## Section: Scientific Foundations

### Sensitivity Analysis - Quantification of Uncertainties

Due to the strong non-linearity of geophysical systems and to their chaotic behavior, the dependence of their solutions on external parameters is very complex. Understanding the relationship between model parameters and model solutions is a prerequisite to design better models as well as better parameter identification. Moreover, given the present strong development of forecast systems in geophysics, the ability to provide an estimate of the uncertainty of the forecast is of course a major issue. However, the systems under consideration are very complex, and providing such an estimation is very challenging. Several mathematical approaches are possible to address these issues, using either variational or stochastic tools.

**Variational approach. ** In the variational framework, the sensitivity is the gradient of a response function with
respect to
the parameters or the inputs of the model. The adjoint techniques can therefore be used
for such a purpose.
If sensitivity is sought in the context of a forecasting system assimilating
observations, the optimality system must be
derived. This leads to the study of second-order properties: spectrum and eigenvectors of
the Hessian
are important information on system behavior.

**Global stochastic approach. **Using the variational approach to sensitivity leads to efficient computations of complex code derivatives.
However, this approach to sensitivity remains local because derivatives are generally computed at specific points. The stochastic approach of uncertainty analysis aims at studying global criteria based on a joint probability distribution functions modelling of the problem variables. The obtained sensitivity indices describe the global variabilities of the phenomena. For example, the Sobol sensitivity index is given by the ratio between the output variance conditionally to one input and the total output variance. The computation of such quantities leads to statistical problems. For example, the sensitivity indices have to be efficiently estimated from a few runs, using semi or non-parametric estimation techniques. The stochastic modelling of the input/output relationship is another solution.