Data assimilation with inequality constraints

If values of variables in a numerical model are limited to specified ranges, these restrictions should be enforced when data are assimilated. The simplest option is to assimilate without regard for constraints and then to correct any violations without worrying about additional corrections implied by correlated errors. This paper addresses the incorporation of inequality constraints into the standard variational framework of optimal interpolation with emphasis on our limited knowledge of the underlying probability distributions. Simple examples involving only two or three variables are used to illustrate graphically how active constraints can be treated as error-free data when background errors obey a truncated multi-normal distribution. Using Lagrange multipliers, the formalism is expanded to encompass the active constraints. Two algorithms are presented, both relying on a solution ignoring the inequality constraints to discover violations to be enforced. While explicitly enforcing a subset can, via correlations, correct the others, pragmatism based on our poor knowledge of the underlying probability distributions suggests the expedient of enforcing them all explicitly to avoid the computationally expensive task of determining the minimum active set. If additional violations are encountered with these solutions, the process can be repeated. Simple examples are used to illustrate the algorithms and to examine the nature of the corrections implied by correlated errors. Published by Elsevier Ltd.