Let $C^\infty$ denote the collection of functions $f:\mathbb{R}\to\mathbb{R}$ such that for every positive integer $n$, the $n$-th derivative of $f$ exists. For $f\in C^\infty$ we set

  • $f^{(0)} = f$, and
  • $f^{(n+1)} = \big(f^{(n)}\big)'$ for all non-negative integers $n$.

Is there $f\in C^\infty$ with the following properties?

  1. for all $x\in (-\infty, 0]$ we have $f(x)=0$.
  2. for all non-negative integers $n$ and $x\in (0,\infty)$ we have $f^{(n+1)}(x) > f^{(n)}(x)$
  3. there is a function $h:\mathbb{R}\to\mathbb{R}$ such that for all non-negative integers $n$ and all $x\in\mathbb{R}$ we have $f^{(n)}(x)\leq h(x)$.

(Additional question for curiosity, answering it is not needed for acceptance of answer: can $h$ be chosen to be continous? Or even $h\in C^\infty$?)

  • 2
    $\begingroup$ @GerhardPaseman For 2 only, something based on $e^{cx}$ with $c>1$ would work $\endgroup$ Nov 26, 2018 at 18:17
  • 1
    $\begingroup$ @GerhardPaseman Well there are tricks for 1, actually I don't readily see how to handle 3 $\endgroup$ Nov 26, 2018 at 18:21
  • 6
    $\begingroup$ If $h$ is continuous (or even $L^\infty_{\text{loc}}$), then Taylor's theorem implies that $f$ is everywhere real analytic. So your function cannot possibly exist, when combining (1) and (2). So at the very least near $0$ you must have $h$ is allowed to blow-up. $\endgroup$ Nov 26, 2018 at 19:12
  • 3
    $\begingroup$ @Neal: yes. condition 2 implies $f$ has at most one zero in $(0,\infty)$, similarly $f'$ hast at most one zero in $(0,\infty)$. Suppose $f'$ is not everywhere positive on $(0,\infty)$, there exists then some small interval $(0,a)\subset (0,\frac12)$ on which $f' < 0$. By the mean value theorem we must be able to find $b\in (0,a)$ such that $0> f(b) > f'(b)$, leading to a contradiction. So $f'$ has to be everywhere positive. $\endgroup$ Nov 26, 2018 at 20:30
  • 8
    $\begingroup$ If f' is positive then so are all higher derivatives, and then one can take h to be monotone increasing, hence also in $L^\infty_{loc}$, so it seems Willie's arguments show that such a function does not exist. $\endgroup$
    – Terry Tao
    Nov 26, 2018 at 20:34

2 Answers 2


Combining my comments with that of Terry Tao's:

  1. First we show that $f^{(n)} > 0$ on $(0,\infty)$ for all $n \geq 1$. The argument is given for $f'$, but, extends easily to all $n \geq 1$.
    Start by noticing that $f'(a) = 0 \implies f''(a) > 0$ so that $f'$ changes sign at most once, and that if it is not everywhere positive on $(0,\infty)$ there must be an initial interval $(0,\epsilon)$ on which $f' < 0$. Assume WLOG that $\epsilon < \frac12$. On $(0,\epsilon)$, we have that $f(x) \geq x \inf_{y\in (0,x)} f'(y)$ by the mean value theorem. This is incompatible with $f'(y) > f(y)$. Hence we conclude that $f'$ is always positive on $(0,\infty)$.

  2. Step 1 implies that $f^{(n)}$ is increasing for all $n \geq 1$. Therefore if $h$ is a function as in condition (3) of the question, so is the increasing function $$ \tilde{h}(y) = \inf_{[y,\infty)} h $$ This function is locally bounded, and implies (by Taylor's inequality) that $f$ is real analytic.

  3. Real analytic functions can't be vanishing on $(-\infty,0)$ and be non-trivial.


By Willie Wong's comment and answer, $f^{(n)}>0$ on $(0,\infty)$ for all $n=0,1,\dots$. Hence, $f^{(n)}\ge0$ on $\mathbb R$. So, by Bernstein's theorem on completely monotone functions (used "right-to-left"; see the "footnote" for details), $f(x)=\int_0^\infty e^{tx}\mu(dt)$ for some nonzero nonnegative measure $\mu$ on $[0,\infty)$ and all real $x\le1$. So, $f>0$ on $(-\infty,1]$, which contradicts condition 1 in the OP. Thus, there exists no $f$ satisfying conditions 1 and 2. (Condition 3 is not needed for the non-existence.)

"Footnote": The condition $f^{(n)}\ge0$ on $\mathbb R$ for all $n=0,1,\dots$ implies $(-1)^n g^{(n)}\ge0$ on $[0,\infty)$ for all $n=0,1,\dots$, where the function $g\colon[0,\infty)\to\mathbb R$ is defined by reading $f$ "right-to-left": $g(y):=f(1-y)$ for $y\in[0,\infty)$. So, $g$ is completely monotone and hence, by Bernstein's theorem, $g(y)=\int_0^\infty e^{-ty}\nu(dt)$ for some nonnegative measure $\nu$ and all real $y\ge0$. Moreover, the measure $\nu$ is nonzero, because otherwise we would have $g=0$ (on $[0,\infty)$), that is, $f=0$ on $(-\infty,1]$, which would contradict the condition that $f^{(n)}>0$ on $(0,\infty)$ for all $n=0,1,\dots$. So, $f(x)=g(1-x)=\int_0^\infty e^{-t+tx}\nu(dt)=\int_0^\infty e^{tx}\mu(dt)$ for all real $x\le1$, where $\mu(dt):=e^{-t}\nu(dt)$, and the measure $\mu$ is nonnegative and nonzero.

  • $\begingroup$ This looks like very neat and concise argument but could you explain what do you mean by "right-to-left"? I failed to figure this out from the Wikipedia page, sorry. Definition of complete monotonicity there involves alternating signs for odd/even order derivatives... $\endgroup$ Nov 30, 2018 at 11:17
  • 1
    $\begingroup$ @მამუკაჯიბლაძე : Thank you for your comment. I have added a "footnote" with details on the meaning of the "right-to-left". $\endgroup$ Nov 30, 2018 at 14:38
  • 1
    $\begingroup$ For the record: Functions with $f^{(n)} \ge 0$ are also known as absolutely monotonic functions, see e.g., mathworld.wolfram.com/AbsolutelyMonotonicFunction.html (and connect to CM functions as noted in Iosif's answer above) $\endgroup$
    – Suvrit
    Nov 30, 2018 at 16:56

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.