I am following the paper on xgboost and got stuck at equation 6 ($3^{rd}$ page), where the authors say that for a given tree, "we can compute the optimal weight $w_j^*$ of leaf $j$ by" \begin{split}w_j^\ast = -\frac{G_j}{H_j+\lambda}\end{split} Here, $G_j$ and $H_j$ are the first and second derivatives of a loss function $l$ of the previous ($t-1$) prediction for the set of examples ($I_j$) ending in the given leaf $j$: \begin{split}G_i &= \sum_{i \in I_j}\partial_{\hat{y}_i^{(t-1)}} l(y_i, \hat{y}_i^{(t-1)})\\ h_i &= \sum_{i \in I_j}\partial_{\hat{y}_i^{(t-1)}}^2 l(y_i, \hat{y}_i^{(t-1)}) \end{split}
My question really is, what makes the weight defined in such a way optimal? Does it follow in some obvious way (since it is not explained in the paper) from the general definition of the loss function? \begin{split} \mathcal{L}^{(t)} = \sum_{j=1}^T [G_jw_j + \frac{1}{2}(H_j + \lambda)w_j^2] + \gamma T \end{split} Particularly, I am struggling with the $H_j$ in the denominator. The general gradient boosting algorithm is supposed to fit new functions to the gradient of the loss function, but the second derivative confuses me).
