteqV(G),

\gamma_G(S)\leq\frac{|S|+\rho_G(S)}2 \leq\frac{|S|+\rho(G)}2.
Proof of Lemma 2.2

Let B be the graph on vertex set S in which distinct s_1,s_2\in S are adjacent if and only if N_G[s_1]\capN_G[s_2]\neq\varnothing. Let \{s_1s'_1,\ldots,s_ts'_t\} be a maximum matching of B, and choose d_i\inN_G[s_i]\capN_G[s'_i] for each i. The set

\bigl(S\setminus\{s_1,s'_1,\ldots,s_t,s'_t\}\bigr) \cup\{d_1,\ldots,d_t\}

dominates S, so \gamma_G(S)\leq|S|-t. The unmatched vertices form an independent set in B by maximality of the matching, and hence a 2-packing of G contained in S. Thus |S|-2t\leq\rho_G(S). Combining these inequalities gives \gamma_G(S)\leq|S|-t\leq(|S|+\rho_G(S))/2.

Lemma

Lemma 2.3 (Steiner)

2W7W

For every graph G and every S\subseteqV(G),

\gamma_G(S)\leq\frac{|S|+\rho^{\{2\}}(G)}3.
Used by Proof, Proof
Lemma

Lemma 2.4

1RFE

For every h\inV(H), the set

P_G\bigl(D\cap(V(G)\timesN_H[h])\bigr)

dominates G. In particular, |D\cap(V(G)\timesN_H[h])|\geq\gamma(G).

Used by Proof, Remark 6.2
Proof of Lemma 2.4

Let g\inV(G). By (?), the vertex (g,h) is dominated by some (g',h')\in D with either h'=h and g'\inN_G[g], or g'=g and h'\inN_H(h). In both cases h'\inN_H[h] and g'\inN_G[g].

Setting

Setting 3.1

O4RL

Let G and H be graphs and let P=\{h_1,\ldots,h_r\} be a 2-packing of H, possibly empty. Put

W:=V(H)\setminusN_H[P],\qquad s:=\gamma_H(W),

and let Q=\{h_{r+1},\ldots,h_{r+s}\} be a minimum set of vertices of H dominating W. Define a partition V(H)=\pi_1\mathbin{\dot\cup}\cdots\mathbin{\dot\cup}\pi_{r+s} as follows.

  1. For i\leq r, let \pi_i:=N_H[h_i]. These sets are pairwise disjoint because P is a 2-packing, and their union is N_H[P].

  2. For i>r, choose \pi_i\subseteqN_H[h_i]\cap W so that \pi_{r+1},\ldots,\pi_{r+s} partition W. Such sets exist because Q dominates W: assign each w\in W to one index i>r for which w\inN_H[h_i].

The parts \pi_1,\ldots,\pi_r are called packing parts, and the remaining parts are called leftover parts. If s=0, then W is empty and there are no leftover parts; if r=0, then W=V(H) and Q is a minimum dominating set of H.

For i\in[r+s], let D_i:=D\cap(V(G)\times\pi_i), and for g\inV(G) let C_g^i:=\{g\}\times\pi_i. The cell C_g^i is vertically dominated if every vertex of C_g^i has a neighbor in

D\cap\bigl(\{g\}\times(V(H)\setminus\pi_i)\bigr),

and put

L_i:=\{g\inV(G):C_g^i\text{ is vertically dominated}\}.

Finally, split D according to layers:

a:=|D\cap(V(G)\times P)|, \quad b:=|D\cap(V(G)\times(N_H[P]\setminus P))|, \quad w:=|D\cap(V(G)\times W)|.

Thus

|D|=a+b+w,\qquad \sum_{i\leq r}|D_i|=a+b,\qquad \sum_{i>r}|D_i|=w.
Lemma

Lemma 3.2

S9VO

For every i\in[r+s], the set P_G(D_i) dominates V(G)\setminus L_i. Consequently,

\gamma_G(V(G)\setminus L_i)\leq|D_i|.
Uses Setting 3.1 · Used by Proof, Proof
Proof of Lemma 3.2

Let g\inV(G)\setminus L_i. Since C_g^i is not vertically dominated, some (g,h)\in C_g^i has no neighbor in D\cap(\{g\}\times(V(H)\setminus\pi_i)). The vertex (g,h) is dominated by some (g',h')\in D. By (?), either g'=g and h'\inN_H[h], in which case h'\in\pi_i by the choice of h, or h'=h\in\pi_i and g'\inN_G[g]. In both cases (g',h')\in D_i and g\inN_G[g'].

Lemma

Lemma 3.3

YL9V

For every i\leq r, we have L_i=\varnothing, and hence \gamma(G)\leq|D_i|.

Proof of Lemma 3.3

Let i\leq r and g\inV(G). By (?), the neighbors of (g,h_i) in the H-fiber at g are the vertices (g,h') with h'\inN_H(h_i)\subseteqN_H[h_i]=\pi_i. Thus (g,h_i)\in C_g^i has no neighbor in \{g\}\times(V(H)\setminus\pi_i), so C_g^i is not vertically dominated. The second assertion follows from Lemma 3.2.

Lemma

Lemma 3.4

BVHK
\sum_{i=r+1}^{r+s}|L_i|\leq |D|-a=b+w.
Proof of Lemma 3.4

For g\inV(G), let m_g be the number of indices i>r with g\in L_i. Then \sum_{i>r}|L_i|=\sum_gm_g, and it is enough to show

m_g\leq |D_g|-|D_g\cap(\{g\}\times P)| \qquad\text{for every }g\inV(G).

Fix g and put X:=P_H(D_g)\setminus P. We claim that

Y:=X\cup\{h_i:i>r,\ g\notin L_i\}

dominates W in H. Let u\in W, and let i>r be the index with u\in\pi_i. If g\notin L_i, then h_i\in Y dominates u because \pi_i\subseteqN_H[h_i]. If g\in L_i, then the cell C_g^i is vertically dominated, so (g,u) has a neighbor (g,h')\in D with h'\notin\pi_i. By (?), h'\inN_H(u). Since u\in W=V(H)\setminusN_H[P], no neighbor of u lies in P. Hence h'\notin P, so h'\in X\subseteq Y and h' dominates u.

Since s=\gamma_H(W),

s\leq |Y|\leq |X|+(s-m_g),

and therefore m_g\leq|X|. Finally, P_H is injective on D_g, so

|X|=|D_g|-|D_g\cap(\{g\}\times P)|,

which proves (Proof). Summing over g gives the result.

Lemma

Lemma 3.6 (Second count on the packing blocks)

S2JP
\begin{aligned} a+\frac b2&\geq r\left(\gamma(G)-\frac{\rho(G)}2\right), \\ a+\frac b3&\geq r\left(\gamma(G)-\frac{\rho^{\{2\}}(G)}3\right). \end{aligned}
Proof of Lemma 3.6

Fix i\leq r, write p=h_i, and let

A_p:=\{g\inV(G):D\cap(\{g\}\timesN_H(p))\neq\varnothing\}.

If g\notin A_p, then by (?) the vertex (g,p) is dominated by some (g',p)\in D with g'\inN_G[g]. Hence P_G(D\cap(V(G)\times\{p\})) dominates V(G)\setminus A_p, and subadditivity gives

\gamma(G) \leq \gamma_G(V(G)\setminus A_p)+\gamma_G(A_p) \leq |D\cap(V(G)\times\{p\})|+\gamma_G(A_p).

Moreover, |A_p|\leq|D\cap(V(G)\timesN_H(p))|, since distinct elements of A_p are witnessed by distinct vertices of D. Lemma 2.2 therefore gives

\gamma(G) \leq |D\cap(V(G)\times\{p\})| +\frac{|D\cap(V(G)\timesN_H(p))|+\rho(G)}2.

Using Lemma 2.3 instead gives the analogous inequality with denominator 3 and \rho^{\{2\}}(G) in the numerator.

Now sum over i\leq r. The singletons \{h_i\} are disjoint with union P, while the open neighborhoods N_H(h_i) are pairwise disjoint, disjoint from P, and have union N_H[P]\setminus P. The two summed inequalities are exactly (3.6) and (3.6).

Lemma

Lemma 3.7 (Leftover parts)

9CIJ
\begin{aligned} s\gamma(G)&\leq w+\frac12\sum_{i>r}|L_i|+\frac{s\rho(G)}2, \\ s\gamma(G)&\leq w+\frac13\sum_{i>r}|L_i|+\frac{s\rho^{\{2\}}(G)}3. \end{aligned}
Proof of Lemma 3.7

For i>r, subadditivity, Lemma 3.2, and Lemma 2.2 applied to S=L_i give

\gamma(G) \leq\gamma_G(V(G)\setminus L_i)+\gamma_G(L_i) \leq|D_i|+\frac{|L_i|+\rho(G)}2.

Using Lemma 2.3 gives instead

\gamma(G)\leq|D_i|+\frac{|L_i|+\rho^{\{2\}}(G)}3.

Sum over the s indices i>r and use \sum_{i>r}|D_i|=w.

Remark

Remark 3.8

007A

Adding Lemma 3.3, summed over i\leq r, to (3.7), and using the weaker estimate \sum_{i>r}|L_i|\leq|D|, recovers for r=0 exactly the proof of (?) in [Steiner2026]; similarly, (3.7) recovers (?). The improvements below come from four features: the packing parts contribute \gamma(G) each without penalty; Lemma 3.4 does not charge vertices above P; Lemma 3.6 supplies a second count; and the inequalities of Lemmas 3.5 and 3.7 are kept separate until the final combination.

Remark

Remark 4.1

5L0T

For P=\varnothing, so r=0 and s=\gamma(H), Theorem 1.1 reduces to Brešar’s inequality (?), and Theorem 1.4 reduces to Steiner’s inequality (?). For any 2-packing P, the right-hand side of (1.1) is at least

\frac{r\gamma(G)}3 +\frac{\gamma(H)(2\gamma(G)-\rho(G))}3,

which is increasing in r. Thus a maximum 2-packing is the best choice when nothing beyond r+s\geq\gamma(H) is known about s, although a nonmaximum P with large s can give a better bound.

Remark

Remark 5.1

DR2O

Inequality (1.2) is Brešar’s bound (?) plus \rho(H)\gamma(G)/3. Inequality (1.2) gives \tfrac23\gamma(G)\gamma(H) only when \rho(G)/\gamma(G)=\rho(H)/\gamma(H); whenever the two factors have different packing-to-domination ratios, one gains a term proportional to the difference.

On the other hand, (1.2) does not contain the classical packing bound

\gamma(G\mathbin{\square}H) \geq\max\{\rho(G)\gamma(H),\rho(H)\gamma(G)\}.

For example, if \rho(H)=\gamma(H), then (1.2) reads

\gamma(G\mathbin{\square}H) \geq\gamma(G)\gamma(H)-\frac{\rho(G)\gamma(H)}3,

whereas the packing bound gives \gamma(G\mathbin{\square}H)\geq\gamma(G)\gamma(H). The two bounds should therefore be used together.

Proposition

Proposition 6.1

PP7D

Let \Phi(x_1,x_2,y_1,y_2) be the maximum of all functions in (?)–(?) and their mirror images. Then

\min\Phi=\frac23

over the box 0<x_1\leq x_2\leq1, 0<y_1\leq y_2\leq1. The minimum is attained at every point

(x_1,x_2,y_1,y_2)=\left(p,\frac12,p,\frac12\right), \qquad 0<p\leq\frac13.

Moreover, at each of these points the linear program obtained from Lemmas 3.4–3.7, with \gamma(G)=\gamma(H)=1, r=y_1, s=1-y_1, \rho(G)=x_1, and \rho^{\{2\}}(G)=2x_2, also has optimal value 2/3.

Uses Lemma 3.4, Lemma 3.7 · Used by Remark 6.2
Proof of Proposition 6.1

By (?) and its mirror image,

\Phi\geq\frac{2+|x_1-y_1|}{3}\geq\frac23

throughout the box. At (p,\tfrac12,p,\tfrac12), the functions in (?)–(?) take the values

\frac23;\qquad \frac{2+\tfrac32p}{4}\leq\frac23;\qquad \frac{1+p}{2}\leq\frac23;\qquad \frac{2-p}{3},\ \frac12,\ p,\ \frac12,

all at most 2/3 for 0<p\leq1/3. The mirror images have the same values, so \Phi=2/3 there.

For the linear program, Lemma 3.5, (3.6), (3.6), and (3.7)–(3.7) combined with Lemma 3.4 give the constraints

\begin{aligned} a+b&\geq r,\\ a+\frac b2&\geq r\left(1-\frac{x_1}{2}\right),\\ a+\frac b3&\geq r\left(1-\frac{2x_2}{3}\right),\\ \frac32w+\frac b2&\geq s\left(1-\frac{x_1}{2}\right),\\ \frac43w+\frac b3&\geq s\left(1-\frac{2x_2}{3}\right), \end{aligned}

with a,b,w\geq0 and objective a+b+w. At r=x_1=p, s=1-p, and x_2=1/2, put

a=p-p^2, \qquad b=p^2, \qquad w=\frac23(1-p)\left(1-\frac p2\right)-\frac{p^2}{3}.

Then a+b=p and a+b/2=p-p^2/2, so the first two constraints hold with equality. Also a+b/3=p-\tfrac23p^2\geq\tfrac23p for p\leq1/2. The fourth constraint holds with equality by the choice of w, and the fifth reduces to 3p^2-6p+2\geq0, which holds for p\leq1-1/\sqrt3. All variables are nonnegative for 0<p\leq1/3, and

a+b+w =p+\frac23(1-p)\left(1-\frac p2\right)-\frac{p^2}{3} =\frac23.

The optimal value is at least 2/3 by the proof of Theorem 1.1, so it equals 2/3.

Remark

Remark 6.2 (The aggregate linear program)

COPM

The proof of Theorem 1.1 uses only Lemmas 3.4, 3.5, (3.6), and (3.7). If one minimizes a+b+w over the corresponding aggregate constraints, including a,b,w\geq0, the optimal value is

\max\left\{ r\gamma(G), \frac{r\gamma(G)}3 +\frac{(r+s)(2\gamma(G)-\rho(G))}{3} \right\}.

Indeed, the second term in (6.2) is the dual lower bound used in the proof of Theorem 1.1, while the first is Lemma 3.5. When

s(2\gamma(G)-\rho(G))\geq r\rho(G),

the second term is the larger one and an optimal solution is

a=r(\gamma(G)-\rho(G)), \qquad b=r\rho(G), \qquad w=\frac23s\left(\gamma(G)-\frac{\rho(G)}2\right)-\frac13r\rho(G).

All four aggregate inequalities are then tight. If (6.2) fails, the exact LP value is r\gamma(G); for example, one may take w=0 and choose b=s(2\gamma(G)-\rho(G))\leq r\rho(G) and a=r\gamma(G)-b.

At the extremal points used in Proposition 6.1, condition (6.2) holds. There the packing parts are negligible, (1.2) is \tfrac23\gamma(G)\gamma(H) up to lower-order terms, and the loss comes from applying Lemma 2.2 to the sets L_i of leftover parts, where it gives \gamma_G(L_i)\lesssim |L_i|/2. Lemma 2.3 improves the coefficient 1/2 to 1/3 but pays \rho^{\{2\}}(G)/3 per part, which explains why (?) does not help when \rho^{\{2\}}(G)\approx\gamma(G). The Hou–Lu bound (?) likewise gives only about 1/2 in this regime.

Thus a further improvement requires an estimate strong when \rho(G)\ll\gamma(G) and \rho^{\{2\}}(G)\approx\gamma(G). In this regime the half-integral packing furnished by a maximum 2-packing function has value \rho^{\{2\}}(G)/2\approx\gamma(G)/2; this is a lower bound for the fractional domination number, but it does not by itself determine that parameter. The hypothetical extremal dominating set described by the linear program covers only about \gamma(G) of the |V(G)| fibers in each G-layer above a vertex of N_H(p); the domination of those layers, expressed by Lemma 2.4, is information not used by the aggregate argument.

Remark

Remark 6.3

78FF

Proposition 6.1 concerns only the optimization problem in which x_1,x_2,y_1,y_2 are treated as free parameters subject to x_1\leq x_2 and y_1\leq y_2. It does not assert that (\tfrac14,\tfrac12,\tfrac14,\tfrac12), or any point with \Phi=2/3, is realized or asymptotically realized by pairs of graphs. Additional universal relations among \rho, \rho^{\{2\}}, and \gamma would restrict the box and could raise the certified constant.

Source and licence

The text on this page is from arXiv:2607.01109v3 by Mohsen Aliabadi and Elliot Krop, under the licence CC BY 4.0. The corpus made these changes: it read the items out of the LaTeX source, it expanded the macros of the paper, it put each proof after its statement, and it converted the text to HTML. The formulas are the TeX of the paper, and KaTeX renders them. A formula or a reference can look different from the PDF. The page is published under CC BY 4.0.

The record of each item is in analysis/objects/2607/2607.01109.jsonl.