Nonlinear error propagation law (Q1363452)

Let the function \(f:\mathbf R^{n}\longrightarrow \mathbf R\) be real analytic in a domain \(\mathcal R\subset \mathbf R^{n}\); the value of \(f\) at an unknown point \(\mu \in \mathcal R\) is to be determined. Let \(\hat \mu \) be an estimator of \(\mu \) such that \(E\hat \mu =\mu \). Set \( \xi =\frac {\partial f}{\partial x}(\hat \mu )\), \(\Sigma = E\left ((\hat \mu -\mu )(\hat \mu -\mu )'\right )\), then \(\tilde V= \xi '\Sigma \xi \) is an estimator of \(\operatorname {Var}[f(\hat \mu )]\). Under various assumptions on \(f\) and the law of \(\hat \mu \) the author derives several explicit formulae for bias and variance of the estimator \(\tilde V\). In particular, the case of polynomial \(f\) and \(\hat \mu \) normal is considered.