Let $\mathcal{R} : \mathbb{R} \times \mathbb{R} \to \mathbb{R}$ denote the reproducing kernel of a Hilbert space $\mathcal{H}$ of functions on $[0,1]$ endowed with inner product $ \langle \cdot, \cdot \rangle_{\mathcal{H}}$. For my particular setting I am able to show that $$ \sup_{x \in [0,1]} ||\mathcal{R}(x, \cdot)||_{\mathcal{H}} \leq C,$$ for some $C>0$. Now let us imagine that we have a set of $n$ values $t_j$ in $[0,1]$, from which we evaluate $\mathcal{R}(t_i, t_j)$ and form the corresponding matrix $\mathbf{R}$.

What I am wondering is, if we can also establish a similar bound for the largest eigenvalue of this positive-semidefinite matrix. That is, if for a vector $\mathbf{a} \in \mathbb {R}^n$ we can have $$ \sum_j^n \sum_i^n a_i a_j \mathcal{R}(t_i, t_j) \leq B C \sum_{j=1}^n a_j^2,$$ for another constant $B$.

All help is greatly appreciated, thank you.


Using a property of reproducing kernels, and convexity of the squared norm:

\begin{aligned} \sum_i^n \sum_j^n a_i a_j \mathcal{R}(t_i, t_j) &= \sum_i^n \sum_j^n a_i a_j <\mathcal{R}(t_i, \cdot), \mathcal{R}(t_j, \cdot)>_\mathcal{H} \\ &= <\sum_i^n a_i \mathcal{R}(t_i, \cdot), \sum_j^n a_j \mathcal{R}(t_j, \cdot)>_\mathcal{H}\\ &= \left\Vert a \right\Vert_1^2 \left\Vert \sum_i^n \frac{a_i}{\left\Vert a \right\Vert_1} \mathcal{R}(t_i, \cdot) \right\Vert^2_\mathcal{H}\\ &\leq \left\Vert a \right\Vert_1^2 \sum_i^n \frac{a_i}{\left\Vert a \right\Vert_1} \left\Vert \mathcal{R}(t_i, \cdot) \right\Vert^2_\mathcal{H} \\ & \leq \left\Vert a \right\Vert_1^2 C^2 \\ & \leq \frac{\left\Vert a \right\Vert_2^2}{(\min_{\Vert x \Vert_1=1} \Vert x \Vert_2)^2} C^2 \end{aligned}

  • $\begingroup$ How exactly do you get the last inequality? I think it is wrong. $\endgroup$ – JohnK Oct 9 '20 at 10:55
  • $\begingroup$ By subadditivity property of the norm, and the scalability. See points 1 and 2 in the definition here: wiki. $\endgroup$ – karamel-kitty Oct 9 '20 at 13:09
  • $\begingroup$ But this is a square-norm so this is not correct. $\endgroup$ – JohnK Oct 9 '20 at 13:26
  • $\begingroup$ I have updated my answer. It's not the 2-norm as you wanted but the 1-norm that appears naturally. $\endgroup$ – karamel-kitty Oct 11 '20 at 7:16
  • $\begingroup$ While this fixes the previous mistake, this is not what the question is about. $\endgroup$ – JohnK Oct 11 '20 at 8:58

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.