Let me first introduce the restricted setting in which this question has a nice answer. I came up with this when messing around with a homework problem in a PDE class a couple years back.

Let $\phi \in C^2[0,1]$. It induces a function $\phi_*: L^2[0,1] \to L^2[0,1]$, $f \mapsto \phi\circ f$. One can ask whether $\phi_*$ is continuous with respect to the weak topology on $L^2[0,1]$: Does $f_n \rightharpoondown f \Rightarrow \phi(f_n) \rightharpoondown \phi(f)$?.

By looking at the function $f(x)=r\cos(n\pi x)+x_0$, one can get a necessary condition on $\phi$, that $\phi''(x_0)=0$ for all $x_0$. I.e. $\phi(x)=ax+b$. This is obviously sufficient too. Here's how you get the condition:

Assuming the continuity of $\phi_*$ with respect to the weak topology: $$\int_{-r+x_0}^{r+x_0}\frac{\phi(z)}{\sqrt{r^2-(z-x_0)^2}}=\int_{[0,1]} \phi(x_0+r\cos(n\pi x))dx$$

$$=\langle \text{weak limit of } \phi(x_0+r\cos(n \pi x)),1 \rangle $$

$$=\langle \phi(\text{weak limit of } (r\cos(n \pi x)+x_0)),1\rangle=\langle \phi(x_0),1\rangle=\phi(x_0).$$ So if $\phi''(x_0)>0$ for some $x_0$, one can find a small ball $B_r(x_0)$ where $\phi''>0$ and the extreme LHS won't be the extreme RHS. Similarly $\phi''$ is not less than 0 anywhere.

My question for stackexchange is whether this is something that is well studied? It has to be. It is such a simply phrased question. If that is too vague, my question for you is what happens when one removes the differentiability conditions on $\phi$, and one looks at the induced function on $L^2[\text{non compact euclidean space}]$. This argument doesn't even work in higher dimensions so I am at a loss to see how to generalize what I did above.

  • 2
    $\begingroup$ Note that the weak topology of a infinite dimensional Hilbert space $H$ is not metrizable (it is only its restriction on the unit ball), and for a map $H\to H$ continuity wrto the weak topology is not a priori equivalent to the sequential continuity ($f_n \rightharpoondown f \Rightarrow \phi(f_n) \rightharpoondown \phi(f)$) . $\endgroup$ Oct 2, 2016 at 9:07
  • 1
    $\begingroup$ Also, for the composition $f\to \phi \circ f$ to be defined on $L^2[0,1]$, you need $\phi$ to be defined on the whole $\mathbb{R}$ (not just $[0,1]$), together with some $|\phi(t)|\le a|t|+b$ for all $t$. $\endgroup$ Oct 2, 2016 at 9:19

2 Answers 2


Consider the sequence $f_k:[0,1]^n\to\mathbb{R}$ defined by $f_k(x_1,\dots,x_n):=(-1)^{\sum_{j=1}^n \lfloor kx_j \rfloor}$: it converges weakly to $0$. For any $\phi:\mathbb{R}\to\mathbb{R}$, and for any $a$ and $b$ in $\mathbb{R}$, there holds: $$\phi\circ (a+bf_k)=\frac{\phi(a+b)+\phi(a-b)}{2}+\frac{\phi(a+b)-\phi(a-b)}{2}\, f_k$$ so assuming $f\to \phi\circ f$ is sequentially continuous wrto the weak topology of $L^2[0,1]^n$ and taking weak limits we get $$\phi (a)=\frac{\phi(a+b)+\phi(a-b)}{2}.$$

Thus, testing the weak continuity along all sequences $a+bf_k$, we conclude that $\phi$ must be an affine function.


Adding to the answer of Pietro:

Exercise 4.20 in Brezis' "Functional Analysis, Sobolev Spaces and Partial Differential Equations" also tells you that weakly continuous Nemytskii operators on $L^p(0,1)$ are already necessarily affine linear. In fact, the proof uses so-called Rademacher's functions, which are closely related to the sequence $f_k$ defined by Pietro in his answer (luckily, Brezis provides solutions for his exercises :-)).


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.