Theorems explore flow matching with von Mises-Fisher posteriors, revealing structural limitations in multimodal scenarios.
Please don't cite it. The author found that the proof of some important theorems have some ambiguity, so the correctness is questionable. Disclosure: The proof verification involves AI-assistance. We study flow matching on the Fisher-Rao simplex Δᵈ⁻¹, realized isometricallyas the positive orthant Sᵈ⁻¹₊ of the round sphere via xᵢ↦√xᵢ. Aposterior over completions pₜ(x₁ xₜ) modeled by a single von Mises-Fisher(vMF) law is unimodal by construction, so it cannot represent a genuinely multimodalposterior: under the flow-matching loss it is driven toward a single mean direction thatneed not track any true mode. We give two exact, non-asymptotic results. First, for asingle vMF component, we compute the posterior-averaged log map in closed form up to anexplicit O(1/κ²) remainder, replacing an earlier probabilistic concentrationargument and recovering its κ→∞ limit as the special case $K=1$. Second,for a K≥2-component mixture-vMF posterior, we prove that its marginal velocity fieldcannot equal any single-target field log(·)(μ⁾ on Sᵈ⁻¹₊: exactly,with no error terms, in the limit κₖ→∞. Multimodality is therefore astructural obstruction that no single-target field can remove. We do not address herethe Fisher-Rao boundary-explosion problem near ᵈ⁻¹₊ or the relationship tothe α-Flow family.
No takes yet. Share an insight, caveat, or question.
Le-Rong Hsu (2026) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: