|
sigma
1.0.0
|
This page covers the theory behind Sigma's Uncertain class. For the set-valued representations, which enclose a result rather than estimating its distribution, see Interval Arithmetic, Affine Arithmetic, and Taylor Models.
Like the uncertainties and Measurements.jl libraries, Sigma is based on the theory of linear error propagation. Here we will provide a brief summary of that theory.
Assume a sequence of \( n \) variables \( \left(a_{i}\right)_{i=1}^{n} \) whose elements are defined as \( a_{i} = \bar{a}_{i} \pm \sigma_{a_{i}} \), where \( \bar{a}_{i} \) is the mean value of \( a_{i} \) and \( \sigma_{a_{i}} \) is the standard deviation of \( a_{i} \). \( F(A) \) is a function of \( A \), where \( A = \{a_{i} : i \in \mathbb{N}, 1 \leq i \leq n\} \). The uncertainty of \( F(A) \) can be determined as
\[ \normalsize \sigma_{F} = \sqrt{ \sum_{i=1}^{n} \left( \left( \left. \frac{\partial F}{\partial a_{i}} \right|_{a_{i}=\bar{a}_{i}} \sigma_{a_{i}} \right)^2 + 2 \sum_{j=i+1}^{n} \left( \left(\frac{\partial F}{\partial a_{i}}\right)_{a_{i}=\bar{a}_{i}} \left(\frac{\partial F}{\partial a_{j}}\right)_{a_{j}=\bar{a}_{j}} \sigma_{a_{i}a_{j}} \right) \right) } \]
where \( \sigma_{a_{i}a_{j}} \) is the covariance of \( a_{i} \) and \( a_{j} \). These covariances can be eliminated from the above equation if the uncertainties of the variables are independent from one another, which is a requirement imposed here. As such, the uncertainty of \( F(A) \) when the members of \( A \) are independent from one another is
\[ \normalsize \sigma_{F} = \sqrt{ \sum_{i=1}^{n} \left( \left. \frac{\partial F}{\partial a_{i}} \right|_{a_{i}=\bar{a}_{i}} \sigma_{a_{i}} \right)^2 } \]
Next, we consider a set \( B = \{x, y\} \) where \( x = a_{i} + a_{j} \) and \( y = a_{j} - a_{k} \), i.e. the elements of \( B \) are dependent on a subset of the independent variables in set \( A \). Given the function \( G(B) \), the uncertainty of \( G \) can be determined by application of the chain rule to relate the independent variables to \( G \) through their relationships with the dependent variables
\[ \normalsize \sigma_{G} = \sqrt{ \left( \left( \frac{\partial G}{\partial x} \frac{\partial x}{\partial a_{i}} \right)_{a_{i}=\bar{a}_{i}} \sigma_{a_{i}} \right)^2 + \left( \left( \frac{\partial G}{\partial x} \frac{\partial x}{\partial a_{j}} + \frac{\partial G}{\partial y} \frac{\partial y}{\partial a_{j}} \right)_{a_{j}=\bar{a}_{j}} \sigma_{a_{j}} \right)^2 + \left( \left( \frac{\partial G}{\partial y} \frac{\partial y}{\partial a_{k}} \right)_{a_{k}=\bar{a}_{k}} \sigma_{a_{k}} \right)^2 } \]