This might be a simple question but what is $\frac{\partial g_{\mu\nu}}{\partial g^{\sigma\rho}}$, with proof please if possible.
3 Answers
Hint $$\frac{\partial}{\partial g^{\sigma\rho}}(g_{\mu\nu}g^{\nu\lambda})=\frac{\partial}{\partial g^{\sigma\rho}}(\ \delta_\mu^\lambda)=0$$ You can use the product rule on the first term to get $$\frac{\partial g_{\mu\nu}}{\partial g^{\sigma\rho}} g^{\nu\lambda}+ g_{\mu\nu} \frac{\partial g^{\nu\lambda}}{\partial g^{\sigma\rho}}=0$$ For a general tensor we have $$\frac{\partial g^{\nu\lambda}}{\partial g^{\sigma\rho}}=\delta^\nu_\sigma\delta^\lambda_\rho$$ but since the metric tensor is symmetric the individual components are not independent.For example $$\frac{\partial g^{31}}{\partial g^{13}}=\frac{\partial g^{13}}{\partial g^{13}}=1$$ Accounting for this symmetry we get $$\frac{\partial g^{\nu\lambda}}{\partial g^{\sigma\rho}}=\tfrac 1 2(\delta^\nu_\sigma\delta^\lambda_\rho+\delta^\nu_\rho\delta^\lambda_\sigma).$$ To see where this factor one half comes from go to this answer.
Finally we get
$$\frac{\partial g_{\mu\nu}}{\partial g^{\sigma\rho}}=-g_{\mu\rho}g_{\sigma\nu}$$ for general tensors and $$\frac{\partial g_{\mu\nu}}{\partial g^{\sigma\rho}}=-\tfrac 1 2(g_{\mu\rho}g_{\sigma\nu}+g_{\mu\sigma}g_{\rho\nu})$$ for the metric tensor

- 17,187
-
$ - \frac{1}{2} ( g_{\mu\sigma} g_{\nu\rho} + g_{\nu\sigma} g_{\mu\rho})$ ? – ODE Apr 12 '20 at 09:58
-
I added more to the hint – AccidentalTaylorExpansion Apr 12 '20 at 19:05
-
$-g_{\mu\rho}g_{\nu\sigma}$ ? can you please give me the answer ? – ODE Apr 13 '20 at 10:46
-
That's correct (using what I said). I forgot that the metric is symmetric so I added more. I also added the solution explicitly for later reference. – AccidentalTaylorExpansion Apr 13 '20 at 13:50
-
1As pointed out by @ODE, your expression for the derivative is wrong, since for example it leads to $\partial g^{00}/\partial g^{00}=2$. – Giorgio Comitini Apr 13 '20 at 13:52
-
@GiorgioComitini I forgot about the diagonals, I'll update it and hope I haven't made any further misstakes. – AccidentalTaylorExpansion Apr 13 '20 at 13:55
For fun, here's how to do it in a coordinate-free manner :
The inverse metric is the matrix inverse of the metric, so that, if we vary our metric, we can use the binomial identity :
\begin{equation} (g + h)^{-1} = g^{-1} - g^{-1} (I + hg^{-1})^{-1} h g^{-1} \end{equation}
In our case :
\begin{eqnarray} \frac{\delta_h g^{-1}}{\delta g} &=& \lim_{\varepsilon \to 0} \frac{(g + \varepsilon h)^{-1} - g^{-1}}{\varepsilon} \\ &=& - \lim_{\varepsilon \to 0} \frac{g^{-1} (I + \varepsilon hg^{-1})^{-1} \varepsilon h g^{-1}}{\varepsilon} \\ &=& - \lim_{\varepsilon \to 0} \frac{\varepsilon g^{-1} h g^{-1} + \mathcal{O}(\varepsilon^2)}{\varepsilon} \\ &=& -g^{-1} h g^{-1} \end{eqnarray}
Therefore, the variation in the direction of a perturbation $h$ is $g^{-1} h g^{-1}$ (in more physicky notation, $\delta g^{-1} = g^{-1} \delta g g^{-1}$). If we pick a specific direction such as one of the component, this means
\begin{equation} \delta g^{ab} = -g^{ac} \delta g_{cd} g^{db} \end{equation}
This is pretty much what is expected. As the inverse of the inverse metric is the metric, the relation holds the other way around, as well.

- 16,329
-
Maybe you should say "indices-free" rather than "coordinate-free". A matrix representation for $g$ requires a choice of coordinates. – Giorgio Comitini May 26 '20 at 09:45
-
Uh and you missed two minuses in the last equation and in the last row of the second equation. – Giorgio Comitini May 26 '20 at 09:47
For a metric tensor the answer should be
$$\frac{\partial g_{\mu\nu}}{\partial g^{\sigma\rho}}=-\frac{n(\sigma\rho)}{2}\left(g_{\mu\sigma}g_{\nu\rho}+g_{\nu\sigma}g_{\mu\rho}\right),$$
where $n(\sigma\rho)=2-\delta_{\sigma\rho}$ and $\delta_{\sigma\rho}$ is the Kronecker delta, i.e. the function is 1 if $\sigma=\rho$ and 2 if $\sigma\neq\rho$.