See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/349823598

Integer Programming Formulations for Minimum Spanning Tree Interdiction

Article in INFORMS Journal on Computing · March 2021

DOI: 10.1287/ijoc.2020.1018

CITATIONS READS
15 480

3 authors, including:

Ningji Wei Jose L. Walteros

Texas Tech University University at Buffalo, The State University of New York
15 PUBLICATIONS 42 CITATIONS 40 PUBLICATIONS 497 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Ningji Wei on 08 March 2021.

The user has requested enhancement of the downloaded file.

Integer Programming Formulations for Minimum Spanning Tree Interdiction
Ningji Wei, Jose L. Walteros∗,
Department of Industrial and Systems Engineering
University at Buffalo
[email protected] [email protected]

Foad Mahdavi Pajouh

Department of Management Science and Information Systems
University of Massachusetts Boston
[email protected]

Abstract
We consider a two-player interdiction problem staged over a graph where the leader’s objective is to minimize
the cost of removing edges from the graph so that the follower’s objective, i.e., the weight of a minimum
spanning tree in the residual graph, is increased up to a predefined level r. Standard approaches for graph
interdiction frame this type of problems as bi-level formulations, which are commonly solved by replacing the
inner problem by its dual to produce a single level reformulation. In this paper we propose an alternative
integer program derived from the combinatorial structure of the follower’s solution space and prove that this
new formulation yields a stronger linear relaxation than the bi-level counterpart. We also study the convex
hull of the feasible solutions of the problem and identify several families of facet-defining inequalities that
can be used to strengthen the proposed integer program. We then proceed by introducing an alternative
formulation defined by a set of so-called supervalid inequalities that may exclude feasible solutions, albeit
solutions whose objective value is not better than that of an edge cut of minimum cost. We discuss several
computational aspects required for an efficient implementation of the proposed approaches. Finally, we per-
form an extensive set of computational experiments to test the quality of our approaches by solving a large
collection of real-life and randomly generated instances with various configurations, analyzing and comparing
the benefits of each model, and also identifying further enhancements.

Keywords: Network Interdiction, Minimum Spanning Tree, Minimum Edge Blocker Problems, Bi-level
Optimization.

1 Introduction and Motivation

We study an attacker-defender interdiction problem staged over a graph where two decision makers—an at-
tacker who plays first and a defender who plays second—sequentially solve interrelated problems that optimize
conflicting objective functions. In this particular context, the objective of the attacker is to minimize the cost
of inflicting a given disruption level to the defender’s objective by removing edges from the graph. In turn, the
objective of the defender is to identify a minimum spanning tree that remains available in the residual graph.
Graph interdiction has been an ongoing research endeavor for both academicians and practitioners for
many years (Corley and Sha 1982, Israeli and Wood 2002, Kennedy et al. 2011). Despite having their root
in military applications (Ghare et al. 1971, Wood 1993), these models have recently gained traction in many
other areas thanks to their ability of informing decision makers about vulnerabilities on infrastructure and/or
∗
Corresponding author. Phone: 716-645-8876; Fax: 716-645-3302; Address: 413 Bell Hall, Buffalo, NY 14260; Email: jose-
[email protected]

1
operational networks. Therefore, it is common to find graph interdiction applications in areas as diverse as
homeland security (Grubesic et al. 2008, Houck et al. 2004), evacuation planning (Matisziw and Murray 2009),
immunization strategies (Tao et al. 2005), energy systems (Salmeron et al. 2004), communications (Dinh et al.
2012) and transportation (Jenelius et al. 2006), among others.
As attacker-defender problems keep permeating the applied literature, many researchers have focused on
developing general models to study the disruption of different structures and operational properties of graphs
that are amenable for modeling a wide variety of applications (Lozano and Smith 2016, Chestnut and Zenklusen
2017). Among all these models, the ones designed to disrupt optimal solutions of network flow problems, such
as shortest paths, maximum flow, and minimum cost flow problems (Corley and Sha 1982, Israeli and Wood
2002, Kennedy et al. 2011, Lim and Smith 2007, Matisziw and Murray 2009, Wollmer 1964, Wood 1993) are
the ones that have received the most attention. These problems have gained great notoriety because the wide
range of applications that can be modeled with them, but also because most of such underlying network flow
problems admit linear formulations that permit applying traditional dualize-and-combine solution strategies
(Cormican et al. 1998).
A second area that has gained some traction lately deals with a specific type of problems often called critical
elements detection problems that study vulnerabilities of graphs with respect to connectivity and cohesiveness
properties, such as the number of connected vertex pairs, the size of the largest connected component, and the
number of connected components (Addis et al. 2013, Arulselvan et al. 2009, Di Summa et al. 2012, Dinh et al.
2012, Oosten et al. 2007, Shen et al. 2012, Veremyev et al. 2014). Applications of these problems are often
used to find vulnerable elements in communication networks, or—in social network analysis—to identify “key
players” used to propagate ideas like advertising campaigns over social networks.
Lastly, other variations that are worth mentioning are interdiction problems aimed to disrupt the sizes or
relative weight of some topological subsets of vertices or edges, like spanning trees, dominating sets, central
vertices (e.g., the 1-median or 1-center vertices), matchings, independent sets, cliques, and vertex covers (Bazgan
et al. 2010, 2011, Frederickson and Solis-Oba 1999, Furini et al. 2019, Mahdavi Pajouh et al. 2014, 2015,
Zenklusen 2010).
Among the different structures mentioned above, we focus our attention on interdicting minimum spanning
trees. In general, spanning trees are among the most widely studied graph structures having both theoretical
and practical properties often used to derive insights in many problem settings (Gabow et al. 1986, Magnanti
and Wolsey 1995). Given their ability to identify low-cost connected subgraphs on targeted networks, spanning
trees are often used for designing computer, telecommunications, transportation and infrastructure networks
(Graham and Hell 1985); for building optimal circuits (Ohlsson et al. 2004); for handwriting recognition (Tapia
and Rojas 2004), image registration and feature extraction in computer vision (Suk and Song 1984); and for
social networks analysis (Ströele A. Menezes et al. 2008). Modeling spanning tree interdiction games can
therefore help to analyze graph resiliency properties in a wide variety of contexts.
Traditional attacker-defender interdiction problems are typically modeled from two different perspectives
depending on the way in which the interaction between the attacker’s and defender’s objectives is defined.
From the first perspective, the attacker is provided with an interdiction budget for attacking elements in the
graph—either vertices or edges—and its objective is to design an optimal attack that maximizes the overall
disruption on the defender’s objective. Alternatively, from the second perspective, the attacker minimizes the
cost of attacking the graph so as to guarantee that a minimum required disruption level is inflicted over the
defender’s objective. In this paper we develop solution methods to tackle the latter version.

1.1 Previous Work

Early approaches for attacking minimum spanning trees on graphs focused on finding a single edge whose
deletion results in the largest increase of the spanning tree’s weight. This variation is often called the most
vital edge with respect to the minimum spanning tree problem and is proven to be solvable in polynomial time
(Hsu et al. 1991). Several approaches, including sequential (Hsu et al. 1991, Iwano and Katoh 1993), randomized
(Shen 1995), and parallelized (Suraweera et al. 1995) algorithms have been proposed over the years to tackle

2
this particular variation. These algorithms are mainly adaptations of traditional methods for finding minimum
spanning trees that are also able to identify the most vital edge while preserving their original complexity.
The natural extension that attempts to find k-most vital edges for k > 1 has received significantly more
attention over the past few years and, contrary to the single edge version, it has been proven to be NP-hard in
the strong sense (Lin and Chern 1993, Frederickson and Solis-Oba 1999). Among all the techniques that have
been developed to solve this problem, algorithms with an approximation guarantee and exact approaches have
been the preferred choices by most researchers.
Frederickson and Solis-Oba (1999) were the first to propose an approximation algorithm for this problem
that yields a proven O(log k) approximation for any given k > 1. Furthermore, for the case in which there is
a cost associated with the removal of each edge and a deletion budget, their algorithm produces a O(log m)
approximation (here, n and m represent the number of vertices and edges of the graph, respectively). Almost
two decades later, Zenklusen (2015) was the first to discover a O(1)-approximation algorithm for the general cost
version of this problem. This 14-approximation algorithm was the first approach proven to yield a constant
approximation. More recently, Linhares and Swamy (2017) improved this approximation ratio to 4, after
introducing some further enhancements to Zenklusen’s algorithm.
As for exact approaches, Liang and Shen (1997) were the first to develop an exact algorithm based on total
enumeration for the k-most vital edges problem with an overall running time of O(nk+1 ). Subsequently, Bazgan
et al. (2012) proposed an improved combinatorial branch-and-bound algorithm that runs in O(nk log α(2(n −
1), n)) time for any fixed k, where α(·, ·) is the inverse of the Ackerman function. They also introduced the
first integer program inspired by similar formulations used to solve other interdiction problems (Cormican et al.
1998).
A closely related variation of the k-most vital edges for the spanning tree problems is called the minimum
edge blocker spanning tree problem, which is the problem of finding the minimum number of edges whose
removal from G results in a graph with spanning trees with a weight no less than a constant r. The relationship
between this and the k-most vital edge problem has been studied by Bazgan et al. (2013). The attacker-defender
game that we study in this paper is a generalization of this edge blocker problem.
In addition to the above discrete versions in which the attacker completely removes some edges, there is
also another continuous variation where the attacker attempts to find a minimum cost increase of the edge
weights so that the value of all minimum spanning trees is equal to some target value. Frederickson and Solis-
Oba (1999) studied the case in which the cost of increasing the weights is linear and nondecreasing. They
proved that this version of the problem is solvable in polynomial time and provided an algorithm that runs
in O(n3 m2 log(n2 /m)) time. Baı̈ou and Barahona (2008) proposed a linear programming formulation for this
problem that can be used to tackle the case in which the costs are nondecreasing piecewise-linear.

1.2 Our Contributions

Despite the wide interest in these problems, most of the available literature is focused on the k-vital edge case
instead of the more general version that also accounts for edge removal costs. Similarly, with the exception of
Bazgan et al. (2013), most efforts have been concentrated in the version that maximizes the increment of the
weight of the minimum spanning tree given a predefined attacking budget, and very little attention has given to
the version that minimizes the interdiction costs given a desired disruption level. Furthermore, there have been
only a few approaches that use integer programming techniques and, to the best of our knowledge, there has
not been any study that analyzes the convex hull of the corresponding solution space for this problem. In this
paper, we are interested in addressing some of the aforementioned gaps by developing an integer programming
framework aimed to optimally solve minimum spanning tree interdiction problems. A summary of the four
main contributions of this paper follows:

1. We provide a detailed study of the problem’s solution space including: a characterization of the set
of feasible strategies for the attacker; an integer programming formulation derived directly from such a
characterization; a polyhedral analysis of the convex hull of feasible solutions; a collection of facet-defining

3
 inequalities used to strengthen the given integer program; and an analytical comparison of the strength
of other traditional models with respect to the proposed integer program.

2. We provide an alternative formulation defined by a set of so-called supervalid inequalities that may exclude
feasible solutions, albeit solutions whose objective value is not better than that of an edge cut of minimum
cost. We show that this alternative formulation is not only stronger than the first formulation, but also
requires fewer constraints to be defined.

3. We develop efficient algorithms to solve these formulations under a traditional branch-and-cut framework
and provide the specific details required for the proper implementation of our methods. We also made
our implementation publicly available to provide other researchers with a direct benchmark.

4. We test our models on a wide variety of networks arising from real-life applications as well as on randomly
generated graphs to compare their performance and identify potential improvements.

The rest of the paper is organized as follows. In Section 2, we provide a general definition of the problem
and discuss its computational complexity. In Section 3, we describe two mathematical formulations, the first
derived from a dualization approach commonly used to formulate interdiction problems, and the second derived
from the characterization of the set of feasible strategies for the attacker. In Section 4, we provide a polyhedral
study of the convex hull of the feasible solutions. In Section 5, we exploit some fundamental properties of
the feasible solutions to produce a strong formulation defined by a set of so-called supervalid inequalities. In
Section 6, we study the computational performance of the proposed formulations. Finally, in Section 7, we
provide our concluding remarks. To streamline the discussion of the paper, we compiled the proofs of all the
non-trivial mathematical statements introduced hereafter in the Online Supplement, Appendix A.

2 Problem Definition
We work with a simple, nonempty, connected graph G = (V, E), where V is a set of n vertices and E ⊆ V2

a set of m edges. We will refer to each edge by its index e = 1, . . . , m or by its unordered pair of endpoints
{i, j}. We associate edge set E with two rational vectors w = [we ]e∈E and c = [ce ]e∈E , each comprising what
E
we denote the weights and costs of the edges, respectively.
E
P A weight function P w(·) : 2 → Q and cost function
c(·) : 2 → Q are defined accordingly, so that w(S) = e∈S we and c(S) = e∈S ce , for any given edge subset
S ⊆ E. A subset of edges S ⊆ E is called an edge cut of G if graph G(V, E \ S) is disconnected. The size of the
smallest edge cut of a graph G is known as its edge connectivity and is denoted by λ(G). Similarly, the cost
of a minimum-cost edge cut in G with respect to the cost vector c is denoted by λc (G). The value of λc (G),
and consequently of λ(G), can be found in polynomial time for any graph G by solving a sequence of max-flow
problems between the different pair of vertices in G (Gomory and Hu 1961).
We say that H is a subgraph of G, if the vertices and edges in H, denoted by V (H) and E(H), are subsets
of V and E, respectively. A subgraph T of G is called a tree if every pair of vertices in V (T ) is connected by a
unique path in T . Furthermore, T is called a spanning tree if it includes all the vertices in G, i.e., V (T ) = V . We
use ΩG to denote the set of all spanning trees in G. For a given edge set S ⊆ E, let ΩG [S] := {T ∈ ΩG |S ⊆ E(T )}
be the set of all the spanning trees in G that contain all the edges in S. Whenever set S is a singleton (i.e.,
S = {e} for a given e ∈ E), we simply use ΩG [e]. The weight of a spanning tree T ∈ ΩG is denoted by w(E(T )),
or simply w(T ). We call the spanning trees of minimum and maximum weight minimum spanning trees and
maximum spanning trees, and denote their weight as wG and wG , respectively. Both minimum and maximum
spanning trees of G can be computed in polynomial time (Ahuja et al. 1993). If graph G is disconnected, we
assume that ΩG = ∅, and consequently say that both wG and wG are infinity.
We provide a summary of the notation used throughout the paper as a reference table in the Online
Supplement, Appendix B.

4
2.1 Attacking Spanning Trees
Formally, the spanning tree interdiction problem that we study in this paper is defined as follows. Given a
graph G = (V, E), cost and weight vectors c and w in Qm , respectively, and a scalar r ∈ Q, the problem is
aimed to find a subset of edges S ⊆ E such that the weight of any spanning tree in G(V, E \ S) is at least r
and the cost c(S) is minimum. That is,

min c(S) (1)

S⊆E
s.t. min {w(T )} ≥ r. (2)
T ∈ΩG(V,E\S)

Importantly, other interdiction problems with a similar structure for which the attacking cost c is equal to
1 are often called in the literature Edge Blocker Problems (Bazgan et al. 2013, Mahdavi Pajouh et al. 2014,
2015). We will resort however to the term spanning tree interdiction (as in Zenklusen (2015)) given the general
association with many other attacker-defender games of similar nature, where non-unitary vertex/edge removal
costs are introduced.
Before addressing its computational complexity, we will discuss some basic characteristics of the problem.
First, we will assume without loss of generality that the edge weights are nonnegative (i.e., w ≥ 0). Note that
any instance with negative weights can be trivially transformed into an instance with nonegative weights; one
just need add a sufficiently large constant M to all the weights to make them positive and then set r + (n − 1)M
as the interdiction bound. Furthermore, we will also assume without loss of generality that the edge costs are
positive (i.e., c > 0). Notice that any edge with a negative deletion cost will be part of any optimal solution
and since removing any edge with deletion cost zero does not affect the objective function, one might as well
directly remove it from G. Thus, we can set any edge with non-positive cost as part of the attacker’s strategy
and solve the resulting problem in which all the edges have a positive cost.
Second, for a given edge e ∈ E, notice that if the minimum weight among all the spanning trees in ΩG [e]
is greater than or equal to r, edge e is in fact irrelevant for the attacker and will never be selected as part of
any optimal solution. Then, one may equivalently solve the problem over G(V, E \ {e}) instead. Furthermore,
checking whether there exists an irrelevant edge can be trivially done in polynomial time by performing m calls
to a minimum spanning tree algorithm. We will assume then that set E contains no irrelevant edges.
Third, independently of the value of r, we note that the problem is always feasible because removing any
subset of edges that disconnects the graph (i.e., an edge cut) is sufficient to raise the weight of the spanning
trees in the resulting graph to infinity. As a consequence, the optimal objective function value of the problem
is bounded above by λc (G). Furthermore, if the value of r is greater than the largest weight of a spanning tree
in ΩG , the optimal solution of the problem is λc (G) because the set of feasible solutions will simply be the set
of edge cuts of G. Similarly, if the value of r is less than or equal to the weight of a minimum spanning tree
in ΩG , the empty set is a feasible solution to this problem. To avoid the aforementioned trivial cases, we will
henceforth assume that the required disruption level r is strictly sandwiched between wG and wG .

2.2 Computational Complexity

In this section, we discuss the complexity of spanning tree interdiction problems and briefly address the rela-
tionship with other well-studied variations. We begin by defining the decision (or recognition) version of this
problem.

minimum spanning tree interdiction (Decision)

Instance: A graph G = (V, E), a non-negative rational vector of edge weights w, a positive rational vector of
edge costs c, and two rational numbers r and b (denoted by hG(V, E), w, c, r, bi).

Question: Is there a subset of edges S ⊆ E with edge cost c(S) less than or equal to b such that the weight
of a minimum spanning tree in graph G(V, E \ S) is greater than or equal to r?

5
Theorem 1 The minimum spanning tree interdiction problem is strongly NP-complete (Bazgan et al. 2013,
Frederickson and Solis-Oba 1999).

The NP-completeness of this problem should come at no surprise because it is clearly a generalization of the
decision version of the k-most vital edges for the minimum spanning tree (Frederickson and Solis-Oba 1999)
and the edge blocker spanning tree problems (Bazgan et al. 2013), which both assume c = 1. Since both
problems are known to be NP-hard, even for graphs with weights restricted to 0 or 1, the decision version of
the minimum spanning tree interdiction problem is NP-complete too. Furthermore, the fact that this problem
is NP-complete directly renders the two optimization versions we discussed previously NP-hard as well.
When the costs of removing the edges are all equal to one (i.e., c = 1), Bazgan et al. (2013) also proved
that both of these optimization problems are polynomial-time equivalent, which is expected given that both
problems arise from the same decision problem. Naturally, this observation can be easily extended for the case
in which c ∈ Qm by noting that if an efficient algorithm exists for solving any of the two variations, one can
immediately solve instances of the other via binary search.

3 Mathematical Formulations
In this section we study two formulations for the proposed spanning tree interdiction problem. We begin our
discussion by presenting a formulation that is derived directly from (1)–(2). This formulation is an adaptation
from one originally introduced by Bazgan et al. (2012) to model the k-most vital edges variation of the prob-
lem, and is solved via a reformulation technique commonly used to tackle interdiction problems in which the
defender’s decisions are modeled as linear programs (Cormican et al. 1998, Lim and Smith 2007, Wood 1993).

3.1 Bi-level Mixed-Integer Programming Formulation

The minimum spanning tree problem is known to admit several compact, linear extended formulations that
can be used directly to model the defender’s problem in constraint (2). Here, we consider the following network
design type of formulation (Magnanti and Wolsey 1995) that is commonly used to characterize the spanning
trees in a graph G:
X
min we ye (3)
e∈E
X X
l l
fij − fji = bli ∀i ∈ V, l ∈ V \ {k} (4)
{j:{i,j}∈E} {j:{i,j}∈E}
l l
ye ≥ fij + fji ∀e = {i, j} ∈ E, l ∈ V \ {k} (5)
X
ye = n − 1 (6)
e∈E
l l
fij , fji ≥0 ∀{i, j} ∈ E, l ∈ V \ {k} (7)
ye ≥ 0 ∀e ∈ E, (8)

where y = [ye ]e∈E represents the incidence vector of the spanning tree to be selected. In this setting, a vertex
k ∈ V is arbitrarily chosen to be the root of the spanning tree and a commodity l is defined for every vertex
l ∈ V \ {k}. One unit of each commodity l, which is assumed to emanate from the root vertex k, is required to
be delivered to vertex l. This is modeled by a parameter bli that takes the values of 1, −1, and 0 whenever i = k,
i = l, and i ∈ V \ {k, l}, respectively. The commodities are allowed to flow through edge e = {i, j} ∈ E in both
directions—i to j and j to i—provided that edge e belongs to the selected spanning tree (i.e., ye = 1). For a
given edge {i, j} ∈ E, the flow decision variable fijl indicates the flow of commodity l from i to j and the balance
flow constraints (4) ensure that all the commodities reach their corresponding destination. Constraints (6) limit
the number of edges to be selected to n − 1, constraints (5) restrict the flow being pushed through the edges
that do not belong to the selected spanning tree, and constraints (7) and (8) define the nonegative nature of
the variables. Notice that the balance flow constraints guarantee that a path exists from vertex k to every

6
other vertex i ∈ V \ {k} and the requirement of having exactly n − 1 edges selected guarantees that y indeed
represents a spanning tree of G. Therefore, since the objective function (3) minimizes the sum of the weights
of the selected spanning tree edges, the resulting formulation yields a minimum spanning tree in G.

Remark 1 Formulation (4)–(8) is known to be integral, which implies that the y variables will always take
the value of either 0 or 1 without the need of enforcing integrality constraints (Magnanti and Wolsey 1995).
Moreover, the set of extreme points of this formulation corresponds to solutions (y, f ) where y is the incidence
vector of a spanning tree T ∈ ΩG and f is a set of flows over T satisfying balance flow constraints (4).

Given an interdiction strategy S ⊆ E for the attacker, let x = [xe ]e∈E be the incidence vector of edge set
S. Since the defender edge selection is also limited by the attacker’s strategy, the formulation must then be
adapted to ensure the defender does not select any of the edges in S. A common practice (Cormican et al.
1998) to model this requirement is to alter the objective function (3) by adding a sufficiently large weight Me to
every edge e ∈ E that is interdicted by the attacker. Then, the resulting minimum spanning tree formulation
for the defender becomes:
X
min (we + Me xe )ye (9)
e∈E
s.t (6)–(8),

whose corresponding dual is:

X X
max bli αil + (n − 1)γ (10)
i∈V l∈V \{k}
X
s.t. βel + γ ≤ we + Me xe ∀e ∈ E (11)
l∈V \{k}

αil − αjl − βel ≤ 0 ∀e ∈ E, l ∈ V \ {k} (12)

αjl − αil − βel ≤0 ∀e ∈ E, l ∈ V \ {k} (13)
βel ≥0 ∀e ∈ E, l ∈ V \ {k}. (14)
Here, αil , βel ,
and γ represent the dual variables associated with constraints (4), (5), and (6), respectively,
and (11), (12), and (13) are the dual constraints associated with variables ye , fijl and fjil , respectively, for all

e ∈ E, l ∈ V \ {k}. It should be noted that the feasible space of this dual formulation is not empty, as the
solution where all the dual variables are set to zero is feasible. Furthermore, since by the weak duality theorem
any feasible solution to (11)–(14) yields a lower bound on the objective functionPof thePprimal problem (i.e., the
weight of a minimum spanning tree in the residual graph given x), by enforcing i∈V l∈V \{k} bli αil +(n−1)γ ≥
r, one can then replace the defender’s problem in (2) with a constraint set derived from its dual counterpart
obtaining the single-level reformulation:
X
(EXT): min ce xe (15)
e∈E
s.t. (11)–(14)
X X
bli αil + (n − 1)γ ≥ r (16)
i∈V l∈V \{k}

xe ∈ {0, 1} ∀e ∈ E, (17)

which has an O(nm) number of variables and constraints. Since this formulation is defined over a higher
dimensional space given by dual variables α, β, and γ, we will also refer to this formulation as the mixed-integer
extended formulation (EXT).
Importantly, a careful choice of the weights Me , for all e ∈ E has a strong impact on the quality of
this formulation. Ideally, one should aim for the smallest possible value for such weights to obtain a tight
representation. Since the interdiction criteria requires an optimal solution of (9) to be at least r, it is possible
to identify a set of lower bounds on the values of Me , for e ∈ E that guarantee the validity of the resulting
formulation.

7
Proposition 1 Setting the values of Me , for all e ∈ E so that
  
  X 
Me ≥ max 0, max r− we0 , ∀e ∈ E, (18)
 T ∈ΩG [e]  
e0 ∈E(T )

results in a valid EXT formulation for the minimum spanning tree interdiction problem.

Finding a tight set of values for Me for e ∈ E according to Proposition 1 can then be achieved in O(n2 log n)
time by repeatedly applying Prim’s algorithm (Ahuja et al. 1993), each time fixing edge e ∈ E to be part of
the selected spanning tree.

3.2 Critical Spanning Trees Formulation

We begin this section by characterizing the set of feasible solutions of the spanning tree interdiction problem
and then describe an integer programming formulation that results directly from it.

Definition 1 (Critical Spanning Tree) Given a rational value r, so that wG < r < wG , a spanning tree T
of graph G is said to be critical if w(T ) < r. We use ΩrG to denote the set of all critical spanning trees of graph
G.

The set of critical spanning trees given the interdiction threshold r is naturally the collection of targets the
attacker must focus the interdiction efforts on. This observation yields the following direct characterization of
the feasible solutions to the minimum spanning tree interdiction problem, whose proof is omitted as it follows
directly from the definition.

Theorem 2 A given edge set S ⊆ E is a feasible strategy for the attacker if and only if S ∩ T 6= ∅, for all
T ∈ ΩrG .

This characterization yields the following set-covering type of model for which x = [xe ]e∈E is the incidence
vector of edge set S. In this paper we refer to this model as the critical spanning tree formulation (CST), and
consequently will refer to the constraints in set (20) as critical spanning tree constraints.
X
(CST): min ce x e (19)
e∈E
X
s.t xe ≥ 1, ∀T ∈ ΩrG (20)
e∈E(T )

xe ∈ {0, 1}, ∀e ∈ E. (21)

As stated in Theorem 2, constraint set (20) ensures that each critical spanning tree in the graph is effectively
interdicted by the attacker; otherwise, the defender can then freely select any of such critical spanning trees
left intact therein.
Set covering formulations, like (19)–(21), that are designed to restrict the existence of some type of graph
structures from a given solution space appear to be quite common in the literature. See for instance the work by
Chvátal and Cook (1990) who use a similar set covering formulation to block a specific type of edge structures
called whips to raise the domination number of a graph; the set covering formulation by Ben-Ameur et al.
(2015) that blocks connected components with more than k vertices to solve the k-separator problem; or the
cut-set inequalities that are often used to break subtours when solving traveling salesman problems (Nemhauser
and Wolsey 1988).
For the particular case of graph interdiction, the critical spanning tree constrains (20) are analogous to the
supervalid type-I inequalities first introduced in the seminal paper by Israeli and Wood (2002) for interdicting
shortest paths. One important difference however is that for the particular type of graph interdiction problem
that we solve here (i.e., we minimize the cost of achieving a given disruption level over the defender’s objective),
constraints (20) are in fact valid as they naturally describe the attacker’s feasible space.

8
3.3 Comparing the Strength of the Linear Relaxations
A common approach to analyze the quality of an integer formulation is by studying the strength of its linear
relaxation (Vielma 2015). In this section we compare the linear relaxations of the extended formulation (EXT)
and the critical spanning tree formulation (CST). Performing a direct comparison between these formulations
raises some challenges because the extended formulation is defined over a higher dimensional space due to the
additional variables that are introduced by the dual of the inner spanning tree model. Interestingly, we will
show that it is possible to obtain a full description of the projection of the extended formulation into the space
of variables x. We will then use such a projection as the base for our comparison.
2
Theorem 3 Let Pext = {(x, α, β, γ) ∈ [0, 1]m × Rn × Rnm × R, satisfying (11)–(14),(16)} be the polytope
described by the linear relaxation of the extended formulation, with a proper value for Me , ∀e ∈ E satisfying
(18). Then, the projection projx (Pext ) onto the x-space corresponds to the polyhedron described by the following
inequalities:
X X
Me x e ≥ r − we , ∀T ∈ ΩG (22)
e∈E(T ) e∈E(T )

0 ≤ xe ≤ 1, ∀e ∈ E. (23)

An interesting observation derived from Theorem 3 is summarized by the following corollary.

Corollary 1 Let Pcst = {x ∈ [0, 1]m , x satisfying (20)} be the polytope described by the linear relaxation of
the critical spanning tree formulation, then

1. Pcst ⊆ projx (Pext ).

2. Pcst = projx (Pext ) if and only if all critical spanning trees have the same weight.

In conclusion, the critical spanning tree formulation always yields a tighter relaxation for the non-trivial
instances of the problem.

3.4 Separating the Critical Spanning Tree Constraints

The size of the critical spanning tree in formulation (19)–(21) is directly given by the critical spanning trees in
ΩrG . Depending on the graph’s density and the weight vector w, it is possible for this set to grow exponentially
large. Thus, to solve this formulation, some critical spanning tree constraints must be generated sequentially
when needed. In this section, we discuss the complexity of separating such constraints. We observe that
separation is polynomial-time solvable for integer values of x, but show that separation is at least weakly
NP-hard for the general case.
The separation problem for the critical spanning tree formulation can be stated as follows.

Critical spanning tree constraint separation

Instance: A graph G = (V, E), a non-negative integer vector of rational edge weights w, a vector x ∈ [0, 1]m ,
and a rational number r.

Question: Is there a spanning tree T ∈ ΩrG such that e∈T xe < 1?

P

Theorem 4 Let x be a candidate solution for the attacker. Then, separating the critical spanning tree con-
straints is:

1. NP-hard if x ∈ [0, 1]m

2. solvable in O(m log n) if x ∈ {0, 1}m

9
 Interestingly, the NP-hardness of the separation algorithm also implies the following corollary, whose proof
follows directly from Theorem 1 in Carr and Lancia (2002) because the critical spanning tree formulation does
not admit a compact separation.

Corollary 2 Unless P=NP, there is no polynomial-time constructible extended formulation for the polytope
Pcst described by the critical spanning tree constraints.

In Section 4, we will show that most of the critical spanning trees constraints induce facets for the minimum
spanning tree interdiction polytope. This result and the one of Corollary 2 provide some additional evidence
of the strength of the proposed formulation.
In practice, given that the size of set ΩrG may grow exponentially with respect to the size of the graph, we
solve formulation (19)–(21) via branch and cut, generating critical spanning tree constraints as lazy constraints
(i.e., we call the separation subroutine whenever there is an integer candidate solution for the attacker). We
notice however that the constrained minimum spanning tree problem admits a polynomial-time approximation
scheme (PTAS) (Ravi and Goemans 1996) that can be adapted to potentially separate fractional solutions as
well. In our computational experiments, we limit our implementation and exclusively separate integer solutions.
Importantly, each round of the separation subroutine may yield several violated critical spanning tree
constraints. This is particularly crucial at early stages of the branch and bound scheme, when the algorithm
has yet to add enough constraints to yield a good lower bound. In fact, during the separation stage one
may identify a set of k minimum spanning trees, for any given k (Katoh et al. 1981) and add the constraints
associated with spanning trees among those that are deemed critical.
A typical implementation can benefit from a simpler greedy approach to potentially find several critical
spanning trees whose corresponding inequality is violated. Given a candidate integer attacker solution x, for
which S is the associated set of edges being targeted, on can first find the minimum spanning tree T left in
G(V, E \ S). Second, among the edges not in T or S, find the edge e of minimum weight and add it to T
creating a cycle. Third, find the edge e0 of maximum weight in the cycle such that if e0 is replaced by e in T
the resulting new spanning tree is also critical. This process can be repeated until no further critical spanning
trees can be found.

4 Polyhedral Analysis
In this section, we study the convex hull of the incidence vectors of all edge subsets S ⊆ E that yield feasible
solutions for the attacker and discuss some important properties of such polytope.

4.1 Basic Properties of the Convex Hull

Let P denote the convex hull of the incidence vectors of all feasible solutions, for any given r > 0. We begin
our discussion by proving that P is a full-dimensional polytope and that the upper-bound constraint on each
decision variable (xe ≤ 1, for each edge e ∈ E) induces a trivial facet of P. We then identify necessary and
sufficient conditions under which the non-negativity constraints (xe ≥ 0, for each edge e ∈ E) and the critical
spanning tree constraints are facet inducing for P. These results are presented in the following theorem.

Theorem 5 Given a simple, nonempty, connected graph G = (V, E), a weight vector w ∈ Qm so that there are
no irrelevant edges in E (i.e., min{w(T )| T ∈ ΩG [e]} < r, ∀e ∈ E), and a scalar r ∈ Q, so that wG < r < wG ,
the following statements are true.

(1) P is a full-dimensional polytope.

(2) Given an edge e ∈ E, inequality xe ≤ 1 induces a facet of P.

(3) Given an edge e ∈ E, inequality xe ≥ 0 induces a facet of P if and only if n ≥ 4.

10
 (4) Given a critical spanning tree T ∈ ΩrG , the corresponding critical spanning tree inequality
P
e∈E(T ) xe ≥1
induces a facet of P if and only if:

(a) there is no edge e ∈ E such that if added to T , it completes a Hamiltonian cycle in G, or

(b) if such an edge exists, at least one of the spanning trees contained in the resulting Hamiltonian cycle
has a weight greater than r.

4.2 Other Sets of Facet-Defining Inequalities

In this section, we study a broad family of facet defining inequalities of P that could be used to strengthen
formulation (19)–(21). We analyze the complexity of the corresponding separation algorithm and provide a
brief discussion about how to use some of these inequalities in practice.

Definition 2 (Cactus) A cactus is defined as a connected graph in which every edge belongs to at most one
cycle. For a given graph G = (V, E), a cactus H that contains all the vertices in V is called a spanning cactus
of G. The rank of a given cactus H is defined as the number of cycles contained in H; we will refer to a cactus
of rank k, simply as a k-cactus.

In what follows, we deal exclusively with spanning cacti of a graph G; thus, we will drop the spanning
categorization when referring to those, unless further clarifications are required.
Given a k-cactus H of graph G, let E(H) be the set of all edges in H; clearly, |E(H)| = n + k − 1.
Furthermore, for any order given to the k cycles of H, let Cl (H), ∀l = 1, . . . , k be the set of edges comprising
the lth cycle of H, C(H) = kl=1 Cl (H), and F (H) = E(H) \ C(H) be the set (possibly empty) composed of
S
the edges that do not belong to any cycle in H. Notice that removing one edge from cycle Cl (H), for any
l = 1, . . . , k, reduces H into a (k − 1)-cactus; removing any two edges from Cl (H) Q disconnects H; as well as
removing any edge from set F (H). It is easy to see that H is in fact the union of kl=1 |Cl (H)| spanning trees
of G, each of those created by removing one of the edges from each subset Cl (H), ∀l = 1, . . . , k.
Figure 1 provides an example of a 3-cactus H with 18 vertices and 20 edges. In this example, there are
three cycles, C1 (H) = {e3 , e6 , e7 }, C2 (H) = {e8 , e9 , e10 , e11 }, and C3 (H) = {e15 , e16 , e17 , e18 , e19 , e20 }, hence,
the cactus is of rank 3. The set C(H) = {e3 , e6 , e7 , e8 , e9 , e10 , e11 , e15 , e16 , e17 , e18 , e19 , e20 } is simply the union of
cycles C1 (H), C2 (H) and C3 (H), and Qkset F (H) = {e1 , e2 , e4 , e5 , e12 , e13 , e14 } is its complement. It is also easy
to verify that H contains a total of l=1 |Cl (H)| = 3 × 4 × 6 = 72 spanning trees.

e5
e4

e16
e6

e 17
e3

5
e1

e2 C1 e8 e14
C3
e7
e1

C2
e 20

8
e1
0
e9

e1

e11 e1 e19
2

3
e1

Figure 1: An example of a 3-cactus.

Consider a k-cactus H of a given graph G = (V, E). For each edge set Cl (H), with l = 1, . . . , k, let
Ĉl (H) ⊆ Cl (H) be a set that includes each edge e ∈ Cl (H) for which ΩrG(V,E(H)\{e}) is not empty. That is, edge
e ∈ Cl (H) belongs to Ĉl (H) if, after removing edge eSfrom H, at least one critical spanning tree T remains in
ˆ
the resulting (k − 1)-cactus. We also denote C(H) = kl=1 Ĉl (H) to be the set of all these edges.

11
Theorem 6 Given a simple, nonempty, connected graph G = (V, E), a weight vector w ∈ Qm , a scalar r ∈ Q
so that wG < r < wG , and a k-cactus H of G, with k ≥ 1, the inequality
X X X
xe + 2 xe + 2 xe ≥ 2 (24)
e∈Ĉ(H) e∈C(H)\Ĉ(H) e∈F (H)

induces a facet of P, if and only if |Ĉl (H)| ≥ 3, ∀l = 1, . . . , k.

In what follows, we will refer to these cuts as cactus inequalities.

Cactus graphs are essentially a generalization of other well-known structures. Indeed, notice that a tree is a
cactus of rank 0, and a 1-tree (i.e., a connected graph containing exactly one cycle) (Held and Karp 1970), also
called a pseudotree, is a cactus of rank 1. In our computational experiments we give special attention to the
cactus inequalities associated with 1-trees, for those are easier to separate compared to inequalities for cacti of
higher ranks.
We now discuss the separation problem for the cactus inequalities, which is formally stated below. Theorem
4 provided a proof that separating the critical spanning tree constraints is NP-hard. Since a tree is a cactus of
rank 0, separating cactus inequalities is in general NP-hard. The proof of the following theorem (see Appendix
A in the Online Supplement) goes a step further and presents a simple construction that is used to show that
separating inequalities associated with 1-cacti is NP-hard too. It is easy to see that such a construction can be
extended for cactus inequalities of any given rank k.

cactus inequality separation

Instance: A graph G = (V, E), a non-negative integer vector of rational edge weights w, a vector x ∈ [0, 1]m ,
and a rational number r.

Question: Is there a k-cactus H of G with k ≥ 1 and Ĉl (H) ≥ 3, ∀l = 1, . . . , k for which inequality (24) is
violated?

Theorem 7 Let x ∈ [0, 1]m be a candidate solution for the attacker. Then, the separation problem for the
cactus inequalities is NP-hard.

Despite the fact that the separation algorithm for the cactus inequalities is NP-hard, in our implementation
we tested the following greedy integer separation approach tailored for cacti of rank one (i.e., 1-trees). Given
a candidate integer attacker solution x, for which S is the associated set of edges being targeted, we first
find a minimum spanning tree T left in G(V, E \ S). Among the edges not in T or S, we find the edge e of
minimum weight and add it to T creating a 1-tree H. Then, for each edge e0 ∈ C1 (H), we add e to Ĉ1 (H) if
w(T ) + we + we0 < r. If |Ĉ1 (H)| ≥ 3, we add the resulting cactus inequality.
As will be described in Section 6, we observed in our experiments that separating 1-tree inequalities too
often as described above can be relatively slow, potentially hindering the efficacy of the algorithm, particularly
when applied over dense graphs. Thus, our implementation was adapted so that the search for this type of
inequalities is only conducted with a predefined probability.

5 Critical Edge Sets Formulation

In this section, we exploit some fundamental properties of the the problem’s feasible solutions to produce a
stronger variation of the critical spanning tree formulation. This alternative formulation is defined by a new
type of structures referred to as critical edge sets, which can be used to produce a set of so-called supervalid
inequalities (cf. Israeli and Wood (2002)) that exclude feasible solutions for the attacker, albeit solutions whose
objective value is not better than the minimum cost edge cut λc (G). We show that the linear relaxation of this

12
formulation is not only contained by the linear relaxation of (20)–(21), but also requires fewer constraints to
be defined.
We will discuss several properties of this formulation including details about the complexity of the separation
algorithm for its constraints. Interestingly, despite requiring an additional computational effort, the separation
algorithm can take advantage of an efficient enumeration scheme that exploits some hereditary properties of
the critical edge sets. Based on the results of our computational experiments (see Section 6), this approach is
the one that produced the best results.
It is important to mention that the supervalid inequalities that we will describe in this section are by design
different to the ones firstly introduced in Israeli and Wood (2002). As mentioned in Section 3, the inequalities
developed in their paper (particularly the inequalities of type I) are analogous to the critical spanning tree
constraints we discuss here, albeit defined over a set of st-paths instead of spanning trees. As we will explain in
the following section, the set of supervalid inequalities that we propose come from the fact that there exists a
natural bipartition of the feasible solutions for the attacker into solutions that are edge cuts and solutions that
are not. Our inequalities are somehow aimed to restrict the search for optimal solutions over the latter set.

Definition 3 (Critical edge set) Given an integer value r, so that wG < r < wG , a subset of edges K ⊆ E
is called a critical edge set if ΩG [K] ⊆ ΩrG and ΩG [K] 6= ∅. We use ΛrG and Λ̃rG to denote the collection of all
critical edge sets and the set of all minimal critical edge sets under inclusion, respectively.

Clearly, a critical edge set K ∈ ΛrG contains no cycles because otherwise no spanning tree would contain
all the edges in K. Also, the edge set E(T ) of any critical spanning tree T ∈ ΩrG is a critical edge set too,
for ΩG [E(T )] = {T }. In fact, it is a maximal critical edge sets under inclusion because any proper superset of
E(T ) contains cycles. Moreover, notice that for all T ∈ ΩrG , there is a K ∈ Λ̃rG such that K ⊆ E(T ), as either
set E(T ) or a subset of it belongs to Λ̃rG .

ΩrG ΩG \ ΩrG
(2, 2)
)
(1, 1)

(3, 2)
,3
(5

(4, 3) T1 T2 T3 T4 T5 T6 T7 T8
G
Λ̃rG

{e1 , e2 } {e1 , e3 }

Figure 2: An example of critical spanning trees and critical edge sets.

For illustrative purposes, consider the example presented in Figure 2, where we depict a graph G composed
of n = 4 vertices and m = 5 edges. Next to each edge, we provide the pair (e, we ) that indicates the index
and weight of the corresponding edge, respectively. We assume for this example that r = 7. The graph
contains eight spanning trees, ΩG = {Ti }8i=1 , among which four are critical, i.e., ΩrG = {T1 , T2 , T3 , T4 }. The
set of critical edge sets is ΛrG = {E(T1 ), E(T2 ), E(T3 ), E(T4 ), {e1 , e2 }, {e1 , e3 }}, and the set of minimal critical
edge sets is then Λ̃rG = {{e1 , e2 }, {e1 , e3 }}. Notice for instance that edge set {e3 } is not critical because
ΩG [e3 ] = {T1 , T2 , T3 , T5 , T8 } and both T5 and T8 are not critical. Similarly, edge set {e3 , e4 } is not critical
because ΩG [{e3 , e4 }] = {T2 , T5 }, but T5 is not critical. In contrast, edge set {e1 , e2 } is critical because all trees
in ΩG [{e1 , e2 }] = {T1 , T4 } are critical.

13
 The following theorem provides an interesting characterization of the set of feasible solutions of the inter-
diction problem with respect the the critical edge sets.
Theorem 8 An edge set S ⊆ E is a feasible solution for the attacker if and only if S is an edge cut or
S ∩ K 6= ∅, for all K ∈ Λ̃rG .
Interestingly, Theorem 8 defines a partition of the feasible strategies for the attacker into two sets: the
solutions that are edge cuts and the ones that are not. As mentioned in Section 2, since the objective value
of any edge cut is bounded below by λc (G), to find the optimal solution of the interdiction problem, one may
find first a minimum cost edge cut (which can be done in polynomial time), and then focus on searching for
the best non-edge-cut feasible strategy. Based on this observation, we use the result from Theorem 8 to define
the following formulation, named the critical edge set formulation (CES). We will show that we can use this
integer program to find a minimum cost non-edge-cut solution for the attacker.
X
(CES): min ce x e (25)
e∈E
X
s.t xe ≥ 1, ∀K ∈ Λ̃rG (26)
e∈K
xe ∈ {0, 1}, ∀e ∈ E. (27)

Importantly, the critical edge set formulation is not valid for the original problem because its solution space may
omit some feasible solutions that are edge cuts. Nevertheless, an optimal solution of the minimum spanning
tree interdiction problem, based on Theorem 8, corresponds to either an optimal solution of (25)–(27) or an
edge cut of minimum cost. For this reason, if while solving the critical edge set formulation, one finds that a
lower bound for the formulation is greater than or equal to λc (G), then one can stop the search and use any
minimum cost edge cut as the optimal strategy for the attacker.
Consider the example shown in Figure 3, which presents a side-by-side comparison of the critical spanning
tree and the critical edge set formulations for the instance depicted in Figure 2. Notice that solution S = {e3 , e4 }
is a feasible solution for the attacker, as it is an edge cut, but it is not feasible for the critical edge set formulation.

CST Formulation: CES Formulation:

P P
min ce xe min ce xe
e∈E e∈E
s.t. x1 + x2 + x3 ≥ 1 s.t. x1 + x3 ≥ 1

x1 + x3 + x4 ≥ 1 x1 + x2 ≥ 1

x1 + x3 + x5 ≥ 1 xe ∈ {0, 1}, ∀e ∈ E

x1 + x2 + x4 ≥ 1

xe ∈ {0, 1}, ∀e ∈ E

Figure 3: A comparison of the critical spanning tree (CST) and the critical edge set (CES) formulations for
the example depicted in Figure 2.

Remark 2 The quality of the critical edge set formulation strongly depends on whether there exist irrelevant
edges in graph G. We recall from Section 2.1 that an edge e ∈ E is called irrelevant if the weight of all the
spanning trees in ΩG [e] (i.e., all the spanning trees that contain e) is greater than r. Having irrelevant edges in
the graph may prevent some edge sets to be critical, which will in turn affect the overall strength of the critical
edge set formulation. Therefore, one should guarantee that all the irrelevant edges are removed before solving
the interdiction problem.

14
 For an illustration, consider again the example depicted in Figure 2. Suppose that we create a new graph
G0 by adding the second diagonal edge e6 to G (i.e., the edge that is left for the graph to be complete) so that
w6 = 7. Since the weight of e6 is equal to r, no spanning tree that contains this edge is critical; therefore, e6 is
irrelevant. Clearly, solving the minimum spanning tree interdiction problem in both G and G0 yields the same
solution. However, notice that for graph G0 , the edge sets {e1 .e2 } and {e1 , e3 } are no longer critical because the
spanning trees induced by the edges {e1 .e2 , e6 } and {e1 , e3 , e6 } are not critical, which implies that Λ̃rG0 = ΩrG0 .
In conclusion, the critical edge set formulation gets weakened in the presence of e6 .

5.1 A Characterization of the Solution Space of the Critical Edge Set Formulation
The result given by Theorem 8 guarantees that solving the CES formulation coupled with a global minimum cost
edge-cut incumbent solution is a valid approach for tackling the proposed minimum spanning tree interdiction
problem. However, this theorem provides no further information about the properties of this new formulation
and its relation to the CST formulation. In what follows, we will study these by analyzing the solution space of
the CES formulation and provide an analytical description of why this formulation is in general stronger. We
consider this analysis one of the key results of this paper.
First, the following theorem provides an exact characterization of this solution space.

Theorem 9 An edge set S ⊆ E is a feasible solution for the critical edge set formulation if and only if S is a
superset (not necessarily proper) of any non-edge-cut feasible strategy for the attacker.

This theorem implies that the CES formulation is tight with regards to excluding edge-cut solutions. That
is, most of the edge-cut solutions are removed by the CES constraints, and those remaining in the solution
space are all dominated by some non-edge-cut solutions. In essence, the CES formulation takes advantage of
the global min-edge-cut incumbent, and focus the solution search procedure in a smaller space that consists
of only non-edge-cut strategies. A direct consequence of this is that the solution space of CES formulation
is contained within the solution space of CST formulation. Interestingly, this also offers another perspective
about the origin of critical edge sets: they are induced by the bipartition of the solution space resulting from
the edge-cut property.
Next, the following theorem establishes a relation between the minimal critical edge sets and the critical
spanning trees.

Theorem 10 For any minimal critical edge set K ∈ Λ̃rG , there exists a spanning tree T ∈ ΩG [K] that does not
contain any other minimal critical edge set K 0 ∈ Λ̃r , where K 6= K 0 .

This relation allows us to build an injective function from the set of minimal critical edge sets Λ̃rG to the
set of critical spanning trees ΩrG . Hence, we have the following corollary.

Corollary 3 |Λ̃rG | ≤ |ΩrG |.

Notice that Λ̃rG and ΩrG are the sets of constraints for the CES and CST formulations, respectively. Therefore,
the solution space of CES formulation is not just contained within the CST formulation, it also requires fewer
inequalities to be described. This is quite interesting—and perhaps counter-intuitive—because, in general,
formulations that are stronger often require more variables/constraints to be represented. Moreover, in practice,
the number of minimal critical edge sets needed in Λ̃rG for solving a problem can be much smaller than |ΩrG |,
since usually a small-sized minimal critical edge set can be contained by many critical spanning trees. This
is of great importance for practical purposes, for it implies that the CES formulation does not just produce
stronger inequalities, but also requires fewer of them to solve the problem.

15
5.2 Separating the Critical Edge Set Constraints
It should be clear that separating a critical edge set constraint from a fractional solution is NP-Hard, for
a separation in polynomial time implies that the critical spanning tree constraints can also be separated in
polynomial time, which contradicts Theorem 4. For an integer solution, we will show that the separation
complexity is O(α(n, m)n log n).
To design an efficient algorithm to separate the critical edge set constraints, we first consider the partial
ordered set (poset) defined by (↓ ΩrG , ⊆), where ↓ ΩrG represents the lower closure of ΩrG , i.e., ↓ ΩrG = {K ⊆
E(T ) | T ∈ ΩrG }, and ⊆ indicates that the poset is ordered by inclusion. Since any subset of elements of this
poset has a greatest lower bound (i.e., a meet) the poset is also a meet-semilattice, which means that the poset
is closed under intersection. We will show that this meet-semilattice structure yields a natural enumeration
scheme for all the critical edge sets. The following theorem provides some important properties of this meet-
semilattice.

Theorem 11 The following statements about the poset (↓ ΩrG , ⊆) are true:

1. The least element in (↓ ΩrG , ⊆) is the empty set.

2. The maximal elements in (↓ ΩrG , ⊆) are all the critical spanning trees.

3. (ΛrG , ⊆) is a subposet embedded in (↓ ΩrG , ⊆).

In this meet-semilattice, we call the empty set the root, each maximal element a leaf, and each maximal
chain that emanates from ∅ and reaches a critical spanning tree a branch. For example, the meet-semilattice of
(↓ ΩrG , ⊆) for the graph G of Figure 2 is illustrated in Figure 4. Notice, that the bold nodes and arcs compose
the subposet (ΛrG ,⊆) that is embedded in this meet-semilattice.

{e1 } {e2 } {e3 } {e4 } {e5 }

{e1 , e2 } {e1 , e3 } {e1 , e4 } {e1 , e5 } {e2 , e3 } {e3 , e4 } {e3 , e5 }

{e1 , e2 , e3 } {e1 , e3 , e4 } {e1 , e3 , e5 }

Figure 4: The meet-semilattice of (↓ ΩrG , ⊆) for the example depicted in Figure 2.

Before the construction, the algorithm sorts all the edges so that at each level of the meet-semilattice, the
weights of the nodes of the meet-semilattice are increasing from left to right. Clearly, the size of this meet-
semilattice is very large even for small sized instances; therefore, in our implementation, we construct portions
of it, when needed. At the beginning of the algorithm, we begin with the root. Once the formulation returns
an integer candidate solution associated with edge set S, the algorithm creates the leftmost unexplored branch
that does not intersect S. Since all the branches are sorted from left to right, this branch corresponds to one
of a critical spanning tree of minimum weight that is not blocked by S. If S is not a feasible solution, then
there exists a critical edge set on this branch, and the separation algorithm proceeds to search for the minimal
set on this branch. Once such edge set is found, the corresponding critical edge set constraint is added to the
constraint pool of the formulation.

16
 At each node of the branch, with edge set K, the algorithm needs to compute a maximum spanning
tree restricted on K to check whether it is a critical edge set or not. We use a variant of the Kruskal’s
algorithm, which is of complexity O(m log n). However, the dominant part of Kruskal’s algorithm is the edge
sorting; once the edges are sorted, a maximum spanning tree can be found with time complexity O(nα(n, m))
using a sophisticated disjoint-set data structure (Tarjan and Van Leeuwen 1984, Tarjan 1979). Therefore, the
verification complexity at each node is O(nα(n, m)), since all the edges have been sorted in advance.
If we execute the aforementioned branching method naively, the overall complexity of separating a critical
edge set is O(α(n, m)n2 ). However, the following proposition ensures that a binary search based branching
scheme can be applied.

Proposition 2 Let β be the function from the ordered set (↓ ΩG , ⊆) to (R, ≤) defined as β(K) = maxT ∈ΩG [K] w(T ),
then β is order reversing.

Given a branch, this proposition helps the binary search to determine which half to proceed. Suppose
the algorithm is at a node of an edge set K with a search scope from K0 to K1 in a given branch. Clearly,
K0 ⊆ K ⊆ K1 . If β(K) < r, then K is a critical edge set, so we update the search scope to the subchain
between K0 and K in order to find a smaller one; otherwise, K is not a critical edge set, so the new search
scope is the subchain between K and K1 . With this branching method, the complexity of separating a critical
edge set which is minimal on the branch is O(α(n, m)n log n).
The last thing to notice is that the critical edge set, which is minimal on each given branch produced by
the above method, is not necessarily an element in Λ̃rG . But, since all the minimal critical edge sets would be
eventually enumerated from this semilattice, the correctness of our implementation still holds.

6 Computational Experiments
The computational experiments reported in this section were conducted on a computing cluster equipped with
twelve nodes, each having a 12-core Intel Xeon E5-2620 v3 2.4GHz processor, 128 GB of RAM, and running
Linux x86 64, CentOS 7.2. All the formulations and algorithms were implemented in C++, compiled with
GCC 6.3.0, and solved using the commercial Optimizer Gurobi 8.0. Each instance was solved by a single node
of the cluster under a time limit of 7, 200 seconds. The data sets are provided in the online supplements.
We compare the performance of all the formulations studied in this paper: (1) the mixed-integer extended
formulation derived from the bi-level representation of the problem (see Section 3.1), which we refer to as EXT;
(2) the critical spanning tree formulation (see Section 3.2), which we call CST; (3) the critical spanning tree
formulation strengthened with cactus inequalities of rank 1 (see Section 4.2), which we named CST-1; and (4)
the critical edge set formulation (see Section 5), referred to as CES. The CST, CST-1, and CES formulations are
all implemented using the proposed sequential cut-generation method, due to the large number of constraints.
In our implementation, both CST and CST-1 are initialized with a minimum spanning tree constraint, and the
CES is initialized with a constraint of critical edge set, which is contained in a minimum spanning tree. Notice
that one may warm-start the CST and CST-1 formulations by introducing as many constraints associated with
critical spanning trees as one may desire. For instance, one can solve an instance for the k-minimum spanning
trees (Katoh et al. 1981) for any given k and add the constraints associated with spanning trees among those
that are deemed critical.
We use three main criteria for our tests: the percentage of instances solved to optimality, the average
runtime required to solve the problem, and the average optimality gap. Furthermore, we also compare the cut
generation subroutines for the CST and CES formulations.

6.1 Test Instances

We generated a total of 155 connected graphs of different sizes and topologies using four types of random graph
generator models: small-world (Watts and Strogatz 1998), Erdös-Renyi (Erdös and Rényi 1959), binomial

17
(Gilbert 1959), and Barabási-Albert (Albert and Barabási 2002). For the size of the graphs, we set n =
40, 80, 160, 200. Furthermore, we adjusted the parameters of each graph generator model using three different
settings in order to produce three sets of graphs of similar densities for each value of n (except for n = 40, for
which we only used two). In the following tables, we use the average number of edges of the graphs (m̄) to
represent the densities. For each combination of n, graph density, and graph generator model, we created three
different graphs and use a uniform distribution from 20 to 80 to generate the weights (w) and costs (c) of the
edges. The data sets associated with all the 155 graphs are available as online supplemental material.
In addition to the density of the graph, the complexity of the instances also depends on the values chosen
for r, as this parameter dictates the cardinality of ΩrG . We solve all instances with four different levels for r.
More specifically, we pick r = (wG − wG ) × γ, for γ = {0.05, 0.3, 0.7, 0.95}, where γ is referred to as the
interdiction requirement.
Notice that since the weights of the graphs are randomly generated, varying the interdiction requirement
as a percentage of the difference between the weights of the minimum and maximum spanning trees gives us
the possibility of aggregating instances that are expected to be of similar difficulty, rather than by directly
changing the values for r. In general, for a fixed value of γ, the value of r increases with respect to the number
of vertices and the minimum degree of the graphs. Thus, larger instances tend to be more challenging for most
formulations, not only because of their size, but also because of the interdiction requirement. For the different
instances that we generated, we ensured that, even when taking the lowest value of γ = 0.05, the resulting
values of r avoid trivial cases where the set of the critical spanning trees ΩrG is empty or simply too small, thus
providing a good spectrum for the performance comparisons. In summary, each of the four formulations was
tested over a set of 155 (graphs) × 4 (interdiction requirements) = 620 instances.
To present the computational results, we aggregate all the instances that are expected to be of similar
complexity based on the size of the graph in terms of n and m̄, as well as the interdiction requirement γ. We
call such a triplet (n, m̄, γ) a graph configuration. All the tables provided in the following sections showcase
the averages over all the instances of the same graph configuration for the different metrics. Interestingly,
we observed no major difference in the difficulty of the different instances with respect to the type of graph
generator model used. Thus, the averages shown in the tables are not disaggregated by such factor.
Whenever we present the average of the optimality gaps, we evaluated them as (U B − LB)/LB × 100,
where UB and LB stand for the best upper and lower bounds identified by the formulations, respectively.
Furthermore, if a formulation was not able to find the optimal solution of an instance within the time limit, we
set the corresponding running time as 7,200 seconds.
In addition to the randomly generated graphs, for the purpose of demonstrating the performance of our
algorithms on real-life graphs, we also collected 12 instances from Network Repository (2019) comprising infras-
tructure, power, and road networks. The data sets of these instances are also available as online supplemental
material. As mentioned before, because of their sparsity, this type of instances are significantly easier to solve;
therefore, we only tested two of our formulations on these instances.
Finally, since any minimum-cost edge cut is a feasible solution for the attacker, we first identified one in
polynomial time (Gomory and Hu 1961) and feed it to all the formulations as an initial incumbent solution.
Thus, all the approaches started with the same upper bound. For the particular case of the CES formulation,
we did not add it as an incumbent because such a solution is likely to be infeasible for that formulation (see
Theorem 9); however, we used its objective value as a stopping criterion, as described in Section 5. Identifying
a minimum-cost edge cut solution is significantly faster than solving the interdiction problem. Even for the
largest and densest instances, it takes less than 0.1 seconds to compute. Since all the formulations use λc (G)
as a valid upper bound, we report the time required to compute it separately from the overall running times.

6.2 Results for the CST and CST-1 Formulations

We begin our discussion by analyzing the effect that introducing cactus inequalities of rank 1 has on the efficiency
of the critical spanning tree formulation. Among all the four formulations we compare in our experiments, the
CST and the CST-1 are clearly the most similar ones, as the latter is simply the CST formulation strengthened

18
with some extra facet-defining inequalities. Since the CST-1 formulation may have a stronger lower bound
thanks to the extra cuts, one might expect it to obtain better results than the plain CST formulation.
However, we note that in practice the implementation of the separation algorithm for these cuts has a
strong impact on the formulation’s performance. In general, when applied too often, the overhead required to
separate and add the cactus inequalities may overcome the benefit that they produce. Thus, instead of naively
separating them at every branch-and-bound node, we generate them randomly with a probability of ρ. That
is, at each time after separating a critical spanning tree constraint, there is a ρ chance that a cactus inequality
is separated. Tuning this probability is crucial for the performance of the CST-1 algorithm, as adding too
few inequalities would not make much of a difference, but adding too many would make the linear relaxation
grow significantly, thus taking longer to solve. In our experiments, we observed a good performance when
selecting ρ = 1/3. This method is commonly used in the literature to improve the performance of cutting plane
algorithms (Fischetti et al. 2017).
Table 1 reports the results for these two formulations; we highlight in bold the best results obtained for
each of the metrics. As expected, for most of the graph configurations, CST-1 outperforms CST in all the three
criteria.
An interesting observation is that the total number of cuts added with both formulations is relatively similar,
but the impact of adding the cactus inequalities on the overall performance is notable and often results in a
much faster execution. See for example the cases for the graph configurations where n = 80 and m̄ = 160;
n = 160 and m̄ = 318; n = 200 and m̄ = 396, for which the CST-1 performs much better.
Since the CST formulation was outperformed in most of the instances by the CST-1 formulation, we exclude
it from all further comparisons.

6.3 EXT, CST-1, and CES Performance Comparison

In this section, we continue our discussion by analyzing the computational results of the three remaining
formulations. Table 2 reports the averages of the three metrics we use for our analysis. As before, we highlight
in bold the best results obtained for each graph configuration.
A glance at Table 2 shows that the CES formulation outperforms the other two in all three metrics. The
percentage of instances solved by CES is close to 90%, which is almost 20% higher than that of the other two
formulations. The average running time of CES is also significantly shorter—about 70 % faster than the times
of the other two formulations. In fact, in several occasions, it takes only a few seconds for CES to solve several
instances compared to the several minutes it takes EXT and CST-1. Also, the average gap is noticeable smaller,
exceeding 15% in just a few cases, whereas gaps larger than 30% are common for the other two formulations.
A further analysis of this data helps us draw the following conclusions. First, in terms of the percentage
of instances solved, EXT performance is similar to that of CST-1, as both have solved around 68% of the
instances. However, with respect to the average running times and optimality gaps, EXT is the worst among
all three formulations. The fact that EXT has a similar solving percentage as CST-1 can be directly explained
by the fact that EXT performs better on small instances. In fact, the CST-1 formulation outperforms EXT in
all occasions except for the cases where the graphs are small and both m̄ and γ are large. This behavior can
be partially explained by the fact that the CST-1 formulation yields a tighter linear relaxation than EXT (See
Corollary 1) and also by the fact that the number of variables comprising EXT is O(nm), which can become
quite large for some of the big instances.
Second, the CES formulation drastically outperforms CST-1, except for the cases where γ is small. An
interesting explanation for this difference in performance is that the CES formulation removes a large subset of
fractional solutions that are feasible for the linear relaxation of the CST-1 formulation. This effect is exacerbated
when γ is large, in which case the critical edge set constraints tends to become significantly stronger than the
critical spanning tree constraints (see Corollary 3). This effect can be observed from for the different values of
γ in Table 2. Notice that both the average running times and gaps decrease significantly for CES whenever γ
increases. Conversely, when γ is very small, the number of critical spanning trees |ΩrG | tends to be also small,
so CST-1 occasionally performs better because its separation subroutine is faster.

19
Table 1: Computational results for the CST and CST-1 formulations. For each graph configuration, we show
the time required to identify a minimum-cost edge cut (λc (G)), the percentage of instances solved to optimality
(Opt), the average run time in seconds (time), the average optimality gap (Gap), the number of cuts generated
(Num. cuts), and the total time in seconds spent separating inequalities (Sep. time).

Configuration Opt (%) Time Gap (%) Num. Cuts Sep. Time
n m̄ γ λc (G) CST CST-1 CST CST-1 CST CST-1 CST CST-1 CST CST-1
40 78 0.05 0.00 100 100 0.02 0.04 0.0 0.0 39 51 0.00 0.01
0.30 0.00 100 100 0.30 0.24 0.0 0.0 616 549 0.07 0.06
0.70 0.00 100 100 0.31 0.24 0.0 0.0 616 549 0.07 0.06
0.95 0.00 100 100 0.32 0.24 0.0 0.0 616 549 0.07 0.06
118 0.05 0.00 100 100 0.50 0.26 0.0 0.0 579 459 0.08 0.05
0.30 0.00 79 96 1, 521.26 1,163.71 1.3 0.6 33, 557 32, 419 32.36 22.19
0.70 0.00 79 96 1, 520.60 1,166.55 1.3 0.6 33, 591 32, 401 32.17 22.00
0.95 0.00 79 96 1, 520.42 1,156.39 1.3 0.6 33, 530 32, 539 31.83 22.38
80 160 0.05 0.01 100 100 0.41 0.49 0.0 0.0 460 513 0.10 0.09
0.30 0.01 100 100 417.60 87.04 0.0 0.0 10, 151 7, 774 7.77 2.37
0.70 0.01 100 100 429.02 81.89 0.0 0.0 10, 151 7, 774 7.87 2.37
0.95 0.01 100 100 427.80 83.81 0.0 0.0 10, 151 7, 774 8.11 2.35
240 0.05 0.02 71 75 2, 183.99 1,943.23 10.1 9.8 35, 350 41, 248 39.01 26.64
0.30 0.02 50 54 3, 879.27 3,682.90 15.0 14.7 63, 256 76, 053 65.18 46.34
0.70 0.02 50 54 3, 877.28 3,683.06 15.0 14.7 62, 657 76, 121 64.99 46.59
0.95 0.02 50 54 3, 880.59 3,680.04 15.0 14.7 62, 990 76, 196 64.79 46.36
305 0.05 0.04 83 83 1, 753.58 1,469.49 4.0 2.9 38, 062 40, 842 31.56 18.93
0.30 0.04 50 50 4, 460.88 3,864.57 18.1 17.2 76, 276 77, 701 78.58 48.50
0.70 0.04 50 50 4, 490.00 3,870.72 18.1 17.2 76, 217 78, 267 79.64 48.66
0.95 0.04 50 50 4, 497.59 3,851.17 18.1 17.2 76, 190 77, 742 79.78 48.07
160 318 0.05 0.05 100 100 482.46 97.28 0.0 0.0 10, 843 11, 076 11.24 6.15
0.30 0.05 90 90 1, 184.63 789.86 4.7 4.8 17, 259 18, 296 22.61 16.58
0.70 0.05 90 90 1, 188.60 790.20 4.7 4.8 17, 202 18, 181 22.85 16.49
0.95 0.05 90 90 1, 193.56 790.62 4.7 4.8 17, 218 18, 287 23.25 16.45
476 0.05 0.06 40 48 4, 554.39 4,021.91 24.1 22.5 54, 125 64, 388 80.05 62.79
0.30 0.06 40 48 4, 559.88 4,019.90 24.1 23.4 54, 282 63, 315 78.73 62.93
0.70 0.06 40 48 4, 547.75 4,029.75 24.1 23.4 54, 169 63, 153 78.81 63.10
0.95 0.06 40 48 4, 551.50 4,025.95 24.1 23.4 53, 996 63, 254 79.33 63.12
624 0.05 0.06 0 0 7,200.00 7,200.00 42.8 44.2 78, 794 92, 249 116.04 86.68
0.30 0.06 0 0 7,200.00 7,200.00 42.8 44.1 79, 571 92, 678 116.18 86.07
0.70 0.06 0 0 7,200.00 7,200.00 42.9 44.1 78, 600 93, 037 116.05 86.94
0.95 0.06 0 0 7,200.00 7,200.00 42.9 44.1 78, 782 92, 626 114.94 85.85
200 398 0.05 0.08 90 100 777.55 280.76 0.9 0.0 10, 719 14, 264 12.99 11.17
0.30 0.08 90 100 773.93 273.86 0.9 0.0 10, 779 14, 264 12.61 10.87
0.70 0.08 90 100 771.35 276.13 0.9 0.0 10, 858 14, 264 12.41 10.77
0.95 0.08 90 100 774.78 273.65 0.9 0.0 10, 688 14, 264 12.55 11.15
596 0.05 0.09 48 48 3,858.18 3, 904.13 30.2 31.0 42, 346 47, 730 65.39 55.38
0.30 0.09 48 48 3,857.94 3, 906.96 30.2 31.0 42, 244 47, 816 64.90 55.44
0.70 0.09 48 48 3,854.05 3, 911.56 30.2 31.1 42, 179 47, 784 64.78 55.56
0.95 0.09 48 48 3,855.45 3, 903.94 30.2 31.0 42, 347 47, 849 64.63 55.32
784 0.05 0.09 0 0 7,200.00 7,200.00 63.2 64.0 71, 315 82, 216 122.47 112.42
0.30 0.09 0 0 7,200.00 7,200.00 63.2 64.0 71, 608 82, 111 120.90 112.43
0.70 0.09 0 0 7,200.00 7,200.00 63.2 64.0 71, 476 82, 241 120.26 113.03
0.95 0.09 0 0 7,200.00 7,200.00 63.2 64.0 71, 090 82, 118 122.40 112.14

20
Table 2: Computational results for the EXT, CST-1, and CES formulations. For each graph configuration,
we show the time required to identify a minimum-cost edge cut (λc (G)), the percentage of instances solved to
optimality (Opt), the average run time in seconds (time), and the average optimality gap (Gap).

Configration Opt (%) Time Gap (%)

n m̄ γ λc (G) EXT CST-1 CES EXT CST-1 CES EXT CST-1 CES
40 78 0.05 0 100 100 100 3.41 0.04 0.08 0 0 0
0.30 0 100 100 100 5.33 0.24 0.13 0 0 0
0.70 0 100 100 100 5.29 0.24 0.04 0 0 0
0.95 0 100 100 100 5.34 0.24 0.01 0 0 0
118 0.05 0 100 100 100 33.01 0.26 250.12 0 0 0
0.30 0 100 96 100 38.19 1, 163.71 89.29 0 1 0
0.70 0 100 96 100 21.28 1, 166.55 0.38 0 1 0
0.95 0 100 96 100 20.86 1, 156.39 0.01 0 1 0
80 160 0.05 0.01 100 100 100 430.38 0.49 2.74 0 0 0
0.30 0.01 100 100 100 108.36 87.04 4.39 0 0 0
0.70 0.01 100 100 100 53.92 81.89 0.36 0 0 0
0.95 0.01 100 100 100 39.48 83.81 0.01 0 0 0
240 0.05 0.02 67 75 63 3, 663.23 1,943.23 2, 821.78 22 10 12
0.30 0.02 100 54 79 648.94 3, 682.90 1, 566.82 0 15 5
0.70 0.02 100 54 100 477.78 3, 683.06 30.72 0 15 0
0.95 0.02 100 54 100 373.60 3, 680.04 0.02 0 15 0
305 0.05 0.04 83 83 67 4, 425.12 1,469.49 3, 463.12 5 3 13
0.30 0.04 100 50 83 1, 936.39 3, 864.57 1,818.04 0 17 4
0.70 0.04 100 50 100 2, 328.33 3, 870.72 12.19 0 17 0
0.95 0.04 100 50 100 2, 241.41 3, 851.17 0.02 0 17 0
160 318 0.05 0.05 80 100 90 3, 845.33 97.28 848.19 18 0 3
0.30 0.05 100 90 100 2, 295.34 789.86 25.62 0 5 0
0.70 0.05 90 90 100 2, 174.50 790.20 0.70 9 5 0
0.95 0.05 90 90 100 1, 973.65 790.62 0.02 8 5 0
476 0.05 0.06 24 48 48 5, 876.53 4, 021.91 3,829.95 64 22 18
0.30 0.06 32 48 68 5, 479.02 4, 019.90 2,448.93 52 23 9
0.70 0.06 40 48 100 5, 512.48 4, 029.75 64.36 44 23 0
0.95 0.06 40 48 100 5, 009.95 4, 025.95 0.04 39 23 0
624 0.05 0.06 0 0 20 7, 200.00 7, 200.00 6,066.43 95 44 35
0.30 0.06 0 0 40 7, 200.00 7, 200.00 4,409.31 94 44 21
0.70 0.06 0 0 100 7, 200.00 7, 200.00 274.48 90 44 0
0.95 0.06 0 0 100 7, 200.00 7, 200.00 0.06 100 44 0
200 398 0.05 0.08 60 100 100 4, 377.91 280.76 645.62 38 0 0
0.30 0.08 50 100 100 4, 222.71 273.86 12.33 36 0 0
0.70 0.08 60 100 100 3, 493.65 276.13 2.66 21 0 0
0.95 0.08 60 100 100 4, 136.79 273.65 0.02 20 0 0
596 0.05 0.09 28 48 48 5, 597.36 3, 904.13 3,757.74 67 31 26
0.30 0.09 44 48 56 4, 706.76 3, 906.96 3,270.92 51 31 18
0.70 0.09 36 48 92 5, 462.50 3, 911.56 1,109.46 59 31 2
0.95 0.09 40 48 100 5, 762.72 3, 903.94 0.08 53 31 0
784 0.05 0.09 0 0 0 7,200.00 7,200.00 7,200.00 97 64 53
0.30 0.09 0 0 20 7, 200.00 7, 200.00 5,789.52 100 64 37
0.70 0.09 0 0 75 7, 200.00 7, 200.00 2,124.91 100 64 1
0.95 0.09 0 0 100 7, 200.00 7, 200.00 0.16 100 64 0
Average 0.03 68 68 87 2, 964.13 2, 569.97 1,051.13 27 15 5

21
 Third, one additional explanation of why the CES formulation performs significantly better for larger values
of γ is that whenever the interdiction requirement becomes relatively large, it is likely for the optimal solutions
of the instances to be minimum-cost edge cuts (i.e., the interdiction requirement is so high that the best option
for the attacker is to disconnect the graph). Despite all formulations starting with a minimum-cost edge cut as
an incumbent solution, it takes significantly less time for the CES formulation to close the gap thanks to its
stronger constraints.

6.4 CST and CES Constraint Separation Comparison

We now study the performance of separation subroutines for the CST and CES formulations in terms of the
total number of cuts generated and the strength of such cuts. To this end, we analyze three different metrics:
the number of cuts generated by each separation subroutine; the average size of the cuts generated (i.e., the
number of variables with nonzero coefficient); and the separation time ratio, which we calculate as the time
spent by the separation subroutine over the total running time taken by the optimizer to solve each instance.
We summarize these results in Table 3.
Table 3: Computational results of the cut generation subroutines for the CST and CES formulations. For
each graph configuration, we show the average number of constraints added by the separation subroutine
(Constraints), the average size of the constraints (size), and the average separation time ratio (Sep. Ratio).

Configuration Constraints Size Sep. Ratio Configuration Constraints Size Sep. Ratio
n δ γ CST CES CST CES CST CES n δ γ CST CES CST CES CST CES
40 78 0.05 39 109 39 36 6 55 160 318 0.70 17, 202 217 159 59 22 98
0.30 616 251 39 28 28 59 0.95 17, 218 7 159 7 21 85
0.70 616 54 39 16 26 75 476 0.05 54, 125 69, 188 159 146 9 30
0.95 616 16 39 2 27 10 0.30 54, 282 48, 798 159 109 9 49
118 0.05 579 11, 453 39 37 21 45 0.70 54, 169 11, 702 159 75 9 86
0.30 33, 557 8, 872 39 28 13 53 0.95 53, 996 20 159 7 9 85
0.70 33, 591 627 39 18 13 80 624 0.05 78, 794 102, 209 159 145 2 6
0.95 33, 530 15 39 2 12 13 0.30 79, 571 81, 787 159 106 2 20
80 160 0.05 460 2, 111 79 73 34 61 0.70 78, 600 41, 180 159 85 2 70
0.30 10, 151 1, 923 79 55 21 79 0.95 78, 782 30 159 7 2 72
0.70 10, 151 233 79 32 21 90 200 398 0.05 10, 719 14, 803 199 179 24 79
0.95 10, 151 10 79 4 21 91 0.30 10, 779 2, 312 199 142 25 89
240 0.05 35, 350 60, 714 79 74 19 27 0.70 10, 858 663 199 78 25 99
0.30 63, 256 39, 271 79 56 13 45 0.95 10, 688 8 199 9 24 97
0.70 62, 657 6, 386 79 38 13 85 596 0.05 42, 346 53, 423 199 182 16 42
0.95 62, 990 17 79 4 13 86 0.30 42, 244 54, 065 199 136 16 48
305 0.05 38, 062 76, 541 79 73 4 16 0.70 42, 179 31, 639 199 83 17 75
0.30 76, 276 49, 249 79 54 3 38 0.95 42, 347 36 199 9 16 93
0.70 76, 217 7, 838 79 41 3 88 784 0.05 71, 315 99, 195 199 181 2 5
0.95 76, 190 12 79 4 3 67 0.30 71, 608 91, 791 199 132 2 11
160 318 0.05 10, 843 21, 253 159 143 22 68 0.70 71, 476 78, 270 199 102 2 58
0.30 17, 259 3, 401 159 112 21 86 0.95 71, 090 73 199 9 2 82

As one might expect, by observing the total number of cuts added and separation time ratio of both
approaches, it is easy to conclude that the separation subroutine for the CST formulation is much faster, as it
only requires finding spanning trees of small weight, whereas the CES separation algorithm requires exploring
the branches of the semilattice structure described in Section 5. Indeed, the separation time ratio for CES is
in average about four times the one of CST.
Furthermore, the results shown in Table 3 also provide a clear explanation why the CES formulation typically
outperforms the CST formulation in almost all the graph configurations, despite having to compute a more
complex separation subroutine. First, the experiments indicate that the size of the constraints—measured by
the number of non-zeros—of the CES formulation is always smaller than those of CST, often by a large margin.
As expected, this trend becomes significantly more pronounced for larger values of γ. Notice for example that
when γ = 0.70, the size of the critical edge set constraints is about half the size of the critical spanning tree
constraints; and, when γ = 0.95, the size of the critical edge set constraints becomes less than 6% the size of

22
the critical spanning tree constraints, making the former significantly stronger. Second, it can be observed that
the average number of constraints added to the CES formulation is much smaller than the ones added to CST,
which is also expected given the result from Corollary 3. The number of cuts generated by the CES separation
subroutine is only larger than the one for CST when γ is small, but decreases rapidly as γ becomes larger.
With constraints of such smaller sizes, it is natural to expect the CES’s feasible set to be significantly smaller
than the one of CST. A natural observation derived from this trend is that whenever the value of γ grows, a
large portion of the feasible solutions for the attacker tend to be edge cuts, which are part of the solution space
of CST, but by design are excluded form the one of CES.

6.5 Results on Real-Life Networks

All the experiments presented so far were conducted on randomly generated graphs. In general, real-life
infrastructure graphs have interesting properties that are worth studying. In Table 4, we show the results
of running the CST-1 and CES algorithms over 12 real-life networks collected from the Network Repository
(2019). These are instances that have more vertices than the random graphs we used before, but in general
are sparser. For these instances, we also generate the weight on each edge using the same uniform distribution
ranging from 20 to 80, but we assume the interdiction costs are one. Also, we set the interdiction percentage γ
at the lowest level 0.05. The results are presented in Table 4. Interestingly, due to the sparsity of the graphs,
despite their size, both formulations are able to solve these instances significantly faster than the randomly
generated graphs. We notice that CST-1 is slightly faster than CES over all the instances. This is consistent
with the previous analysis, since when γ is small, CST-1 tends to have a better performance.

Table 4: Computational results for the CES formulation for solving several real-life instances (time in seconds).

Network type Instance n m Avg. Degree γ r Time CST-1 Time CES

inf-euroroad 2, 641 3, 303 2.50 0.05 50, 542 0.16 0.83
inf-openflights 662 1, 568 4.74 0.05 88, 446 0.07 0.36
Infrastructure
inf-power 494 1, 080 4.37 0.05 20 042, 61 0.06 0.49
inf-USAir97 1, 174 1, 442 2.46 0.05 79, 46 0.01 0.06
power-1138-bus 5, 300 13, 571 5.12 0.05 49, 115 0.03 0.12
power-494-bus 1, 174 1, 442 2.46 0.05 21, 840 0.01 0.03
Power power-662-bus 4, 9377 6, 587 2.67 0.05 26, 525 0.02 0.05
power-bcspwr10 4, 936 6, 584 2.67 0.05 192, 378 0.12 0.66
power-US-Grid 2, 939 15, 687 10.68 0.05 20 045, 29 0.06 0.48
road-euroroad 1, 138 2, 596 4.56 0.05 48, 906 0.02 0.1
Road
road-minnesota 327 2, 113 12.92 0.05 109, 363 0.12 0.92

7 Concluding Remarks
In this paper, we discussed several integer programs for solving a minimum spanning tree interdiction problem
where the objective of the attacker is to minimize the cost of raising the weight of the minimum spanning tree
in the residual graph to a given threshold by removing a set of edges.
We performed a polyhedral analysis of the convex hull of feasible solutions for the problem, identifying
several families of facet defining inequalities, and designing the corresponding algorithms to separate them. We
also developed a third formulation composed by a family of so-called supervalid inequalities that may remove
several feasible solutions, but retain at least one optimal solution. The constraints of this third formulation
result from a new class of structures we called critical edge sets. We proved several fundamental properties of
such structures and of the resulting formulation.
Our experiments showed that a systematic generation of cactus inequalities of rank one can help to improve
the performance of the critical spanning tree formulation. Furthermore, they also evidenced that the critical
edge set formulation outperforms the other two in most of the instances we tested.
Finally, we believe that identifying other families of valid inequalities, as well as developing efficient sep-
aration algorithms of cactus inequalities of higher ranks can be beneficial in practice. All the developments

23
of this paper are aimed towards exact solution approaches for the spanning tree interdiction problem. We
believe, however, that exploring heuristic algorithms can further complement our approaches and may help
to potentially solve other large and challenging instances. Furthermore, it may be interesting to test whether
other types of critical structures, like the critical edge sets, can be found in other graph interdiction problems
to produce similar types of supervalid inequalities.

24
A Mathematical Proofs
Proposition 1 Setting the values of Me , for all e ∈ E so that
  
  X 
Me ≥ max 0, max r− we0 , ∀e ∈ E, (18)
 T ∈ΩG [e]  
e0 ∈E(T )

results in a valid EXT formulation for the minimum spanning tree interdiction problem.

Proof. By the weak duality theorem, constraint (16) ensures that for any spanning tree T ∈ ΩG whose weight
is less than r, at least one edge e ∈ E(T ) must be removed by the attacker (i.e., there is at least one edge
e ∈ E(T ) for which xe = 1). Thus, to guarantee the validity of (15)–(17), it suffices to show that for the given
bounds, any spanning tree T in G that includes an edge e for which xe = 1 has an objective value in (9) of no
less than r. To this end, assume for a contradiction that
P for a given feasible
P attacking strategy S ⊆ E, there
is a spanning tree T ∈ ΩG so that E(T ) ∩ S 6= ∅ and e∈E(T ) we + e∈E(T )∩S Me < r. Then, for any edge
e ∈ E(T ) ∩ S,  
X X  X 
Me ≤ Me0 < r − we0 ≤ max r− we0 ,
0 0
T 0 ∈ΩG [e]  0 0

e ∈E(T )∩S e ∈E(T ) e ∈E(T )

which contradicts (18).P Furthermore, this bound is tight for edge e ∈ E, if there is a spanning tree T ∈
arg maxT 0 ∈ΩG [e] {r − e0 ∈E(T 0 ) we0 }, so that w(T ) < r and S ∩ T = {e}. 

2
Theorem 3 Let Pext = {(x, α, β, γ) ∈ [0, 1]m × Rn × Rnm × R, satisfying (11)–(14),(16)} be the polytope
described by the linear relaxation of the extended formulation, with a proper value for Me , ∀e ∈ E satisfying
(18). Then, the projection projx (Pext ) onto the x-space corresponds to the polyhedron described by the following
inequalities:
X X
Me x e ≥ r − we , ∀T ∈ ΩG (22)
e∈E(T ) e∈E(T )

0 ≤ xe ≤ 1, ∀e ∈ E. (23)

Proof. Let S ⊆ E be an interdiction strategy for the attacker, with x = [xe ]e∈E being the incidence vector of
edge set S. From (10)–(14) we have that S is a feasible solution for the attacker if the following system is not
empty:
X X
− bli αil − (n − 1)γ ≤ −r
i∈V l∈V \{k}
X
βel + γ ≤ we + Me xe ∀e ∈ E
l∈V \{k}

αil − αjl − βel ≤ 0 ∀e ∈ E, l ∈ V \ {k}

αjl − αil − βel ≤0 ∀e ∈ E, l ∈ V \ {k}
βel ≥0 ∀e ∈ E, l ∈ V \ {k}.
By the Farkas’ theorem of alternatives, the above system has a solution if and only if there is no solution
(z, y, f ), where z ∈ R, y ∈ Rn and f ∈ R(2m)×(n−1) , that satisfies the following system:
X
(we + Me xe )ye − rz < 0 (28)
e∈E
X X
l l
fij − fji − bli z = 0 ∀i ∈ V, l ∈ V \ {k} (29)
{j:{i,j}∈E} {j:{i,j}∈E}
l l
ye − fij − fji ≥0 ∀e = {i, j} ∈ E, l ∈ V \ {k} (30)

25
 X
ye − (n − 1)z = 0 (31)
e∈E
z≥0 (32)
l l
fij , fji ≥0 ∀{i, j} ∈ E, l ∈ V \ {k} (33)
ye ≥ 0 ∀e ∈ E, (34)

Claim 1 The set of extreme rays of the cone defined by (29)–(34) is characterized by the spanning trees of
graph G.

Proof. First, it should be clear that the cone is centered at the origin (z, y, f ) = (0, 0, 0) and from (29)
and (31), having z = 0 results in y = 0 and f = 0, which implies that, for all the rays of the given cone,
z > 0. Second, note that setting z = 1 results in the network design formulation (6)–(8), which is known to
be integral and whose extreme points characterize the set of spanning trees in G (see Remark 1). Finally, the
value of z can be seen as a positive multiplier for the right hand side of constraints (6) and (7). Thus, any ray
satisfying (29)–(34) with z = ẑ > 0 is essentially the ray (1, y, f ) that characterizes a spanning tree T ∈ ΩG
after being scaled by ẑ.

By the Minkowski-Weyl representation theorem, all the solutions of cone (29)–(34) can be produced as conic
combinations of its extreme rays. Thus, system (28)–(34) has no solution if
X
(we + Me xe ) ≥ r ∀T ∈ ΩG ,
e∈T

proving the result. 

Corollary 1 Let Pcst = {x ∈ [0, 1]m , x satisfying (20)} be the polytope described by the linear relaxation of
the critical spanning tree formulation, then

1. Pcst ⊆ projx (Pext ).

2. Pcst = projx (Pext ) if and only if all critical spanning trees have the same weight.
P
Proof. For part 1, we argue that given any x ∈ Pcst , the inequality e∈E(T ) Me xe ≥ r − w(T ) holds for all
/ ΩrG , as for
T ∈ ΩG . First, it is clear that the corresponding constraint is trivially satisfied for any tree T ∈
r
those trees, r < w(T ). Now, if T ∈ ΩG , then we have
X
Me xe ≥ min Me ≥ r − w(T ),
e∈T
e∈E(T )

where the first inequality stands from the fact that x ∈ Pcst and the second inequality is derived directly from
the definition of Me in (18).
For part 2, it is enough to show that Pcst ⊇ projx (Pext ) if and only if all critical spanning trees have the
same weight. To prove sufficiency, we note that if all the critical spanning trees have the same weight (i.e.,
w(T ) = wG , ∀T ∈ ΩrG ), then from (18) we obtain

Me = r − min w(T 0 ) = r − wG ,
T 0 ∈ΩG [e]

for all e ∈ E. Now, for any x ∈ projx (Pext ) we have

X
Me xe ≥ r − w(T ) = r − wG ), ∀T ∈ ΩrG .
e∈E(T )

26
Dividing by Me on both sides yields e∈T xe ≥ 1, ∀T ∈ ΩrG , which implies that x ∈ Pcst .
P
To prove necessity, we begin by showing that Pcst ⊇ projx (Pext ) implies that

Me = r − w(T ), ∀T ∈ ΩrG , e ∈ E(T ). (35)

To prove this statement, assume for a contradiction that Pcst ⊇ projx (Pext ) and that there exists a T0 ∈ ΩrG
and an edge e0 ∈ E(T0 ) for which Me0 6= r − w(T0 ). From (18), this implies that Me0 > r − w(T0 ). We now
show that in such a case, it is possible to construct a solution x̄ ∈ projx (Pext ), so that x̄ ∈
/ Pcst . The proposed
construction is as follows: 
(r − w(T0 )) /Me0 e = e0 ,

x̄e = 0 e ∈ E(T0 ) \ {e0 },

1 e ∈ E \ E(T0 ).


First, to see that x̄ ∈ projx (Pext ) note that for any T ∈ ΩrG , if T 6= T0 , there is some edge e1 ∈ E(T ) for which
x̄e1 = 1, then from (18) we obtain
X
Me x̄e ≥ Me1 x̄e1 = Me1 ≥ r − w(T ).
e∈E(T )

On the other hand, if T = T0 , we have

X
Me xe = Me0 xe0 = Me0 (r − w(T0 ))/Me0 = r − w(T0 ),
e∈E(T0 )
P
which proves that x̄ satisfies (22)–(23). Furthermore, observe that for this solution, e∈T0 x̄e = (r−w(T0 ))/Me0 <
1 implying that x̄ ∈ / Pcst . Thus, Pcst 6⊇ projx (Pext ), a contradiction.
We continue the proof by considering any pair of critical spanning trees T1 , T2 ∈ ΩrG . Notice that that if there
exists some edge ê ∈ E for which ê ∈ E(T1 ) ∩ E(T2 ), then from (35) we have that Mê = r − w(T1 ) = r − w(T2 ),
implying that w(T1 ) = w(T2 ). Now, suppose E(T1 ) ∩ E(T2 ) = ∅; then, since the spanning trees are the basis of
the graph matroid, by the Strong Basis Exchange Property of matroids (Brualdi 1969), there exists a pair of
edges e1 ∈ E(T1 ) and e2 ∈ E(T2 ) such that the subgraphs T10 and T20 induced by the edge sets E(T1 )\{e1 }∪{e2 }
and E(T2 ) \ {e2 } ∪ {e1 }, respectively, are spanning trees. Without loss of generality, assume that w(T10 ) ≤ w(T1 )
(otherwise w(T20 ) ≤ w(T2 ) and the following result applies for T20 ); then, T10 must also be a critical spanning tree
and since T10 shares at least one edge with T1 and T2 , form (35) we have that w(T10 ) = w(T1 ) = w(T2 ), proving
the desired result. 

Theorem 4 Let x be a candidate solution for the attacker. Then, separating the critical spanning tree con-
straints is:

1. NP-hard if x ∈ [0, 1]m

2. solvable in O(m log n) if x ∈ {0, 1}m

Proof. The membership in NP should be obvious. The first statement follows by noting that this separation
problem is closely related to the constrained minimum spanning tree problem, which is known to be weakly
NP-hard (Aggarwal et al. 1982, Ravi and Goemans 1996). The decision version of this problem is given below.

Constrained minimum spanning tree (Decision)

Instance: A graph G = (V, E), a non-negative integer vector of edge weights v = [ve ]e∈E , a non-negative
integer vector of lengths ` = [le ]e∈E , and two integer positive numbers R and L.
P P
Question: Is there a spanning tree T in G such that e∈T ve ≤ R and e∈T le < L?

27
 Given an instance of the constrained minimum spanning tree problem, one can easily construct a critical
spanning tree separation instance by setting xe = min{1, le /L}, ∀e ∈ E, r = R − 1, and w = v, which is
possible given that constant L > 0. This yields the separation problem NP-hard.
To prove that the integral separation is polynomial-time solvable, it suffices to see that finding a minimum
spanning tree, according to the weight vector w in graph G(V, E \ S), where S = {e ∈ E, xe = 1} yields a most
violated critical spanning tree constraint, in case the weight of such spanning tree is less than r. Generating
G(V, E \ S) can be done in O(m) and solving the spanning tree problem can be done in O(m log n) (Ahuja
et al. 1993). 

Theorem 5 Given a simple, nonempty, connected graph G = (V, E), a weight vector w ∈ Qm so that there are
no irrelevant edges in E (i.e., min{w(T )| T ∈ ΩG [e]} < r, ∀e ∈ E), and a scalar r ∈ Q, so that wG < r < wG ,
the following statements are true.

(1) P is a full-dimensional polytope.

(2) Given an edge e ∈ E, inequality xe ≤ 1 induces a facet of P.

(3) Given an edge e ∈ E, inequality xe ≥ 0 induces a facet of P if and only if n ≥ 4.

(4) Given a critical spanning tree T ∈ ΩrG , the corresponding critical spanning tree inequality
P
e∈E(T ) xe ≥1
induces a facet of P if and only if:

(a) there is no edge e ∈ E such that if added to T , it completes a Hamiltonian cycle in G, or

(b) if such an edge exists, at least one of the spanning trees contained in the resulting Hamiltonian cycle
has a weight greater than r.

Proof.
(1) To show that P is a full-dimensional polytope, it suffices to prove that there are m+1 affinely independent
points in P. Notice that edge sets E \ {e}, ∀e ∈ E, along with edge set E are feasible solutions for the attacker
because removing each of those edge sets disconnects graph G. It can be easily verified that the incidence
vectors of these feasible solutions are affinely independent; therefore, P is a full-dimensional polytope.

(2) For any given edge e ∈ E, notice that the face of P that is induced by inequality xe ≤ 1 contains the
following m feasible solutions, the complete edge set E and the edge sets E \ {e0 }, ∀e0 ∈ E \ {e}. Since the
corresponding incidence vectors of these solutions are affinely independent, the dimension of such a face is
m − 1; therefore, it is a facet of P.

(3) To show sufficiency, we prove that for any given edge e ∈ E, the face induced by inequality xe ≥ 0 is
of dimension m − 1, whenever n ≥ 4. To this end, notice that sets E \ {e} and E \ {e, e0 }, ∀e0 ∈ E \ {e} are
feasible solutions for the attacker because removing the those edge sets leaves at the most two edges in the
graph, making the resulting graph disconnected. Observe that for the graph to remain connected, there must
exists a spanning tree with n − 1 ≥ 3 edges, since by assumption n ≥ 4. Furthermore, it is easy to see that
such m solutions belong to the face and are affinely independent.
To show necessity, it is trivial to see by inspection that xe ≥ 0 only induces a facet of P for a connected graph
G with n ≤ 3 in cases where wG = wG ≥ r, or whenever there is an irrelevant edge in G, which are both trivial
instances ruled out by definition.

(4) To show sufficiency, it is enough to prove that given a critical spanning tree TP∈ ΩrG , satisfying either
(a) or (b), the dimension of the face induced by the critical spanning tree inequality e∈E(T ) xe ≥ 1 is m − 1.
We will show that it is possible to construct m affinely independent solutions contained in such a face. The
proposed construction is as follows. First, consider the n − 1 solutions, given by one edge e ∈ E(T ) and all the
edges in E \ E(T ). Clearly, such solutions lie in the given face and are feasible alternatives for the attacker, as

28
those disconnect the graph. Second, consider the solution composed of all the edges in E \ E(T ) except for one
edge e0 and one edge e ∈ E(T ) that is selected depending on e0 as follows: since adding edge e0 to spanning
tree T produces a cycle, if T satisfies condition (a), then select as edge e one of the edges in E(T ) that is not
in the cycle. Alternatively, if T satisfies condition (b), then select as edge e one of the edges in E(T ) so that
the tree induced by (E(T ) \ {e}) ∪ {e0 } has a weight greater than r. This procedure is then repeated for every
edge e0 ∈ E \ E(T ), thus generating the other m − n + 1 solutions. Notice that all of them lie on the given face
and are feasible for the attacker, as the residual graph becomes disconnected whenever T satisfies (a), or the
residual graph left is a spanning tree of weight greater than r whenever T satisfies (b). It can be easily verified
that the incidence vectors associated with such solutions are affinely independent.
To show necessity, assume for a contradiction that there is a spanning tree T ∈ ΩrG and an edge e0 ∈ E \ E(T )
so that E(T ) ∪ {e0 } yields a Hamiltonian cycle in G for which all the n spanning trees contained in it are in ΩrG .
First, notice that we can take all the n critical spanning tree inequalities associated with the spanning trees
contained in the Hamiltonian cycle and apply the Chvátal-Gomory rounding procedure (Chvátal 1973, Gomory
1963) to them in order to generate a new valid inequality. To this end, we simply add such inequities and since
each edge in the Hamiltonian cycle appears in n − 1 spanning trees, after dividing by n − 1 and rounding up
the resulting values, we obtain the following valid inequality
 
X n
xe + xe0 ≥ = 2. (36)
n−1
e∈E(T )

Then, notice that adding the upper bound inequality −xe0 ≥ −1 to (36) produces the original critical spanning
tree inequality, which implies that such a constraint can be generated by a linear combination of other valid
inequalities for the problem and therefore does not induce a facet of P. This concludes the proof. 
Theorem 6 Given a simple, nonempty, connected graph G = (V, E), a weight vector w ∈ Qm , a scalar r ∈ Q
so that wG < r < wG , and a k-cactus H of G, with k ≥ 1, the inequality
X X X
xe + 2 xe + 2 xe ≥ 2 (24)
e∈Ĉ(H) e∈C(H)\Ĉ(H) e∈F (H)

induces a facet of P, if and only if |Ĉl (H)| ≥ 3, ∀l = 1, . . . , k.

Proof. We begin by proving that the cactus inequality (24) is valid for P. To this end, we apply the Chvátal-
Gomory rounding procedure for all the critical spanning tree inequalities associated with the critical spanning
trees in H. For each edge e ∈ E(H), let πe and π̄e be the number of critical spanning trees in H that contain
and not contain edge e, respectively. Since the total number of critical spanning trees in H is |ΩrH |, we have
that πe + π̄e = |ΩrH |, for every e ∈ E(H). Furthermore, notice that πe = |ΩrH |, for all e ∈ F (H) as well as
ˆ
for all e ∈ C(H) \ C(H). To see this, observe that each edge e ∈ F (H) must be part of any spanning tree of
H because removing e disconnects H. Similarly, each e ∈ C(H) \ C(H) ˆ must exist in all the critical spanning
trees of H because those trees are generated by removing one edge from each Ĉl (H), ∀l = 1, . . . , k. If one of
ˆ
the edges in C(H) \ C(H) were to be missing from one of the critical spanning trees, that would imply that the
tree does not span V because in such a case there must be a cycle Ĉl (H) for some l ∈ {1, . . . , k} for which two
edges are not in the tree.
Now, after adding all the inequalities associated with the critical spanning trees in H, dividing by π =
maxe∈Ĉ(H) {πe }, and rounding up the resulting coefficients, we obtain the following expression:
 r  X  |Ωr |   r 
X l πe m X |ΩH | H |ΩH |
xe + xe + xe ≥ . (37)
π π π π
ˆ
e∈C(H) ˆ
e∈C(H)\C(H) e∈F (H)

|ΩrH |
l m
To complete the proof of validity, it suffices to show that 1 < π ≤ 2. Let e0 ∈ arg maxe∈Ĉ(H) {πe }.
First, note that
|ΩrH | πe0 + π̄e0 π̄e0
= =1+ > 1.
π πe0 πe0

29
Then, assume that e0 ∈ Ĉl (H) and, since by definition |Ĉl (H)| ≥ 3, let e00 ∈ Ĉl (H), so that e00 6= e0 . It should
be clear that πe00 ≥ π̄e0 because both edges cannot be missing at the same time from any spanning tree, so if e0
is not in a spanning tree, e00 must be in such a tree. Furthermore, observe that π = πe0 ≥ πe00 ≥ π̄e0 , proving
π̄
that πe00 ≤ 1.
e
We now proceed to prove the main result of the theorem. To show sufficiency, it is enough to prove that for
the given k-cactus H, the dimension of the face induced by (24) is m − 1. For this purpose, we will show that
it is possible to construct m affinely independent, feasible solutions contained in such a face. The proposed
construction is divided into the following three steps:
First, let e(1) , e(2) , and e(3) be the first three edges of Ĉl (E), based on the index given to those edges in
E. For each edge e ∈ Ĉl (H) \ {e(1) } and each cycle l ∈ 1, . . . , k, construct the solution given by the edge set
{e(1) , e} ∪ (E \ E(H)). Also, consider the solution associated with edge set {e(2) , e(3) } ∪ (E \ E(H)). This
procedure results in a total of |C(H)| ˆ solutions that lie on the given face. Note that all of these solutions are
feasible for the attacker because they disconnect G.
Second, consider all the |E(H) \ C(H)| ˆ ˆ
solutions composed of one edge e from either F (H) or C(H) \ C(H)
and all the edges in E \ E(H). Notice that if e ∈ F (H), then the resulting solution disconnects G and, if
e ∈ C(H) \ Ĉ(H), then the corresponding solution leaves a subgraph in which all the spanning trees left have
a weight at least r. Therefore, all of these solutions are feasible for the attacker and belong to the given face.
Third, consider the solutions composed of all the edges in E \ E(H), except for one edge e, and two edges e0
and e00 from a cycle Ĉl (H), so that l, e0 and e00 are selected depending on edge e as follows: (1) if edge e shares
both endpoints with edges in F (H), then select as e0 and e00 any two edges in Ĉl (H), for any l ∈ {1, . . . , k}.
(2) If edge e shares both endpoints, say vertices i and j, with edges of cycle Cl (H), for a given l ∈ {1, . . . , k}
(i.e., if e is a cord of cycle Cl (H)), then it is easy to see that cycle Cl (H) is composed of two paths connecting
vertices i and j and, since |Ĉl (H)| ≥ 3, one of these two paths must contain at least two edges in Ĉl (H). For
this case then, select as e0 and e00 any two edges in Ĉl (H) that belong to the same path. (3) If edge e shares
exactly one endpoint, say i, with edges in F (H) and the other endpoint, say j, with edges in Cl (H), for a given
l ∈ {1, . . . , k}, then there must be a path in H connecting vertices i and j that passes through a vertex p that
is the endpoint of both an edge in F (H) and an edge in Cl (H) (i.e., such a path goes from i to p via edges in
F (H) and from p to j via edges in Cl (H)). Furthermore, since vertices j and p are both endpoints of edges
in cycle Cl (H), there are two paths connecting such vertices in the cycle. For this case then, similarly as in
(2), select as e0 and e00 any two edges in Ĉl (H) that belong to the same path. (4) If edge e shares exactly one
endpoint, say i, with edges in Cl (H) and the other endpoint, say j, with edges in Cl0 (H), then there must be
a path in H connecting vertices i and j that passes through a vertex p 6= i that is the endpoint of an edge
in Cl (H) (i.e., such a path goes from i to p via edges in Cl (H) and from p to j via edges in F (H) ∪ Cl0 (H).
Furthermore, since vertices j and p are both endpoints of edges in cycle Cl (H), there are two paths connecting
such vertices in the cycle. For this case then, similarly as in (2) and (3), select as e0 and e00 any two edges in
Ĉl (H) that belong to the same path. Notice that this third step produces |E \ E(H)| solutions that belong to
the face and are specifically constructed to disconnect graph G; thus, are feasible solutions for the attacker.
For any given graph G having a k-cactus H in which Ĉl (H) ≥ 3, ∀l = 1, . . . , k, the above procedure generates
ˆ
|C(H)| + |E(H) \ C(H)| ˆ + |E \ E(H)| = m feasible solutions for the attacker. Moreover, from the way they
are constructed, it can be easily verified that the incidence vectors associated with such solutions are affinely
independent; therefore, cactus inequality (24) induces a facet of P.
Finally, to show necessity, we consider the two remaining trivial cases. First, notice that if |Ĉl (H) = 1| for
some l ∈ {1, . . . , k}, then the corresponding cactus inequality (24) for H is a linear combination of the cactus
inequality associated with the (k − 1)-cactus that results from removing the only edge e ∈ Ĉl (H) from H and
the lower bound xe ≥ 0; thus, it does not induce a facet of P. Second, if |Ĉl (H) = 2| for some l = 1, . . . , k|,
then the corresponding cactus inequality (24) for H is a linear combination of the two cactus inequalities for
the two (k − 1)-cacti that result from removing each of the two edges in Ĉl (H), respectively, and the two lower
bounds of such edges. 

30
Theorem 7 Let x ∈ [0, 1]m be a candidate solution for the attacker. Then, the separation problem for the
cactus inequalities is NP-hard.

Proof. The membership in NP should be obvious. Now, the following is a trivial transformation from the
critical spanning tree constraint separation problem, stated in Section 3, to the cactus inequality separation
problem. Given an instance for the critical spanning tree constraint separation problem defined by a graph
G = (V, E), w, x, and r, we construct a graph Ĝ by adding two dummy vertices v 0 and v 00 to G, one edge
between v 0 and a vertex v ∈ V , one edge between v 00 and v, and one edge between v 0 and v 00 . Hence, Ĝ =
(V ∪ {v 0 , v 00 }, E ∪ {{v, v 0 }, {v, v 00 }, {v 0 , v 00 }}). We create a weight vector ŵ ∈ Qm+3 by letting ŵe = we , ∀e ∈ E
and ŵ{v,v0 } = ŵ{v,v00 } = ŵ{v0 ,v00 } = 0. Similarly, we construct a vector x̂ ∈ [0, 1]m+3 by letting x̂e = xe , ∀e ∈ E
and x̂{v,v0 } = x̂{v,v00 } = x̂{v0 ,v00 } = 0. Finally, we say that r̂ = r. The procedure then results on an instance for
the cactus inequality separation problem defined by Ĝ, ŵ, x̂, and r̂.
It is easy to see that every spanning tree T in G is associated with a 1-cactus H in Ĝ whose cycle C1 (H)
is composed of edges {v, v 0 }, {v, v 00 }, {v 0 , v 00 } and whose set F (H) = E(T ). Clearly, if T ∈ ΩrG , then H is a
1-cactus for which Ĉl (H) = Cl (H) P and |Ĉl (H)| = 3. Moreover, since P x̂e = 0, ∀e ∈ C(H), then the resulting
cactus inequality is violated (i.e., 2 e∈F (H) xe < 2), if and only if e∈E(T ) xe < 1. 

Theorem 8 An edge set S ⊆ E is a feasible solution for the attacker if and only if S is an edge cut or
S ∩ K 6= ∅, for all K ∈ Λ̃rG .

Proof. The result is trivial for the case where S is an edge cut because removing S leaves the graph disconnected;
thus, we will focus on the alternative. To prove sufficiency, we recall from Theorem 2 that a solution S ⊆ E is
feasible for the attacker if it contains at least one edge from each critical spanning tree T ∈ ΩrG . Let T be any
of such trees. Since E(T ) contains some K ∈ Λ̃rG , then S ∩ E(T ) ⊇ S ∩ K 6= ∅, proving the result.
To show necessity, we prove the contrapositive. Suppose that a strategy S is not an edge cut and there
exists some K ∈ Λ̃rG , whose intersection with S is empty. Clearly, graph G(V, E \ S) is connected since S is not
an edge cut. Moreover, since K ⊆ (E \ S) and the fact that spanning trees are the bases of the graphic matroids
of connected graphs, by the augmentation property of matroids (Oxley 2006), there must exist a spanning tree
T in G(V ; E \ S) so that E(T ) contains K (i.e., T ∈ ΩG [K]). By definition all the spanning trees in ΩG [K] are
critical; therefore, T must be critical, which implies that S is not a feasible strategy. 
The Definition 4 and Proposition 3 below are for the proofs of Theorem 9 and 10.
Q
Definition 4 (Dependent Product) Given an indexed family of sets {Ai }i∈I , the dependent product i∈I Ai
is defined as the collection of all the indexed sets {ai ∈ Ai }i∈I . Equivalently, it is the space of all the choice
functions that select
Q one element from each set Ai . Since the sets in {Ai }i∈I are not necessarily disjoint, the
ordered sets in i∈I Ai may have repeated elements.

Proposition 3 Given a spanning tree T and a cycle C such that E(T ) ∩ C 6= ∅, then, for any e ∈ E(T ) ∩ C
there exists an edge e0 ∈ C \ E(T ) such that the cycle that is formed by adding e0 to T contains e.

Proof. Let e = {u, v} be an arbitrary edge in E(T )∩C, then the rest edges in the cycle form a path P = C \{e}
that connects u and v. Removing e from E(T ) disconnects T into two trees Tu and Tv that contain the endpoints
u and v, respectively. Let V (Tu ) and V (Tv ) be their sets of vertices; then, all the edges in E across these two
vertex sets compose an edge cut of G. Since the path P connects u ∈ V (Tu ) and v ∈ V (Tv ), there must exist
an edge e0 = {u0 , v 0 } ∈ P that is also in the cut. Now, we show that this choice of e0 has the desired properties.
First, e0 6= e because e is not an edge in P by definition. Since both u and u0 are in V (Tu ), and Tu is a tree,
there is an unique path in Tu that connects u and u0 . Similarly, there is an unique path in Tv that connects v
and v 0 . Then, these two paths along with e and e0 form a cycle in E(T ) ∪ {e0 } that contains e. 

Theorem 9 An edge set S ⊆ E is a feasible solution for the critical edge set formulation if and only if S is a
superset (not necessarily proper) of any non-edge-cut feasible strategy for the attacker.

31
Proof. To prove sufficiency, suppose the edge set S is the superset of a non-edge-cut feasible strategy for the
attacker S 0 . Then, by Theorem 8, S 0 ∩ K 6= ∅, for all K ∈ Λ̃rG ; therefore, both S and S 0 satisfy (26).
Now, let {Xi }i∈I be the family of all minimal non-edge-cut feasible strategies for the attacker, in no particular
order. To prove necessity, assume for a contradiction that S is a feasible solution for the critical edge set
formulation, but is not a superset of any non-edge-cutQ feasible strategy for the attacker. Then, XiQ\ S 6= ∅ for
all i ∈ I, which implies that the dependent product i∈I (Xi \S) is not empty. Clearly, every set L ∈ i∈I (Xi \S)
has a non-empty intersection with all non-edge-cut feasible solutions and L ∩ S = ∅, by construction.

Claim 2 If S is a feasible solution for the critical edge set formulation,

Q but it is not a superset of any non-
edge-cut feasible strategy for the attacker, then every set L ∈ i∈I (Xi \ S) has cycles.

Proof. First, note that ΩG [L] ⊆ ΩrG ; otherwise, there is a spanning tree T ∈ ΩG [L] which is non-critical
and therefore, E \ E[T ] is a non-edge-cut feasible strategy for the attacker, which is a contradiction since
L ∩ (E \ E[T ]) = ∅. Also, notice that L ∈ / ΛrG , i.e., L cannot be a critical edge set, otherwise S ∩ L 6= ∅ because
S is a feasible solution for the formulation. This forces ΩG [L] = ∅ by the definition of the critical edge sets,
which implies that some edges in L must form cycles.
We proceed to prove that the Q result of this claim leads to a contradiction by induction on the Q number of
cycles in any arbitrary set L ∈ i∈I (Xi \ S). Here, the induction basis states that there is an L ∈ i∈I (Xi \ S)
that containsQno cycles, which directly leads to a contradiction. Now for the inductive step, suppose that there
Q L ∈ i∈I (Xi \ S) that contains n + 1 cycles, we will show that in such a case, there exists another set
is a set
L0 ∈ i∈I (Xi \ S) that contains at the most n cycles. First, since set Xi is minimal, for all i ∈ I, for each edge
e ∈ Xi , there must exists a critical spanning tree Ti ∈ ΩrG such that E[Ti ]∩Xi = {e}; otherwise, if all the critical
spanning trees that contain {e} also contain more edges from Xi , then Xi \ {e} is still a feasible strategy, which
contradicts the minimality of Xi . Now, for an arbitrary cycle C ⊆ L, and an edge e ∈ arg maxe0 ∈C {we0 }, there
a spanning tree T ∈ ΩrG , such that E[T ] ∩ L = {e}. Then, by Proposition 3, there is another edge e0 ∈ C \ E[T ]
that forms a cycle with T that contains e. Since w(e) ≤ w(e0 ) by our choice of e, the new tree T 0 for which
E[T 0 ] = E[T ] \ {e} ∪ {e0 } has a smaller weight than the original tree T ; thus, is critical too. Let Xi \ S be the
set where e is selected from to compose L, then either e0 ∈ Xi or not. For the latter case, since E[T ] ∩ Xi = {e}
by our choice of T , and e0 ∈ / Xi , then E(T 0 ) ∩ Xi = ∅, which is a contradiction because it implies Xi is not
a feasible solution. For the former case, since e0 ∈ Xi , we can construct another set L0 ∈ i∈I (Xi \ S) by
Q
replacing e with e0 from set Xi \ S (clearly e0 is not in S since it has been selected from (Xj \ S)) and keeping
other elements the same. Now, either e is not in L0 , or it is still there because some other Xk \ S (k 6= i, k 6= j)
also select ei to compose L. In the former case, the cycle C has been removed along with the edge e, so there
are fewer cycles. Then, by induction, it leads to a contradiction. For the latter, we can repeat the previous
argument for any set Xk that also selected edge e until we reach the desired contradiction. This procedure
eventually terminates since L only contains a finite number of elements. 

Theorem 10 For any minimal critical edge set K ∈ Λ̃rG , there exists a spanning tree T ∈ ΩG [K] that does not
contain any other minimal critical edge set K 0 ∈ Λ̃r , where K 6= K 0 .

Proof. We prove the theorem by contradiction. Suppose there is a minimal critical edge set K ∈ Λ̃rG such
that all trees in ΩG [K] contain at least one critical edge set from Λ̃rG \ {K}. First, suppose that the spanning
trees in ΩG [K] are listed in no particular order according to an index set I, so that ΩG [K] = {Ti }i∈I . Then, by
assumption, we can directly create a list {Ki }i∈I so that Ki is a minimal critical edge set in Λ̃rG \ {K} contained
by spanning tree Ti , for all i ∈ I. The elements of list {Ki }i∈I
Q may not necessarily unique, but, since each Ki
is minimal
Q and different from K, the dependent product set i∈I (Ki \ K) is nonempty. Moreover, any element
L ∈ i∈I (Ki \ K) is an edge cut. For otherwise, G(V, E \ L) would still be connected and contain K, then
there must exist a spanning tree T in G(V, E \ L) that contains K. Note that E(T ) ∩ L = ∅, so it does not
contain any Ki by the construction of L, which Q is a contradiction.
Now, we show that the non-emptiness of i∈I (Ki \ K) leads to a contradiction by induction on the number
of minimal edge cuts that are contained in L. Here, the induction basis states that there are no minimal edge

32
cuts in L, which is a constradiction from the last paragraph. Now, for the inductive step, suppose L contains
n + 1 minimal edge cuts. Take any minimal edge cut D ⊆ L, and pick an edge e ∈ D with the highest weight,
and let Ki \ K be the set where e is picked from to produce L. Since Ki is a minimal critical edge set, there
must exist some non-critical spanning tree T that contains Ki \ {e}, so that e ∈ / E(T ). Clearly, the intersection
of E(T ) with D contains some other edge e0 since D is a minimal edge cut. Now, either e0 belongs to Ki \ K or
not. For the latter case, E(T ) ∪ {e} forms a cycle that contains both e and e0 , and by our choice, w(e) ≥ w(e0 ),
so the tree formed by the edge set E(T 0 ) = E(T ) \ {e0 } ∪ {e} is still non-critical. However, this contradicts the
fact that T 0 contains the critical edge set Ki . For the former case, we can construct a new L0 ∈ i∈I (Ki \ K)
Q
by replacing e with e0 in the set Ki \ K and keep other choices the same. Now, either L0 still contains e because
some other Kj \ K also selects e to compose L, or it does not. In the latter case, the minimal edge cut D
is not a subset of L0 , so there are fewer minimal edge cuts, then by induction, it leads to a contradiction. In
the former case, we repeat the previous argument for any set Kj that also contains edge e until we reach the
desired contradiction. This procedure terminates since L only contains a finite number of edges. 

Theorem 11 The following statements about the poset (↓ ΩrG , ⊆) are true:

1. The least element in (↓ ΩrG , ⊆) is the empty set.

2. The maximal elements in (↓ ΩrG , ⊆) are all the critical spanning trees.

3. (ΛrG , ⊆) is a subposet embedded in (↓ ΩrG , ⊆).

Proof. For part 1, the empty set is a least element since it is contained by all other sets; and it is known that
in any poset, the least element is unique given it exists. Part 2 is trivial, since ↓ ΩrG consists all the subsets
of critical spanning trees. For part 3, every critical edge set by definition is a subset of some critical spanning
tree, so it is embedded in the meet-semilattice (↓ ΩrG , ⊆) and the ordering on the set ΛrG is naturally inherited
from the poset. 

Proposition 2 Let β be the function from the ordered set (↓ ΩG , ⊆) to (R, ≤) defined as β(K) = maxT ∈ΩG [K] w(T ),
then β is order reversing.

Proof. If K ⊆ K 0 , then by definition ΩG [K] ⊇ ΩG [K 0 ], so the maximum in a subset must be less or equal to
the original set. 

33
B Notation Table

Symbol Definition
G = (V, E) graph G with vertex set V and edge set E
n number of vertices of given graph
m number of edges of given graph
V (H) vertices in subgraph H
E(H) edges in subgraph H
w(e) weight of edge e
w(K) total weight of edges in edge set K
w(H) total weight of edges in subgraph H
λ(G) edge connectivity of graph G
λc (G) value of a min-cost edge cut of graph G with respect to cost vector c
w(G) weight of a maximum spanning tree in G
w(G) weight of a minimum spanning tree in G
ΩG all spanning trees in graph G
ΩG [e] all spanning trees in graph G that contains edge e
ΩG [K] all spanning trees in graph G that contains edge set K
all spanning trees in graph G that have weight less than r, i.e., the set
ΩrG
of all critical spanning trees
ΛrG all critical edge sets in graph G with respect to r
Λ̃rG all minimal (with respect to set inclusion) critical edge sets in ΛrG
↓S the lower closure of S, i.e., ↓ S = {K ⊆ S | ∀S ∈ S}

34
References
Addis B, Summa MD, Grosso A (2013) Identifying critical nodes in undirected graphs: Complexity results and polynomial
algorithms for the case of bounded treewidth. Discrete Applied Mathematics 161(16–17):2349–2360.
Aggarwal V, Aneja Y, Nair K (1982) Minimal spanning tree subject to a side constraint. Computers & Operations Research
9(4):287 – 296.
Ahuja RK, Magnanti TL, Orlin JB (1993) Network Flows: Theory, Algorithms, and Applications (New Jersey, USA:
Prentice Hall).
Albert R, Barabási AL (2002) Statistical mechanics of complex networks. Rev. Mod. Phys. 74:47–97.
Arulselvan A, Commander CW, Elefteriadou L, Pardalos PM (2009) Detecting critical nodes in sparse graphs. Computers
and Operations Research 36(7):2193–2200.
Baı̈ou M, Barahona F (2008) A linear programming approach to increasing the weight of all minimum spanning trees.
Networks: An International Journal 52(4):227–234.
Bazgan C, Toubaline S, Tuza Z (2011) The most vital nodes with respect to independent set and vertex cover. Discrete
Applied Mathematics 159(17):1933–1946.
Bazgan C, Toubaline S, Vanderpooten D (2010) Complexity of determining the most vital elements for the 1-median
and 1-center location problems. Wu W, Daescu O, eds., Combinatorial Optimization and Applications, volume 6508,
237–251 (Springer Berlin Heidelberg).
Bazgan C, Toubaline S, Vanderpooten D (2012) Efficient determination of the k most vital edges for the minimum spanning
tree problem. Computers & Operations Research 39(11):2888–2898.
Bazgan C, Toubaline S, Vanderpooten D (2013) Critical edges/nodes for the minimum spanning tree problem: complexity
and approximation. Journal of Combinatorial Optimization 26(1):178–189.
Ben-Ameur W, Mohamed-Sidi MA, Neto J (2015) The k-separator problem: polyhedra, complexity and approximation
results. Journal of Combinatorial Optimization 29(1):276–307.
Brualdi RA (1969) Comments on bases in dependence structures. Bulletin of the Australian Mathematical Society
1(2):161–167.
Carr RD, Lancia G (2002) Compact vs. exponential-size lp relaxations. Operations Research Letters 30(1):57 – 65.
Chestnut SR, Zenklusen R (2017) Interdicting structured combinatorial optimization problems with 0, 1-objectives. Math-
ematics of Operations Research 42(1):144–166.
Chvátal V (1973) Edmonds polytopes and a hierarchy of combinatorial problems. Discrete Mathematics 4(4):305 – 337.
Chvátal V, Cook W (1990) The discipline number of a graph. Discrete mathematics 86(1-3):191–198.
Corley H, Sha DY (1982) Most vital links and nodes in weighted networks. Operations Research Letters 1(4):157–160.
Cormican KJ, Morton DP, Wood RK (1998) Stochastic network interdiction. Operations Research 46(2):184–197.
Di Summa M, Grosso A, Locatelli M (2012) Branch and cut algorithms for detecting critical nodes in undirected graphs.
Computational Optimization and Applications 53(3):649–680.
Dinh TN, Xuan Y, Thai MT, Pardalos PM, Znati T (2012) On new approaches of assessing network vulnerability:
Hardness and approximation. IEEE/ACM Transactions on Networking 20(2):609–619.
Erdös P, Rényi A (1959) On random graphs, I. Publicationes Mathematicae (Debrecen) 6:290–297.
Fischetti M, Ljubić I, Sinnl M (2017) Redesigning benders decomposition for large-scale facility location. Management
Science 63(7):2146–2162.
Frederickson GN, Solis-Oba R (1999) Increasing the weight of minimum spanning trees. Journal of Algorithms 33(2):244–
266.
Furini F, Ljubić I, Martin S, San Segundo P (2019) The maximum clique interdiction problem. European Journal of
Operational Research 277(1):112 – 127.
Gabow HN, Galil Z, Spencer T, Tarjan RE (1986) Efficient algorithms for finding minimum spanning trees in undirected
and directed graphs. Combinatorica 6(2):109–122.
Ghare P, Montgomery DC, Turner W (1971) Optimal interdiction policy for a flow network. Naval Research Logistics
Quarterly 18(1):37–45.
Gilbert EN (1959) Random graphs. The Annals of Mathematical Statistics 30(4):1141–1144.
Gomory RE (1963) An algorithm for integer solutions to linear programs. Graves R, Wolfe P, eds., Recent advances in
mathematical programming, volume 6508, 269–302 (McGrawHill, New York).

35
Gomory RE, Hu TC (1961) Multi-terminal network flows. Journal of the Society for Industrial and Applied Mathematics
9(4):551–570.
Graham RL, Hell P (1985) On the history of the minimum spanning tree problem. Annals of the History of Computing
7(1):43–57.
Grubesic TH, Matisziw TC, Murray AT, Snediker D (2008) Comparative approaches for assessing network vulnerability.
International Regional Science Review 31(1):88–112.
Held M, Karp RM (1970) The traveling-salesman problem and minimum spanning trees. Operations Research 18(6):1138–
1162.
Houck DJ, Kim E, O’Reilly GP, Picklesimer DD, Uzunalioglu H (2004) A network survivability model for critical national
infrastructures. Bell Labs Technical Journal 8(4):153–172.
Hsu LH, Jan RH, Lee YC, Hung CN, Chern MS (1991) Finding the most vital edge with respect to minimum spanning
tree in weighted graphs. Information Processing Letters 39(5):277–281.
Israeli E, Wood RK (2002) Shortest-path network interdiction. Networks 40(2):97–111.
Iwano K, Katoh N (1993) Efficient algorithms for finding the most vital edge of a minimum spanning tree. Information
Processing Letters 48(5):211–213.
Jenelius E, Petersen T, Mattsson LG (2006) Importance and exposure in road network vulnerability analysis. Transporta-
tion Research Part A: Policy and Practice 40(7):537–560.
Katoh N, Ibaraki T, Mine H (1981) An algorithm for finding k minimum spanning trees. SIAM Journal on Computing
10(2):247–255.
Kennedy KT, Deckro RF, Moore JT, Hopkinson KM (2011) Nodal interdiction. Mathematical and Computer Modelling
54(11-12):3116–3125.
Liang W, Shen X (1997) Finding the k most vital edges in the minimum spanning tree problem. Parallel Computing
23(13):1889–1907.
Lim C, Smith JC (2007) Algorithms for discrete and continuous multicommodity flow network interdiction problems. IIE
Transactions 39(1):15–26.
Lin KC, Chern MS (1993) The most vital edges in the minimum spanning tree problem. Information Processing Letters
45(1):25–31.
Linhares A, Swamy C (2017) Improved algorithms for MST and metric-tsp interdiction http://arxiv.org/abs/1706.
00034. (Accessed Apr 03, 2019).
Lozano L, Smith JC (2016) A backward sampling framework for interdiction problems with fortification. INFORMS
Journal on Computing 29(1):123–139.
Magnanti TL, Wolsey LA (1995) Optimal trees. Handbooks in operations research and management science 7:503–615.
Mahdavi Pajouh F, Boginski V, Pasiliao EL (2014) Minimum vertex blocker clique problem. Networks 64(1):48–64.
Mahdavi Pajouh F, Walteros JL, Boginski V, Pasiliao EL (2015) Minimum edge blocker dominating set problem. European
Journal of Operational Research 247(1):16–26.
Matisziw TC, Murray AT (2009) Modeling s-t path availability to support disaster vulnerability assessment of network
infrastructure. Computers and Operations Research. 36:16–26.
Nemhauser GL, Wolsey LA (1988) Integer and Combinatorial Optimization (New York, NY, USA: Wiley-Interscience).
Network Repository (2019) Network repository. https://networkrepository.com/ (Accessed Apr 03, 2019).
Ohlsson H, Gustafsson O, Wanhammar L (2004) Implementation of low complexity fir filters using a minimum spanning
tree. Proceedings of the 12th IEEE Mediterranean Electrotechnical Conference, volume 1, 261–264.
Oosten M, Rutten JHGC, Spieksma FCR (2007) Disconnecting graphs by removing vertices: a polyhedral approach.
Statistica Neerlandica 61(1):35–60.
Oxley JG (2006) Matroid theory, volume 3 (Oxford University Press, USA).
Ravi R, Goemans MX (1996) The constrained minimum spanning tree problem. Karlsson R, Lingas A, eds., Algorithm
Theory — SWAT’96, 66–75 (Berlin, Heidelberg: Springer Berlin Heidelberg).
Salmeron J, Wood KR, Baldick R (2004) Analysis of electric grid security under terrorist threat. Power Systems, IEEE
Transactions on 19(2):905–912.
Shen H (1995) Improved parallel algorithms for the most vital edge of a graph with respect to minimum spanning tree.
Technical Report (Dept. of Computer Sci., Abo Akademi Univ. Finland).

36
Shen S, Smith JC, Goli R (2012) Exact interdiction models and algorithms for disconnecting networks via node deletions.
Discrete Optimization 9(3):172–188.
Ströele A Menezes V, Tadeu da Silva R, Ferreira de Souza M, Oliveira J, de Mello CER, Moreira de Souza J, Zimbrão
G (2008) Mining and analyzing organizational social networks using minimum spanning tree. Meersman R, Tari Z,
Herrero P, eds., On the Move to Meaningful Internet Systems: OTM 2008 Workshops, 18–19 (Berlin, Heidelberg:
Springer Berlin Heidelberg).
Suk M, Song O (1984) Curvilinear feature extraction using minimum spanning trees. Computer Vision, Graphics, and
Image Processing 26(3):400 – 411.
Suraweera F, Maheshwari P, Bhattacharya P (1995) Optimal algorithms to find the most vital edge of a minimum spanning
tree. Technical Report CIT-95-21 (School of Computing and Information Technology, Griffith University).
Tao Z, Zhongqian F, Binghong W (2005) Epidemic dynamics on complex networks. Progress in Natural Science 16(5):452–
457.
Tapia E, Rojas R (2004) Recognition of on-line handwritten mathematical expressions using a minimum spanning tree con-
struction and symbol dominance. Lladós J, Kwon YB, eds., Graphics Recognition. Recent Advances and Perspectives,
329–340 (Berlin, Heidelberg: Springer Berlin Heidelberg).
Tarjan RE (1979) A class of algorithms which require nonlinear time to maintain disjoint sets. Journal of computer and
system sciences 18(2):110–127.
Tarjan RE, Van Leeuwen J (1984) Worst-case analysis of set union algorithms. Journal of the ACM (JACM) 31(2):245–281.
Veremyev A, Prokopyev OA, Pasiliao EL (2014) An integer programming framework for critical elements detection in
graphs. Journal of Combinatorial Optimization 28(1):233–273.
Vielma JP (2015) Mixed integer linear programming formulation techniques. Siam Review 57(1):3–57.
Watts DJ, Strogatz SH (1998) Collective dynamics of ‘small-world’networks. nature 393(6684):440.
Wollmer R (1964) Removing arcs from a network. Operations Research 12(6):934–940.
Wood RK (1993) Deterministic network interdiction. Mathematical and Computer Modeling 17(2):1–18.
Zenklusen R (2010) Matching interdiction. Discrete Appl. Math. 158(15):1676–1690.
Zenklusen R (2015) An o(1)-approximation for minimum spanning tree interdiction. 2015 IEEE 56th Annual Symposium
on Foundations of Computer Science, 709–728.

37

View publication stats