We work in $\mathbb R^n$ and denote $r$ the distance to the origin. Define the smooth approximations of the identity $$\phi_\delta = \frac1{\delta^n \int_{\mathbb R^n}\exp\left( \frac{-1}{1-r^2} \right)dx} \cdot \begin{cases} \exp \left( \frac{-1}{1-(r/\delta)^2} \right)dx &: r < \delta \\ 0 &: r \geq \delta \end{cases} $$ so that $\phi_\delta * f \to f$ as $\delta \to 0$, locally uniformly for all smooth $f$. (The factor is just to normalize the $\phi_\delta$.)
Let $D$ be a differential operator on $\mathbb R^n$, say a monomial in the $\partial/\partial x_i$.
What is a useful sufficient condition under which $(D\phi_\delta) * f \to Df$ ? Is there a better choice of the $\phi_\delta$ for which this holds? (It can depend on $f$.)
Fruitless ideas: (1) I'm unable to do something with the usual criterion for converges of derivatives because the $D\phi_\delta$ certainly don't converge uniformly. (2) Power series behave well under differentiation, but there are no power series here.
I'm actually interested in the more general case of convolution on a Riemannian symmetric space $S$, where the approximate identities are functions of the geodesic distance: $$\int_S \phi_\delta(d(y,x))f(x)dx$$ This here is a special case. I hope it will help to do the general case.