# Direct sum of reproducing kernel Hilbert spaces (RKHS)

I am currently diving into the theory of reproducing kernel Hilbert spaces and am just at the beginning of understanding the background of reproducing kernels. I have stumbled upon the following theorem firstly published by N. Aronszajn in Theory of reproducing kernels, 1950. It states:

If $$K_i(x,y)$$ is the reproducing kernel of class $$F_i$$ with norm $$\Vert . \Vert_i$$ for $$i=1,2$$ respectively, then $$K(x,y)=K_1(x,y)+K_2(x,y)$$ is the reproducing kernel of the class $$F=\lbrace f=f_1+f_2 \vert f_1 \in F_1, f_2 \in F_2 \rbrace$$ with norm $$\Vert f \Vert = \min \lbrace \Vert f_1 \Vert_1 + \Vert f_2 \Vert_2 \mid f_1 \in F_1, f_2 \in F_2 \rbrace.$$

I completely understand the given proof of this statement but I currently still have problems with the intuition behind it. For my understanding what we do to prove this statement is to use the Hilbert space $$F_1 \times F_2$$ with inner product $$\langle (f_1,f_2),(g_1,g_2) \rangle = \langle f_1,g_1 \rangle _1 + \langle f_2,g_2 \rangle _2$$ and then filter away the nullspace of the addition operator $$(f,g)\mapsto f+g$$ for the case that $$F_1 \cap F_2 \neq \lbrace 0 \rbrace$$. At this point I am wondering why the direct sum (not necessarily the orthogonal direct sum) $$F_1 + F_2 = \lbrace f_1 + f_2 | f_1 \in F_1, f_2 \in F_2 \rbrace$$ with the above defined scalar product (and induced norm $$\Vert . \Vert = \Vert . \Vert_1 + \Vert . \Vert_2$$) doesn't yield the RHKS for the kernel $$K=K_1 +K_2$$.

I suspect that something with the reproducing properties has to go wrong in this case but I cannot figure out what exactly it is, since we have \begin{align} i)\quad &K(.,y)=K_1(.,y)+K_2(.,y) \in F_1 + F_2 \quad \forall y \\ ii) \quad & \langle f_1+f_2,K(.,y) \rangle = \langle f_1+f_2,K_1(.,y)+K_2(.,y) \rangle \\ &= \langle f_1,K_1(.,y) \rangle _1+ \langle f_2,K_2(.,y) \rangle _2 = f_1(y)+f_2(y) \quad \forall y. \end{align} Thank you in advance!

• What do you mean with direct sum $F_1+F_2$, when $F_1 \cap F_2 \neq \lbrace 0 \rbrace$? Apr 13, 2022 at 17:47
• @JoséCFerreira: thank you for your reply! I actually mean the "normal" or "standard" direct sum $F_1+F_2= \lbrace f_1 + f_2 | f_1 \in F_1, f_2 \in F_2$ which is also sometimes denoted by $F_1 \oplus F_2$ but I wanted to highlight that I don't necessarily mean the orthogonal sum $F_1 \oplus F_2$ where $F_1 \cap F_2 = \lbrace 0 \rbrace$. With this case Aronszajn points out that the norm of the RKHS is simplified to $\Vert f \Vert = \Vert f_1 \Vert_1 + \Vert f \Vert_2$. My question is where does this norm fail when $F_1 \cap F_2 \neq \lbrace 0 \rbrace$. Apr 14, 2022 at 6:11

If $$F_1\neq F_2$$ (as a normed space) and $$F_1 \cap F_2 \neq \lbrace 0 \rbrace$$ (as sets), you can choose functions $$f=f_1+f_2\in F_1$$, $$g=g_1+f_2\in F_2$$, with $$f_2\in F_1 \cap F_2$$.
The expression $$\Vert h \Vert = \Vert h_1 \Vert_1 + \Vert h_2 \Vert_2$$ (on $$F_1+F_2$$) is not well defined to $$h=f+g=h_1+h_2$$, because $$h_1$$ and $$h_2$$ can be writen as $$h_1=f_1+2f_2$$ and $$h_2=g_1$$ or $$h_1=f_1$$ and $$h_2=g_1+2f_2$$, for instance.
Perhaps you find useful results searching for "$$\|h\|= \|h_1 \|_1 + \| h _2\|_2$$" on SearchOnMath.
• Thank you so much - this was the point I've overseen! However, just as a formality I would say the norm is not well-defined since in your example above $h=f+g=h_1+h_2$ where $h_1=f_1+2f_2$ and $h_2=g_2$ since $g_2$ doesn't necessarily have to be an element of $F_1$. But nevertheless this is the argument which was missing for my understanding! Apr 21, 2022 at 8:06