One-Sided Crossing Minimization (OSCM)

• OSCM is a combinatorial optimization problem that minimizes edge crossings in two-layer bipartite graph drawings by determining an optimal vertex order.
• It includes variants like local crossing minimization, edge-weighted adaptations, and metro-line crossing minimization, addressing both theoretical and practical challenges.
• Recent approaches leverage exact, parameterized, quantum, and heuristic methods, making OSCM crucial for applications in network, circuit, and hierarchical graph designs.

One Sided Crossing Minimization (OSCM) is a central combinatorial optimization problem in graph drawing, particularly within the context of two-layer (bipartite) layouts used in hierarchical or layered visualizations. Given a bipartite graph $G = (A \cup B, E)$ with a fixed linear order $\pi_A$ on $A$, OSCM asks for a linear order $\pi_B$ of $B$ that minimizes the number of edge crossings in a standard two-level drawing where all edges are realized as straight-line segments. The problem is foundational for reducing visual complexity in applications ranging from network visualization and circuit design to layered graph drawing frameworks such as Sugiyama’s methodology. Subtle variations include local crossing minimization (minimizing per-edge crossings), edge-weighted versions, and online settings with irrevocable vertex placements. Recent work spans exact algorithms, parameterized and quantum approaches, sophisticated heuristics, and advances in complexity analysis, with ongoing research addressing both theoretical limits and scalable practical solutions.

1. Formal Definition and Problem Variants

OSCM is formulated as follows: Given a bipartite graph $G = (A \uplus B, E)$ and a fixed ordering $\pi_A: A \rightarrow [n_0]$, find a permutation $\pi_B: B \rightarrow [n_1]$ that minimizes the total number of crossings defined by pairs of edges $(a_1,b_1),(a_2,b_2)$ such that

$$(\pi_A(a_1) < \pi_A(a_2) \wedge \pi_B(b_1) > \pi_B(b_2)) \vee (\pi_A(a_1) > \pi_A(a_2) \wedge \pi_B(b_1) < \pi_B(b_2)).$$

The classical objective is total edge crossings, but variants target the local crossing number (maximum crossings per edge) (Gutowski et al., 30 Sep 2025), weighted versions, or restrict the host graphs (e.g., to trees or paths) (Dobler, 2023).

A related but distinct body of work considers OSCM in the context of “periphery constraints” or metro-line crossing minimization (MLCM-P) (Fink et al., 2013, Okamoto et al., 2013), one-sided tanglegram layouts (Dobler et al., 2023), slotted/online orderings (Burjons et al., 2022), and block crossing minimization (Dobler et al., 2023).

2. Computational Complexity and Theoretical Insights

OSCM is NP-hard even for highly restricted instances: this holds for general bipartite graphs (Fink et al., 2013), for trees (Dobler, 2023), almost-trees (Bannister et al., 2013), and for local crossing minimization (Gutowski et al., 30 Sep 2025). The reduction techniques often trace to canonical hard problems such as 3-Partition or adaptations from specialized bipartite cases (e.g., unions of 4-stars) (Dobler, 2023).

Recently, a precise link between OSCM and width parameters was formalized: if the cutwidth of the bipartite graph (the maximal edge cut induced by a prefix of $\pi_A$0) is $\pi_A$1, the total number of unsuited pairs (pairs whose mutual order is not fixed by local structure) is at most $\pi_A$2 (Rauch et al., 17 Jan 2025). This directly controls the effective hardness, as exact solutions require resolving only the relative positions for these unsuited pairs, yielding an exponential-time algorithm in $\pi_A$3 when $\pi_A$4 is constant, a significant improvement over the $\pi_A$5 brute force.

In the case of trees, a claimed polynomial-time solution was refuted via a concrete counter-example showing that feedback arc set approaches can yield cycles, invalidating the assumed reduction to topological sort (Dobler, 2023).

For block crossing versions in one-sided tanglegrams, the problem is NP-complete for complete binary trees and for non-binary trees (Dobler et al., 2023), while constant-factor approximations (2.25-approximate for block crossings), as well as FPT algorithms parameterized by the number of blocks, are available.

3. Exact, Parameterized, and Efficient Algorithms

Classical Approaches: The problem is instantiated as a mixed-integer program (MIP) (Rauch, 2024) or as a binary linear program with cycle constraint relaxations (Jünger et al., 2024); both build on variants of the linear ordering polytope. A key formulation minimizes

$\pi_A$6

subject to transitivity constraints on the $\pi_A$7 variables and $\pi_A$8. Here $\pi_A$9 encodes that $A$0 precedes $A$1 in $A$2, and $A$3 represents the number of crossings induced by ordering $A$4 before $A$5.

A crucial efficiency gain is obtained through decomposition: nodes in $A$6 are partitioned into components based on open intervals defined by their $A$7-neighbors; components with disjoint intervals can be solved independently (Jünger et al., 2024). Additional rules prune variables by fixing orders where crossing counts are unambiguous (e.g., $A$8 and $A$9 implies $\pi_B$0 must precede $\pi_B$1).

Branch-and-Cut and Reduction Techniques: Modern solvers (e.g., MPPEG’s PACE 2024 winner (Jünger et al., 2024)) use branch-and-cut over the binary linear ordering, separating short and long cycle constraints, leveraging initial heuristics (e.g., Kernighan-Lin 2), and Möbius-ladder constraints to efficiently tighten relaxations. Data reduction, interval decomposition, and fixed variable elimination are mainstays.

Kernelization and FPT: For k-almost-trees or graphs of cyclomatic number $\pi_B$2 (Bannister et al., 2013), kernelization reduces the instance to a core of $\pi_B$3 nodes, so all permutations can be enumerated in $\pi_B$4 (or further improved using fast matrix multiplication algorithms). For graphs of bounded treewidth, Courcelle’s theorem guarantees FPT algorithms when crossing number or treewidth is parameter (Bannister et al., 2014), though the implied constants are large.

For metro-line crossing minimization (MLCM-P/OSCM in paths), dynamic programming parameterized by multiplicity $\pi_B$5 yields $\pi_B$6-time algorithms (Okamoto et al., 2013).

4. Heuristic, Approximation, and Online Algorithms

Heuristic approaches are integral due to the problem’s NP-hardness:

• Barycenter and Median heuristics: The barycenter heuristic orders $\pi_B$7 by the mean position of its neighbors in $\pi_B$8 (guaranteed $\pi_B$9-approximate), while the median algorithm achieves a 3-approximation (tight), even for local crossing minimization (Rauch, 2024, Gutowski et al., 30 Sep 2025). For the local OSCM variant, the median heuristic with refined tie-breaking produces solutions with maximum per-edge crossings at most three times the optimum; this is best possible for its class (Gutowski et al., 30 Sep 2025).
• Probabilistic Variants: A probabilistic median (randomly selecting between neighbor indices) achieves an expected approximation factor of about 1.4664 (Rauch, 2024).
• Evolutionary Algorithms: Recent studies empirically demonstrate that jump-based evolutionary algorithms (EA)—where a vertex is “jumped” to a new position in the permutation—dominate both traditional heuristics and simpler permutation mutations (e.g., swaps or exchanges). With optimized selection of “acceptable” moves (not increasing crossing count), EAs achieve near-optimal solutions up to $B$0 faster than standard variants (Baumann et al., 2024).
• Star Insertion and Greedy Approaches: For general crossing minimization, iterative star insertion and greedy selection based on local crossing contributions provide scalable strategies, particularly in dense settings (Clancy et al., 2018, Rauch et al., 17 Jan 2025).
• Approximation for Metro-Line Crossing Minimization: OSCM as MLCM-P can be reduced to Min 2CNF Deletion, permitting an $B$1-approximation (Fink et al., 2013). Similar 2SAT-based approaches apply structurally to OSCM.

Online Algorithms: In the slotted online OSCM model, vertices must be irrevocably placed in “slots” upon arrival. No algorithm achieves a constant competitive ratio for general degrees, but for 2-regular graphs tight bounds of $B$2 on competitive ratio are proven, with algorithms leveraging "propagation arrows” to minimize crossings (Burjons et al., 2022).

5. Structural and Parameterized Results

Structural analysis reveals that the essential complexity of OSCM (number of admissible solutions) is governed not by all $B$3 pairs but rather by the number of unsuited pairs (both mutual crossing numbers nonzero). For cutwidth $B$4, the number of unsuited pairs is at most $B$5 (Rauch et al., 17 Jan 2025). Thus, for graphs with bounded cutwidth, the combinatorial search is exponentially smaller.

Moreover, part-to-part ordering constraints yield transitive tournaments, allowing for greedy algorithms that progressively order vertex pairs by increasing crossing involvement—a mechanism that underpins fast approximation albeit with currently unproven constant-factor guarantees for total crossings (Rauch et al., 17 Jan 2025).

Block crossing minimization in tanglegrams requires more sophisticated combinatorial reasoning (transpositions, breakpoints, blocks), with dynamic programming yielding constant-factor approximation algorithms in polynomial time and FPT algorithms for bounded number of block crossings (Dobler et al., 2023).

6. Advanced and Emerging Techniques: Quantum and Parallel Algorithms

Quantum algorithms provide the first sub-$B$6 exponential-time methods for exact OSCM in general bipartite graphs (Caroppo et al., 2024). By expressing OSCM as a “set problem” with a natural recurrence, one obtains an $B$7-time, polynomial-space quantum algorithm leveraging quantum dynamic programming and quantum minimum finding. A QRAM-free variant achieves $B$8-time while using only polynomial space, outperforming the best classical methods on instances with $B$9 crossings.

Parallelization of existing dynamic programs and FPT routines yields substantial speedups. For bitmask-based DP (state-space partitioned by subset size) a nearly $G = (A \uplus B, E)$0 speedup is attained on 16-core machines (Popa et al., 28 Sep 2025). The main bottlenecks are synchronization (shared mutex/queues in FPT search trees), memory contention (large DP tables), and imbalanced state space layers; addressing these is the focus of current work.

Recent solvers (e.g., for PACE 2024 (Jünger et al., 2024, Rauch, 2024)) combine advanced branch-and-cut, reduction, and heuristics to scale to practical instances; data reduction, interval-based decomposition, and incremental constraint generation are central for tractability.

7. Applications and Future Directions

OSCM and its variants are core subproblems in hierarchical (Sugiyama) graph drawing frameworks, VLSI layout, network visualizations, phylogenetic tanglegrams, and circuit wiring design. Minimizing per-edge crossings is especially relevant in settings where edge readability is critical (e.g., flow diagrams, social networks, biochemical pathways).

Open research questions include:

• Whether improved constant-factor approximation algorithms or PTAS can be designed for both total and local crossing minimization (Gutowski et al., 30 Sep 2025).
• The parameterized complexity for important width measures (treewidth, cutwidth) and the tightness of current FPT bounds.
• The trade-offs and practical integration of quantum and parallel approaches for large-scale, real-world networks.
• Structural characterizations beyond suited/unsuited pairs, particularly those enabling more refined kernelization or fast exact methods.

Continuous methodological innovation—across kernelization, branch-and-cut, heuristic acceleration, memorization reduction, and quantum/parallel computing—remains at the forefront of OSCM research, providing both theoretical challenges and practical solutions in graph drawing and combinatorial optimization.