Coagulation-Fragmentation Duality of Infinitely Exchangeable Partitions from Coupled Mixed Poisson Species Sampling Models

Abstract: Jim Pitman's (1999) celebrated coagulation-fragmentation duality for the PD($\alpha$,$\theta$) family of laws of Pitman and Marc Yor (1997) has resisted generalization beyond its canonical setting. We resolve this by introducing a novel, four-part coupled process built upon the Poisson Hierarchical Indian Buffet Process (PHIBP), a framework we recently developed for modeling microbiome species sampling across multiple ($J \ge 1$) groups. This provides a tractable generalization of the duality to processes driven by arbitrary subordinators, yielding explicit laws for both single-group and the previously uncharacterised multi-group structured partitions. The static, fixed-time partitions are revealed to be a single projection of an inherently dynamic, four-component coupled system. This new construction simultaneously defines: (i) the fine-grained partition, (ii) its coagulation operator, (iii) a forward-in-time system of coupled, time-homogeneous fragmentation processes in the sense of Jean Bertoin (2006), and (iv) a dual, backward-in-time structured coalescent. All four components are governed by the same underlying compositional structure, which yields their exact compound Poisson representations and, for the coalescent, drives simultaneous, across-group merger events. The chief result is a general coagulation-fragmentation duality that generalizes Pitman's duality in two fundamental directions. First, it holds for arbitrary driving subordinators, defining a vast new family of single-group dualities. Second, it operates in the previously uncharacterised multi-group ($J$-group) setting. This work establishes a unified and tractable framework for a rich class of partition-valued dynamics, revealing structural connections between the theories of fragmentation and coalescence.