I asked the following question a few days ago: Approximation of a $C^{\infty}_c$ function by tensor products However, I then realised that I actually need a stronger result in my proof.

As in the previous question, I suppose that $f \in C^{\infty}_c ( \mathbb{R}^2 )$, i.e. $f$ is a $C^{\infty}$ function with compact support defined on $\mathbb{R}^2$. We know that $f$ can be approximated by a sequence of functions $(f_n)$ in $C^{\infty}_c ( \mathbb{R}) \otimes C^{\infty}_c ( \mathbb{R})$ in the infinity norm, where each $f_n$ is of the form

$$ f_n( x,y) = \sum_{i=1}^{k_n} a^{(n)}_i (x) b^{(n)}_i (y), $$

for some functions $a^{(n)}_i, b^{(n)}_i \in C^{\infty}_c ( \mathbb{R})$, $1 \leq i \leq k_n$.

$\textbf{However, I am interested in fixing the tensor ranks}$ $\{k_n \}_{n \in \mathbb{N}}$ $\textbf{to be a constant}$ $K$, $\textbf{i.e.}$ $k_n = K$, $\textbf{for each}$ $n$.

Therefore, I first wonder whether it is possible to construct such a sequence of functions $(f_n)_{n \in \mathbb{N}}$ which approximates $f$ in the infinity norm AND satisfies this property of constant tensor rank $K$.

If yes, can we further require $(f_n)$ to satisfy the requirement of uniform bound on the $C^1$ norms of the component functions of $x$ and $y$, as in the previous question, i.e. $$ \sup_{n \in \mathbb{N}, 1 \leq i \leq K} \big\| a^{(n)}_i \big\|_{\infty} < + \infty, \quad \quad \sup_{n \in \mathbb{N}, 1 \leq i \leq K} \big\| b^{(n)}_i \big\|_{\infty} < + \infty, $$ and $$ \sup_{n \in \mathbb{N}, 1 \leq i \leq K} \big\| \big( a^{(n)}_i \big)' \big\|_{\infty} < + \infty, \quad \quad \sup_{n \in \mathbb{N}, 1 \leq i \leq K} \big\| \big( b^{(n)}_i \big)' \big\|_{\infty} < + \infty?$$

I actually found out that there is a wide range of literature on the topic of approximation by tensor products with constant tensor ranks, e.g.




In particular, the second link (the paper by Temlyakov) gives $L^p$ estimates of the error incurred by approximation with tensor products with constant tensor rank. However, I can't find any paper that addresses the aforementioned issue of a uniform bound on the $C^1$ norms of the component functions. Since I am not in the area of approximation theory, any help/references would be greatly appreciated!

  • 1
    $\begingroup$ Not sure if I am mistaken, but if you evaluate $ f$ on a matrix of $K+1$ by $K+1$ points and get nonzero determinant, then approximation by rank $K$ tensor is impossible, even in the sup norm, as all rank $K$ tensors evaluated on such a matrix of points will give a matrix of rank(!) at most K. $\endgroup$
    – Fan Zheng
    Nov 29 '16 at 2:52

Fixing the rank of the approximating tensors yields closed subsets: More precisely: Given two Banach spaces $X$ and $Y$ and $n\in\mathbb N$ the set $\otimes_n(X,Y)=\lbrace \sum_{i=1}^n x_i\otimes y_i: x_i\in X, y_i\in Y\rbrace$ is closed in $X\tilde{\otimes}_\varepsilon Y$. (This is exercise 12.7 in the book Tensor Products and Operator Ideals of Defant and Floret, they refer to an article of Valdivia A class of locally convex spaces without C-webs, Ann. Inst. Fourier 32 (1982), 261-269). Since the $\varepsilon$-tensor product for spaces of continuous functions corresponds to uniform approximation this should give a negative answer to your question.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.