Ancestral State Reconstruction under Localized Evolutionary Change: Error Patterns and Their Control by Tree Partitioning

Abstract

Localized evolutionary changes can distort ancestral state reconstruction (ASR) because conventional continuous-trait methods generally treat branchwise changes as draws from a shared background process. A disproportionately large change confined to a single branch therefore creates localized model misspecification, but how the resulting errors vary among reconstruction paradigms and phylogenetic positions remains poorly understood. Using controlled simulations on idealized and empirical mammalian phylogenies, we characterized the error patterns produced by single-branch large changes (SBLCs) across five Brownian-motion-based ASR implementations. On intact trees, errors were highly structured and method dependent: the recursive local method propagated distortion mainly upstream, whereas global methods produced multidirectional effects involving both SBLC boundary nodes, neighboring lineages, and descendants of the affected branch. We then evaluated patching, a conditional robustness procedure that partitions the phylogeny at a candidate branch, applies the same ASR method independently to the resulting subtrees, and estimates the upstream boundary node state by branch-length-weighted interpolation. Patching reduced errors near and away from the affected branch and lowered tree-wide error across trait-increasing and trait-decreasing SBLCs, a broad range of sensitivity analyses, and configurations with two co-occurring SBLCs, while introducing little additional error when no SBLC was present. Under the reference upsizing SBLC, for instance, it cut the maximum-likelihood reconstruction error at the upstream boundary node from 56.20 to 0.03 (mean true state, 4.98), matching the stable-process benchmark. In cetacean body-length evolution, patching candidate branches identified by a Levy-process model generally shifted conventional estimates toward stable-process estimates and nearby fossil references. Under an Ornstein-Uhlenbeck process with strong post-SBLC mean reversion, however, neither patching nor stable-process reconstruction recovered downstream ancestral states once the phenotypic signature of the SBLC had largely disappeared from descendant tip values, revealing an identifiability limit of terminal-only inference. Localized evolutionary change therefore creates structured, method-dependent reconstruction errors, and tree partitioning provides a complementary diagnostic of the robustness of ancestral inferences to candidate branch-localized changes.