Maximum Robust Flow: Optimization & Algorithms

• Maximum Robust Flow (MRF) is a robust optimization model that computes a path-flow maximizing the guaranteed surviving flow under any k-arc deletion.
• The computational complexity shifts from polynomial-time solvable for k=1 to NP-hard and even Σ2P-hard for k ≥ 2, highlighting significant algorithmic challenges.
• Advanced techniques like primal–dual iterative separation and column-and-cut generation offer effective approximations and practical scalability in large networks.

Maximum Robust Flow (MRF) is the central robust optimization model for characterizing flows in capacitated networks subject to arc failures. In the classical setting, an $s$–$t$ flow is assigned to paths in a directed graph $G=(V,E)$, and an adversary may remove up to $k$ arcs, destroying all flow traversing any removed arc. The MRF problem is to compute a path-flow maximizing the guaranteed surviving flow under any such adversarial $k$-arc deletion. This problem lies at the intersection of combinatorial optimization, network interdiction, and robust optimization, and has attracted significant research interest due to its algorithmic challenges, application to critical infrastructure resilience, and connections to robust and adaptive flows, interdiction games, and rerouting models.

1. Formal Definition and Canonical LP Formulation

Given a directed network $G=(V,E)$ with source $s\in V$, sink $t\in V$, and nonnegative arc capacities $u_e$ ($e\in E$), let $t$0 denote the set of all simple $t$1–$t$2 paths. For integer $t$3, the set of $t$4-failure scenarios is $t$5. Let decision variables $t$6 be the flow assigned to path $t$7, and $t$8 represent the worst-case flow loss. The path-based MRF problem is

$t$9

where $G=(V,E)$0 in the optimal solution is the worst-case total flow destroyed by any $G=(V,E)$1-arc failure.

This formulation can be cast as a compact bilinear LP if and only if both $G=(V,E)$2 and $G=(V,E)$3 are “small,” but in general both these sets are exponentially large in $G=(V,E)$4 and $G=(V,E)$5.

In the general robust optimization context, the problem seeks

$G=(V,E)$6

where $G=(V,E)$7 is the path-flow polytope under arc capacities.

2. Complexity Landscape and Approximation

The combinatorial and computational complexity of MRF displays a sharp dichotomy with respect to the parameter $G=(V,E)$8 (number of arc failures):

Setting	Complexity Status	Reference
$G=(V,E)$9	Polynomial-time (efficient parametric LP)	(Disser et al., 2017), Aneja–Nair (2001)
fixed $k$0	NP-hard for any constant $k$1	(Matuschke, 9 Nov 2025)
$k$2 part of input	Strongly NP-hard; $k$3-hard	(Matuschke, 9 Nov 2025, Disser et al., 2017)
Integral MRF, $k$4	$k$5-hard (integral version)	(Matuschke, 9 Nov 2025)

The transition from tractability at $k$6 to NP-hardness for $k$7 is notable: for $k$8, the problem reduces to trading between total flow and the maximal load on any single arc, solvable using a parametric sweep over arc capacities and $k$9 max-flow calls. For $k$0, (Matuschke, 9 Nov 2025) establishes that even for constant $k$1 the adversarial $k$2-arc deletion makes the problem NP-complete (fractional) and $k$3-hard (integral). When $k$4 is input, MRF becomes $k$5-hard via a reduction from the Boolean hierarchy.

Polynomial-time solvability remains only for special cases: e.g., in unit-capacity networks or for edge-based (as opposed to path-based) robust flow models (Biefel et al., 2022).

Approximation algorithms have been developed. The best known guarantee is $k$6, specifically a $k$7-approximation (Mies et al., 2017), and a $k$8-approximation via $k$9-route flows [Baffier-Robins-Zeghal]. There is currently no hardness of approximation lower bound beyond NP-hardness.

3. Algorithmic Approaches: Primal–Dual Iterative Separation

Direct application of LP or MIP solvers is infeasible due to exponential numbers of variables and constraints. (Mies et al., 2017) develops an effective column-and-cut generation (“simultaneous primal–dual separation”) scheme:

• Restricted master problem: Start with small pools $G=(V,E)$0 (paths), $G=(V,E)$1 (scenarios). Solve the LP restricted to these sets.
• Column generation: Use dual multipliers to price (find) new path-columns with positive reduced cost. The pricing subproblem reduces to maximizing reduced cost over all $G=(V,E)$2–$G=(V,E)$3 paths, with rewards and penalties according to dual variables; this is an (NP-hard) combinatorial optimization, but efficiently solvable heuristics suffice in practice for moderate $G=(V,E)$4.
• Cut generation: Given a candidate flow, solve the network interdiction problem (inner minimization) to find a worst-case $G=(V,E)$5-arc failure scenario; this separation problem is also NP-hard, but can be managed via heuristic or enumerative approaches for small $G=(V,E)$6.

The cycle continues, alternately adding violating scenarios (primal cuts) and beneficial new paths (dual columns), until primal and dual bounds coincide. In practice, only a small number of iterations (typically $G=(V,E)$7–$G=(V,E)$8) and a modest number of scenarios ($G=(V,E)$9) are needed to obtain optimal or near-optimal flows for networks of thousands of arcs.

Initialization and heuristics: The initial path pool can be seeded using the Bertsimas–Sim relaxed flow—a nominal flow penalized by $s\in V$0 (maximum arc load)—which is near-optimal in most tested instances.

Empirical performance: The approach handles large random and structured networks, demonstrates stable convergence with robust lower/upper bounds, and supports further customization to limit scenario generation per iteration or diversify the search (minor penalties for scenario reuse).

4. Hybrid Robust–Stochastic MRF and Generalization Error Bounds

(Mies et al., 2017) proposes an extension interpolating between classical robust and stochastic models:

• Failure (or attack) scenarios are grouped into classes $s\in V$1 with assigned probabilities $s\in V$2.
• The objective is $s\in V$3, where $s\in V$4 denotes the surviving flow under scenario $s\in V$5.
  • If $s\in V$6 and $s\in V$7 is the set of all $s\in V$8-failures, this recovers the standard MRF.
  • If $s\in V$9 and each $t\in V$0 is the empirical probability of scenario $t\in V$1, the model is fully stochastic.

The same primal–dual separation mechanics apply classwise, with $t\in V$2 variables tracking classwise worst-case loss.

A generalization-error bound is established: Letting $t\in V$3 be the sample size of historical failure data and $t\in V$4 the empirical class frequencies, the bound on the hybrid robust value is of Rademacher-type,

$t\in V$5

where $t\in V$6 is explicit and diminishes with increasing $t\in V$7 and class sizes.

A min–max theorem guarantees that, under convexity and concavity conditions, the hybrid model optimizes the worst-case expected value over distributions consistent with class weights.

5. Extensions to General Robust Linear Optimization

The primal–dual separation technique and the hybrid robust–stochastic framework generalize to settings with any convex feasible region $t\in V$8 and concave scenario payoffs $t\in V$9. The only essential requirement is the convexity of $u_e$0 and concavity of each scenario utility in $u_e$1. Both the min–max equivalence and the generalization-error bounds carry over, making this approach broadly applicable beyond robust flows (Mies et al., 2017).

6. Empirical Findings and Practical Recommendations

Computational studies on various network families (series, NETGEN, R-MAT) with $u_e$2 demonstrate that typically $u_e$3–$u_e$4 primal–dual separation iterations suffice. The number of generated paths per run is usually less than $u_e$5; the number of scenario cuts reaches up to $u_e$6 in the largest instances. The path-pricing mixed integer programs (MIPs) require under 10 seconds, with primal-cut MIPs (interdiction) almost always below 1 second per iteration.

A summary of practical insights:

Recommendation	Justification from Computational Study
Initialize path pool with Bertsimas–Sim H	Empirically optimal or near-optimal in most tested networks
Generate up to $u_e$7 scenarios per round	Balances number of scenarios vs. pricing solution time
Apply minor penalties for cut/column reuse	Increases diversity and convergence stability
Solver scalability	Thousands of arcs and hundreds of scenarios in minutes

Where the Bertsimas–Sim heuristic is not optimal, the optimality gap is modest ($u_e$8–$u_e$9\%).

7. Connections, Generalizations, and Open Problems

MRF is an anchor problem within broader robust network optimization. Key variations include:

• Edge-based MRF: Instead of path-based flows, assign flow on arcs under weak conservation in every scenario, making the problem polynomial-time solvable for all $e\in E$0 (Biefel et al., 2022).
• General subpath-based models: Unify path and edge models, strictly extend optimal robust value, but are NP-hard except for $e\in E$1 (Biefel et al., 2022).
• Reroutable flows: Allow interrupted flow to be “rerouted” after arc failures—leading to a distinct set of tractability and approximation characteristics, including min-cut duals bounding the flow–cut gap by a factor of $e\in E$2 (Matuschke et al., 2017).
• Dynamic robust flows: Over time, with uncertain travel times and delays, robust flow problems are typically NP-hard; temporally repeated flows can exhibit unbounded suboptimality compared to optimal dynamic policies (Gottschalk et al., 2016).

Recent hardness results resolve long-standing open questions: MRF is NP-hard even for $e\in E$3 (Matuschke, 9 Nov 2025), answering the complexity status for small fixed $e\in E$4 that had previously been open due to gaps in prior hardness reductions (Disser et al., 2017).

Open problems include: obtaining constant-factor approximation algorithms for general $e\in E$5; characterizing the support size of optimal robust flows; and designing practical heuristics and algorithms for MRF variants in operational settings, especially for large-scale, dynamic, or hybrid uncertainty regimes.

---

MRF constitutes a central case of robust network optimization under discrete scenario uncertainty, is now known to be tractable only for special cases (small or highly structured $e\in E$6), and remains an active area for algorithmic and modeling advances in both theory and application.