A Diameter-Revealing Proof of the Bondy-Lovasz Lemma

We present a strengthened version of a lemma due to Bondy and Lov'asz. This lemma establishes the connectivity of a certain graph whose nodes correspond to the spanning trees of a 2-vertex-connected graph, and implies the k=2 case of the Gy\H{o}ri-Lov'asz Theorem on partitioning of k-vertex-connected graphs. Our strengthened version constructively proves an asymptotically tight O(|V|^2) bound on the worst-case diameter of this graph of spanning trees.

💡 Research Summary

The paper revisits a classical lemma due to Bondy and Lovász that underlies the k = 2 case of the Győri–Lovász partition theorem. The lemma concerns a graph G_T whose vertices are all spanning trees of a given 2‑vertex‑connected graph G, rooted at a distinguished vertex a. Two spanning trees T and T′ are adjacent in G_T if their intersection contains a tree on |V| − 1 vertices that includes a. While the original lemma guarantees that G_T is connected, the known bound on its diameter is exponential, leaving open whether a polynomial‑length path always exists between any two trees.

The authors present a strengthened version that proves, constructively, that the diameter of G_T is at most O(|V|²) for any 2‑vertex‑connected graph G and any choice of root a. Moreover, they show that this bound is asymptotically tight by constructing a family of graphs G_k for which the distance between two particular spanning trees is Ω(|V_k|²).

The core of the upper‑bound proof relies on the concept of an st‑ordering of a biconnected graph, a well‑known tool that can be computed in polynomial time (Eberhardt, Lempel et al.). Choosing an arbitrary edge incident to a, the authors compute an st‑ordering v₁ = a, v₂,…,v_n. Using this ordering they define a canonical spanning tree T⁺: vertex v_n is made a child of a, and every other vertex becomes a child of its highest‑numbered neighbor. T⁺ is uniquely defined and serves as a convenient “starting point”.

To reach an arbitrary target tree T′, the authors introduce a sequence of intermediate “milestones” S₁,…,S_n, where each S_k is a connected subgraph of T′ containing a and S_{k+1} adds exactly one new vertex. For each milestone they construct a spanning tree T_k that contains S_k as a subtree. The transition from T_k to T_{k+1} is performed by a two‑phase leaf‑re‑attachment process:

1. Forward phase – vertices outside S_k are processed in increasing order of the st‑numbering. Each vertex (except a distinguished vertex v*_k) is detached from its current parent and attached to its lowest‑numbered neighbor; v*_k is attached to a specially chosen neighbor u*_k and the process stops.
2. Backward phase – the vertices re‑attached in the forward phase (except v*_k) are processed in reverse order and re‑attached to their highest‑numbered neighbor, essentially undoing the forward moves.

Because every detachment occurs when the vertex is a leaf, each step changes exactly one leaf’s parent, which matches the adjacency definition of G_T. The forward phase uses O(n) moves, the backward phase another O(n), and there are n milestones, yielding a total of O(n²) moves. All operations involve only adjacency lists and the pre‑computed st‑ordering, so the entire sequence can be generated in polynomial time.

For the lower bound, the authors define G_k with 4k + 1 vertices arranged as a chain of 4‑cycles sharing edges, plus two extra edges from the root a = v₀ to v₁ and v₂. Two spanning trees T_Ak (a simple path) and T_Bk (alternating edges) are considered. By analyzing the earliest index t_i at which edge e_i appears/disappears along any sequence of adjacent trees, they prove that the indices must strictly decrease, forcing at least Σ_{i=0}^{k-1}