When is greedy exact for weighted minimum 3-cut?
The greedy algorithm for weighted minimum 3-cut first takes a global minimum-weight cut and then splits the cheaper side. It is a tight 4/3-approximation. The question is when it is exact in general weighted graphs.
A case analysis based on posimodularity shows that greedy(S) ≤ max(λ + d(A), W(A,B,C)) for every 3-partition and every global minimum-weight cut. The remaining hard instances, where all pair weights are below λ, are handled with merge lemmas for Gomory–Hu sides. A 7-vertex counterexample shows that Gomory–Hu sides alone do not suffice for Kapoor's algorithm, and the argument extends to minimum k-cut. The general theorems are formalized in Lean 4 for all n, and the counterexample is verified by kernel evaluation.
Greedy is exact whenever some optimum has two parts joined by weight at least λ. This extends the Hochbaum–Shmoys theorems to all weighted graphs and recovers the 4/3 bound. The 7-vertex graph has a unique Gomory–Hu tree and a unique optimum with no part equal to a tree side. Lean proofs and Python reproduction scripts are provided, and random and adversarial computational tests found no violations of the proved statements.
