4
$\begingroup$

Suppose $H$ is the reproducing kernel Hilbert space on a space $X$ with reproducing kernel $K$. If, say, $K - c$ is a positive definite kernel for some $c>0$ then $H$ contains the constant functions on $X$. But this condition is not necessarily easy to verify.

Are there other known sufficient conditions that imply that a reproducing kernel Hilbert space on a space $X$ contains the constant functions? In particular I'm interested in the case in which $X$ is a compact metric space and $K(x,y) = e^{-d(x,y)}$. (This kernel is not necessarily positive definite; I'm interested in those $X$ for which it is.) Certain special cases, like subsets of Euclidean space or spheres, I can handle with Fourier analysis via the above condition, but the more general case is unclear.

$\endgroup$
4
  • $\begingroup$ Your edits appeared after I posted my answer, but I don't think you want $\delta_{xy}$ in the condition you give, rather you subtract off the kernel that is identically $1$ on $X\times X$. Also, for most of the kernels I'm familiar with this condition is not so hard to check; do you have a particular kernel that you're looking at? $\endgroup$
    – Mike Jury
    Jun 4, 2013 at 15:14
  • $\begingroup$ Mike: You're right, I was sloppy about which basis I was thinking about. I'll also add in the specific kernel I'm interested in. $\endgroup$ Jun 4, 2013 at 15:17
  • $\begingroup$ There is another condition in terms of the geometry of the embedding $X\to H$ via $e:x\to k_x$ (assuming the normalization $k(x,x)\equiv 1$), though I'm not sure it's any easier to check: The space $H$ contains the constant functions if and only if the image of $X$ under $e$ lies in an affine hyperplane of $H$. (If $w\in H$ is constant on $X$, then the $k_x$'s lie in an affine hyperplane orthogonal to $w$, and conversely.) Though maybe this is the conclusion you want rather than a condition you want to check... $\endgroup$
    – Mike Jury
    Jun 4, 2013 at 16:59
  • $\begingroup$ Mike: Your guess is correct --- this is actually the conclusion I want. $\endgroup$ Jun 4, 2013 at 18:27

1 Answer 1

5
$\begingroup$

Yes. A (fairly standard) fact about RKHS's is that a function $f:X\to \mathbb C$ belongs to the RKHS (with kernel $k$) if and only if there is a $c>0$ such that the Hermitian kernel $$ k(x,y) -c^2f(x)\overline{f(y)} $$ is positive semidefinite; the norm of $f$ is then equal to $1/c$ for the best possible $c$. Thus the constants belong to the RKHS if and only if $$ k(x,y)-c^2 $$ is positive semidefinite for some positive $c$.

$\endgroup$
4
  • $\begingroup$ Thanks for that. As you noted, I added an edit while you were writing your answer to point out that I know this condition, but I wasn't positive that it was necessary as well as sufficient. $\endgroup$ Jun 4, 2013 at 15:25
  • $\begingroup$ Do you know if this fact applies to real valued RKHSs too? Also, do you know where I might find a reference for this? $\endgroup$
    – user27182
    Apr 25, 2021 at 12:15
  • $\begingroup$ Do you have a reference for this? $\endgroup$
    – user27182
    Feb 24, 2023 at 11:58
  • $\begingroup$ Theorem 4.15 here: math.uh.edu/~vern/rkhs.pdf $\endgroup$
    – user27182
    Mar 8, 2023 at 22:42

Your Answer

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

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