To apply the CDF to a generic level 2 atmospheric product (primarily vertical profiles of temperature or constituents), the product must have specific characteristics. These requirements ensure that the CDF algorithm can be correctly applied and that the results are physically meaningful.

Six Fundamental Requirements

  1. Optimal estimation technique — The product must be retrieved using the optimal estimation method.
  2. Full state vector — The complete state vector \widehat{\mathbf{x}}_{i} must be available.
  3. Profile definition — Profiles in the state vector must be defined in terms of concentrations or partial columns.
  4. Vertical grid — The vertical grid (pressure or height, levels or layers) on which the profiles are defined must be available.
  5. A priori state vector — The complete a priori state vector \mathbf{x}_{ai} must be available, with profile components defined on the same grid as the retrieved profiles.
  6. Characterisation matrices — At least two of the following three matrices must be available:
    • The Averaging Kernels Matrix \mathbf{A}_{i} (for the full state vector)
    • The Variance-Covariance Matrix of total errors \mathbf{S}_{i} (for the full state vector)
    • The a priori Variance-Covariance Matrix \mathbf{S}_{ai} (for the full state vector)

If only two of the three matrices in requirement 6 are available, the third can be calculated using the relationships below.

Deriving Missing Characterisation Matrices

The three matrices in requirement 6 are related by the following equations. Use these to compute the missing matrix when only two are available.

\mathbf{S}_{i} = \left( \mathbf{I} - \mathbf{A}_{i} \right) \cdot \mathbf{S}_{ai}
(P1)
\mathbf{S}_{ai} = \left( \mathbf{I} - \mathbf{A}_{i} \right)^{- 1} \cdot \mathbf{S}_{i}
(P2)
\mathbf{A}_{i} = \left( \mathbf{I} - \mathbf{S}_{i} \cdot {\mathbf{S}_{ai}}^{- 1} \right)
(P3)

Note on equation (P2) and (P3): Both inverses refer to matrices that are typically invertible. However, the noise covariance matrix \mathbf{S}_{ni}, if available instead of the total error covariance \mathbf{S}_{i}, is more problematic because the inverse of the averaging kernel matrix is often singular. The relationship \mathbf{S}_{ni} = \mathbf{A}_{i} \cdot \mathbf{S}_{i} can be safely calculated in the forward direction, but the inverse transformation is not generally possible.