Abstract
Given an uncountable, compact metric space X, we show that there exists no reproducing kernel Hilbert space that contains the space of all continuous functions on X.
Similar content being viewed by others
1 Introduction
Given a compact metric space (X, d), we denote the space of all continuous functions \(f:X\rightarrow \mathbb {R}\) by C(X). If \(k:X\times X\rightarrow \mathbb {R}\) is a continuous reproducing kernel, see e.g. [1] and [18, Chapter 4] for an introduction, then it is well-known that its reproducing kernel Hilbert space (RKHS) H satisfies \(H\subset C(X)\). Such RKHSs have a variety of applications in different fields, and in particular in statistics [4] and machine learning [5, 14, 18] they have become very popular. In these applications, it is often desirable that H is sufficiently large. One way to express this is the notion of universality [12, 17], which assumes that H is dense in C(X), as well as the closely related notion of characteristic kernels [13, 15], which can be used to define distances between probability measures [16]. While universality is sufficient to establish basic learning guarantees for RKHS-base learning algorithms, see e.g. [18, Chapters 6–8], it is often not fully satisfactory in the sense that one would ideally have even larger RKHS to work with, see e.g. [2, Section 2.3] and [9]. This raises the question, whether, for example, there exist RKHSs H, for which the inverse inclusion \( C(X)\subset H\) holds, see also [10] for a discussion of this question in view of the distances between probability measures mentioned above. Here we note that if X is infinite, such an RKHS H cannot have a continuous kernel since otherwise we would have \(H\subset C(X) \subset H\) and the norms of H and C(X) would be equivalent by a double application of the closed graph theorem. In other words, C(X) would be isomorphic to a Hilbert space, which is false, see e.g. [11, Example 1.11.25].
Obviously, this reasoning does not exclude RKHSs H with a discontinuous kernel and \(C(X) \subset H\). However, the following theorem shows that no such RKHSs exist in essentially all situations of practical interest.
Theorem 1.1
Let (X, d) be a compact metric space. If X is uncountable, then there exists no RKHS H on X such that \(C(X) \subset H\).
Finally, recall that so-called reproducing kernel Banach spaces (RKBS) have recently gained attraction, see e.g. [3, 8], partially because they somewhat naturally appear in the analysis of neural networks. In a nutshell, an RKBS E is a Banach space of functions \(X\rightarrow \mathbb {R}\) such that all point evaluation functionals are continuous on E. Obviously, every RKHS is an RKBS and C(X) is an example of an RKBS that is not an RKHS if X is infinite. Consequently, every RKHS can be embedded into an RKBS while Theorem 1.1 shows that the converse is false. In this sense, the class of RKBSs is strictly richer than the class of RKHSs.
2 Proof
Proof of Theorem 1.1
Let us assume that there was an RKHS H on X such that \(C(X) \subset H\). Denoting the kernel of H by k, we define
for all \(n\in \mathbb {N}\). Since \(X= \bigcup _{n\ge 1} X_n\), we then find an \(m\in \mathbb {N}\) such that \(X_{m}\) is uncountable. We write \(X_*:= {\overline{X_{m}}}\) for the closure of \(X_{m}\) in X. Moreover, we define
Then it is well known, see e.g. [7, p. 195], that for every \(f\in C_{\textrm{uc}}(X_{m})\), there is a unique, uniformly continuous \({\hat{f}}:X_*\rightarrow \mathbb {R}\) such that the restriction \({\hat{f}}_{|X_{m}}\) of \({\hat{f}}\) to \(X_{m}\) satisfies \({\hat{f}}_{|X_{m}} = f\). Since \(X_*\) is compact, \({\hat{f}}\) is bounded, and hence f is bounded. Consequently, we can equip \(C_{\textrm{uc}}(X_{m})\) with the supremum-norm \(\Vert \cdot \Vert _\infty \). In addition, each \({\hat{f}}\in C(X_*)\) is just the continuous extension of \(f\in C_{\textrm{uc}}(X_{m})\) from the dense subset \(X_{m}\subset X_*\) to \(X_*\) and using this, we can easily verify that this extension operation gives an isometric and linear operator
Moreover, the compactness of \(X_*\) ensures that every \(f\in C(X_*)\) is uniformly continuous, and hence this extension operator is surjective and thus actually an isometric isomorphism. In particular, \(C_{\textrm{uc}}(X_{m})\) is a Banach space. In addition, [11, Example 1.11.25] shows that \(C(X_*)\) is not reflexive since \(X_*\) is infinite, and consequently, \(C_{\textrm{uc}}(X_{m})\) is not reflexive, either.
Let us now recall from e.g. [1, p. 351] that the RKHS of the restricted kernel \(k_m:=k_{|X_{m}\times X_{m}}\) is given by
Since \(k_m\) is bounded by construction, namely \(k(x,x) \le m\) for all \(x\in X_{m}\), we know that the inclusion map \({{\,\textrm{id}\,}}:H_m\rightarrow \ell _\infty (X_{m})\) is well-defined and continuous with
for all \(h\in H_m\), see e.g. [18, Lemma 4.23], where \(\ell _\infty (X)\) denotes the space of all bounded functions \(X\rightarrow \mathbb {R}\).
Our next goal is to show \(C_{\textrm{uc}}(X_{m})\subset H_m\). To this end, we fix an \(f\in C_{\textrm{uc}}(X_{m})\). Since X is, as a metric space, a normal space, see e.g. [6, Theorem 2.6.1], Tietze’s extension theorem, see e.g. [6, Theorem 2.6.4], then gives a \(g\in C(X)\) with \(g_{|X_*} = {\hat{f}}\). By our initial assumption \(C(X)\subset H\), we then have \(g\in H\), and hence we find the desired
For later use, we note that the just established inclusion \(C_{\textrm{uc}}(X_{m})\subset H_m\) in combination with (2.1) yields
for all \(f\in C_{\textrm{uc}}(X_{m})\).
We now show that the inclusion map \({{\,\textrm{id}\,}}: C_{\textrm{uc}}(X_{m})\rightarrow H_m\) is continuous. To this end, we fix a sequence \((f_n)\subset C_{\textrm{uc}}(X_{m})\) that converges to some \(f\in C_{\textrm{uc}}(X_{m})\) with respect to \(\Vert \cdot \Vert _\infty \) and for which there exists an \(h\in H\) with \(f_n \rightarrow h\) with respect to \(\Vert \cdot \Vert _H\). Now, \(\Vert f_n-f \Vert _\infty \rightarrow 0\) implies \(f_n(x) \rightarrow f(x)\) for all \(x\in X\). In addition, convergence in an RKHS norm implies pointwise convergence by the very definition of RKHSs, and hence we have \(f_n(x) \rightarrow h(x)\) for all \(x\in X\). Consequently, we have \(f=h\), and the closed graph theorem therefore shows that the inclusion map \({{\,\textrm{id}\,}}: C_{\textrm{uc}}(X_{m})\rightarrow H_m\) is continuous. For the constant
we thus have \(\Vert f \Vert _{H_m} \le K \Vert f \Vert _\infty \) for all \(f\in C_{\textrm{uc}}(X_{m})\).
By combining the latter estimate with (2.2), we now conclude that for all \(f\in C_{\textrm{uc}}(X_{m})\), we have
In other words, \( \Vert \cdot \Vert _{H_m}\) is an equivalent norm on \(C_{\textrm{uc}}(X_{m})\), and since \(\Vert \cdot \Vert _\infty \) is complete on \(C_{\textrm{uc}}(X_{m})\), so is \( \Vert \cdot \Vert _{H_m}\). Consequently, \(C_{\textrm{uc}}(X_{m})\) is isomorphic to a Hilbert space and thus reflexive. This contradicts our previous finding that \(C_{\textrm{uc}}(X_{m})\) is not reflexive. \(\square \)
Finally, we like to remark that the following theorem can be shown analogously.
Theorem 2.1
Let X be an uncountable set. Then there exists no RKHS H on X such that \(\ell _\infty (X) \subset H\).
References
Aronszajn, N.: Theory of reproducing kernels. Trans. Amer. Math. Soc. 68, 337–404 (1950)
Bach, F.: Breaking the curse of dimensionality with convex neural networks. J. Mach. Learn. Res. 18, 1–53 (2017)
Bartolucci, F., De Vito, E., Rosasco, L., Vigogna, S.: Understanding neural networks with reproducing kernel Banach spaces. Appl. Comput. Harmon. Anal. 62, 194–236 (2023)
Berlinet, A., Thomas-Agnan, C.: Reproducing Kernel Hilbert Spaces in Probability and Statistics. Kluwer, Boston (2004)
Cucker, F., Zhou, D.X.: Learning Theory: An Approximation Theory Viewpoint. Cambridge University Press, Cambridge (2007)
Dudley, R.M.: Real Analysis and Probability. Cambridge University Press, Cambridge (2002)
Kelley, J.L.: General Topology. D. Van Nostrand, Toronto (1955)
Lin, R.R., Zhang, H.Z., Zhang, J.: On reproducing kernel Banach spaces: generic definitions and unified framework of constructions. Acta Math. Sin. (Engl. Ser.) 38, 1459–1483 (2022)
Liu, F., Shi, L., Huang, X., Yang, J., Suykens, J.A.K.: Generalization properties of hyper-RKHS and its applications. J. Mach. Learn. Res. 22, 140 (2021)
Loomis, S.P., Crutchfield, J.P.: Topology, convergence, and reconstruction of predictive states. Phys. D 445, Paper No. 133621, 15 pp. (2023)
Megginson, R.E.: An Introduction to Banach Space Theory. Springer, New York (1998)
Micchelli, C.A., Xu, Y., Zhang, H.: Universal kernels. J. Mach. Learn. Res. 7, 2651–2667 (2006)
Muandet, K., Fukumizu, K., Sriperumbudur, B., Schölkopf, B.: Kernel mean embedding of distributions: a review and beyond. Found. Trends Mach. Learn. 10, 1–141 (2017)
Schölkopf, B., Smola, A.J.: Learning with Kernels. MIT Press, Cambridge (2002)
Sriperumbudur, B.K., Fukumizu, K., Lanckriet, G.R.G.: Universality, characteristic kernels and RKHS embedding of measures. J. Mach. Learn. Res. 12, 2389–2410 (2011)
Sriperumbudur, B.K., Gretton, A., Fukumizu, K., Schölkopf, B., Lanckriet, G.R.G.: Hilbert space embeddings and metrics on probability measures. J. Mach. Learn. Res. 11, 1517–1561 (2010)
Steinwart, I.: On the influence of the kernel on the consistency of support vector machines. J. Mach. Learn. Res. 2, 67–93 (2001)
Steinwart, I., Christmann, A.: Support Vector Machines. Springer, New York (2008)
Acknowledgements
The author would like to thank Daniel Winkle for bringing up this question and Andrey Kharitenko for a valuable discussion that lead to a more elegant proof.
Funding
Open Access funding enabled and organized by Projekt DEAL.
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Steinwart, I. Reproducing kernel Hilbert spaces cannot contain all continuous functions on a compact metric space. Arch. Math. 122, 553–557 (2024). https://doi.org/10.1007/s00013-024-01976-0
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00013-024-01976-0