Would multi-modal posteriors cause different distortion patterns than unimodal ones across stages?

Your question asks whether the shape of a posterior distribution—specifically whether it has one peak (unimodal) or multiple peaks (multimodal)—affects how distortions accumulate across sequential inference stages.

The archive directly addresses this. When normalizing flows are used as flexible priors in iterative Bayesian workflows, multimodal posteriors do indeed cause different distortion patterns than unimodal ones [AG-2025.09-1431]. Specifically, the study found that when posteriors from one analysis are used to inform priors for the next stage, "in cases with pronounced multi-modality or dataset tension, distortions may arise" — whereas unimodal target distributions were successfully reproduced across stages [AG-2025.09-1431]. This matters because it means sequential Bayesian pipelines (common in physics) can silently accumulate errors when the true posterior has multiple competing solutions.

The mechanism appears related to how normalizing flows—which learn flexible prior distributions—handle complex structure. When a posterior splits into distinct modes (say, two physically plausible but incompatible parameter configurations), passing that through to the next stage requires the flow to capture those correlations accurately. Flows trained on unimodal posteriors do this reliably using KL divergence loss, but multimodal cases introduce what the paper calls "distortions" that propagate forward [AG-2025.09-1431].

A related principle applies more broadly in Bayesian inference: using single-component models on data that mixture of multiple models introduces systematic bias, and multi-component approaches mitigate it [AG-2024.11-573]. This suggests that recognizing and explicitly modeling multiple modes—rather than treating the posterior as having a single peak—is crucial for maintaining accuracy across inference stages.

The preprints don't provide detailed quantification of *how much* the distortion patterns differ between unimodal and multimodal cases, only that they do differ and that practitioners should be cautious.