When normalized=FALSE this function computes the generalized forecast error variance decomposition proposed by Pesaran and Shin (1998) which takes the form:
$$
\alpha _{ij}^{g}(h) = \frac{\sigma_{ii}^{-1}\sum_{l=0}^{h-1}(\mathbf{e'}_{i}\Theta _{l}\Sigma_{\varepsilon}\mathbf{e}_{j})^{2}}{\sum_{l=0}^{h-1}(\mathbf{e'}_{i}\Theta _{l}\Sigma _{\varepsilon }\Theta'_{l}\mathbf{e}_{i})}, \quad i,j = 0,1,2\ldots, K
$$
Where \(\mathbf{\Theta}_{l}\), are the coefficients matrix of the MA representation of the VAR model, \(\mathbf{\Sigma}_{\varepsilon}\)
is the variance matrix of the reduced-form error vector \(\varepsilon\), \(\sigma_{ii}\) is the standard deviation of the error term for the
\(ith\) equation and \(e_{i}\) and \(e_{j}\) are selection vectors with ones as the ith element and zeros elsewhere.
If normalized=TRUE (the default value) then g.fevd computes:
$$
\tilde{a}_{ij}^{g}(h) = \frac{a_{ij}^{g}(h)} {\sum_{j=1}^{K} a_{ij}^{g}(h)}
$$
This fact implies the normalization is simply each entry of the generalized fevd divided by the its corresponding row sum.