Inverse of the derivative of a function evaluated at tensorial value

65 Views Asked by At

If the value $\lambda$ is a function of a tensor $\textbf{C}=[C_{ij}]$,

$$\lambda=\tilde{\lambda}(\textbf{C})$$

$$\textbf{C}=\tilde{\textbf{C}}(\lambda)$$

then let us define $\textbf{D}=[D_{ij}]$ such

$$D_{ij}=\dfrac{\partial\tilde{\lambda}}{C_{ij}}$$

and also $\textbf{E}$ such

$$\textbf{E}=\dfrac{\partial \tilde{\textbf{C}}}{\partial \lambda}$$

then how valid is it to say

$$\textbf{D}\textbf{E}=\textbf{I}$$

where $\textbf{I}$ is the identity matrix. If it is proved I appreciate any support.

All values are in $\mathbb R$.

1

There are 1 best solutions below

0
On

Assume that somehow, we have calculated the gradients $\frac{\partial\lambda}{\partial C}$ and $\frac{\partial C}{\partial\lambda}$

Then consider the differentials of each variable $$\eqalign{ d\lambda &= \frac{\partial\lambda}{\partial C} : dC \cr dC &= \frac{\partial C}{\partial\lambda} \, d\lambda \cr }$$ where a colon represents the trace/Frobenius product, i.e. $\,\,A:B={\rm tr}(A^TB)$

Now substitute the first differential into the second, and vice versa $$\eqalign{ d\lambda &= \frac{\partial\lambda}{\partial C} :\Bigg(\frac{\partial C}{\partial\lambda} \, d\lambda\Bigg) = \Bigg(\frac{\partial\lambda}{\partial C} :\frac{\partial C}{\partial\lambda}\Bigg) \, d\lambda \cr dC &= \frac{\partial C}{\partial\lambda} \, \Bigg(\frac{\partial\lambda}{\partial C} : dC\Bigg) = \Bigg(\frac{\partial C}{\partial\lambda} \, \frac{\partial\lambda}{\partial C}\Bigg) : dC \cr }$$ The terms in parenthesis on the RHS must be the identity elements for the respective product operations. $$\eqalign{ \frac{\partial\lambda}{\partial C} :\frac{\partial C}{\partial\lambda} &= \frac{\partial\lambda}{\partial\lambda} = 1 \cr \frac{\partial C}{\partial\lambda} \, \frac{\partial\lambda}{\partial C} &= \frac{\partial C}{\partial C} = {\mathcal E} \cr }$$ where the 4th order tensor ${\mathcal E}$ has components $${\mathcal E}_{ijkl} = \delta_{ik}\,\delta_{jl}$$ There is no combination of these gradients which yields the identity matrix.

However, if we were considering a vector $v$ instead of a matrix $C$, then the identity matrix can be obtained as $$\eqalign{ \frac{\partial v}{\partial\lambda} \, \frac{\partial\lambda}{\partial v} &= \frac{\partial v}{\partial v} = I \cr }$$