[Paper Review] Regular networks are determined by their trees

Stephen J. Willson|arXiv (Cornell University)|Feb 17, 2009

Interconnection Networks and Systems17 references3 citations

TL;DR

This paper proves that regular phylogenetic networks are uniquely determined by the set of rooted trees they display. It introduces a constructive algorithm, MaximumProperChild, that reconstructs a regular network from its displayed trees by iteratively identifying clusters and their proper children using tree consistency conditions, establishing that Tr(N) = Tr(M) implies N = M for regular networks.

ABSTRACT

A rooted acyclic digraph N with labelled leaves displays a tree T when there exists a way to select a unique parent of each hybrid vertex resulting in the tree T. Let Tr(N) denote the set of all trees displayed by the network N. In general, there may be many other networks M such that Tr(M) = Tr(N). A network is regular if it is isomorphic with its cover digraph. This paper shows that if N is regular, there is a procedure to reconstruct N given Tr(N). Hence if N and M are regular networks and Tr(N) = Tr(M), it follows that N = M, proving that a regular network is uniquely determined by its displayed trees.

Motivation & Objective

• To establish conditions under which a phylogenetic network can be uniquely reconstructed from the collection of trees it displays.
• To address the problem that non-regular networks may share the same displayed trees, leading to non-unique reconstruction.
• To prove that regular networks—defined by unique clusters and path-based cluster inclusion—are uniquely determined by their displayed trees.
• To develop a constructive algorithm that recovers the full network structure from Tr(N), the set of displayed trees.
• To clarify the biological and mathematical conditions under which network reconstruction from gene trees is possible and unique.

Proposed method

• Define a regular network as one where each vertex has a unique cluster of descendants and cluster inclusion corresponds exactly to directed path existence.
• Introduce the concept of a 'proper tree' for a cluster C: a tree in Tr(N) where C appears and its parent is also a cluster in N.
• For each cluster C, compute ProperTr(C) as the set of trees in Tr(N) where C appears with a parent that is a cluster in N.
• Identify ProperCh(C) as the set of children of C in any proper tree, then extract MaxProperCh(C) as the maximal such children sets.
• Construct the network incrementally by adding vertices corresponding to MaxProperCh(C) and arcs from C to each maximal child, ensuring consistency with the original network structure.
• Prove that the resulting network M_k is isomorphic to the original regular network N, thus reconstructing N from Tr(N).

Experimental results

Research questions

• RQ1 Can a phylogenetic network be uniquely reconstructed from the set of trees it displays?
• RQ2 Under what conditions is the set of displayed trees Tr(N) sufficient to recover the original network N?
• RQ3 Does the regularity condition—unique clusters and path-based cluster inclusion—guarantee unique reconstruction from Tr(N)?
• RQ4 Is there a constructive algorithm that can recover a regular network from Tr(N) alone?
• RQ5 What happens when regularity fails? Can non-regular networks with identical Tr(N) exist?

Key findings

• The MaximumProperChild algorithm successfully reconstructs a regular network N from its set of displayed trees Tr(N).
• If two regular networks M and N have the same set of displayed trees (Tr(M) = Tr(N)), then M and N are isomorphic (M = N), proving uniqueness.
• The procedure fails on non-regular networks, as demonstrated by Figure 1, where two distinct networks M and N share the same Tr(N) = {b, c}.
• The method relies on identifying 'proper trees' for each cluster C, ensuring that only trees where the parent of C is a valid cluster in N are considered.
• The algorithm correctly identifies maximal children of each cluster by analyzing their appearance across proper trees, avoiding spurious children from non-proper trees.
• The final reconstructed network M_k is isomorphic to the original network N, with vertices and arcs matching exactly the clusters and parent-child relationships in N.