Minimum Spanning Tree Interdiction

• MST interdiction is a network problem where selective edge removal forces the residual minimum spanning tree to be as expensive as possible.
• A key methodological insight is the use of replacement edges and single-edge sensitivity measures, with Δ(e) quantifying the cost increase from interdiction.
• Approximation and parameterized approaches leverage graph connectivity, matroid theory, and greedy strategies to balance budget constraints with maximum MST cost increase.

Minimum Spanning Tree Interdiction is a network interdiction problem in which one removes edges from an undirected weighted graph so as to make the minimum spanning tree (MST) of the residual graph as expensive as possible. In its classical budgeted form, the input consists of an undirected graph $G=(V,E)$, nonnegative edge weights, positive interdiction costs on edges, and a budget $B$; the task is to choose an interdiction set $R\subseteq E$ with total interdiction cost at most $B$ that maximizes the weight of an MST in $(V,E\setminus R)$ (Zenklusen, 2015). The topic sits at the intersection of network robustness, combinatorial optimization, matroid theory, approximation algorithms, and sensitivity analysis. Closely related formulations include the decision problem $b$-Most Vital Edges in MST, exact single-edge vulnerability analysis via replacement edges, and more recent objectives that maximize the increase in MST weight rather than the absolute post-interdiction MST weight (Guo et al., 2014, Ostrovsky et al., 25 Jul 2025).

1. Problem formulations and objective functions

A standard optimization formulation is

$$\max\{\operatorname{val}(R)\mid R\subseteq E,\ c(R)\le B\},$$

where $\operatorname{val}(R)$ denotes the weight of an MST in $(V,E\setminus R)$ (Zenklusen, 2015). In the parameterized-complexity literature, the same theme appears as the decision problem $b$-Most Vital Edges in MST ($B$0-MVE): given an edge-weighted undirected graph $B$1, integers $B$2 and $B$3, determine whether there exists $B$4 with $B$5 such that

$B$6

where $B$7 is the weight of a minimum spanning tree of $B$8 (Guo et al., 2014).

A later line of work distinguishes between maximizing the total post-interdiction MST weight and maximizing the increase caused by interdiction. For a removal set $B$9, the increase-based objective is

$R\subseteq E$0

with associated budget and profit formulations

$R\subseteq E$1

This distinction is algorithmically significant because approximation guarantees for $R\subseteq E$2 do not automatically transfer to $R\subseteq E$3 (Ostrovsky et al., 25 Jul 2025).

Formulation	Objective	Salient feature
Classical budgeted interdiction	$R\subseteq E$4	Maximizes post-interdiction MST weight
$R\subseteq E$5-MVE decision problem	$R\subseteq E$6	Unit interdiction costs
Profit maximization	$R\subseteq E$7	Maximizes increase over baseline MST
Budget minimization	$R\subseteq E$8	Meets a target increase

The treatment of disconnected residual graphs varies by formulation. In the parameterized and profit-based formulations, if interdiction disconnects the graph then the MST value is conventionally taken to be $R\subseteq E$9, so disconnecting the graph is a valid and maximally harmful interdiction outcome (Guo et al., 2014, Ostrovsky et al., 25 Jul 2025). By contrast, the 14-approximation framework assumes the budget is not sufficient to disconnect the graph, formally

$B$0

so every budget-feasible interdiction set leaves the graph connected (Zenklusen, 2015). This difference is fundamental when comparing results across papers.

MST interdiction is also naturally matroidal. Because spanning trees are precisely the bases of the graphic matroid, MST interdiction is a special case of matroid interdiction: the leader removes elements under a knapsack constraint, and the follower computes a minimum-weight basis disjoint from the interdiction set (Weninger et al., 2024).

2. Single-edge interdiction, replacement edges, and most-vital-edge analysis

The most elementary interdiction primitive is the failure or deletion of a single MST edge. Given a weighted undirected graph $B$1, an MST $B$2, and a tree edge $B$3, removing $B$4 splits $B$5 into two connected components $B$6 and $B$7. A replacement edge for $B$8 is a non-tree edge $B$9 whose endpoints lie in different components of $(V,E\setminus R)$0. The minimum replacement edge is

$(V,E\setminus R)$1

Equivalently, $(V,E\setminus R)$2 is the lightest non-tree edge crossing the cut induced by deleting $(V,E\setminus R)$3, or the lightest non-tree edge whose fundamental cycle with $(V,E\setminus R)$4 contains $(V,E\setminus R)$5 (Bader et al., 2019).

This immediately yields the exact one-edge interdiction value. If $(V,E\setminus R)$6 is the MST after deleting $(V,E\setminus R)$7, then

$(V,E\setminus R)$8

so the increase in MST cost is

$(V,E\setminus R)$9

Hence the single-edge interdiction value of $b$0 is exactly the replacement gap $b$1, and the most vital edge is the MST edge maximizing that quantity (Bader et al., 2019).

The replacement-edge problem therefore gives a complete sensitivity profile for all single-edge failures in the MST. The 2019 replacement-edge algorithm computes all $b$2 values in

$b$3

given the MST and the non-tree edges sorted in nondecreasing order by weight. The method improves on Tarjan’s $b$4 bound in the presorted setting and is described as the first practical linear-time algorithm for that case (Bader et al., 2019).

Its structure has three ingredients. First, it roots the MST and computes DFS parent pointers and interval labels $b$5, $b$6, enabling ancestry tests of the form

$b$7

Second, it scans non-tree edges in increasing weight order, so the first non-tree edge whose fundamental cycle contains a tree edge must be that tree edge’s minimum replacement by the cut property. Third, it uses the Gabow–Tarjan disjoint-set structure for a known union tree to compress already-labeled tree paths, ensuring that each tree edge is traversed at most once (Bader et al., 2019).

The scope is exact but narrow. The method solves the complete single-tree-edge sensitivity profile and the most-vital-edge problem in overall $b$8 time once the MST and sorted non-tree edges are available. It does not solve general multi-edge interdiction with budgets or arbitrary interdiction costs (Bader et al., 2019).

3. Approximation algorithms and the multilevel structure of MST interdiction

The classical approximation breakthrough is the first $b$9-approximation for MST interdiction, specifically a 14-approximation (Zenklusen, 2015). The input model is an undirected loopless multigraph with nonnegative edge weights $$\max\{\operatorname{val}(R)\mid R\subseteq E,\ c(R)\le B\},$$0, positive interdiction costs $$\max\{\operatorname{val}(R)\mid R\subseteq E,\ c(R)\le B\},$$1, and budget $$\max\{\operatorname{val}(R)\mid R\subseteq E,\ c(R)\le B\},$$2, and the objective is to maximize the MST weight after deleting a budget-feasible edge set (Zenklusen, 2015).

The key difficulty is that MST interdiction is inherently multilevel. After rounding positive edge weights down to powers of two, the MST value can be written in terms of connectivity across all weight thresholds: $$\max\{\operatorname{val}(R)\mid R\subseteq E,\ c(R)\le B\},$$3 for $$\max\{\operatorname{val}(R)\mid R\subseteq E,\ c(R)\le B\},$$4, where $$\max\{\operatorname{val}(R)\mid R\subseteq E,\ c(R)\le B\},$$5 is the number of connected components (Zenklusen, 2015). Creating one additional component at threshold $$\max\{\operatorname{val}(R)\mid R\subseteq E,\ c(R)\le B\},$$6 increases the MST weight by $$\max\{\operatorname{val}(R)\mid R\subseteq E,\ c(R)\le B\},$$7, so the problem is not a single cut problem but a layered connectivity problem.

The 14-approximation combines three components. First, rounding edge weights to powers of two incurs at most a factor-2 loss. Second, the algorithm computes a geometric upper bound on the optimum via a Pareto frontier of cost versus interdiction value, using parametric submodular minimization of

$$\max\{\operatorname{val}(R)\mid R\subseteq E,\ c(R)\le B\},$$8

Third, from an over-budget removal set $$\max\{\operatorname{val}(R)\mid R\subseteq E,\ c(R)\le B\},$$9 with good efficiency $\operatorname{val}(R)$0, it extracts a budget-feasible interdiction set through a hierarchical “removal pattern” defined over laminar component partitions. The central conversion guarantee is

$\operatorname{val}(R)$1

which is strengthened to

$\operatorname{val}(R)$2

A 7-approximation on the rounded instance and the factor-2 rounding loss yield the final factor $\operatorname{val}(R)$3 (Zenklusen, 2015).

A newer approximation line targets the increase-based objective $\operatorname{val}(R)$4 directly rather than $\operatorname{val}(R)$5 (Ostrovsky et al., 25 Jul 2025). The central structural fact is that $\operatorname{val}(R)$6 is a monotone nondecreasing, nonnegative supermodular set function. The budget problem

$\operatorname{val}(R)$7

thus becomes minimizing a linear cost subject to a supermodular covering constraint, while the profit problem becomes maximizing a monotone supermodular function under a knapsack constraint (Ostrovsky et al., 25 Jul 2025).

The algorithmic primitive is the partial cut

$\operatorname{val}(R)$8

together with the certified lower bound

$\operatorname{val}(R)$9

for any edge $(V,E\setminus R)$0 crossing the cut $(V,E\setminus R)$1. If $(V,E\setminus R)$2 has cost $(V,E\setminus R)$3 and profit $(V,E\setminus R)$4, the paper proves the existence of partial cuts $(V,E\setminus R)$5 such that

$(V,E\setminus R)$6

This graph-theoretic relaxation motivates a batch-greedy algorithm (Ostrovsky et al., 25 Jul 2025).

The resulting guarantees separate two objectives that had previously been conflated. For budget minimization, if there exists a solution of cost $(V,E\setminus R)$7 and profit $(V,E\setminus R)$8, the algorithm returns a set meeting profit $(V,E\setminus R)$9 with cost $b$0, and the bound is explicitly less than $b$1 after doubling search (Ostrovsky et al., 25 Jul 2025). For strict-budget profit maximization, if optimum profit under budget $b$2 is $b$3, the algorithm returns a budget-feasible set with profit at least

$b$4

The same work also identifies a polynomial-time tractable island: the minimum-cost way to increase the MST by any positive amount can be solved exactly in

$b$5

where $b$6 is the time for a minimum $b$7-$b$8 cut computation (Ostrovsky et al., 25 Jul 2025).

4. Parameterized complexity and exact algorithms

The parameterized picture is sharp but incomplete. For the decision problem $b$9-MVE, parameterization by the target post-interdiction MST weight $B$00 is W[1]-hard: $B$01 The hardness holds even when edge weights are restricted to $B$02, via a reduction from Minimum $B$03-way Edge Cut (Guo et al., 2014).

A positive result exists for parameter $B$04, the number of deleted edges, but only in the binary-weight case. When edge weights are restricted to $B$05, $B$06-MVE is fixed-parameter tractable with respect to $B$07 (Guo et al., 2014). The central structural insight is that, in $B$08-weighted graphs, the MST weight is controlled by the number of connected components in the zero-weight subgraph. This leads to a decomposition strategy: remove all weight-1 edges, compute per-component fragmentation tables using an FPT algorithm for Minimum $B$09-Way Edge Cut, and combine the local tradeoffs by enumerating additive partitions of the budget. The proof also uses the MST-specific claim that in a $B$10-edge-connected graph, an optimal solution contains no weight-1 edge (Guo et al., 2014).

The tractability boundary remains open for general nonnegative integer weights. The paper explicitly states that $B$11-MVE is FPT with parameter $B$12 for weights $B$13 or $B$14, while “The case with integer positive weights remains open” (Guo et al., 2014).

For exact solution, a recent matroid-interdiction framework gives a new combinatorial branch-and-bound algorithm for MST interdiction as the graphic-matroid case of a more general basis interdiction problem (Weninger et al., 2024). The objective is written as

$B$15

where $B$16 and $B$17 is the set of bases of the matroid disjoint from $B$18. For graphic matroids, the follower’s basis is an MST in $B$19 (Weninger et al., 2024).

The upper bound driving the exact algorithm is a dynamic program based on the discrete derivative

$B$20

that is, the increase in follower value caused by interdicting one additional element. For graphic matroids, this derivative is bounded by solving a sequence of minimum $B$21-$B$22 cut problems, which capture whether earlier interdictions can destroy all cheaper $B$23-$B$24 alternatives to a given edge (Weninger et al., 2024). The branch-and-bound recursion then branches only on edges of the current lexicographically smallest MST, reflecting the fact that interdicting nonbasis edges does not immediately change the follower solution.

The algorithm is exact, asymptotically faster than the previously best known exact enumerative MST interdiction algorithm up to polylogarithmic factors, and computationally strong: it solved all available instances from the literature, and in many cases reduced the best running time from hours to seconds (Weninger et al., 2024). The same study reports an average optimality gap of $B$25 for the greedy lower bound and $B$26 for the strengthened root upper bound, with worst observed gaps $B$27 and $B$28, respectively (Weninger et al., 2024).

5. Matroidal generalizations and related perturbation models

A broader interdiction framework studies structured maximization problems with $B$29-objectives, including maximum weight forest and maximum weight independent set in a matroid (Chestnut et al., 2015). Its general form is

$B$30

under assumptions such as $B$31, $B$32-down-closedness of $B$33, and box-$B$34-DI solvability (Chestnut et al., 2015). The main guarantee is a pseudoapproximation tradeoff: for any $B$35, the algorithm returns either a budget-feasible $B$36-approximation, or a solution whose budget is exceeded by at most $B$37 while achieving value at least as good as the optimum under the original budget (Chestnut et al., 2015).

For MST interdiction, however, this connection is indirect rather than direct. The framework covers maximum weight forests through the graphic matroid independent-set viewpoint, whereas classical MST interdiction is a minimization-over-bases problem rather than a down-closed maximization problem. A careful reading therefore supports the statement that the paper is relevant through matroid structure and forest interdiction, but does not provide a dedicated theorem for classical minimum spanning tree interdiction (Chestnut et al., 2015).

A different but closely related perturbation model replaces edge deletion by edge imposition. Isoart and Régin study the cost increase incurred when a spanning tree is required to contain specified non-tree edges (Isoart et al., 2019). For a fixed MST $B$38 and a set $B$39 of imposed non-tree edges, the constrained MST $B$40 satisfies

$B$41

where $B$42 is the replacement cost of imposing $B$43 into $B$44 (Isoart et al., 2019). For a single imposed edge, the bound is exact: $B$45 This is not standard interdiction, but it is a constructive perturbation analogue: instead of deleting cheap options, one compels the follower to use edges that are suboptimal for the original MST (Isoart et al., 2019).

These adjacent viewpoints matter because they isolate common mechanisms: cycle exchanges, replacement costs, cut structure, and lower bounds on cumulative degradation. A plausible implication is that MST interdiction is best understood not as a single problem family, but as a cluster of perturbation models linked by the geometry of the graphic matroid.

6. Assumptions, edge cases, and conceptual boundaries

Several recurring assumptions delimit what current results do and do not establish. The linear-time replacement-edge algorithm assumes that the MST is already given and that non-tree edges are sorted by weight. This is natural after Kruskal’s algorithm, but the linear bound does not directly improve prior general bounds without presorting (Bader et al., 2019).

Bridges are a fundamental edge case. If a tree edge is also a bridge of the underlying graph, then there is no replacement edge, $B$46 cannot be reconnected by any non-tree edge, and the most vital edge is considered undefined in the single-edge sensitivity convention of the replacement-edge paper (Bader et al., 2019). By contrast, in formulations that allow disconnected residual graphs and assign them value $B$47, such edges represent maximally successful interdiction actions (Guo et al., 2014, Ostrovsky et al., 25 Jul 2025).

Another important boundary is between single-edge and multi-edge interdiction. Replacement-edge methods compute exact one-edge failure penalties

$B$48

but they do not directly extend to multiple simultaneous deletions, where replacement interactions are nontrivial (Bader et al., 2019). Likewise, parameterized tractability for budget $B$49 is known only for $B$50-weighted graphs, not for general positive integer weights (Guo et al., 2014).

A frequent misconception is that approximation guarantees for the total post-interdiction MST weight automatically approximate interdiction gain. The increase-based analysis shows this is false: if the baseline MST is already heavy, a constant-factor approximation to $B$51 may correspond to negligible profit $B$52 (Ostrovsky et al., 25 Jul 2025). Another misconception is that general matroid-interdiction frameworks immediately resolve classical MST interdiction. The forest-based pseudoapproximation framework does not do so directly because its nominal problem is a down-closed maximization problem, whereas MST interdiction is a minimization-over-bases problem (Chestnut et al., 2015).

Tie structure also matters. With equal edge weights there may be multiple MSTs and multiple minimum replacement edges, and the vulnerability profile can depend on which MST is supplied to a sensitivity algorithm (Bader et al., 2019). This suggests that some interdiction statements are best interpreted relative to a chosen baseline MST unless uniqueness is guaranteed.

Taken together, the literature supports a layered view. Exact single-edge sensitivity is well understood and can be computed in linear time under presorted input. Budgeted multi-edge interdiction admits constant-factor and logarithmic approximations, but with different objective functions and connectivity conventions. Parameterized tractability is established only in restricted settings. Exact solution is possible through modern branch-and-bound methods that exploit graphic-matroid structure and min-cut-based bounds. The resulting field is technically mature in several special directions, but still fragmented across objective choices, graph connectivity conventions, and the gap between single-edge sensitivity and full budgeted interdiction (Zenklusen, 2015, Weninger et al., 2024).