Let $W^{2,2}(\Omega_i)$, $\Omega_i = [-1,1]$, $i = 1,\ldots,d$ be the Sobolev spaces of twice weakly differentiable, square integrable functions. Let further $\otimes_a$ denote the algebraic tensor product (in other words, the non-topological one) and $\Omega = \Omega_1 \times \ldots \times \Omega_d = [-1,1]^d$.

All functions in the space $\mathcal{W} := W^{2,2}(\Omega_1) \otimes_a \ldots \otimes_a W^{2,2}(\Omega_d)$ are continuous, i.e. $\mathcal{W} \subset C^0(\Omega)$. The induced norm $\|\cdot\|_{\mathcal{W}}$ is stronger than the norm $\|\cdot\|_{2,2}$ on $W^{2,2}(\Omega)$.

  1. Is $\mathcal{W}$ dense in $W^{2,2}(\Omega)$ (with respect to the norm $\|\cdot\|_{2,2}$). Is hence $\overline{\mathcal{W}}^{\|\cdot\|_{2,2}} = W^{2,2}(\Omega)$?

  2. Is the semi norm $|\cdot|_{2,2}$, defined via $|u|_{2,2}^2 := \sum_{|\alpha| = 2} \|D^{\alpha} u\|_{L^2}^2$, equivalent to $\|\cdot\|_{2,2}$ on the quotient space of functions that only differ by a constant function?

  3. Let $r \in \mathbb{N}$, $$\mathcal{W}_r := \bigcap_{\mu = 1}^{d-1} \{ \phi \in \mathcal{W} \mid \exists g_i \in W^{2,2}(\Omega_1) \otimes_a \ldots \otimes_a W^{2,2}(\Omega_\mu), \exists h_i \in W^{2,2}(\Omega_{\mu+1}) \otimes_a \ldots \otimes_a W^{2,2}(\Omega_d): \phi(x) = \sum_{i = 1}^r g_i(x_1,\ldots,x_\mu) \cdot h_i(x_{\mu+1},\ldots,x_d), \forall x \in \Omega \}$$ the space of (tensor train) rank $r$ functions in $\mathcal{W}$. Then this space $\mathcal{W}_r$ is closed under $\|\cdot\|_{2,2}$. Let further $p_i \in \Omega, \ i = 1,\ldots,m$, for some $m \in \mathbb{N}$, and $y \in \mathbb{R}^m$. Is the problem $$u^\ast = \mathrm{argmin}_{u \in \mathcal{W}_r} \sum_{i = 1}^m (u(p_i) - y_i)^2 + |u|_{2,2}^2$$ hence well defined (based on the this propert of $\mathcal{W}_r$, and does there exist such a minimizer (regardless of uniqueness)?

I strongly believe that all three assertion hold true, but I am unaware of easy ways to prove such. As this is a bit out of topic for me, I also have a bit of trouble judging what counts as rigorous argument. I have seen variants of these questions, but could not find a specific enough answer.

I have found a remark about the validity of $1.$ in a tensor calculus book, but no proof is provided. Point $2.$ should be provided by the Poincaré inequality, although it needs to be applied on two levels, and the devil is a bit in the detail. The first part of item $3.$ requires knowledge about tensor formats, so it might be hard to answer, but just under the assumption that $\mathcal{W}_r$ is closed, the second part is still of great interest to me. I of course appreciate any help regarding only any one of the three questions.



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.