Normalization before sensitivity analysis


I have a very general question: Is normalization necessary before global sensitivity analysis? Let’s say I have a differential equation that takes parameters x_n. But parameters have widely different range of values they can take, let’s say x_1 is a subset of [0,1] whereas x_2 is a subset of [10, 1000]. Is some kind of normalization necessary? And if so, how can one implement this in GlobalSensitivity.jl?


The answer is also very general: it depends on the method. For instance, some methods support categorical variables for which normalization doesn’t make much sense.

I will let the maintainers of GlobalSensitivity.jl discuss the specific methods implemented and their requirements.

Kind of depends on the method, though it’s all already re-scaled to the output variable. It basically calculates the variance explained over a parameter space. The percentage of the variance explained in the output with respect to input x does not really care about the relative size of the space of x, just how much y varies as you change x. Internally, many of the methods re-scale down to a [0,1]^n grid internally anyways.

1 Like