We examine the theoretical foundation of the method to estimate error covariances based on analysis residuals in observation space, also known as the Desroziers method. Our analysis also includes a method based on a posteriori diagnostics of variational analysis schemes. A mathematical analysis of convergence is carried out with a simplified regular observation network, where we identify stable and unstable fixed‐point solutions, examine their rate of convergence and the conditions for convergence to the truth and compare it with maximum‐likelihood estimation. It is shown that the estimation of variance parameters converges to the truth if the error correlation is specified correctly. The convergence is much faster with the analysis increment method than with the method designed for variational analysis schemes. We also propose a combination of the Desroziers scheme and the maximum‐likelihood estimation method that could be used to estimate the spatial correlation length‐scale of observation errors. The estimation of both full observation and background‐error covariances matrices derived entirely from observation‐based residuals does not change the gain matrix, but the estimation of either one matrices is well‐defined. An analysis of the estimation of the full observation error covariance matrix using a regular observation network reveals that if all eigenvalues of the prescribed background‐error covariance are smaller than their corresponding innovation covariance eigenvalue, then the estimated error covariance converges to the truth but only if the prescribed background error is correctly specified. In the case where some eigenvalues of the background‐error covariance greatly exceed those of the innovation covariance, the convergent observation‐error matrix may become rank‐deficient. Its inverse does not exist.

We compare two optimized chemical data assimilation systems, one based on the ensemble Kalman filter (EnKF) and the other based on four-dimensional variational (4D-Var) data assimilation, using a comprehensive stratospheric chemistry transport model (CTM). This work is an extension of the Belgian Assimilation System for Chemical ObsErvations (BASCOE), initially designed to work with a 4D-Var data assimilation. A strict comparison of both methods in the case of chemical tracer transport was done in a previous study and indicated that both methods provide essentially similar results. In the present work, we assimilate observations of ozone, HCl, HNO₃, H₂O and N₂O from EOS Aura-MLS data into the BASCOE CTM with a full description of stratospheric chemistry. Two new issues related to the use of the full chemistry model with EnKF are taken into account. One issue is a large number of error variance parameters that need to be optimized. We estimate an observation error variance parameter as a function of pressure level for each observed species using the Desroziers method. For comparison purposes, we apply the same estimate procedure in the 4D-Var data assimilation, where both scale factors of the background and observation error covariance matrices are estimated using the Desroziers method. However, in EnKF the background error covariance is modelled using the full chemistry model and a model error term which is tuned using an adjustable parameter. We found that it is adequate to have the same value of this parameter based on the chemical tracer formulation that is applied for all observed species. This is an indication that the main source of model error in chemical transport model is due to the transport. The second issue in EnKF with comprehensive atmospheric chemistry models is the noise in the cross-covariance between species that occurs when species are weakly chemically related at the same location. These errors need to be filtered out in addition to a localization based on distance. The performance of two data assimilation methods was assessed through an 8-month long assimilation of limb sounding observations from EOS Aura MLS. This paper discusses the differences in results and their relation to stratospheric chemical processes. Generally speaking, EnKF and 4D-Var provide results of comparable quality but differ substantially in the presence of model error or observation biases. If the erroneous chemical modelling is associated with moderately fast chemical processes, but whose lifetimes are longer than the model time step, then EnKF performs better, while 4D-Var develops spurious increments in the chemically related species. If, however, the observation biases are significant, then 4D-Var is more robust and is able to reject erroneous observations while EnKF does not.

The formulation of the background error covariances represented in the spectral space is discussed in the context of univariate assimilation relying on a grid point model, leaving out all the aspects of balances between the different control variables needed in meteorological assimilation. The spectral transform operations are discussed in the case of a spherical harmonics basis and we stress that there is no need for an inverse spectral transform and of a Gaussian grid. The analysis increments are thus produced directly on the model grid. The practice of producing analysis increments on a horizontal Gaussian grid and then interpolating to an equally spaced grid is also shown to produce a degradation of the analysis. The method discussed in this paper allows the implementation of separable and non-separable spatial correlations. The separable formulation has been implemented in the Belgian Assimilation System for Chemical ObsErvations (BASCOE) and its impact on the assimilation of O₃ observed by the Michelson Interferometer for Passive Atmospheric Sounding (MIPAS) is shown. To promote the use of this method by other non-meteorological variational systems and in particular chemistry, the Fortran code developed is made available to the community.

We examine how passive and active observations are useful to evaluate an air quality analysis. By leaving out observations from the analysis, we form passive observations, and the observations used in the analysis are called active observations. We evaluated the surface air quality analysis of O₃ and PM_2.5 against passive and active observations using standard model verification metrics such as bias, fractional bias, fraction of correct within a factor of 2, correlation and variance. The results show that verification of analyses against active observations always give an overestimation of the correlation and an underestimation of the variance. Evaluation against passive or any independent observations display a minimum of variance and maximum of correlation as we vary the observation weight, thus providing a mean to obtain the optimal observation weight. For the time and dates considered, the correlation between (independent) observations and the model is 0.55 for O₃ and 0.3 for PM_2.5 and for the analysis, with optimal observation weight, increases to 0.74 for O₃ and 0.54 for PM_2.5. We show that bias can be a misleading measure of evaluation and recommend the use of a fractional bias such as the modified normalized mean bias (MNMB). An evaluation of the model bias and variance as a function of model values also show a clear linear dependence with the model values for both O₃ and PM_2.5.

The effects of model discretization errors on the propagation of error covariance have a more complex nature than the effect on the state variable. The analysis is carried out for the advection transport equation, where the continuous (in space and time) propagation of the related error covariance function can be written, solved and compared with the discrete model applied to the covariance matrix. The numerical analysis of the problem is carried out with a 1D-problem, but is also illustrated with a 3D chemical transport model (CTM) used for chemical data assimilation. It is shown that variance loss (compared to the continuous propagation solution) depends on the covariance function itself as well as the numerical discretization scheme. The variance loss is particularly sensitive to the correlation length. In a simple first-order discretization, an analytical expression is obtained and is used to derive an analytical expression for inflation. Experiments show, for example, that following an analysis over a dense network of observations (with spatially uncorrelated errors) a significant variance loss occurs in the propagation step. With the (variance) inflation scheme, we are able to restore the variance lost at each grid point, and at each timestep, during the entire integration. The variance inflation scheme applied to an EnKF can be formulated to change the variance spread of the ensemble or to act directly on the state.