On the approximability of robust spanning tree problems

In this paper the minimum spanning tree problem with uncertain edge costs is discussed. In order to model the uncertainty a discrete scenario set is specified and a robust framework is adopted to choose a solution. The min-max, min-max regret and 2-stage min-max versions of the problem are discussed. The complexity and approximability of all these problems are explored. It is proved that the min-max and min-max regret versions with nonnegative edge costs are hard to approximate within $O(\log^{1-\epsilon} n)$ for any $\epsilon>0$ unless the problems in NP have quasi-polynomial time algorithms. Similarly, the 2-stage min-max problem cannot be approximated within $O(\log n)$ unless the problems in NP have quasi-polynomial time algorithms. In this paper randomized LP-based approximation algorithms with performance ratio of $O(\log^2 n)$ for min-max and 2-stage min-max problems are also proposed.

💡 Research Summary

The paper investigates the Minimum Spanning Tree (MST) problem under cost uncertainty within a robust optimization framework. The authors consider a connected undirected graph G = (V, E) together with a finite set of K scenarios Γ = {S₁,…,S_K}. Each scenario S assigns a cost c_e^S to every edge e ∈ E. Two classic robust criteria are studied: (i) the min‑max objective, which seeks a spanning tree T minimizing the worst‑case total cost across all scenarios, formally OPT₁ = min_{T∈Φ} max_{S∈Γ} Σ_{e∈T} c_e^S; and (ii) the min‑max regret objective, which minimizes the maximum deviation of a tree’s cost from the optimal MST cost in each scenario, i.e., OPT₂ = min_{T∈Φ} max_{S∈Γ} ( Σ_{e∈T} c_e^S – C⁎(S) ), where C⁎(S) denotes the optimal MST cost under scenario S.

The authors first establish strong inapproximability results for both problems when edge costs are non‑negative. By a reduction from the Label‑Cover problem, they construct instances where the optimal value is either 1 (when the underlying Label‑Cover instance is satisfiable) or at least g = Θ(log N) (when it is not). The construction uses a gadget that replaces each edge of the bipartite Label‑Cover graph with a small component containing “label edges” and “dummy edges”. Scenarios are defined so that any spanning tree that selects g pairwise label‑distinct edges incurs cost g, while a tree that respects a total labeling of size 1 incurs cost at most 1. Because the size of the resulting robust MST instance grows only polynomially in the size of the original Label‑Cover instance, any polynomial‑time algorithm achieving an approximation factor better than O(log^{1‑ε} n) for any ε > 0 would imply NP ⊆ DTIME(n·polylog n), contradicting widely‑believed complexity assumptions. Consequently, both the min‑max and min‑max regret MST problems are not approximable within O(log^{1‑ε} n) unless NP has quasi‑polynomial time algorithms.

When both positive and negative edge costs are allowed, the situation worsens dramatically. The authors reduce from a variant of 3‑SAT (called 3‑SAT‑T) to the min‑max MST problem. Each clause is represented by a small subgraph of five vertices and six edges; edges corresponding to literals have cost 0 in the base scenario, while all other edges have cost –1. For each pair of contradictory literals a special scenario raises the cost of the two corresponding literal edges to a large positive value (4m‑1) and leaves all other edges at –1. If the formula is satisfiable, there exists a spanning tree whose maximum cost over all scenarios is non‑positive; otherwise any spanning tree incurs a positive maximum cost. This yields a gap that cannot be bridged by any constant‑factor approximation, even on series‑parallel graphs. Hence, with unrestricted sign costs the min‑max MST problem becomes completely inapproximable unless P = NP.

The paper also introduces a two‑stage robust MST model. In the first stage, a subset E₁ of edges is purchased at known deterministic costs. After the scenario S is revealed, a second‑stage set E_S² of additional edges is bought at scenario‑dependent costs, completing a spanning tree T_S = E₁ ∪ E_S². The objective is to minimize the worst‑case total cost: OPT₃ = min_{E₁, {E_S²}} max_{S∈Γ} ( Σ_{e∈E₁} c_e + Σ_{e∈E_S²} c_e^S ). Using a similar Label‑Cover reduction, the authors prove that this two‑stage problem cannot be approximated within any constant factor unless NP ⊆ DTIME(n·log log n), and it also admits an O(log n) lower bound.

Despite these negative results, the authors present positive algorithmic contributions. They design randomized approximation algorithms based on linear programming (LP) relaxations for both the single‑stage min‑max MST and the two‑stage min‑max MST. The approach solves a fractional LP that simultaneously captures all scenarios, then applies a randomized rounding scheme inspired by the method of conditional expectations and Chernoff bounds. The rounding ensures that, with high probability, the cost incurred in any scenario exceeds the fractional LP value by at most an O(log n) factor. Since the LP optimum itself is at most the optimal integer solution, the overall expected approximation ratio becomes O(log² n). The same technique extends to the two‑stage setting, yielding an O(log² n) randomized algorithm for the robust two‑stage MST.

In summary, the paper establishes that robust MST problems are intrinsically hard to approximate: non‑negative costs already lead to logarithmic inapproximability, while allowing negative costs makes the problem completely inapproximable. It also shows that the two‑stage robust version inherits similar hardness. On the algorithmic side, the authors provide the first non‑trivial approximation algorithms for these problems, achieving an O(log² n) performance guarantee via LP‑based randomization. These results close a long‑standing open question about constant‑factor approximability of robust MSTs and open new avenues for robust network design under uncertainty.