Second order partial derivative for multivariate function, derivation

Consider a function:

\[f\left(x_{1}\left(\lambda_{1},\lambda_{2},..,\lambda_{M}\right),x_{2}\left(\lambda_{1},\lambda_{2},..,\lambda_{M}\right),..,x_{N}\left(\lambda_{1},\lambda_{2},..,\lambda_{M}\right)\right)\]

The first partial derivative is given as:

\[\frac{\partial}{\partial\lambda_{i}}f\left(x_{1},x_{2},..,x_{N}\right)=\frac{\partial f}{\partial x_{1}}\frac{\partial x_{1}}{\partial\lambda_{i}}+\frac{\partial f}{\partial x_{2}}\frac{\partial x_{2}}{\partial\lambda_{i}}+...+\frac{\partial f}{\partial x_{N}}\frac{\partial x_{N}}{\partial\lambda_{i}}\]

The above equation can be formulated as:

\[\frac{\partial}{\partial\lambda_{i}}f\left(x_{1},x_{2},..,x_{N}\right)=\sum_{k}^{N}\frac{\partial f}{\partial x_{k}}\frac{\partial x_{k}}{\partial\lambda_{i}}\]

Now consider the second derivative:

\[\frac{\partial^{2}}{\partial\lambda_{j}\partial\lambda_{i}}f\left(x_{1},x_{2},..,x_{N}\right)=\frac{\partial}{\partial\lambda_{j}}\left(\frac{\partial f}{\partial x_{1}}\frac{\partial x_{1}}{\partial\lambda_{i}}+\frac{\partial f}{\partial x_{2}}\frac{\partial x_{2}}{\partial\lambda_{i}}+...+\frac{\partial f}{\partial x_{N}}\frac{\partial x_{N}}{\partial\lambda_{i}}\right)\]

This is equal to:

\[\frac{\partial^{2}}{\partial\lambda_{j}\partial\lambda_{i}}f\left(x_{1},x_{2},..,x_{N}\right)=\frac{\partial}{\partial\lambda_{j}}\left(\frac{\partial f}{\partial x_{1}}\frac{\partial x_{1}}{\partial\lambda_{i}}\right)+\frac{\partial}{\partial\lambda_{j}}\left(\frac{\partial f}{\partial x_{2}}\frac{\partial x_{2}}{\partial\lambda_{i}}\right)+...+\frac{\partial}{\partial\lambda_{j}}\left(\frac{\partial f}{\partial x_{N}}\frac{\partial x_{N}}{\partial\lambda_{i}}\right)\]

Since all of the terms should be treated the same way, lets just focus on the first term. Lets define \(f_{x_{1}}=\frac{\partial f}{\partial x_{1}}\) and \(x_{1,\lambda_{i}}=\frac{\partial x_{1}}{\partial\lambda_{i}}\):

\[\frac{\partial}{\partial\lambda_{j}}\left(\frac{\partial f}{\partial x_{1}}\frac{\partial x_{1}}{\partial\lambda_{i}}\right)=\frac{\partial}{\partial\lambda_{j}}\left(f_{x_{1}}x_{1,\lambda_{i}}\right)\]

Now by using the product rule:

\[\frac{\partial}{\partial\lambda_{j}}\left(\frac{\partial f}{\partial x_{1}}\frac{\partial x_{1}}{\partial\lambda_{i}}\right)=\frac{\partial f_{x_{1}}}{\partial\lambda_{j}}x_{1,\lambda_{i}}+f_{x_{1}}\frac{\partial x_{1,\lambda_{i}}}{\partial\lambda_{j}}\]

It can be seen that the first term is equal to the first order partial derivative:

\[\frac{\partial f_{x_{1}}}{\partial\lambda_{j}}=\sum_{l}^{N}\frac{\partial f_{x_{1}}}{\partial x_{l}}\frac{\partial x_{l}}{\partial\lambda_{j}}\]

Thus:

\[\frac{\partial}{\partial\lambda_{j}}\left(\frac{\partial f}{\partial x_{1}}\frac{\partial x_{1}}{\partial\lambda_{i}}\right)=\sum_{l}^{N}\frac{\partial f_{x_{1}}}{\partial x_{l}}\frac{\partial x_{l}}{\partial\lambda_{j}}x_{1,\lambda_{i}}+f_{x_{1}}\frac{\partial x_{1,\lambda_{i}}}{\partial\lambda_{j}}\]

Now by back-inserting the definitions:

\[\frac{\partial}{\partial\lambda_{j}}\left(\frac{\partial f}{\partial x_{1}}\frac{\partial x_{1}}{\partial\lambda_{i}}\right)=\sum_{l}^{N}\left[\frac{\partial^{2}f}{\partial x_{l}\partial x_{1}}\frac{\partial x_{l}}{\partial\lambda_{j}}\right]\frac{\partial x_{1}}{\partial\lambda_{i}}+\frac{\partial f}{\partial x_{1}}\frac{\partial^{2}x_{1}}{\partial\lambda_{j}\partial\lambda_{i}}\]

Thus, the final equation for the second order partial derivative for a multivariate function:

\[\frac{\partial^{2}}{\partial\lambda_{j}\partial\lambda_{i}}f\left(x_{1},x_{2},..,x_{N}\right)=\sum_{k}^{N}\sum_{l}^{N}\left[\frac{\partial^{2}f}{\partial x_{l}\partial x_{k}}\frac{\partial x_{l}}{\partial\lambda_{j}}\right]\frac{\partial x_{k}}{\partial\lambda_{i}}+\sum_{k}^{N}\frac{\partial f}{\partial x_{k}}\frac{\partial^{2}x_{k}}{\partial\lambda_{j}\partial\lambda_{i}}\] If you enjoyed this post you can donate a coffee , if you like :)