The ErrorCovarianceToolbox Application

This application is the core application of SABER, the place of choice to test a background error matrix outside of a Variational application. It is also used for covariance calibration.

More precisely, this application:

  • Creates a SABER error covariance;

  • Runs a Dirac test (optional);

  • Generates random samples through randomization (optional).

This is briefly detailed in the next sections.

Creation of a SABER error covariance

In addition of the mere creation of each SABER block in the SABER error covariance, this includes (for each SABER block):

  • the block calibration, if yaml key calibration is provided at the SABER block level (cf. Calibration of a SABER error covariance model);

  • an adjoint test, if adjoint test is specified to true at the SABER block level, or, otherwise, if adjoint test is specified to true at the covariance level;

  • an inverse test, if inverse test is true at the covariance level, unless skip inverse test is true at the block level;

  • a square root test for the central block, if square-root test is true at the covariance level.

More information on SABER block testing is given in the SABER block testing page.

Dirac test

The Dirac test consists in applying the error covariance to a Dirac input or to a sum of Dirac inputs. This test is performed by the application if the dirac key is given. The value of this key is the list of coordinates of the Dirac input(s).

The output increment is written to file if the dirac output key is given.

If the diagnostic points key is given, the output values at the Dirac points and at a set of diagnostic points are output to the Test log.

If the covariance profile key is given, a 1D covariance profile is output to the Test log or written to file.

Randomization

Randomization is used to generate random increments that are normally distributed, with zero mean and with the SABER covariance as covariance matrix. Randomization runs if no Dirac test has been run or if the randomization size key is given at the covariance level. The output increments are written to file if the output states and/or output increments keys are given.