2
$\begingroup$

As it is know local sensitivity analysis attempt to quantify the local impact of input factors on the model, through partial derivatives: a derivative of the outputs accordingly to the inputs, when the last ones are slightly perturbed. Many methods are used to estimates this derivatives: Direct method, Green Function Method, etc.

However, the input-output relationship is generally assumed to be linear (from Saltelli textbooks on the subject). Why it is assumed to be linear in a local neighborhood, even the whole system may not be linear. I guess it may have a relation with Taylor expansion expression. I am trying to understand the real reason behind this assumption. Thank you.

Improve this question
$\endgroup$
6
  • 3
    $\begingroup$ There's no need to invoke Taylor: "locally linear" means the same thing as differentiable. As soon as you frame the sensitivity analysis in terms of partial derivatives you have, a fortiori, assumed local linearity. $\endgroup$
    – whuber ♦
    Dec 5, 2012 at 17:27
  • $\begingroup$ Okay thanks a lot. I guess this answer to the question. However, the local sensitivity analysis is reduced only to analysis via derivates, I am not aware of other techniques. Is it correct therefore to assume this linearity, knowing that local analysis is carried on systems which may not be linear. Thus, the shape of output function can be completly unknown (even the continuity is not sure), as some hard optimization problems. $\endgroup$
    – omar
    Dec 5, 2012 at 18:17
  • 3
    $\begingroup$ If the system is not even continuous, you will have a hard time computing partial derivatives! You might want to check whether people aren't actually computing finite differences--which work regardless of differentiability or continuity--and are perhaps only calling them derivatives. $\endgroup$
    – whuber ♦
    Dec 5, 2012 at 18:18
  • $\begingroup$ Yes, Finite-Differences are used to approximate these derivatives, but I think they are slower and less accurate. $\endgroup$
    – omar
    Dec 5, 2012 at 18:39
  • 1
    $\begingroup$ Your last two comments are confusing: the penultimate one assumes the system is differentiable while the ultimate one asks about non-differentiable methods, even though the penultimate complains that these are "slower and less accurate." It would be nice if in your question you would clarify what assumptions you want to make and precisely what kinds of alternatives you are looking for. $\endgroup$
    – whuber ♦
    Dec 5, 2012 at 21:05

1 Answer 1

Reset to default
0
$\begingroup$

A local sensitivity analysis method makes the assumption that a model is linear. There are alternative techniques which do not make this assumption.

The Method of Morris, as implemented in the open-source Python package SALib, computes multiple finite differences at many points in the model input-space. These are then averaged to give a more complete (but still crude) understanding of the sensitivity of the model outputs to inputs.

More comprehensive variance-based methods, such as those proposed by Saltelli and others (also implemented in the Python package mentioned earlier) are more computationally demanding, but give even more information about the sensitivity of a model to its inputs, including interactions between pairs of inputs.

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.