The perceptual and spatial architecture of Mullerian mimicry in Heliconius Butterflies

Mullerian mimicry theory predicts reciprocal convergence among defended species toward shared warning signals, yet the extent, symmetry, and perceptual basis of this convergence remain poorly quantified. Using deep-learning-based phenotyping combined with neural networks calibrated to avian predator and butterfly visual systems, we quantify mimetic similarity across 56 subspecies of Heliconius erato and H. melpomene. Mimetic phenotypes show broad clustering consistent with recognized mimicry rings, but vary continuously within. Phenotypic similarity is weakly structured by phylogeny and strongly organized geographically, consistent with repeated local convergence. Spatial overlap among co-mimics is heterogeneous, suggesting that frequency-dependent selection operates at variable spatial scales across mimicry communities. Models parameterized to different visual systems reveal that trait salience and discriminatory performance depend on the observer, indicating that apparent imperfect mimicry may arise from receiver-specific perceptual weighting. Asymmetric learning between lineages further suggests that convergence may not always be reciprocal, consistent with one species evolving towards another in advergence. Together, these results recast mimicry as a perceptually structured and spatially dynamic continuum rather than a set of discrete, mutually reinforcing rings, highlighting how community composition and sensory ecology shape the evolution of adaptive resemblance.