3

Let (fn) be a sequence of twice differentiable functions on R such that for each n there exists some xnR such that:

  • fn is strictly convex on (,xn),
  • fn is strictly concave on (xn,+).

Suppose also that fn uniformly converges to a twice differentiable function f.

It was conjectured that then the sequence (xn) will be convergent.

This conjecture was disproved.

The OP then asked in a comment whether the additional condition that f not vanish on any nonempty interval can help.

Below it will be shown that the answer is still negative, even with the latter additional condition.

In another answer, it will be shown how conditions on the f_n's or on f can be modified to make the sequence (x_n) convergent.

\endgroup
4
  • 3
    \begingroup Why didn't you include these refined examples in your answer to the original post? Posting questions with an immediate answer is, to say the least, not the usual thing in MO. \endgroup Commented Oct 4, 2023 at 13:03
  • 1
    \begingroup @JochenWengenroth : (i) The answer to the original conjecture was complete and thus needed no additions. (ii) The refinements presented on the current page were not in my possession at the time of the previous answer. (iii) These refinements involve different ideas. (iv) We should not encourage the OP's moving the goalposts. (v) Every time when one posts a question, the system itself automatically and immediately "invites" the poster to do something like "share your knowledge, in a question-and-answer style". I too can see nothing bad in doing so. Are you satisfied with this response? \endgroup Commented Oct 4, 2023 at 13:50
  • \begingroup I completely agree with (iv). I have just read the "share your knowledge" details -- and I was indeed quite surprised because this is quite unusual for MO. Honestly, I am not completely convinced that this is a case where it is better to post a new question than to modify the answer, but I have no strong objections. I was just surprised. \endgroup Commented Oct 4, 2023 at 15:47
  • \begingroup @JochenWengenroth : I agree with you that "shar[ing] [one's] knowledge in a question-and-answer style" is not very common on MO. However, isn't sharing knowledge in whatever appropriate way a worthy goal? Oftentimes I do modify my answers, to address users' comments or correct my own errors or misunderstandings. However, in this case, for reasons listed in my previous comment (and also because of the length of the latter answers), "shar[ing] [...] knowledge in a question-and-answer style" seemed much more appropriate. \endgroup Commented Oct 4, 2023 at 16:09

2 Answers 2

4
\begingroup

Let b be the continuously differentiable bump function defined by the formula \begin{equation} b(u):=60u_+^2(1-u)_+^2 \end{equation} for real u, where u_+:=\max(0,u). So, b is supported on the interval [0,1]. For real x, let \begin{equation} B(x):=-1+\int_{-\infty}^x b(u)\,du =\begin{cases} -1 & \text{ if }x\le 0, \\ 1 &\text{ if } x>1, \\ 12 x^5-30 x^4+20 x^3-1 & \text{otherwise.} \end{cases} \end{equation} Then B is a twice continuously differentiable nondecreasing function such that |B|\le1.

In what follows, assume that n\ge3. Then x^2-B(x/n)\ge0 for all real x. So, we can define \begin{equation} f_n(x):=\sqrt{x^2-B(x/n)}. \end{equation} Let also \begin{equation} f(x):=\sqrt{x^2+1}; \end{equation} here and in what follows, x is a real number, sometimes subject to additional conditions. Note that f is a smooth function and f'' does not vanish on any nonempty interval.

If x\le0, then f_n(x)=f(x).

If 0<x=o(n), then B(x/n)\to-1 and hence f_n(x)\to f(x) (as n\to\infty) uniformly over all x such that 0<x=o(n).

If n=O(x), then x\to\infty. Therefore and because |B|\le1, we have f_n(x)\to f(x) uniformly over all x such that n=O(x).

Thus, f_n(x)\to f(x) uniformly in all real x.

Next, f_n is twice continuously differentiable, f''_n(x)>0 if x\le0, f''_n(x)<0 if x\ge n, and f_n(x)>0 and \begin{equation} f''_n(x)=\frac{(r(u)-1/n^2)q(u)}{f_n(x)^3} \end{equation} if x\in(0,n), where u:=x/n\in(0,1), \begin{equation} r(u):=\frac{p(u)}{q(u)}, \end{equation} \begin{equation} p(u):=1 - 20 u^3 + 90 u^4 - 72 u^5, \end{equation} \begin{equation} q(u):=60 (1-u) u \left(9 u^6-27 u^5+25 u^4-5 u^3-2 u+1\right)>0. \end{equation} So, the sign of f''_n(x) for x\in(0,n) is the same as the sign of r(u)-1/n^2.

Note that r'(u)<0 for u\in(u_*,1), where u_*:=\frac{3 +\sqrt3}6\in(0,1).
Also, \min\{r(u)\colon u\in(0,u_*]\}=\frac2{205}\, (45 - 16\sqrt3)=0.168\ldots>1/n^2 (since n\ge3). Also, r(1-)=-\infty. So, there is some u_n\in(u_*,1)\subset(0,1) such that r(u)-1/n^2>0 for u\in(0,u_n) and r(u)-1/n^2<0 for u\in(u_n,1). Letting now \begin{equation} x_n:=nu_n, \end{equation} we see that f''_n(x)>0 if x<x_n and f''_n(x)<0 if x>x_n.

So, all the conditions on f and the f_n's hold.

However, x_n\ge nu_*\to\infty. \quad\Box


For an illustration, here are the graphs \{(x,f_3(x))\colon-3\le x\le7\} (above) and \{(u,r(u))\colon0<u<1\} (below):

enter image description here

enter image description here

\endgroup
1
  • \begingroup Thank you for this very clever counter-example! \endgroup
    – NancyBoy
    Commented Oct 4, 2023 at 16:19
1
\begingroup

Here it will be shown how conditions on the f_n's or on f can be modified to make the sequence (x_n) convergent.

First of all, the condition that the f_n's be twice differentiable is of no help; so, this condition will be dropped.

Suppose that for each n there exists some x_n\in\mathbb{R} such that:

  • f_n is semi-strongly convex on (-\infty,x_n) in the sense that f_n is convex on (-\infty,x_n) and f_n(x)/|x|\to\infty as x\to-\infty;

  • f_n is semi-strongly concave on (x_n,\infty) in the sense that f_n is concave on (x_n,\infty) and f_n(x)/x\to-\infty as x\to\infty.

Note that, if f_n is strongly convex on (-\infty,x_n), then f_n is semi-strongly convex on (-\infty,x_n); similarly, if -f_n is strongly convex on (x_n,\infty), then f_n is semi-strongly concave on (x_n,\infty).

Suppose also that (f_n) uniformly converges to a function f, which is not affine on any nonempty interval.

Then the sequence (x_n) is convergent.


Indeed, suppose first that the sequence (x_n) is unbounded. Then, using a left-right symmetry and passing to a subsequence, without loss of generality (wlog) assume that x_n\to\infty. It follows that the limit function f is convex on \mathbb R. So, f(x)\ge a+bx for some real a,b and all real x. Therefore and in view of semi-strongly concavity of f_n on (x_n,\infty), for each n we have f_n(x)-f(x)\to-\infty as x\to\infty, which contradicts the uniform convergence of (f_n) to f.

So, the sequence (x_n) is bounded. Take any subsequence (x_{n_k}) of the sequence (x_n) converging to a limit x_*. Then f is convex on (-\infty,x_*) and concave on (x_*,\infty). Since f is not affine on any nonempty interval, the point x_* is uniquely determined by f. So, any converging subsequence of the sequence (x_n) converges to the same limit x_*. Since the sequence (x_n) is bounded, we conclude that x_n\to x_*, as claimed.


This reasoning shows that the following is true as well:

Suppose that for each n there exists some x_n\in\mathbb{R} such that:

  • f_n is convex on (-\infty,x_n);

  • f_n is concave on (x_n,\infty).

Suppose also that (f_n) converges pointwise (not necessarily uniformly) to a function f, which is not affine on any nonempty interval, not convex on \mathbb R, and not concave on \mathbb R.

Then the sequence (x_n) is convergent.

Details on the latter assertion, requested by Gaetano: Suppose first that the sequence (x_n) is unbounded. Then, using a left-right symmetry and passing to a subsequence, wlog assume that x_n\to\infty. It follows that the limit function f is convex on \mathbb R, which contradicts the assumption that f is not convex on \mathbb R.

So, the sequence (x_n) is bounded. Take any subsequence (x_{n_k}) of the sequence (x_n) converging to a limit x_*. Then f is convex on (-\infty,x_*) and concave on (x_*,\infty). Since f is not affine on any nonempty interval, the point x_* is uniquely determined by f. So, any converging subsequence of the sequence (x_n) converges to the same limit x_*. Since the sequence (x_n) is bounded, we conclude that x_n\to x_*, as claimed.

\endgroup
10
  • \begingroup Thank you@IosifPinelis for your anwser. However, I can't see how can you avoid the semi-strongly convexity/concavity hypothesis in your reasonning (last point). Thank you! \endgroup
    – NancyBoy
    Commented Oct 4, 2023 at 16:21
  • 1
    \begingroup @Gaetano : We only used the semi-strong convexity/concavity to contradict the conclusion that f is convex on \mathbb R. But in the latter case, considered at the end of the answer, f is assumed outright to be neither convex on \mathbb R nor concave on \mathbb R. \endgroup Commented Oct 4, 2023 at 16:28
  • \begingroup Concerning the line "Since f is not affine on any nonempty interval, the point x^* is uniquely determined by f" ". Could you provide me an hint to prove it ? Thank you. \endgroup
    – NancyBoy
    Commented Oct 4, 2023 at 22:01
  • 1
    \begingroup @Gaetano : I have added details on this. In addtion: NB1: The uniform convergence was only needed to contradict the conclusion that f is convex on \mathbb R. But in the latter case, considered at the end of the answer, f is assumed outright to be neither convex on \mathbb R nor concave on \mathbb R. NB2: If f is the pointwise limit of functions f_n convex on an interval I, then f is convex on I; this follows immediately from the definition of convexity. . \endgroup Commented Oct 6, 2023 at 14:23
  • 1
    \begingroup Even though this answer is much less subtle than the other one, I have accepted it, because of it being positive -- that is, making the original conjecture true, albeit under certain additional conditions. With this said, I appreciate your upvotes of the other answer. \endgroup Commented Oct 6, 2023 at 14:28

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .