.. _sensitivity_ancova:
Sensivity analysis with correlated inputs
-----------------------------------------
The ANCOVA (ANalysis of COVAriance) method, is a variance-based method
generalizing the ANOVA (ANalysis Of VAriance) decomposition for models
with correlated input parameters.
Let us consider a model :math:`Y = h(\vect{X})` without making any
hypothesis on the dependence structure of
:math:`\vect{X} = \{X^1, \ldots, X^{n_X}\}`, a :math:`n_X`-dimensional
random vector. The covariance decomposition requires a functional
decomposition of the model. Thus the model response :math:`Y` is
expanded as a sum of functions of increasing dimension as follows:
.. math::
:label: Model
h(\vect{X}) = h_0 + \sum_{u\subseteq\{1,\dots,n_X\}} h_u(X_u)
:math:`h_0` is the mean of :math:`Y`. Each function :math:`h_u`
represents, for any non empty set :math:`u\subseteq\{1, \dots, n_X\}`,
the combined contribution of the variables :math:`X_u` to :math:`Y`.
Using the properties of the covariance, the variance of :math:`Y` can be
decomposed into a variance part and a covariance part as follows:
.. math::
\begin{aligned}
Var[Y] &=& Cov\left[h_0 + \sum_{u\subseteq\{1,\dots,n_X\}} h_u(X_u), h_0 + \sum_{u\subseteq\{1,\dots,n\}} h_u(X_u)\right] \\
&=& \sum_{u\subseteq\{1,\dots,n_X\}} Cov\left[h_u(X_u), \sum_{u\subseteq\{1,\dots,n_X\}} h_u(X_u)\right] \\
&=& \sum_{u\subseteq\{1,\dots,n_X\}} \left[Var[h_u(X_u)] + Cov[h_u(X_u), \sum_{v\subseteq\{1,\dots,n_X\}, v\cap u=\varnothing} h_v(X_v)]\right]
\end{aligned}
The total part of variance of :math:`Y` due to :math:`X_u` reads:
.. math:: S_u = \frac{Cov[Y, h_u(X_u)]}{Var[Y]}
The variance formula described above enables to define each sensitivity
measure :math:`S_u` as the sum of a :math:`\mathit{physical}` (or
:math:`\mathit{uncorrelated}`) part and a :math:`\mathit{correlated}`
part such as:
.. math:: S_u = S_u^U + S_u^C
where :math:`S_u^U` is the uncorrelated part of variance of :math:`Y`
due to :math:`X_u`:
.. math:: S_u^U = \frac{Var[h_u(X_u)]}{Var[Y]}
and :math:`S_u^C` is the contribution of the correlation of :math:`X_u`
with the other parameters:
.. math:: S_u^C = \frac{Cov[h_u(X_u), \displaystyle \sum_{v\subseteq\{1,\dots,n_X\}, v\cap u=\varnothing} h_v(X_v)]}{Var[Y]}
As the computational cost of the indices with the numerical model
:math:`h` can be very high, it is suggested to approximate the model
response with a polynomial chaos expansion. However, for the sake of
computational simplicity, the latter is constructed considering
:math:`\mathit{independent}` components :math:`\{X^1,\dots,X^{n_X}\}`.
Thus the chaos basis is not orthogonal with respect to the correlated
inputs under consideration, and it is only used as a metamodel to
generate approximated evaluations of the model response and its summands
in :eq:`Model`.
.. math:: Y \simeq \hat{h} = \sum_{j=0}^{P-1} \alpha_j \Psi_j(x)
Then one may identify the component functions. For instance, for
:math:`u = \{1\}`:
.. math:: h_1(X_1) = \sum_{\alpha | \alpha_1 \neq 0, \alpha_{i \neq 1} = 0} y_{\alpha} \Psi_{\alpha}(\vect{X})
where :math:`\alpha` is a set of degrees associated to the :math:`n_X`
univariate polynomial :math:`\psi_i^{\alpha_i}(X_i)`.
Then the model response :math:`Y` is evaluated using a sample
:math:`X=\{x_k, k=1,\dots,N\}` of the correlated joint distribution.
Finally, the several indices are computed using the model response and
its component functions that have been identified on the polynomial
chaos.
.. topic:: API:
- See :class:`~openturns.ANCOVA`
.. topic:: Examples:
- See :doc:`/examples/reliability_sensitivity/sensitivity_ancova`
.. topic:: References:
- [caniou2012]_