---
title: Interval Uncertainty in Arc Capacities
url: https://www.emergentmind.com/topics/interval-uncertainty-in-arc-capacities
type: topic
---

# Interval Uncertainty in Arc Capacities

Interval uncertainty in arc capacities fundamentally alters the landscape of network flow optimization. In the minimum cost network flow problem with interval capacities, each arc $e$ is endowed not with a fixed capacity but with an interval $[\underline u_e, \overline u_e]$ within which its actual capacity may lie. The objective is to characterize and compute the "worst-case" optimal cost over all feasible capacity scenarios. Introducing intervals in capacities transforms the classical polynomially tractable problem into a combinatorial setting that exhibits strongly NP-hard behavior, even on restricted graph classes, and gives rise to nonintuitive phenomena such as the more-for-less paradox. This setting also admits a rich interplay between parametric linear programming, combinatorial optimization, and graph-theoretic structure [2602.01360].

## 1. Problem Formulation and Scenario Polytope

Let $G = (V, E)$ be a directed graph with a source $s$, sink $t$, arc cost vector $c \in \mathbb{R}_+^E$, and a flow requirement $f \in \mathbb{N}$. Classical minimum cost flow problems assume fixed arc capacities $u_e \geq 0$, solving:
\[
\min_{x \in \mathcal{F}(u)} c^\top x, \quad 
\mathcal{F}(u) = \left\{ x \in \mathbb{R}_+^E \mid 
\sum_{j:(i,j) \in E} x_{ij} - \sum_{j:(j,i) \in E} x_{ji}
=
\begin{cases}
f, & i=s \\
-f, & i=t \\
0, & \text{otherwise}
\end{cases}
, \; x_e \leq u_e \; \forall e \right\}
\]
Interval uncertainty replaces each $u_e$ with $\mathcal{U}_e = [\underline u_e, \overline u_e]$. A "scenario" is any $u = (u_e)_{e \in E}$ with $u_e \in \mathcal{U}_e$, and $\mathcal{U}$ denotes the set of scenarios admitting a feasible $s$–$t$ flow of size $f$. The worst-case optimal value is defined as
\[
Z^* = \max_{u \in \mathcal{U}} \min_{x \in \mathcal{F}(u)} c^\top x
\]
That is, an adversary sets the capacities to maximize the minimum achievable cost.

## 2. Computational Complexity and Intractability

The introduction of interval uncertainty renders the robust counterpart strongly NP-hard. Even in the unit-cost, unit-flow case where each interval is $\mathcal{U}_e = [0,1]$, determining $Z^*$ is equivalent to solving the longest–path problem via capacity selection, so the worst-case value is combinatorially hard to compute. Notably, this persists for series–parallel graphs—those with treewidth 2—through a reduction from the knapsack problem. In this construction, lower bounds encode item weights and costs encode values, with $f = b$ as the knapsack capacity. Even when the number of $s$–$t$ paths is $O(n)$, the adversarial capacity scenario selection delivers NP-hardness [2602.01360].

| Graph Class            | Complexity of $Z^*$ Computation      | Reduction Used                                   |
|------------------------|--------------------------------------|--------------------------------------------------|
| General                | Strongly NP-hard                     | Longest-path                                     |
| Series–parallel (tw=2) | Strongly NP-hard                     | Knapsack-to-network-flow                         |

The presence of interval capacity uncertainty thus transforms a polynomial-time flow LP into a robust optimization problem of intractable character.

## 3. Optimization Formulations and Tractable Special Cases

Despite general intractability, exact mixed-integer linear programming (MILP) formulations are possible. Utilizing LP duality and complementary slackness, the max–min structure is encoded via primal flow variables $x_{ij}$, dual node potentials $\pi_i$, and binary indicators for active constraints. The resulting MILP has objective $\max \sum_{(i,j) \in E} c_{ij} x_{ij}$ and includes the scenario selection $u \in [\underline u, \overline u]$ directly as variables, subject to feasibility and dual conditions.

For series–parallel graphs—a class closed under arc creation, series, and parallel composition—a pseudopolynomial dynamic programming algorithm achieves exact computation. The method propagates tables $d(\phi)$ reflecting the worst-case cost of sending $\phi$ units of flow, and accounts for restrictable thresholds and maximal feasible flows per component. Arc creation is trivial; series composition aggregates costs additively; parallel composition entails maximization/minimization over possible flow splits. Parallel steps dominate runtime, yielding an overall complexity of $O(mf^2)$ for $m$ arcs and flow value $f$, constituting pseudopolynomial tractability in specialized cases.

## 4. Structural Characterization of Extremal Scenarios

Let $u^*$ be a capacity scenario achieving $Z^*$, and define $E^{\text{int}} = \{e : u^*_e \in (\underline u_e, \overline u_e)\}$ as the set of arcs at interior values. A core structural result is that the subgraph induced by $E^{\text{int}}$ is acyclic—a forest. This follows from perturbing cycles, which does not reduce the cost but reduces the number of interior arcs, contradicting extremality. Thus, $\lvert E^{\text{int}} \rvert \leq |V| - 1$. This bound is tight; there exist $n$-node graphs with $n-1$ interval arcs, each set to an interior value at the worst case, indicating constructions can maximize the complexity of scenario structure within these graph-theoretic limits.

## 5. The More-for-Less Paradox

A non-classical phenomenon in this context is the more-for-less paradox: increasing the required flow may strictly decrease the worst-case cost, i.e., $Z^*(f+1) < Z^*(f)$ for some $f$. This is characterized structurally by the existence of a negative-cost augmenting $s$–$t$ path in the graph with respect to a worst-case scenario and optimal flow at $f$. Necessity: such a path must exist if the paradox occurs. Sufficiency: if such a negative-cost path is present for any scenario and flow, one can construct intervals enforcing the paradox. Thus, the paradox is present if and only if a negative-cost augmenting path exists.

In the case of directed complete graphs, the characterization strengthens: the existence of any improving undirected $s$–$t$ path (negative cost) is necessary and sufficient. This connection provides a clear combinatorial test in these settings.

| Paradox Condition               | General Graphs        | Complete Graphs                  |
|---------------------------------|-----------------------|----------------------------------|
| Existence criterion             | Negative-cost augment.<br />path exists | Negative-cost undirected $s$–$t$ path exists |

## 6. Cost-Matrix Immunity and Algorithmic Limits

A cost matrix $c$ is said to be immune if no choice of intervals admits the more-for-less paradox. For complete graphs, this is equivalent to the absence of any negative-cost improving path:
\[
c_{si} - \max \{\text{path-cost}(i \to j)\} + c_{jt} \geq 0
\]
for all nodes $i, j$. The immunity region for $c$ thus forms a convex polyhedral cone requiring exponentially many path-type inequalities. However, deciding whether a given $c$ is immune is strongly co-NP-hard, via a reduction from the longest path problem. There is therefore no known polytime certificate for "paradox-free" costs, and verifying robust safety against interval-induced paradoxes is itself provably intractable [2602.01360].

## 7. Synthesis and Theoretical Impact

The introduction of interval uncertainty in arc capacities for minimum-cost flow problems results in a robust optimization problem displaying several distinguished features:

- Even though the deterministic LP for fixed capacities is polynomial-time solvable, interval uncertainty in arc capacities yields a worst-case value that is strongly NP-hard to compute.
- Series–parallel graphs admit pseudopolynomial-time solutions via dynamic programming, which is not extensible to broader graph classes due to inherent hardness.
- In any extremal scenario, the set of arcs at strictly interior capacity bounds is structurally limited to a forest, with an upper bound of $n-1$, matching worst-case constructions.
- The more-for-less paradox is completely characterized by augmenting paths of negative cost, highlighting the subtlety of adversarial capacity selection.
- The challenge of certifying cost-matrix immunity to the paradox is co-NP-hard, indicating a fundamental limit to polytime characterization.

These elements underscore how interval uncertainty causes the underlying network flow problem to become a highly structured intersection of robust linear programming, combinatorial complexity, and intricate graph-theoretic phenomena [2602.01360].

Source: https://www.emergentmind.com/topics/interval-uncertainty-in-arc-capacities