A data assimilation method for log‐normally distributed observational errors

Steven J. Fletcher, Milija Županski

Quarterly Journal of the Royal Meteorological Society · 2006 · 92 citations · 17 references

Concepts

TL;DR

The paper proposes replacing the usual Gaussian assumption in variational data assimilation with log‑normally distributed observational errors and investigates which statistic best characterizes the distribution for univariate and multivariate cases, justifying the use of the mode. The authors derive a cost function, Jacobian, and Hessian for the log‑normal error model, then design a Hessian‑based preconditioner and generalize it for arbitrary cost functions to improve minimization. The Jacobian is shown to vanish at the solution in both model and observation spaces. © 2006 Royal Meteorological Society.

Abstract

Abstract In this paper we change the standard assumption made in the Bayesian framework of variational data assimilation to allow for observational errors that are log‐normally distributed. We address the question of which statistic best describes the distribution for the univariate and multivariate cases to justify our choice of the mode. From this choice we derive the associated cost function, Jacobian and Hessian with a normal background. We also find the solution to the Jacobian equal to zero in both model and observational space. Given the Hessian that we derive, we define a preconditioner to aid in the minimization of the cost function. We extend this to define a general form for the preconditioner, given a certain type of cost function. Copyright © 2006 Royal Meteorological Society

References

17