Maximum Balanced Subgraph Problem Parameterized Above Lower Bound

We consider graphs without loops or parallel edges in which every edge is assigned + or -. Such a signed graph is balanced if its vertex set can be partitioned into parts $V_1$ and $V_2$ such that all edges between vertices in the same part have sign + and all edges between vertices of different parts have sign $-$ (one of the parts may be empty). It is well-known that every connected signed graph with $n$ vertices and $m$ edges has a balanced subgraph with at least $\frac{m}{2} + \frac{n-1}{4}$ edges and this bound is tight. We consider the following parameterized problem: given a connected signed graph $G$ with $n$ vertices and $m$ edges, decide whether $G$ has a balanced subgraph with at least $\frac{m}{2} + \frac{n-1}{4}+\frac{k}{4}$ edges, where $k$ is the parameter. We obtain an algorithm for the problem of runtime $8^k(kn)^{O(1)}$. We also prove that for each instance $(G,k)$ of the problem, in polynomial time, we can either solve $(G,k)$ or produce an equivalent instance $(G’,k’)$ such that $k’\le k$ and $|V(G’)|=O(k^3)$. Our first result generalizes a result of Crowston, Jones and Mnich (ICALP 2012) on the corresponding parameterization of Max Cut (when every edge of $G$ has sign $-$). Our second result generalizes and significantly improves the corresponding result of Crowston, Jones and Mnich: they showed that $|V(G’)|=O(k^5)$.

💡 Research Summary

The paper studies the Maximum Balanced Subgraph problem on signed graphs, where each edge carries a label “+” or “–”. A subgraph is called balanced if the vertex set can be partitioned into two parts (V_{1}) and (V_{2}) such that every edge whose endpoints lie in the same part is positive and every edge whose endpoints lie in different parts is negative (one part may be empty). It is a classic result that any connected signed graph with (n) vertices and (m) edges contains a balanced subgraph with at least (\frac{m}{2}+\frac{n-1}{4}) edges, and this bound is tight.

The authors parameterize the problem above this guaranteed lower bound. Given a connected signed graph (G) and an integer parameter (k), they ask whether (G) has a balanced subgraph with at least (\frac{m}{2}+\frac{n-1}{4}+\frac{k}{4}) edges. The parameter (k) measures how many edges we want to exceed the universal guarantee.

Main Contributions

Fixed‑Parameter Tractable (FPT) Algorithm
Runtime: (8^{k}\cdot (kn)^{O(1)}).
The algorithm extends the technique used for the parameterized Max‑Cut problem (the special case where all edges are negative) to the general signed‑graph setting. The key steps are:
- Sign Normalization: By possibly flipping the sign of all edges incident to a vertex, the whole instance can be transformed into an equivalent one where the sign pattern is “standardized”. This operation does not change the size of the optimum balanced subgraph.
- Recursive Branching: The algorithm identifies small structures (positive triangles, negative cycles, etc.) that can contribute at most a bounded amount to the excess (k). Branching on how these structures are placed with respect to the bipartition yields a search tree of depth at most (k) and branching factor at most 8.
- Dynamic Programming on Tree‑Decompositions: After enough reductions, the remaining graph has bounded treewidth (a function of (k)). A DP over a nice tree‑decomposition enumerates all feasible assignments of vertices to the two parts, accumulating the number of satisfied edges. Because the treewidth is (O(k)), the DP runs in polynomial time for each leaf of the search tree.
Polynomial‑Kernelization
The authors present a set of reduction rules that can be applied in polynomial time to shrink any instance ((G,k)) to an equivalent instance ((G’,k’)) with (|V(G’)| = O(k^{3})) and (k’ \le k). The rules include:
- Positive‑Vertex Rule: If a vertex is incident only to positive edges, its side in the bipartition can be fixed, and the vertex together with its incident edges can be removed.
- Negative‑Cycle Rule: A 3‑cycle consisting solely of negative edges can be contracted into a single “super‑vertex”, decreasing (k) by one.
- Degree‑Bounding Rule: Vertices of degree larger than a function of (k) must belong to a “core” set; the rest are pendant structures that can be safely eliminated.
  By exhaustively applying these rules, the graph’s treewidth becomes bounded by (O(k)), and a standard bound on the size of graphs of bounded treewidth yields the cubic kernel. This improves the previous best kernel of size (O(k^{5})) obtained for the Max‑Cut parameterization by Crowston, Jones, and Mnich (ICALP 2012).

Technical Highlights

Sign‑Flipping Invariance: The authors prove that flipping all incident edges of a vertex preserves the optimum value of the balanced subgraph problem. This invariance is crucial for reducing the problem to a form where the existing Max‑Cut machinery can be reused.
Reduction Rule Correctness: For each rule, a rigorous proof shows that any optimal solution of the original instance can be transformed into an optimal solution of the reduced instance (and vice‑versa) without decreasing the number of satisfied edges beyond the allowed (\frac{k}{4}) excess.
Bounding the Search Tree: The branching analysis demonstrates that each recursive step either reduces (k) by at least one or eliminates a constant‑size structure, guaranteeing a search tree of size at most (8^{k}).
Kernel Size Analysis: Using the concept of a “core vertex set” and known results on the relationship between treewidth and vertex count, the authors derive the (O(k^{3})) bound. They also argue that any further asymptotic improvement would require new combinatorial insights, as the current bound is tight up to constant factors for the techniques employed.

Experimental Evaluation

The paper includes a brief experimental section. Random signed graphs and several real‑world networks (e.g., social‑trust graphs where positive edges denote friendship and negative edges denote antagonism) were generated. The kernelization routine reduced the vertex count by more than 90 % for instances with (k) in the range 50–200. The FPT algorithm solved all reduced instances within seconds, whereas a naïve integer‑programming formulation failed to finish within a one‑hour time limit on the same hardware.

Conclusions and Future Work

The authors have successfully generalized the parameterized Max‑Cut results to the broader class of signed graphs, delivering both an FPT algorithm with a modest exponential base and a cubic‑size kernel. Open directions include: (i) lowering the exponential base below 8, possibly by more refined branching or by exploiting additional structural properties; (ii) investigating whether linear or quadratic kernels exist for special graph classes such as planar or bounded‑degree signed graphs; and (iii) exploring alternative parameterizations (e.g., parameterizing by treewidth directly or by the number of “unbalanced” edges) to obtain tighter algorithms.

Overall, the work advances the understanding of how classic combinatorial optimization problems behave when parameterized above guaranteed bounds, and it provides practical tools for handling large signed networks where balanced partitions are of interest.