2
$\begingroup$

Set-up: Let $\{f_n \}_{n=1}^\infty$ be a sequence of smooth convex functions on $(0,1) \subset \mathbb R$ that converge uniformly to the continuous (not necessarily differentiable) convex function $f$.

It is well known that $f$ is differentiable almost everywhere. It is also known that $\{f'_n \}$ converges pointwise to $f'$ wherever $f'$ exists.

Further, by Alexandrov's theorem, $f$ is also twice differentiable almost everywhere.

My question: Is there any type of convergence of $2$nd derivatives?

Thoughts: Clearly, there is not uniform convergence since at any point where $f'$ does not exists, $f''$ will be infinite (e.g., $f(x) =|x|$). But maybe pointwise a.e. or convergence in measure?

$\endgroup$
3
$\begingroup$

Bad news

The second derivatives need not converge a.e. or even in measure. Indeed, let $g_n(x)$ be the $n$th binary digit of $x$. Then the antiderivatives $G_n(x)=\int_0^x g_n(t)\,dt$ uniformly converge to $x/2$, and second antiderivatives $f_n(x)=\int_0^x G_n(t)\,dt$ uniformly converge to $f(x)=x^2/4$. However, $|g_n-f''|=1/2$ a.e.

(In this example, $f_n$ is not twice differentiable, but one can fix this by smoothing $g_n$ a little; this doesn't really matter for integral estimates.)

Good news

Distributional convergence is inherited by derivatives of all orders. This means that for any test function $\phi\in C_c^\infty((0,1))$ we have $$ \int_0^1 f_n'' \phi\,dx \to \int_0^1 f''\phi\,dx \tag{1}$$ This is not particularly surprising: it's just a restatement of the obvious $$ \int_0^1 f_n \phi''\,dx \to \int_0^1 f \phi''\,dx $$
via integration by parts.

But we can do better: (1) holds for all continuous functions $\phi$ with compact support in $(0,1)$. In other words: interpreting $f_n''$ as a measure (which it is), we have weak convergence of measures on compact subsets of $(0,1)$. Here is why:

  1. The measures are bounded on each interval $[a,b]$ with $0<a<b<1$, since $f_n'(a)$ and $f_n'(b)$ are uniformly bounded (controlled by convexity & uniform convergence).
  2. By Banach-Alaoglu, every subsequence $f_{n_k}''$ has a weakly convergent subsequence $f_{n_{k_j}}''$.
  3. Since weak convergence of measures implies distributional convergence, the limit of $f_{n_{k_j}}''$ is $f''$.
  4. General topology fact: if every subsequence has a subsubsequence converging to the same limit, then the whole sequence converges there.
$\endgroup$

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.