Upload
hacong
View
214
Download
2
Embed Size (px)
Citation preview
PATH COVER NUMBER, MAXIMUM NULLITY, AND ZERO
FORCING NUMBER OF ORIENTED GRAPHS AND OTHER
SIMPLE DIGRAPHS
ADAM BERLINER∗, CORA BROWN† , JOSHUA CARLSON‡ , NATHANAEL COX∗,
LESLIE HOGBEN§ , JASON HU¶, KATRINA JACOBS‖, KATHRYN MANTERNACH∗∗,
TRAVIS PETERS††, NATHAN WARNBERG‡ , AND MICHAEL YOUNG‡
April 30, 2014
Abstract. An oriented graph is a simple digraph obtained from a simple graph by choosing1
exactly one of the two arcs (u, v) or (v, u) to replace each edge {u, v}. A simple digraph describes2
the zero-nonzero pattern of off-diagonal entries of a family of (not necessarily symmetric) matrices.3
Minimum rank of a simple digraph is the minimum rank of this family of matrices; maximum nullity4
is defined analogously. The simple digraph zero forcing number and path cover number are related5
parameters. We establish bounds on the range of possible values of all these parameters for oriented6
graphs, establish connections between the values of these parameters for a simple graph G, for various7
orientations ~G and for the doubly directed digraph←→G , and establish an upper bound on the number8
of arcs in a simple digraph in terms of the zero forcing number.9
Keywords. zero forcing number, maximum nullity, minimum rank, path cover num-10
ber, simple digraph, oriented graph11
AMS subject classifications. 05C50, 05C20, 15A0312
1. Introduction. The maximum nullity and zero forcing number of simple di-13
graphs are studied in [10] and [5]. We study connections between these parameters14
and path cover number, and we study all these parameters for special types of di-15
graphs derived from graphs, including oriented graphs and doubly directed graphs.16
Section 2 considers oriented graphs. We establish a bound on the difference of values17
of the parameters path cover number, maximum nullity, and zero forcing number for18
∗Department of Mathematics, Statistics, and Computer Science, St. Olaf College, Northfield, MN55057 ([email protected]), ([email protected]). Research of N. Cox supported by NSF DMS 0750986†Department of Mathematics, Carleton College, Northfield, MN 55057 ([email protected]).
Research supported by NSF DMS 0750986.‡Department of Mathematics, Iowa State University, Ames, IA 50011 ([email protected]),
([email protected]), ([email protected]). Research of J. Carlson supported by ISU Holl funds.Research of M. Young supported by NSF DMS 0946431.§Department of Mathematics, Iowa State University, Ames, IA 50011, ([email protected]) and
American Institute of Mathematics, 360 Portage Ave, Palo Alto, CA 94306 ([email protected]).¶Department of Mathematics, University of California, Berkeley, CA 94720-3840 (ja-
[email protected]). Research supported by NSF DMS 0750986.‖Department of Mathematics, Pomona College, 610 North College Avenue, Claremont, CA 91711
([email protected]). Research supported by NSF DMS 0750986.∗∗Department of Mathematics and Computer Science, Central College, Pella, IA 50219 (manter-
[email protected]). Research supported by ISU Holl funds.††Natural and Mathematical Science Division, Culver-Stockton College, Canton, MO 63435
1
two orientations of one graph, and determine the range of values of these parameters19
for orientations of paths and cycles, and some of the possible values for tournaments.20
We establish connections between these parameters for a simple graph and its dou-21
bly directed digraph in Section 3. In Section 4, we establish an upper bound on the22
number of arcs of a simple digraph in terms of the zero forcing number. We also show23
that several results for simple graphs fail for oriented graphs, including the Graph24
Complement Conjecture and Sinkovic’s theorem that maximum nullity is at most25
path cover number for outerplanar graphs.26
All graphs and digraphs are taken to be simple. We use G = (V (G), E(G)) to27
denote a graph and Γ = (V (Γ), E(Γ)) to denote a digraph, often using V and E when28
G or Γ is clear. For a digraph Γ and R ⊆ V , the induced subdigraph Γ[R] is the29
digraph with vertex set R and arc set {(v, w) ∈ E : v, w ∈ R}; an analogous definition30
is used for graphs. The subdigraph induced by the complement R is also denoted31
by Γ − R, or in the case R is a single vertex v, by Γ − v. A digraph Γ = (V,E) is32
transitive if for all u, v, w ∈ V , (u, v), (v, w) ∈ E implies (u,w) ∈ E.33
For a digraph Γ = (V,E) having v, u ∈ V and (v, u) ∈ E, u is an out-neighbor of v34
and v is an in-neighbor of u. The out-degree of v, denoted by deg+(v), is the number35
of out-neighbors of v in Γ, and analogously for in-degree, denoted by deg−(v). Define36
δ+(Γ) = min{deg+(v) : v ∈ V } and δ−(Γ) = min{deg−(v) : v ∈ V }. For a digraph Γ,37
the reversal ΓT is obtained from Γ by reversing all the arcs.38
Let G be a graph. A path in G is a subgraph P = ({v1, . . . , vk}, E(P )) where39
E(P ) = {{vi, vi+1} :≤ i ≤ k − 1}; this path is often denoted by (v1, . . . , vk) and its40
length is k−1. We say that a path in G is an induced path if it is an induced subgraph41
of G. A path cover of a graph G is a set of vertex disjoint induced paths that includes42
all vertices of G.43
Now suppose Γ is a digraph. A path in Γ is a subdigraph P = ({v1, . . . , vk}, E(P ))44
where E(P ) = {(vi, vi+1) : 1 ≤ i ≤ k − 1}; this path is often denoted by (v1, . . . , vk),45
its length is k − 1, and the arcs of E(P ) are called path arcs. If (v1, . . . , vk) is a path46
in Γ, v1 is called the initial vertex and vk is the terminal vertex. We say vertex u has47
access to v in Γ if there is a path from u to v. A path (v1, . . . , vk) in Γ is an induced48
path if E does not contain any arc of the form (vi, vj) with j > i + 1 or i > j + 1.49
We note this does not necessarily imply that the path subdigraph is induced, because50
any of the arcs in {(vi+1, vi) : 1 ≤ i ≤ k − 1} are permitted. A path (v1, . . . , vk) in Γ51
is Hessenberg if E does not contain any arc of the form (vi, vj) with j > i + 1. Any52
induced path is Hessenberg but not vice versa. A path cover for Γ is a set of vertex53
disjoint Hessenberg paths that includes all vertices of Γ [10].54
For graphs G and digraphs Γ, the path cover number P(G) or P(Γ) is the minimum55
number of paths in a path cover (induced for a graph, Hessenberg for a digraph) and56
a minimum path cover is a path cover with this minimum number of paths.57
Zero forcing was introduced in [1] for (simple) graphs. We define zero forcing58
for (simple) digraphs as in [10]. Let Γ be a digraph with each vertex colored either59
2
white or blue1. The color change rule is: If u is a blue vertex of Γ, and exactly one60
out-neighbor v of u is white, then change the color of v to blue. In this situation,61
we say that u forces v and write u → v. Given a coloring of Γ, the final coloring is62
the result of applying the color change rule until no more changes are possible. A63
zero forcing set for Γ is a subset of vertices B such that, if initially the vertices of64
B are colored blue and the remaining vertices are white, the final coloring of Γ is all65
blue. The zero forcing number Z(Γ) is the minimum of |B| over all zero forcing sets66
B ⊆ V (Γ).67
For a given zero forcing set B for Γ, we create a chronological list of forces by68
constructing the final coloring, listing the forces in the order in which they were69
performed. Although for a given set of vertices B the final coloring is unique, B need70
not have a unique chronological list of forces. Suppose Γ is a digraph and F is a71
chronological list of forces for a zero forcing set B. A forcing chain is an ordered set72
of vertices (w1, w2, . . . , wk) where wj → wj+1 is a force in F for 1 ≤ j ≤ k − 1. A73
maximal forcing chain is a forcing chain that is not a proper subset of another forcing74
chain. The following results will be used.75
Lemma 1.1. [10] Suppose Γ is a digraph and F is a chronological list of forces76
of a zero forcing set B. Then, every maximal forcing chain is a Hessenberg path that77
starts with a vertex in B.78
For a fixed chronological list of forces F of a zero forcing set B of Γ, the chain79
set is the set of all maximal forcing chains. By Lemma 1.1, the chain set of F is a80
path cover, called a zero forcing path cover and the maximal forcing chains are also81
called forcing paths.82
Proposition 1.2. [10] For any digraph Γ, P(Γ) ≤ Z(Γ).83
A cycle of length k ≥ 3 in a graph G or digraph Γ is a sub(di)graph consisting of84
a path (v1, . . . , vk) and the additional edge or arc {vk, v1} or (vk, v1).85
Lemma 1.3. Suppose P = (v1, . . . , vk) is a Hessenberg path in a digraph Γ. Then86
P is an induced path or Γ[V (P )] contains a (digraph) cycle of length at least 3.87
Proof. Suppose P is not an induced path. Then Γ must contain an arc of the88
form (vi, vj) with j > i + 1 or i > j + 1. Since P is Hessenberg, Γ does not contain89
an arc of the form (vi, vj) with j > i + 1. Thus Γ must contain an arc of the form90
(vi, vj) with i > j + 1. Then (vj , vj+1, . . . , vi, vj) is a (digraph) cycle in Γ[V (P )].91
Let F be a field. For a square matrix A = [aij ] ∈ Fn×n, the digraph of A,92
denoted Γ(A) = (V,E), is the (simple) digraph described by the off-diagonal zero-93
nonzero pattern of the entries: the set of vertices is V = {1, 2, . . . , n} and the set of94
arcs is E = {(i, j) : aij 6= 0, i 6= j}. Note that the value of the diagonal entries of A95
does not affect Γ(A).96
Conversely, given any simple digraph Γ (along with an ordering of the vertices),97
1The early literature uses the color black rather than blue.
3
we may associate with Γ a family of matrices MF (Γ) = {A ∈ Fn×n : Γ(A) = Γ}.98
The minimum rank over F of a digraph Γ is mrF (Γ) = min{rankA : A ∈ MF (Γ)}99
and the maximum nullity over F of Γ is MF (Γ) = max{null A : A ∈ MF (Γ)}. It is100
immediate that that mrF (Γ) + MF (Γ) = n.101
Similarly, symmetric matrices and undirected graphs are associated. For a sym-102
metric matrix A = [aij ] ∈ Fn×n, the graph of A is the (simple) graph G(A) = (V,E)103
with V = {1, 2, . . . , n} and E = {{i, j} : i 6= j and aij 6= 0}. The family of symmetric104
matrices associated with G is SF (G) = {A ∈ Fn×n : AT = A,G(A) = G}, and105
minimum rank and maximum nullity are similarly defined for undirected graphs.106
For the much of this paper, we let F = R and we write S(G),M(Γ),M(Γ), and107
mr(Γ) rather than SR(G),MR(Γ),MR(Γ), and mrR(Γ), etc. If a graph or digraph108
parameter that depends on matrices does not change regardless of the field F , then109
we say that that parameter is field independent; in this case MF (Γ) = M(Γ) for every110
field F .111
Remark 1.4. Clearly mrF (ΓT) = mrF (Γ), and Z(ΓT) = Z(Γ) is known [5].112
Because the reversal of a Hessenberg path is a Hessenberg path, P(ΓT) = P(Γ).113
2. Oriented graphs. In this section we establish results for minimum rank,114
maximum nullity, zero forcing number, and path cover number of oriented graphs.115
Given a graph G, an orientation ~G of G is a digraph obtained by replacing each116
edge {u, v} by exactly one of the arcs (u, v) and (v, u) (so a graph G has 2|E(G)|117
orientations, some of which may be isomorphic to each other).118
2.1. Range over orientations. We consider the range of values of β(~G) over119
all possible orientations, for the parameters β = mr,M,Z,P.120
Theorem 2.1. Suppose β is a positive-integer-valued digraph parameter with the121
following properties for every oriented graph ~G:122
(1) β(~GT) = β(~G).123
(2) If (u, v) ∈ E(~G) and ~G0 is obtained from ~G by replacing (u, v) by (v, u) (i.e.,124
reversing the orientation of one arc), then |β(~G0)− β(~G)| ≤ 1.125
Then for any two orientations ~G1 and ~G2 of the same graph G,126
|β(~G2)− β(~G1)| ≤⌊E(G)
2
⌋.127
Furthermore, every integer between β(~G2) and β(~G1) is attained as β(~G) for some128
orientation ~G of G.129
Proof. Without loss of generality, β(~G2) ≥ β(~G1). Let e = |E(G)|. Because130
~G1 and ~G2 share the same underlying graph, it is possible to obtain ~G2 from ~G1 by131
reversing some of the arcs of ~G1. Let ` be the number of arcs we need to reverse to132
obtain ~G2 from ~G1. By hypothesis, reversing the direction of one arc changes the133
value of β by at most one, so β(~G2) − β(~G1) ≤ `. The number of arcs that must be134
4
reversed to obtain ~GT2 from ~G1 is e − `, so β(~GT
2 ) − β(~G1) ≤ e − `. By hypothesis,135
β(~GT2 ) = β(~G2), so β(~G2)−β(~G1) ≤ b e2c. The last statement follows from hypothesis136
(2) and the fact that we can go from ~G1 to ~G2 by reversing one arc at a time.137
Corollary 2.2. If ~G1 and ~G2 are both orientations of the graph G, then:138
|mr(~G2)−mr(~G1)| ≤ bE(G)
2c,139
|M(~G2)−M(~G1)| ≤ bE(G)
2c,140
|Z(~G2)− Z(~G1)| ≤ bE(G)
2c, and141
|P(~G2)− P(~G1)| ≤ bE(G)
2c.142
Furthermore, every integer between β(~G2) and β(~G1) is attained as β(~G) for some143
orientation ~G of G when β is any of the parameters mr,M,Z,P.144
Proof. The first hypothesis of Theorem 2.1, β(~GT) = β(~G), is established for these145
parameters in Remark 1.4. To show these parameters satisfy the second hypothesis146
of Theorem 2.1, suppose arc (u, v) of ~G is reversed to obtain ~G0 from ~G. In each case,147
the process is reversible, so it suffices to prove β(~G0) ≤ β(~G) + 1.148
For minimum rank, suppose Γ(A) = ~G and rankA = mr(~G). Define B by buu =149
bvv = buv = bvu = −auv and bij = 0 for all other entries of B. Then Γ(A+ B) = ~G0150
and rank(A + B) ≤ rankA + 1. Thus mr(~G0) ≤ mr(~G) + 1. The statement for151
maximum nullity is equivalent.152
For zero forcing number, choose a minimum zero forcing set B and chronological153
list of forces F of ~G. If the force u→ v is in F , then B ∪ {v} is a zero forcing set for154
~G0. If u 6→ v and for some w, v → w is in F , then B ∪ {u} is a zero forcing set for155
~G0. If v does not perform a force and u→ v is not in F , then B is a zero forcing set156
for ~G0. Thus, Z(~G0) ≤ Z(~G) + 1.157
For path cover number, suppose P = {P (1), P (2), . . . , P (k)} is a path cover of ~G158
and |P| = P(~G). If (u, v) is not an arc in one of the paths in P, then P is a path159
cover for ~G0 and P(~G0) ≤ P(~G). So suppose (u, v) is an arc in some path P (`). Then160
we construct a path cover for ~G0 by replacing P (`) by the two paths resulting from161
deleting the arc (u, v). Thus, P(~G0) ≤ P(~G) + 1.162
2.2. Hierarchal orientation. We establish a method for finding an orientation163
~G of a graph G for which P(~G) = P(G). Let P = {P (1), P (2), . . . , P (k)} be any164
path cover of a graph G. A rooted path cover of G, R = {R(1), R(2), . . . , R(k)}, is165
obtained from P by choosing one endpoint as the root of P (i) for each i = 1, . . . , k.166
R is a minimum rooted path cover if |R| = P(G). In the case P is a zero forcing167
path cover of a zero forcing set B, then the root of P (i) is automatically chosen to168
be the unique element of B that is a vertex of P (i). A rooted path cover obtained169
5
from P naturally orders V (P (i)), starting with the root, and we denote this order by170
R(i) = (r(i)1 , r
(i)2 , . . . , r
(i)si ) where si − 1 is the length of P (i). Observe that if a rooted171
path cover is formed from a zero forcing path cover of a zero forcing set, the ordering172
within each rooted path coincides with the forcing order in that path.173
Definition 2.3. Given a rooted path cover R of a graph G, the hierarchal174
orientation_GR of G resulting from R is defined by orienting G as follows:175
(1) Orient R(i) as r(i)1 → r
(i)2 → · · · → r
(i)si ; that is, replace the edge {r(i)j , r
(i)j+1}176
by the arc (r(i)j , r
(i)j+1) for j = 1, . . . , si − 1.177
(2) For any edge between R(i) and R(j) with i < j, orient as i → j; that is, if178
i < j, replace the edge {r(i)`i , r(j)`j} by the arc
(r(i)`i, r
(j)`j
).179
Since by definition, the paths in a path cover of a graph are induced, all the edges of180
G have been oriented by these two rules.181
Observation 2.4. For any rooted path cover R of G, R is a path cover of_GR182
(with each path originating at its root), so P(_GR) ≤ |R|.183
Proposition 2.5. An oriented graph ~G is the hierarchal orientation_GR of G184
for some rooted path cover R of G (not necessarily minimum) if and only if ~G does185
not contain a digraph cycle.186
Proof. Suppose R = {R(1), . . . , R(k)} is rooted path cover of G. Since each path187
R(i) is induced, in order for_GR to have a digraph cycle, V (
_GR) would have to include188
vertices from at least two paths R(i) and R(j), with i < j. But by definition of_GR,189
there are no arcs from vertices in R(j) to vertices in R(i).190
Suppose ~G does not contain a digraph cycle. Then we may order the vertices191
{v1, . . . , vn}, vj does not have access to vi whenever j > i. Then if V (R(i)) = {vi},192
R = {R(1), . . . , R(n)} is a rooted path cover and ~G =_GR.193
Theorem 2.6. Suppose R = {R(1), . . . , R(k)} is a rooted path cover of G and194
_GR is the hierarchal orientation of G resulting from R. Then any path cover for
_GR195
is a path cover for G. If R is a minimum rooted path cover, then P(G) = P(_GR).196
Proof. Let P be a Hessenberg path in_GR. By Proposition 2.5,
_GR does not197
contain a digraph cycle, so by Lemma 1.3, P is an induced path. Thus, any path198
cover for_GR is a path cover for G, and this implies P(G) ≤ P(
_GR). IfR is a minimum199
rooted path cover of G, then P(_GR) ≤ |R| = P(G) ≤ P(
_GR), so P(G) = P(
_GR).200
Example 2.7. Not every orientation ~G having P(~G) = P(G) is a hierarchal201
orientation. The oriented graph ~G shown in Figure 2.1 has P(~G) = 2 = P(G), but ~G202
is not a hierarchal orientation because ~G contains a digraph cycle.203
6
321 4
Fig. 2.1. An oriented graph ~G that is not a hierarchal orientation but has P( ~G) = P(G).
Although for any graph G we can find an orientation so that P(~G) = P(G), this204
is not always the case for zero forcing number or maximum nullity.205
Example 2.8. Consider K4, the complete graph on four vertices. It is well206
known that M(K4) = Z(K4) = 3, whereas we show that for any orientation ~K4 of207
K4, 2 ≥ Z( ~K4) ≥ M( ~K4). If ~K4 contains a directed 3-cycle, then any one vertex on208
the 3-cycle and the remaining vertex form a zero forcing set. If ~K4 has no directed209
3-cycle, then we may order the vertices {u1, u2, u3, u4} where uj does not have access210
to ui whenever j > i. Then, {u1, u3} is a zero forcing set.211
Observation 2.9. If R = {R(1), . . . , R(k)} is a rooted path cover for G, then the212
set of roots {r(1)1 , . . . , r(k)1 } is a zero forcing set of the digraph
_GR, as zero forcing can213
be done in path order along R(k), followed by R(k−1), etc.214
Theorem 2.10. Suppose G is a graph and R is a minimum rooted path cover of215
G. Then Z(_GR) = P(
_GR) = P(G).216
Proof. From Theorem 2.6, Proposition 1.2, Observation 2.9, and the hypotheses,217
P(G) = P(_GR) ≤ Z(
_GR) ≤ |R| = P(G).218
Whenever P(G) = Z(G), we can use a minimum rooted path cover to find an219
orientation of G realizing Z(G) as its zero forcing number.220
Corollary 2.11. Suppose G is a graph such that P(G) = Z(G) and R is a221
minimum rooted path cover of G. Then Z(_GR) = Z(G).222
Because P(T ) = Z(T ) for every (simple undirected) tree T [1], we have the fol-223
lowing corollary.224
Corollary 2.12. If T is a tree, then there exists an orientation ~T of T such225
that Z(~T ) = Z(T ).226
If we allow a path cover that is not a minimum path cover, it is not difficult to find227
a graph and rooted path cover R with P(_GR) < Z(
_GR) (in fact, P(
_GR) < M(
_GR)).228
Example 2.13. Let G be the double triangle graph shown in Figure 2.2(a)229
and consider the rooted path cover of G defined by R = {R(1), R(2), R(3)} where230
V (R(1)) = {1}, V (R(2)) = {3}, and V (R(3)) = {2, 4} with 2 as the root of R(3). The231
hierarchal orientation_GR is shown in Figure 2.2(b).232
7
1
2
3
4
1
2
3
4
(a) (b)
Fig. 2.2. An hierarchal orientation_GR having P(
_GR) < M(
_GR) = Z(
_GR).
Then P(_GR) = 2 because paths (1, 2) and (3, 4) cover all vertices, and the vertices233
1 and 3 must each be initial vertices of any path they are in. Let A =
0 1 0 1
0 1 0 1
0 1 0 1
0 0 0 0
.234
Then Γ(A) =_GR and nullityA = 3. The set {1, 2, 3} is a zero forcing set for
_GR so235
3 ≤ M(_GR) ≤ Z(
_GR) ≤ 3.236
Every graph G we have examined has M(_GR) = P(
_GR) for minimum rooted path237
covers R, but these examples all involve a small number of vertices.238
Question 2.14. Does M(_GR) = P(
_GR) if R is a minimum rooted path cover of239
G?240
2.3. Tournaments. A tournament is an orientation of of the complete graph241
Kn. In this section we consider the possible values of path cover number, maximum242
nullity, and zero forcing number for tournaments.243
Example 2.15. We create an orientation of Kn by labeling the vertices {1, . . . , n}244
and by orienting the edges {u, v} as (u, v) if and only if v < u− 1 or v = u+ 1. The245
resulting orientation is called the Hessenberg tournament of order n, denoted ~K(H)n .246
This is the Hessenberg path on n vertices containing all possible arcs except those247
of the form (u+ 1, u) for 1 ≤ u ≤ n − 1. Since the zero forcing number of any248
Hessenberg path is one, P( ~K(H)n ) = M( ~K
(H)n ) = Z( ~K
(H)n ) = 1. Observe that ~K
(H)n is249
self-complementary.250
Example 2.16. Label the vertices of Kn by {1, . . . , n} and orient the edges {u, v}251
as (u, v) if and only if u < v. The resulting orientation is the transitive tournament,252
denoted ~K(T )n . We show that P( ~K
(T )n ) = Z( ~K
(T )n ) = M( ~K
(T )n ) =
⌈n2
⌉for any n. Let A253
be the adjacency matrix of ~K(T )n , and let D = diag(0, 1, 0, 1, . . . ). Then Γ(A+D) =254
~K(T )n and nullity(A + D) =
⌈n2
⌉because A + D has
⌊n2
⌋duplicate rows and if n is255
odd, an additional row of zeros. The set of odd numbered vertices, B = {1, 3, . . . , },256
is a zero forcing set. Thus,⌈n2
⌉≤ M( ~Kn) ≤ Z( ~Kn) ≤
⌈n2
⌉. Furthermore, from the257
8
definition of ~K(T )n , no more than 2 vertices can be on the same Hessenberg path.258
Proposition 2.17. For any tournament ~Kn, 1 ≤ P( ~Kn) ≤⌈n2
⌉, and for every259
integer k with 1 ≤ k ≤⌈n2
⌉, there is an orientation ~Kn having P( ~Kn) = k.260
For every integer k with 1 ≤ k ≤⌈n2
⌉, there is an orientation ~Kn having Z( ~Kn) = k.261
Proof. For both P and Z, ~K(H)n (Example 2.15) realizes the lower bound and ~K
(T )n262
(Example 2.16) realizes the upper bound. For the upper bound on attainable path263
cover numbers, partition the vertices of ~Kn into⌈n2
⌉sets of size two or one. Each264
pair of vertices and the arc between them forms a path. The assertion that all values265
for P and Z between 1 and⌈n2
⌉are possible follows from Corollary 2.2.266
For n ≤ 7, the transitive tournament ~K(T )n achieves the highest zero forcing267
number; i.e., for all orientations ~Kn, Z( ~Kn) ≤⌈n2
⌉(this has been verified using the268
program [12] written in Sage). But for n = 8, there exists a tournament having max-269
imum nullity greater than that of the transitive tournament, as in the next example.270
Example 2.18.
1
2 3
4
5
67
8
Fig. 2.3. A tournament ~K8 having M( ~K8) = Z( ~K8) = 5 > d 82e.
271
Let ~K8 be the tournament shown in Figure 2.3. Observe that {1, 2, 3, 4, 8} is a272
zero forcing set for ~K8 so Z( ~K8) ≤ 5. The matrix273
A =
0 1 2 1 2 3 1 2
0 0 1 1 1 1 0 0
0 0 1 1 1 1 0 0
0 0 0 0 1 1 1 1
0 0 0 0 1 1 1 1
0 0 0 0 0 0 0 0
0 1 1 0 0 1 0 1
0 1 1 0 0 1 0 1
274
has rankA = 3 and Γ(A) = ~K8, so 5 ≤ M( ~K8). Thus, Z( ~K8) = M( ~K8) = 5 > 4 =275
d 82e. We also show that P( ~K8) = 3: Since {(2, 4, 8), (3, 5, 7), (1, 6)} is a path cover,276
9
P( ~K8) ≤ 3. There are no induced paths of length greater than two in ~K8, so by277
Lemma 1.3 any path of length three or more must have a cycle. Thus vertices 1 and278
6 must be in paths of length at most two. If they are in separate paths in a path279
cover P, then |P| ≥ 3. So assume (1, 6) is a path in P. Since ~K8 − {1, 6} is not a280
(Hessenberg) path, |P| ≥ 3.281
There are also examples of tournaments ~Kn for which M( ~Kn) < Z( ~Kn).282
Proposition 2.19. The tournament ~K7 shown in Figure 2.4 has P( ~K7) = 2,283
M( ~K7) = 3, and Z( ~K7) = 4.
1 2
3
4
56
7
Fig. 2.4. A tournament ~K7 having M( ~K7) = 3 < 4 = Z( ~K7).
284
Proof. Because {(4, 6, 1, 3), (2, 5, 7)} is a path cover for ~K7, and ~K7 is not a285
Hessenberg path, P( ~K7) = 2.286
Next we show M( ~K7) ≤ 3. Suppose Γ(A) = ~K7. The nonzero pattern of A is287
? ∗ ∗ ∗ ∗ 0 0
0 ? ∗ ∗ ∗ 0 0
0 0 ? ∗ ∗ ∗ ∗0 0 0 ? ∗ ∗ ∗0 0 0 0 ? ∗ ∗∗ ∗ 0 0 0 ? ∗∗ ∗ 0 0 0 0 ?
288
where ∗ denotes a nonzero entry and ? may have any real value. Rows 1, 5, and 7 are289
necessarily linearly independent, by considering columns 2, 3, and 6.290
For A to achieve nullity 4, rankA = 3 and thus all the remaining rows must be in291
the span of rows 1, 5, and 7. We show this is impossible, implying that mr( ~K7) ≥ 4292
and M( ~K7) ≤ 3. Once that is done, construct a matrix A with Γ(A) = ~K7 and293
rankA = 4 by setting all nonzero off-diagonal entries to 1 and setting the diagonal294
entries as aii = 0 for i odd and aii = 1 for i even, so M( ~K7) = 3.295
If a11 = 0, then row 3 cannot be expressed as a linear combination of rows 1,296
5, and 7: By considering column 1, the coefficient of row 7 must be zero, which297
10
implies that then the coefficient of row 1 must be zero by considering column 2, but298
by considering column 4, row 3 is not a multiple of row 5. Thus a11 6= 0.299
If a77 6= 0, then row 2 cannot be expressed as a linear combination of rows 1, 5,300
and 7: By considering column 6 the coefficient of row 5 must be zero, which implies301
that the coefficient of row 7 must be zero by considering column 7, but by considering302
column 1, row 2 is not a multiple of row 1 (because a11 6= 0). Thus a77 = 0.303
If a55 = 0, then row 4 cannot be expressed as a linear combination of rows 1, 5,304
and 7: By considering column 3 the coefficient of row 1 must be zero, which implies305
that the coefficient of row 7 must be zero by considering column 1. But row 4 is not306
a multiple of row 5 (because a55 = 0). Thus a55 6= 0.307
Now row 6 cannot be expressed as a linear combination of rows 1, 5, and 7: By308
considering column 3 the coefficient of row 1 must be zero. By considering column309
5 the coefficient of row 5 must be zero. Now by considering column 7, row 6 is not310
a scalar multiple of row 7. Therefore row 6 is not a linear combination of rows 1, 5,311
and 7, rankA ≥ 4, and mr( ~K7) ≥ 4.312
Finally we show that Z( ~K7) = 4. Observe that any zero forcing set must contain a313
vertex from the set {1, 2}: If 1 and 2 are initially colored white, the only vertices that314
can force them are 6 and 7, but 1 and 2 are both out-neighbors of 6 and of 7. Observe315
that any zero forcing set must contain a vertex from the set {6, 7}: If 6 and 7 are316
initially colored white, the only vertices that can force them are 3, 4, and 5, but 6 and317
7 are both out-neighbors of 3, of 4, and of 5. Observe that any zero forcing set must318
contain a vertex from the set {3, 4}: If 3 and 4 are initially colored white, the only319
vertices that can force them are 1 and 2, but 3 and 4 are both out-neighbors of 1 and320
of 2. Observe that any zero forcing set must contain a vertex from the set {4, 5}: If 4321
and 5 are initially colored white, the only vertices that can force them are 1, 2, and 3,322
but 4 and 5 are both out-neighbors of 1, of 2, and of 3. Hence, a zero forcing set must323
contain at least four vertices, unless vertex 4 is the only vertex from {3, 4} and {4, 5}324
selected. However, by inspection the sets {1, 4, 6}, {1, 4, 7}, {2, 4, 6}, {2, 4, 7} are not325
zero forcing sets. The set {1, 2, 4, 6} is a zero forcing set for ~K7, and so Z( ~K7) = 4.326
2.4. Orientations of paths. In this section we consider the possible values of327
path cover number, maximum nullity, and zero forcing number for orientations of328
paths.329
Example 2.20. Starting with the path Pn, label the vertices in path order by330
{1, . . . , n} and orient the edge {i, i + 1} as arc (i, i + 1) for i = 1, . . . , n − 1. The331
resulting orientation is the path orientation of Pn, denoted ~P(H)n . Then P(~P
(H)n ) =332
M(~P(H)n ) = Z(~P
(H)n ) = 1.333
Example 2.21. Starting with the path Pn, label the vertices in path order by334
{1, . . . , n} and orient the edges as follows: Orient {1, 2} as (1, 2). For i = 1, . . . ,⌊n2
⌋−335
1, orient {2i+ 1, 2i} and {2i+ 1, 2i+ 2} as (2i+ 1, 2i) and (2i+ 1, 2i+ 2). If n is odd,336
orient {n− 1, n} as (n, n− 1). The resulting orientation is the alternating orientation337
11
of Pn, denoted ~P(A)n . Note that all odd-numbered vertices have in-degree zero. So338
P(~P(A)n ) = M(~P
(A)n ) = Z(~P
(A)n ) =
⌈n2
⌉because the odd vertices form a minimum zero339
forcing set, there is no directed path of length greater than one, and the adjacency340
matrix A of ~P(A)n has rank
⌊n2
⌋.341
Proposition 2.22. For any oriented path ~Pn, 1 ≤ P(~Pn) ≤ dn2 e, 1 ≤ M(~Pn) ≤342
dn2 e, and 1 ≤ Z(~Pn) ≤ dn2 e. For every integer k with 1 ≤ k ≤⌈n2
⌉, there are (possibly343
three different) orientations ~Pn having P(~Pn) = k, M(~Pn) = k, and Z(~Pn) = k.344
Proof. The proof of the second statement follows from Examples 2.20 and 2.21345
and Corollary 2.2. To complete the proof we show that Z(~Pn) ≤⌈n2
⌉for every346
orientation ~Pn. Apply Corollary 2.2 to the given orientation ~Pn and ~P(H)n . Then347
Z(~Pn)−Z(~P(H)n ) ≤
⌊n−12
⌋, so Z(~Pn) ≤
⌊n−12
⌋+ 1. If n is odd,
⌊n−12
⌋+ 1 = n−1
2 + 1 =348 ⌈n2
⌉. If n is even,
⌊n−12
⌋+ 1 = (n2 − 1) + 1 =
⌈n2
⌉. Therefore Z(~Pn) ≤
⌈n2
⌉.349
2.5. Orientations of cycles. In this section we consider the possible values350
of path cover number, maximum nullity, and zero forcing number for orientations of351
cycles of length at least 4 (since a cycle of length 3 is a complete graph).352
Example 2.23. Starting with the cycle Cn, label the vertices in cycle order by353
{1, . . . , n} and orient the edge {i, i+ 1} as arc (i, i+ 1) for i = 1, . . . , n (where n+ 1354
is interpreted as 1). The resulting orientation is the cycle orientation of Cn, denoted355
~C(H)n . Then P(~C
(H)n ) = M(~C
(H)n ) = Z(~C
(H)n ) = 1.356
Example 2.24. Starting with Cn, label the vertices in cycle order by {1, . . . , n}357
and orient the edges as follows: Orient {1, 2} and {1, n} as (1, 2) and (1, n). For358
i = 1, . . . ,⌊n2
⌋− 1, orient {2i + 1, 2i} and {2i + 1, 2i + 2} as arcs (2i + 1, 2i) and359
(2i + 1, 2i + 2). If n is odd, orient the edge {n − 1, n} as (n, n − 1). The resulting360
orientation is the alternating orientation of Cn, denoted ~C(A)n . If n is odd there is361
one path of length 2, so P(~C(A)n ) =
⌊n2
⌋. Let S be the set of odd numbered vertices362
(with the exception of vertex n if n is odd), so every vertex in S has in-degree zero363
and |S| =⌊n2
⌋. Clearly S ⊆ B for any zero forcing set, and every vertex in S has364
two out-neighbors not in S, so every zero forcing set must have cardinality at least365 ⌊n2
⌋+1. Since S∪{2} is a zero forcing set, Z(~C
(A)n ) =
⌊n2
⌋+1. We construct a matrix366
A ∈ M(~C(A)n ) of nullity
⌊n2
⌋+ 1, showing that M(~C
(A)n ) =
⌊n2
⌋+ 1. Any matrix in367
M(~C(A)n ) has two nonzero off-diagonal entries in every odd row (except n if n is odd)368
and no nonzero off-diagonal entries in every even row. Define a matrix A = [aij ] with369
Γ(A) = ~C(A)n by setting aii = 0 except ann = −1 if n is odd, and in each odd row the370
first nonzero entry is 1 and the second is −1. Then nullA =⌊n2
⌋+ 1.371
Proposition 2.25. Let ~Cn be any orientation of Cn (n ≥ 4). Then 1 ≤ P(~Cn) ≤372 ⌊n2
⌋and for every integer k with 1 ≤ k ≤
⌊n2
⌋, there is an orientation ~Cn having373
P(~Cn) = k. For any orientation of a cycle ~Cn, 1 ≤ M(~Cn) ≤ Z(~Cn) ≤⌊n2
⌋+ 1 and374
for every integer k with 1 ≤ k ≤⌊n2
⌋+1, there are (possibly two different) orientations375
12
~Cn having M(~Cn) = k and Z(~Cn) = k.376
Proof. The proof of the second part of each statement follows from Examples 2.20377
and 2.21 and Corollary 2.2. To complete the proof we show that P(~Cn) ≤⌊n2
⌋and378
Z(~Cn) ≤⌊n2
⌋+ 1 for all orientations ~Cn, by exhibiting a path cover and zero forcing379
set of cardinality not exceeding this bound.380
For path cover number: If n is even, choose two adjacent vertices, cover them381
with one path, and delete them, leaving a path on n − 2 vertices which is an even382
number. By Proposition 2.22, there is a path cover of these n− 2 vertices with n2 − 1383
paths, so there is a path cover of ~Cn having n2 =
⌊n2
⌋paths. If n is odd, then for any384
orientation ~Cn, there is a path on 3 vertices. Cover these vertices with that path and385
delete them, leaving a path on n − 3 vertices (again an even number), which can be386
covered by n−32 paths, and there is a path cover of ~Cn having n−3
2 + 1 =⌊n2
⌋paths.387
For zero forcing number: Delete any one vertex v, leaving a path on n−1 vertices,388
which has a zero forcing set B with |B| =⌈n−12
⌉by Proposition 2.22. Then the set389
B′ := B ∪ {v} is a zero forcing set for ~Cn and |B′| =⌈n−12
⌉+ 1. If n is even,390 ⌈
n−12
⌉+ 1 = n
2 + 1 =⌊n2
⌋+ 1. If n is odd,
⌈n−12
⌉+ 1 = n−1
2 + 1 =⌊n2
⌋+ 1.391
3. Doubly directed graphs. Given a graph G, the doubly directed graph←→G392
of G is the digraph obtained by replacing each edge {u, v} by both of the arcs (u, v)393
and (v, u). In this section we establish results for minimum rank, maximum nullity,394
zero forcing number, and path cover number of doubly directed graphs.395
Proposition 3.1. P(G) = P(←→G ) for any graph G.396
Proof. P(G) is the minimum number of induced paths of G and P(←→G ) is the397
minimum number of Hessenberg paths in←→G . It is enough to show that all Hessenberg398
paths in←→G are induced. Suppose P is a Hessenberg path in
←→G that is not induced.399
Then there exists some arc (vi, vj) ∈ E(←→G ), where i > j+1. But because the digraph400
is doubly directed, (vj , vi) ∈ E(←→G ), which contradicts the definition of a Hessenberg401
path. Therefore, all Hessenberg paths must be induced. Thus, P(G) = P(←→G ).402
Proposition 3.2. Z(G) = Z(←→G ) for any graph G.403
Proof. The color change rule for graphs is that a blue vertex v can force a white404
vertex w if w is the only white neighbor of v. The color change rule for digraphs is405
that a blue vertex V can force a white vertex w if w is the only white out-neighbor406
of v. If G is a graph then for any vertex v ∈ V (G), w is a neighbor of v in G if and407
only if w is an out-neighbor of v in←→G . This means that v forces w in G if and only408
if v forces w in←→G . Thus B is a zero forcing set in G if and only if B is a zero forcing409
set in←→G and Z(G) = Z(
←→G ).410
Observation 3.3. For any graph G, M(G) ≤ M(←→G ), since S(G) ⊆M(
←→G ).411
Corollary 3.4. If G is a graph such that M(G) = Z(G), then M(G) = M(←→G ).412
13
Proof. Z(G) = Z(←→G ) ≥ M(
←→G ) ≥ M(G) = Z(G).413
It was established in [1] that for every tree T , P(T ) = M(T ) = Z(T ), giving the414
following corollary.415
Corollary 3.5. If T is a tree, then P(←→T ) = M(
←→T ) = Z(
←→T ).416
As the following example shows, it is possible to have M(←→G ) > M(G).417
Example 3.6. The complete tripartite graph on three sets of three vertices K3,3,3418
has V1 = {1, 2, 3}, V2 = {4, 5, 6}, V3 = {7, 8.9}, V (K3,3,3) = V1∪V2∪V3 and E(K3,3,3)419
equal to the set of all edges with one vertex in Vi and the other in Vj (i 6= j). It is420
well-known that mr(K3,3,3) = 3. Let J3 be the 3× 3 matrix with all entries equal to421
1, 03 be the 3 × 3 matrix with all entries equal to 0, and let A =
03 J3 −J3J3 03 J3J3 J3 03
.422
Then Γ(A) =←−−→K3,3,3 and rankA = 2. Thus, M(
←−−→K3,3,3) = 7 > 6 = M(K3,3,3).423
The pentasun H5 graph shown in Figure 3.1 has M(H5) = 2 < 3 = P(H5) [4],424
establishing the noncomparability of M and P (because there are many examples of425
graphs G with P(G) < M(G)). The same is true for the doubly directed pentasun.426
v
w
Fig. 3.1. The pentasun H5.
Example 3.7. Theorem 2.8 of [5] describes the cut-vertex reduction method for427
calculating M for directed graphs with a cut-vertex. We compute M(←→H5) = 2 by428
applying the cut-vertex reduction method to vertex v, using the notation found in429
[5]. Because M(H5−w) = Z(H5−w) = 2 and M(H5−{v, w}) = Z(H5−{v, w}) = 2,430
M(←→H5−w) = Z(
←→H5−w) = 2 and M(
←→H5−{v, w}) = Z(
←→H5−{v, w}) = 2, so rv(
←→H5−w) =431
1. Clearly mr(←→H5[{v, w}]) = 1 and mr(
←→H5[{v, w}] − v) = 0, so rv(
←→H5[{v, w}]) = 1.432
The type of the cut-vertex v of a digraph Γ, typev(Γ), is a subset of {C,R}, where433
C ∈ typev(Γ) if there exists a matrix A′ ∈ M(Γ − v) with rankA′ = mr(Γ − v) and434
a vector z in rangeA′ that has the in-pattern of v, and similarly for rows. Thus,435
typev(←→H5[{v, w}]) = ∅, so by [5, Theorem 2.8], rv(
←→H5) = 2. So mr(
←→H5) = mr(
←→H5 −436
{v, w})+mr(←→H5[{w}])+2 = 6+0+2 = 8 and M(
←→H5) = 2. Since P(
←→H5) = P(H5) = 3,437
P(←→H5) > M(
←→H5). It is easy to find an example of a digraph Γ with P(Γ) < M(Γ), e.g.438
Example 2.13, so M and P are noncomparable.439
14
Proposition 3.8. Suppose that both G and←→G have field independent minimum440
rank. Then mr(G) = mr(←→G ) and M(G) = M(
←→G ).441
Proof. Since both G and←→G have field independent minimum rank mr(G) =442
mrZ2(G) and mrZ2(←→G ) = mr(
←→G ). Furthermore, SZ2(G) =MZ2(
←→G ), so443
mr(G) = mrZ2(G) = mrZ2(←→G ) = mr(
←→G ).444
445
The converse of Proposition 3.8 is not true, however.446
Example 3.9. Let G be the full house graph, shown in Figure 3.2. It is well447
known that mrZ2(G) = 3, yet mr(G) = 2 = mr(←→G ).448
1
23
4 5
Fig. 3.2. The full house graph
4. Digraphs in general. In this section, we present some minimum rank, maxi-449
mum nullity, and zero forcing results for digraphs in general, where any pair of vertices450
may or may not have an arc in either direction. We begin with two (undirected) graph451
properties that do not extend to digraphs.452
Sinkovic [11] has shown that for any outerplanar graph G, M(G) ≤ P(G). This is453
not true for digraphs, because it was was shown in Example 2.13 that the outerplanar454
digraph_GR has M(
_GR) = Z(
_GR) = 3 > 2 = P(
_GR), and
_GR is outerplanar (although455
Figure 2.2 is not drawn that way).456
The complement of a graph G = (V,E) (or digraph Γ = (V,E)) is the graph457
G = (V,E) (or digraph Γ = (V,E)), where E consists of all two element sets of458
vertices (or all ordered pairs of distinct vertices) that are not in E. The Graph459
Complement Conjecture (GCC) is equivalent to the statement that for any graph G,460
M(G) + M(G) ≥ |G| − 2. This statement is generalized in [3]: For a graph parameter461
β related to maximum nullity, the Graph Compliment Conjecture for β, GCCβ , is462
β(G) +β(G) ≥ |G| − 2. With this notation, GCC can be denoted GCCM. The Graph463
Compliment Conjecture for zero forcing number, Z(G) + Z(G) ≥ |G| − 2, denoted464
GCCZ, is actually the Graph Complement Theorem for zero forcing [7]. However,465
as the following example shows, GCCZ does not hold for digraphs, and since for any466
15
digraph M(Γ) ≤ Z(Γ), GCCM does not hold for digraphs. A tournament provides a467
counterexample.468
Example 4.1. For the Hessenberg tournament of order n, denoted ~K(H)n ,469
Z( ~K(H)n ) = 1 because ~Hn is a Hessenberg path. Because ~K
(H)n is self-complementary,470
Z( ~K(H)n ) + Z( ~K
(H)n ) = 2,471
but for n ≥ 5, n− 2 = | ~K(H)n | − 2 ≥ 3.472
Some properties of minimum rank for graphs do remain true for digraphs. For a473
graph G, it is well-known that if Kr is a subgraph of G then M(G) ≥ r − 1 (see, for474
example, [2] and the references therein). An analogous result holds true for digraphs,475
although the proof is different than those usually given for graphs.476
Theorem 4.2. Suppose←→Kr is a subgraph of a digraph Γ. Then M(Γ) ≥ r − 1.477
Proof. First, we order the vertices of Γ so that the subdigraph induced on the478
vertices 1, 2, . . . , r is←→Kr. We will construct L ∈ M(Γ) with rankL ≤ n − r + 1,479
where L is partitioned as L =
[A B
C D
]and A ∈ M(Kr). We may choose D ∈480
M(Γ[{r + 1, . . . , n}]) so that rankD = n − r. We now choose C to be any matrix481
with the correct zero-nonzero pattern. Denote the ith column of C by ci and the jth482
column of D by dj . Since D has full rank, there exist coefficients di,1, . . . , di,n−r such483
that ci = di,1d1 + · · ·+ di,n−rdn−r for 1 ≤ i ≤ r.484
Now, we choose B to be any matrix with the correct zero-nonzero pattern and485
denote the jth column of B by bj . Then define E to be the r × r matrix whose ith486
column is equal to di,1b1 + · · · + di,n−rbn−r. Therefore, the matrix L′ =
[E B
C D
]487
has rankL′ = n− r. Let p be a real number greater than the absolute value of every488
entry of E. Define A := E+ pJr, where Jr is the r× r matrix all of whose entries are489
1, so A ∈M(Kr), L =
[A B
C D
]∈M(Γ) and rankL ≤ n− r + 1.490
In [6], the maximum number of edges in a graph G of order n with a prescribed491
zero forcing number k is shown to be kn −(k+12
). We similarly seek to bound the492
number of arcs that a digraph Γ of order n may possess given Z(Γ) = k.493
Theorem 4.3. Suppose Γ is a digraph of order n with Z(Γ) = k. Then,494
|E(Γ)| ≤(n
2
)−(k
2
)+ k(n− 1). (4.1)495
496
Proof. We prove that (4.1) holds for a digraph Γ of order n ≥ k + 1 whenever497
Z(Γ) ≤ k (since for a graph G of order n, Z(G) ≤ n − 1). The proof is by induction498
on n, for a fixed positive integer k. The base case is n = k+ 1, or k = n− 1, and the499
16
inequality (4.1) reduces to500
|E(Γ)| ≤(n
2
)−(n− 1
2
)+(n−1)2 =
n(n− 1)
2− (n− 1)(n− 2)
2+(n−1)2 = n(n−1).501
Any digraph Γ of order n has at most 2(n2
)= n2 − n arcs and thus the claim holds.502
Now assume (4.1) holds true for any digraph of order n−1 that has a zero forcing503
set of cardinality k. Let Γ be a digraph of order n with Z(Γ) ≤ k. Let B be a504
zero forcing set of Γ where |B| = k. Suppose F is a chronological list of forces for505
B and that the first force of F occurs on the arc (v, w). Then, (B \ {v}) ∪ {w} is506
a zero forcing set of cardinality k for Γ − v and the induction hypothesis applies to507
E(Γ−v). In order to determine an upper bound on |E(Γ)| we determine the maximum508
number of arcs incident with v in Γ. Since v forces w first in F , (v, x) ∈ E(Γ) implies509
x ∈ (B \{v})∪{w}. Furthermore, E(Γ) contains at most n−1 arcs of the form (x, v),510
one for each vertex x 6= v. Therefore,511
|E(Γ)| ≤ |E(Γ− v)|+ k + (n− 1) ≤(n− 1
2
)−(k
2
)+ k(n− 2) + k + (n− 1)512
=
(n
2
)−(k
2
)+ k(n− 1).513
514
In the paper [6], the edge bound is used to show that the zero forcing number515
must be at least half the average degree. However, a Hessenberg tournament (cf.516
Example 2.15) has half of all possible arcs and Z( ~Hn) = 1, so the analogous result is517
not true for digraphs, and any correct result of this type for digraphs is not likely to518
be useful.519
For a digraph Γ where Z(Γ) = k, Theorem 4.3 gives an upper bound for the520
number of arcs Γ may possess. However, the proof also suggests that equality is521
achievable in (4.1) when n > k. The following provides a construction of a class of522
digraphs for which (4.1) is sharp.523
Theorem 4.4. Let k be a fixed positive integer. Then for each integer n > k and524
each partition π = (n1, . . . , nk) of n, there exists a digraph Γn,k,π of order n for which525
Z(Γn,k,π) = k, the forcing chains of Γn,k,π have lengths n1, n2, . . . , nk respectively,526
and527
|E(Γn,k,π)| =(n
2
)−(k
2
)+ k(n− 1).528
529
Proof. For 1 ≤ i ≤ k, let Γi be a full Hessenberg path on ni vertices. Among all530
the Γi, there are a total ofk∑i=1
[(ni − 1) +
(ni
2
)]arcs. Within each Γi we denote the531
initial vertex of the Hessenberg path by bi and the terminal vertex of the Hessenberg532
17
path by ti. We define B = {bi : 1 ≤ i ≤ k} and T = {ti : 1 ≤ i ≤ k}, and note that533
B and T will intersect if ni = 1 for some i. To create Γn,k,π we start withk⋃i=1
Γi and534
add arcs between the Γi in the following manner:535
(1) For 1 ≤ j < i ≤ k, we add all arcs from vertices in Γi to vertices in Γj . This536
adds a total of∑i<j
ninj arcs.537
(2) Add all arcs from vertex ti to vertices in other Γj . For each i, this adds538 ∑j 6=i nj = n−ni arcs. Over all i, this adds a total of kn−
∑ki=1 ni = (k−1)n539
arcs.540
Some arcs have been double-counted, which must be reflected in the overall total.541
In particular, arcs from ti to all vertices in Γj (for j < i) have been double-counted.542
For an arc from ti to a vertex v of Γj where v 6= tj , we replace the double-counted543
arc by an arc from v to bi. Therefore, we need only remove from the total count the544
number of arcs from tj to ti for 1 ≤ i < j ≤ k. There are a total of(k2
)such arcs.545
Thus, we have546
|E| =k∑i=1
[(ni − 1) +
(ni2
)]+
∑1≤i<j≤k
ninj + (k − 1)n−(k
2
)547
= n− k +
k∑i=1
(ni2
)+
∑1≤i<j≤k
ninj
+ kn− n−(k
2
)548
=
(n
2
)−(k
2
)+ k(n− 1).549
By Theorem 4.3, we know that Z(Γn,k,π) ≥ k. We claim that B is a zero forcing550
set for Γn,k,π and that a chronological list of forces exists for which the forcing chains551
have lengths n1, . . . , nk respectively. Assume that each vertex of B is blue. Γ1 is a552
Hessenberg path and the only arcs coming from vertices of Γ1 point to vertices in B553
with the exception of arcs coming from t1. Thus, forcing may occur along Γ1, where554
t1 is the last vertex forced. We then proceed to Γ2 and so on through all Γi. When555
we get to Γi, the only arcs coming from vertices of Γi point to vertices in B or to556
the already blue vertices of Γh where h < i, with the exception of arcs coming from557
ti (which is not used to perform a force). So, forcing may occur along Γi until all558
vertices are blue. Therefore B is a zero forcing set for Γn,k,π and Z(Γn,k,π) = k.559
Although the digraph Z(Γn,k,π) achieve equality in the bound (4.3), these are not560
the only digraphs that do so.561
Example 4.5. Let Γ be the digraph of order n = 8 in Figure 4.1. Note that562
{1, 5} is a zero forcing set of Γ. Since deg+(v) ≥ 2 for all vertices v, k = Z(Γ) = 2.563
Also note that Γ has 41 arcs, the same number of arcs as each digraph Γ8,2,π.564
In all of the digraphs Γn,k,π constructed in Theorem 4.4, the forcing process may565
be completed one forcing chain at a time, and we show that this is not true for Γ.566
18
1 2 4
5 6 7 8
3
Fig. 4.1. A digraph with the maximum number of arcs for its zero forcing number.
Since deg+(v) ≥ 3 for all vertices v other than vertex 1, vertex 1 must be contained in567
any minimum zero forcing set of Γ along with one of the two out-neighbors of vertex568
1. Therefore, the only minimum zero forcing sets are B1 = {1, 5} and B2 = {1, 2}. If569
we color the vertices of B1 blue, then the first three forces must occur along the arcs570
(1, 2), (2, 3), and (5, 6), in that order. If we color the vertices of B2 blue, then the first571
three forces must occur along the arcs (1, 5), (2, 3), and (5, 6), in that order. In either572
case, neither of the two forcing chains is completely blue before the forcing process573
must begin on the other. Therefore, Γ is not equal to any of the Γn,2,π constructed574
in Theorem 4.4.575
Although M(Γ) does not necessarily equal Z(Γ) for all digraphs Γ, we get equality576
for all Γn,k,π constructed in the proof of Theorem 4.4.577
Proposition 4.6. If k and n are positive integers where k < n and Γn,k,π is one578
of the digraphs constructed in the proof of Theorem 4.4, then M(Γn,k,π) = Z(Γn,k,π).579
Proof. We adopt the notation and definitions used in the proof of Theorem 4.4.580
By construction, each of the k vertices in T has an arc to every other vertex of Γn,k,π.581
So for all vertices v 6∈ T , deg−(v) ≥ k. Now we consider ti ∈ T . If ti 6= bi, then582
there is an arc to ti from another vertex of Γi. There are also arcs to ti from all583
other vertices of T and therefore deg−(ti) ≥ k. We now consider the case where584
ti = bi. By construction, the subgraph induced on B is←→Kk and thus there is an585
arc to ti from each of the other k − 1 vertices of B. Furthermore, since n > k586
there is a vertex tj ∈ T for which tj 6= bj . By construction, there is also an arc587
from tj to ti and therefore deg−(ti) ≥ k. This implies that δ−(Γn,k,π) ≥ k. Since588
M(Γn,k,π) ≥ max{δ−(Γn,k,π), δ+(Γn,k,π)} [5], k ≤ M(Γn,k,π) ≤ Z(Γn,k,π) ≤ k, and so589
M(Γn,k,π) = Z(Γn,k,π).590
For k = n−1, Γn,k,π is the digraph←→K n, so for n ≥ 4, P(Γn,k,π) =
⌈n2
⌉< n−1 =591
Z(Γn,k,π).592
19
REFERENCES593
[1] AIM Minimum Rank – Special Graphs Work Group (F. Barioli, W. Barrett, S. Butler, S. M.594
Cioaba, D. Cvetkovic, S. M. Fallat, C. Godsil, W. Haemers, L. Hogben, R. Mikkelson,595
S. Narayan, O. Pryporova, I. Sciriha, W. So, D. Stevanovic, H. van der Holst,K. Vander596
Meulen, A. Wangsness). Zero forcing sets and the minimum rank of graphs. Linear Algebra597
and its Applications, 428: 1628–1648, 2008.598
[2] F. Barioli, W. Barrett, S. Fallat, H.T. Hall, L. Hogben, B. Shader, P. van den Driessche, and599
H. van der Holst. Parameters related to tree-width, zero forcing, and maximum nullity of600
a graph. J. Graph Theory, 72: 146–177, 2013.601
[3] F. Barioli, W. Barrett, S. Fallat, H. Hall, L. Hogben, and H. van der Holst. On the graph602
complement conjecture for minimum rank, Linear Algebra and its Applications, 436: 4373–603
4391, 2012.604
[4] F. Barioli, S. Fallat, L. Hogben. Computation of minimal rank and path cover number for605
certain graphs. Linear Algebra and its Applications, 392: 289–303, 2004.606
[5] A. Berliner, M. Catral, L. Hogben, M. Huynh, K. Lied, M. Young. Minimum rank, maximum607
nullity, and zero forcing number for simple digraphs. Electronic Journal of Linear Algebra,608
26: 762–780, 2013.609
[6] S. Butler and M. Young. Throttling zero forcing propagation speed on graphs. Australasian610
Journal of Combinatorics, 57: 65-71, 2013.611
[7] J. Ekstrand, C. Erickson, D. Hay, L. Hogben, and J. Roat. Note on positive semidefinite612
maximum nullity and positive semidefinite zero forcing number of partial 2-trees. Electronic613
Journal of Linear Algebra, 23: 79–97, 2012.614
[8] S. Fallat and L. Hogben. The minimum rank of symmetric matrices described by a graph: a615
survey. Linear Algebra and its Applications, 426: 558–582, 2007.616
[9] S. Fallat and L. Hogben. Minimum Rank, Maximum Nullity, and Zero Forcing Number of617
Graphs. In Handbook of Linear Algebra, 2nd Edition, L. Hogben, Editor, CRC Press, Boca618
Raton, 2014.619
[10] L. Hogben. Minimum rank problems. Lin. Alg. Appl., 432: 1961–1974, 2010.620
[11] J. Sinkovic. Maximum nullity of outerplanar graphs and the path cover number. Linear Algebra621
and its Applications, 432: 2052–2060, 2010.622
[12] N. Warnberg. Zero forcing number of simple digraphs. https://github.com/warnberg/zero_623
forcing_digraph.624
20