@yuggib I am having a really hard time with heat kernels. I'll rant... $\phi_i$ are the eigenfunctions of $\Delta$ and $\mu_i$ the eigenvalues, so $$H(x,y,t)=\sum_{i\ge 1}e^{-\mu_it}\phi_i(x)\phi_i(y).$$ I have uniform convergence on $M\times M\times[a,\infty)$ for any $a>0$. I also have that $$\int \left|\sum_{i=1}^ke^{-\mu_it}\phi_i(x)\nabla\phi_i(y)\right|^2dy$$ converges as $k\to\infty$ uniformly in $(x,t)$. Supposedly we then have that the finite sums converge to $H(x,y,t)$ weakly in
$W^{1,2}$, I think they mean in the $y$ variable, for any $(x,t)$.
It doesn't seem obvious to me.
Maybe one should use Theorem 3 on page 121 of Yosida. Clearly the series is bounded in $W^{1,2}$ because it converges in the $L^2$ part of the norm and the derivative part has a bounded norm.
One can take $C^\infty$ to be a strongly dense subset of the dual. So take $\psi\in C^\infty$, then $$\int\sum_{i=1}^k e^{-\mu_i t}\phi_i(x)\nabla\phi_i(y)\cdot \nabla\psi(y)\,dy.$$
But now one can integrate by parts, noting that the $\phi_i$ are Dirichlet eigenfunctions.
Then one gets something like $$\int \sum_{i\ge 1}e^{-\mu_it}\phi_i(x)\phi_i(y) \Delta \psi(y)\,dy$$ in the limit.
But it's not clear the other term is in $W^{1,2}$, so can one even integrate by parts?
One might be able to use Rellich to show that the kernel is in $W^{1,2}$, but I'm not sure.
@Avantgarde It's a good song
Right, I can use Rellich to get weak convergence of a subsequence in $W^{1,2}$ and strong convergence in $L^2$. But we know the $L^2$ limit, so its weak $W^{1,2}$ limit is $H(x,y,t)$. So it's in $W^{1,2}$ and I can integrate by parts.
It does equal zero on the boundary.