3

These are two possible definitions of antiderivative (integral) incorporating a supposedly natural choice of an integration constant (see this question for further details).

The first one is based on Newton series, interpolated over consecutive derivatives:

$$f^{(-1)}(x)=\sum_{m=0}^{\infty} \binom {-1}m \sum_{k=0}^m\binom mk(-1)^{m-k}f^{(k)}(x)$$

The second one is based on Furier transform:

$$f^{(-1)}(x)=\frac{i}{2\pi}\int_{-\infty}^{+\infty} \frac{e^{- i \omega x}}{\omega} \int_{-\infty}^{+\infty}f(t)e^{i\omega t}dt \, d\omega$$

The question is for a proof that the both definitions coincide exactly (i.e. their values at all points and not only up to a constant) when the both converge.

Without losing the generality it is possible to consider only one point, say, $x=0$, since equality at this point guarantees equality elsewhere.

Note. $f(x)$ is required to be alalytic.

Community
  • 1
Anixx
  • 9,302
  • Are you missing an $i$ in one of your exponentials? Also, how are the integrals and sums supposed to be interpreted if they happen not to converge absolutely? – Joonas Ilmavirta Aug 23 '14 at 16:16
  • @Joonas Ilmavirta 1 - thanks, fixed. 2 - in the most generalized way, for instance, the integral from $-\infty$ to $\infty$ of an odd function should be considered zero. Anyway if there is a proof for only the strictest case, it would be good. Using Dirac Delta if needed is permitted. – Anixx Aug 23 '14 at 16:19
  • 1
    How does your first formula work, even for $s\geq0$ (from your MSE question)? You only evaluate the derivatives at $x$, so the value of $f^{(2)}(x)$, say, can't depend in any way on $f^{(k)}(x)$ for $k\neq2$, but this doesn't seem to be the case. It is also often natural to choose $\binom{-1}{m}=0$, yielding $f^{(-1)}=0$. Can you give references or show that the series does its job for some example function (for any $s$)? – Joonas Ilmavirta Aug 23 '14 at 17:29
  • 1
    @Joonas Ilmavirta It is Newton series https://en.wikipedia.org/wiki/Newton_series#Newton.27s_series where the deltas are taken over the order of derivative rather than the function's argument. – Anixx Aug 23 '14 at 17:59
  • But still, I don't see how your first formula could hold true. Can you use your series to integrate a monomial, for example? – Joonas Ilmavirta Aug 23 '14 at 18:12
  • @Joonas Ilmavirta it will diverge on monomials, unfortunately (except for constant zero that is). Try $2^{-x}$ for instance of non-diverging input. – Anixx Aug 23 '14 at 18:13
  • @Joonas Ilmavirta you can see a similar technique in my answer here: http://mathoverflow.net/questions/17605/how-to-solve-ffx-cosx/44727#44727 – Anixx Aug 23 '14 at 18:18
  • Ok, now I understand, at least partly, why your method works for $s\geq0$. But I still think your $f^{(-1)}$ vanishes identically. But perhaps it's better for me to give up, since I'm making slow progress... – Joonas Ilmavirta Aug 23 '14 at 18:32
  • @Joonas Ilmavirta this may or may not converge for $s<0$. It depends. If converges, it gives the antiderivative. – Anixx Aug 23 '14 at 18:34
  • You seem to be using $e^{ix\xi}$ as the kernel for the Fourier transform, but I've always seen people use $e^{-ix\xi}$. To avoid confusion, do you mind if I change it? – Vectornaut Aug 23 '14 at 19:09
  • @Vectornaut see here for instance: http://reference.wolfram.com/language/ref/FourierTransform.html – Anixx Aug 23 '14 at 19:13
  • Oh, weird! I guess we'll stick to that sign convention, then. Please forgive any sign errors in my answer. – Vectornaut Aug 23 '14 at 19:38

1 Answers1

6

If $f$ is an $L^2$ function, with Fourier transform $\hat{f}$, then the identity you're trying to prove is $$\sum_{m = 0}^\infty \binom{-1}{m} \sum_{k = 0}^m \binom{m}{k} (-1)^{m - k} (-i\xi)^k \hat{f}(\xi) = \frac{1}{-i\xi} \hat{f}(\xi).$$ In other words, you want to show that $$\sum_{m = 0}^\infty \binom{-1}{m} \sum_{k = 0}^m \binom{m}{k} (-1)^{m - k} (-i\xi)^k = \frac{1}{-i\xi}$$ as multiplication operators on some large subspace of $L^2(\mathbb{R})$. Let's call this identity $(\heartsuit)$. The left-hand side can be rewritten as $$\sum_{m = 0}^\infty \binom{-1}{m} (-1 - i\xi)^m.$$ Since $\binom{-1}{m} = (-1)^m$ (Kronenburg 2011), this simplifies to $$\sum_{m = 0}^\infty (1 + i\xi)^m.$$ This is the Taylor series of $\frac{1}{-i\xi}$ at $i$, so we've proven $(\heartsuit)$ as an identity of analytic functions on the unit disk centered at $i$. I think there should be a way to extend this to an identity of multiplication operators, like you wanted, but I don't know what it is.

Readers, if you come up with a way to do this last step, please let me know! I'll make the answer community wiki so you can add it.

Vectornaut
  • 2,224
  • What last step do you mean? – Anixx Aug 23 '14 at 19:38
  • Great! Now it is evident that the Furier formula is a natural consequence of considering integral a -1-th derivative. – Anixx Aug 23 '14 at 19:54
  • 2
    By "this last step," I mean proving $(\heartsuit)$ as an identity of multiplication operators, rather than just an identity of analytic functions near $i$. This is important because you ultimately wanted to apply the inverse Fourier transform to both sides of $(\heartsuit)$. For this to work, both sides of the identity have to be well-defined and equal as $L^2$ functions on $\mathbb{R}$, not just analytic functions on the unit disk around $i$. – Vectornaut Aug 23 '14 at 23:48
  • By the way, the very first identity you wrote is obviously true, it follows from Newton series properties. It is just Newton series expansion of exponential function $g(x)=\hat{f}(\xi)(-i\xi)^x$ taken at $x=-1$. It is obviously true if converges, there is no need for a proof. But what about how the first identity was produced? – Anixx Aug 24 '14 at 11:02
  • Do you mean that if Fourier transforms of two functions are the same, the functions themselves are the same too? – Anixx Aug 24 '14 at 11:09
  • 2
    @Anixx, yes: if two elements of $L^2(\mathbb{R})$ have the same Fourier transform, then they are the same. In fact, the Fourier transform on $L^2(\mathbb{R})$ is invertible. If this fact is new to you, I think you might gain a lot from reading a book or taking a course that introduces the Fourier transform from a mathematical perspective (Reed and Simon's Functional Analysis is my own favorite). – Vectornaut Aug 26 '14 at 09:23
  • @Anixx: What worries me about the very first identity, the one with the $\hat{f}$s, is that I'm pretty sure the left-hand side never converges when $\xi$ is real and nonzero. Since I think you want to prove the very first identity as an identity of $L^2$ functions, this could be a serious problem... – Vectornaut Aug 26 '14 at 09:29
  • do not you limit the functions to belong to $L^2$? So that the identity would not hold for general functions, that may be different by a constant etc? – Anixx Aug 26 '14 at 11:44
  • @Anixx, I'm not sure what you're asking. Could you rephrase your question? – Vectornaut Aug 26 '14 at 22:03
  • can you please make an answewr here? http://math.stackexchange.com/questions/902107/is-there-an-easier-way-to-find-the-natural-integration-constant Write anything you want on the topic, I just do not want the bounty wasted, the current answer by a user in totally ignorant and wrong. – Anixx Aug 27 '14 at 20:09
  • What do u think about this? http://mathoverflow.net/questions/184437/oddify-an-even-function-and-vice-versa-need-a-fourier-transform-based-formula – Anixx Oct 14 '14 at 21:32