A Generalized Unscented Transformation for Probability Distributions.

Area of Science:

• Statistics
• Computational Mathematics
• Data Assimilation

Background:

• The unscented transform (UT) is a method for propagating mean and covariance through nonlinear transformations using sigma points.
• Traditional UT methods often fail with non-Gaussian distributions and significant nonlinearities due to restrictive assumptions.
• Accurate propagation of uncertainty is vital in fields like epidemiology and control systems.

Purpose of the Study:

• To develop a generalized unscented transform (GenUT) that accurately captures higher-order moments of probability distributions.
• To extend the applicability of unscented transforms to non-Gaussian random variables and substantial nonlinearities.
• To demonstrate the utility of GenUT in practical applications such as infectious disease modeling.

Main Methods:

• The generalized unscented transform (GenUT) is developed using 2n+1 sigma points.
• GenUT accurately captures up to the diagonal components of skewness and kurtosis tensors.
• Analytical constraints can be enforced on sigma points, guaranteeing at least second-order accuracy.

Main Results:

• GenUT achieves accurate propagation for a wider range of probability distributions compared to standard UT.
• The method maintains the same number of sigma points as the original UT.
• GenUT is successfully applied to the assimilation of observations in infectious disease modeling, including COVID-19.

Conclusions:

• The generalized unscented transform (GenUT) offers a robust solution for propagating uncertainty in the presence of non-Gaussian distributions and nonlinearities.
• GenUT provides a significant advancement for data assimilation and state estimation in complex dynamical systems.
• This work facilitates more accurate modeling of critical phenomena, such as the spread of infectious diseases.