19
$\begingroup$

Given some sufficiently smooth function $f$ what conditions would be sufficient for its Fourier coefficients, as defined by $$ \hat{f}(n) := \int_{0}^{2\pi}\cos(nx)f(x)\ dx, \quad \text{for } n = 1,2,\ldots, $$ to be monotonic? Given the decay properties of Fourier coefficients, the monotonicity result would translate to $$ |\hat{f}(n)| \geq |\hat{f}(n+1)|, \quad n = 1,2,\ldots. $$ I haven't been able to find any literature regarding this and a result of this nature would be very interesting.

$\endgroup$
5
  • 9
    $\begingroup$ I doubt there will be such a condition in terms of smoothness. Given any smooth function $f$ and a positive integer $n$, the function $f- \hat{f}(n) cos(nx)$ has the same Fourier coefficients as $f$, besides the $n$th, which would be 0. $\endgroup$
    – Itay
    Dec 24, 2020 at 19:11
  • $\begingroup$ Hi @Itay, I'm not necessarily looking for a smoothness condition, but just another general condition for this to hold, e.g. a sufficient decay, has to satisfy condition (X) etc. type of condition. The smoothness assumption was put in place to simplify matters. $\endgroup$
    – spaceman
    Dec 24, 2020 at 19:40
  • 4
    $\begingroup$ @Itay but it's higher derivatives are quite large in a sense. $\endgroup$ Dec 24, 2020 at 19:41
  • $\begingroup$ Are you only interested in even functions $f$, i.e. those satisfying $f(x)=f(2\pi -x)$? If not then your definition of "Fourier coefficient" might not be appropriate $\endgroup$
    – Yemon Choi
    Jan 6, 2021 at 11:55
  • $\begingroup$ Can I just point out that this definition of "Fourier coefficient" looks a bit strange for $f(x)=\sin(x)$, which last time I looked was a perfectly reasonable $2\pi$-periodic smooth function... $\endgroup$
    – Yemon Choi
    Jan 6, 2021 at 18:15

2 Answers 2

24
$\begingroup$

It suffices that $f$ be (the restriction to $[0,2\pi]$ of) a completely monotone real-valued function defined on $[0,\infty)$. Indeed, then for some finite measure $\mu$ on $[0,\infty)$ and all real $x\ge0$ we have $$f(x)=\int_0^\infty\mu(da) e^{-a x},$$ whence for natural $n$ $$\hat f(n)=\int_0^\infty\mu(da) \int_0^{2\pi}dx\,\cos(nx)e^{-a x} =\int_0^\infty\mu(da) \frac{a \left(1-e^{-2 \pi a}\right)}{a^2+n^2},$$ which is obviously decreasing in $n$ (to $0$, by dominated convergence or by the Riemann--Lebesgue lemma).

Note that, if $f(x)\equiv1$ or $f(x)\equiv x$, then $\hat f(n)=0$ for all natural $n$. So, if $f$ has the desired property, then the function $[0,2\pi]\ni x\mapsto a+bx+f(x)$ also has it for any real $a$ and $b$. Also, clearly, if $f$ has the desired property, then do does the function $$[0,2\pi]\ni x\mapsto f^-(x):=f(2\pi-x)$$ -- because $\widehat{f^-}(n)=\hat f(n)$ for all natural $n$. It follows that, if $f$ and $g$ have the desired property, then the function $$[0,2\pi]\ni x\mapsto a+bx+f(x)+g(2\pi-x)$$ also has it for any real $a$ and $b$.

Added:

  1. As noted in a comment by Fedor Petrov, if $f(x)=h(\pi-x)$ for some odd function $h$ and all $x\in[0,2\pi]$, then $\hat f(n)=0$ for all natural $n$.

  2. It follows from this answer by fedja that, if $$f(x)=\int_1^\infty[\mu(dp) x^p+\nu(dp)(2\pi-x)^p]<\infty$$ for some measures $\mu$ and $\nu$ on $[1,\infty)$ and all $x\in[0,2\pi]$, then $f$ has the desired property.

$\endgroup$
4
  • $\begingroup$ Why does $bx+f(x)$ also have this property? $\endgroup$ Dec 25, 2020 at 14:49
  • $\begingroup$ @FedorPetrov : I have added a detail on this. $\endgroup$ Dec 25, 2020 at 15:00
  • $\begingroup$ ah, I see, actually any function of the form $h(x-\pi)$, where $h$ is odd, works $\endgroup$ Dec 25, 2020 at 20:39
  • $\begingroup$ @IosifPinelis Thank you for a wonderful answer, I'm very happy that this led to some interesting discussions and even an extended question. $\endgroup$
    – spaceman
    Dec 28, 2020 at 16:47
3
$\begingroup$

A comment on this problem would be that if $\hat f(n)$ are monotone ( here $f$ any continuous function, not necessarily odd or even, also I assume that $\hat f(n)$ is monotone not $|\hat f(n)|$ ) then one can assume that they are positive. And if Fourier coefficients are real and positive then they must be absolutely convergent, that is $\{\hat f(n)\} \in l_1$. This follows easily from property of Fejer's kernel, i.e. that it is positive operator with integral 1:

$$ \sum_k (1-|k|/n)\hat f(k)exp(ikt) = \int F(t-s)f(s) \le \sup|f|$$ so $1/2 \sum_{k \in (-n/2, n/2)} \hat f(k)) \le \sup|f|$.

$\endgroup$

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.