Conceitos essenciais
Kernel packets (KPs) provide a general framework to construct compactly supported basis functions for Gaussian processes (GPs) driven by stochastic differential equations (SDEs), enabling efficient training and prediction of GP models.
Resumo
The paper presents a general theory for constructing kernel packets (KPs) - a set of compactly supported basis functions - for Gaussian processes (GPs) driven by stochastic differential equations (SDEs).
Key highlights:
- The authors prove that KPs generally exist for GPs defined by SDEs and provide a framework to obtain them.
- KPs are derived from the forward and backward Markov properties of state-space models, in contrast to previous work that used harmonic analysis.
- The minimum number of equations required to construct a minimal KP system is shown to be 2m+1, where m is the order of the SDE.
- The KP basis functions are proven to be linearly independent and can be used to achieve O(n) training time and O(log n) or O(1) prediction time for GP regression.
- The KP framework is extended to handle combined kernels formed by addition and multiplication of individual kernels.
- Examples are provided for the Matérn-3/2 and integrated Brownian motion kernels to illustrate the KP construction.
The proposed KP theory provides a general and efficient approach for GP modeling and inference, with applications in various domains.