Abstract:It is known that the set of lumpable Markov chains over a finite state space, with respect to a fixed lumping function, generally does not form an exponential family of stochastic matrices. In this work, we explore efficiently verifiable necessary and sufficient combinatorial conditions for families of lumpable transition matrices to form exponential families.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is: **On the level of combinatorial properties, determine which reducible families of random matrices can form exponential families**. Specifically, the author explores how to efficiently verify certain conditions so that these families of reducible transition matrices can form exponential families.
### Background and Motivation
Exponential families are of great significance in statistics because they possess some unique inferential properties. For example, exponential families provide sufficient statistics that can compress any number of independent and identically distributed data into a fixed number of values. Moreover, the maximum - likelihood estimator can reach the Cramér - Rao lower bound only when the family of distributions forms an exponential family.
However, for Markov chains, even under a fixed clustering function, the set of reducible Markov chains usually does not form an exponential family. This is in contrast to many other types of families of random matrices (such as the set of reversible random matrices), which always form an exponential family.
### Research Questions
Therefore, the research questions of this paper can be summarized as follows:
- **How to characterize the necessary and sufficient conditions for a reducible family of random matrices to form an exponential family?**
- **Can these conditions be verified within polynomial time?**
### Main Contributions
The author presents the following main results:
1. **Multi - row Merged Block Criterion**:
- **Sufficient Condition**: If there are no multi - row merged blocks in the reducible family \( W_\kappa(Y, E) \), then it is an exponential family.
- **Necessary Condition**: If there are redundant multi - row merged blocks, then this family cannot be an exponential family.
2. **Dimension Criterion**:
- When the reducible family is an exponential family, the dimension of the log - affine envelope of the reducible function modulo the shift function is known. If the dimensions do not match, then this family is not an exponential family, and this property can be verified by a polynomial - time algorithm.
3. **Monotonicity and Stability**:
- The author studies the characteristics of maintaining the exponential family property through basic operations (such as operations on edge sets). For example, if \( W_\kappa(Y, E') \) is an exponential family and \( E\subset E' \), then \( W_\kappa(Y, E) \) is also an exponential family.
### Examples and Applications
The article also provides some specific examples to illustrate the applications of these theoretical results, such as the Hudson extension and the lazy cycle criterion. These examples show that even in the presence of multiple multi - row merged blocks, an exponential family can still be formed in some cases.
### Conclusion
In general, this paper aims to find the conditions for a reducible family of random matrices to form an exponential family by analyzing the combinatorial properties of connection graphs and clustering mappings. This not only helps in understanding the geometric properties of these complex structures but also provides a theoretical basis for model selection in practical applications.