Talk:Fourier series

The Table of common Fourier series would be much clearer

The Table of common Fourier series would be much clearer if *instead* of labeling the columns with "time domain" and "frequency domain", they were labeled with "Function" and "Fourier coefficients".

For several reasons:

1. Above all, that is overwhelmingly how most people think: in terms of functions and their Fourier coefficients. So why be confusing?

2. It is entirely unclear why the column of functions (of whose Fourier coefficients we will soon be told) is labeled with a *domain*, since it is a function, not a domain.

3. There are many, many applications of Fourier series, since they apply to almost any real-valued function that arises in practice. A very large number of those applications are not about either time or frequency.

I have no objection, of course, to mentioning the reasons that the time domain and frequency domain are highly relevant to Fourier analysis, even in this table ... but my point is: Make it easy for people to find words like "function" and "Fourier coefficients", since that is mainly what they will be looking for.

— Preceding unsigned comment added by 98.36.148.11 (talk • contribs)

While I don't love the "time/frequency" metaphor, it is baked into the structure of the article and really does clarify a lot of things here. So I'm opposed to any plan to eliminate it from one part of the article, without a provision for improving the rest of the article without sacrificing clarity of the presentation. Tito Omburo (talk) 10:16, 8 November 2024 (UTC)

Symmetry relations

@Tito Omburo, I see you've reverted my edit regarding the deletion of the duplicate "Symmetry" subsection.

The only difference is the symbol $s$ instead of $f$, the subsections Fourier_series#Symmetry_relations and Fourier_transform#Symmetry are otherwise indistinguishable. I think it's rather presumptuous to state that it's the Fourier transform beteeen Z and S^1.

Could you please elaborate on why you feel the subsections are distinct?

Kind regards, Roffaduft (talk) 12:11, 12 December 2024 (UTC)

The Fourier transform covered in the article Fourier transform is for continuous signals, while the symmetry relations covered here are applied for discrete signals or periodic signals. For example, the Fourier transform is gotten by integrating over R, while Fourier series are gotten by integrating over S^1. (Although often in digital signal processing, the roles of time and frequency are reversed, and the signal is discrete, and the Fourier transform is periodic.) Tito Omburo (talk) 12:39, 12 December 2024 (UTC)

Yes, I understand the difference, but nowhere in the subsection is this distinction being made. It is very cleary a simple copy/paste with no information specific to the Fourier series.

If the symmetry subsection was written as in

• Oppenheim and Schafer (1975) Digital Signal Processing, Prentice–Hall, pp 20–26.

then I wouldn't have deleted it. Roffaduft (talk) 12:46, 12 December 2024 (UTC)

Recent edits

The 11 Dec 2024 version was careful to distinguish between real-valued s(x) and complex-valued s(x). The analysis and synthesis formulas are developed for real-valued s(x). Eq.6, for instance, states

$c_{_{n}}=c_{|n|}^{*},\quad n<0$

which is true in general only for real-valued s(x). There was once a section to explain the generalization to complex s(x). Here is a copy/paste version:

Complex-valued functions

If $s(x)$ is a complex-valued function of a real variable $x,$ both components (real and imaginary part) are real-valued functions that can be represented by a Fourier series. The two sets of coefficients and the partial sum are given by:

$c_{_{Rn}}={\frac {1}{P}}\int _{P}\operatorname {Re} \{s(x)\}\cdot e^{-i{\tfrac {2\pi }{P}}nx}\ dx$ and $c_{_{In}}={\frac {1}{P}}\int _{P}\operatorname {Im} \{s(x)\}\cdot e^{-i{\tfrac {2\pi }{P}}nx}\ dx$

$s_{_{N}}(x)=\sum _{n=-N}^{N}c_{_{Rn}}\cdot e^{i{\tfrac {2\pi }{P}}nx}+i\cdot \sum _{n=-N}^{N}c_{_{In}}\cdot e^{i{\tfrac {2\pi }{P}}nx}=\sum _{n=-N}^{N}\left(c_{_{Rn}}+i\cdot c_{_{In}}\right)\cdot e^{i{\tfrac {2\pi }{P}}nx}.$

Defining $c_{n}\triangleq c_{_{Rn}}+i\cdot c_{_{In}}$ yields:

$s_{_{N}}(x)=\sum _{n=-N}^{N}c_{n}\cdot e^{i{\tfrac {2\pi }{P}}nx}$ Eq.7

This is identical to Eq.5 except $c_{n}$ and $c_{-n}$ are no longer complex conjugates. The formula for $c_{n}$ is also unchanged:

${\begin{aligned}c_{n}&={\tfrac {1}{P}}\int _{P}\operatorname {Re} \{s(x)\}\cdot e^{-i{\frac {2\pi }{P}}nx}\ dx+i\cdot {\tfrac {1}{P}}\int _{P}\operatorname {Im} \{s(x)\}\cdot e^{-i{\tfrac {2\pi }{P}}nx}\ dx\\[4pt]&={\tfrac {1}{P}}\int _{P}\left(\operatorname {Re} \{s(x)\}+i\cdot \operatorname {Im} \{s(x)\}\right)\cdot e^{-i{\tfrac {2\pi }{P}}nx}\ dx\ =\ {\tfrac {1}{P}}\int _{P}s(x)\cdot e^{-i{\tfrac {2\pi }{P}}nx}\ dx.\end{aligned}}$

References

end

I added that section, but later replaced it with a short statement and footnote, which are now gone.
Bob K (talk) 13:30, 13 December 2024 (UTC)

Not sure where to start on this one. First of all, in the definition of the Fourier series (at the beginning of the subsection), $s$ and $c$ are complex. I've added an inline citation to Folland (1992) to support this statement.

The complex-valued s(x) was introduced 14:02, 24 November 2024. Prior to that, we began with real-valued s(x), and maintained it all the way through the exponential form, pointing out that $C_{n}=C_{|n|}^{*},\ n<0,$ as a result. The definition inserted ahead of that created an incongruity that I tried to mitigate with the addition of labels Real valued s(x) and Complex valued s(x).
--Bob K (talk) 02:50, 14 December 2024 (UTC)

Second, the references you've provided do not make any distinction between real- and complex-valued $s(x)$, for good reason. Because there really isn't any.

I chose readily available online references, which do not overtly specify real-valued s(x), but they do define our Eqs (1), (2), and (5), which I believe do require that assumption. That point is also supported at:
McGillem, Clare D.; Cooper, George R. (1984). Continuous and Discrete Signal and System Analysis (2 ed.). Holt, Rinehart and Winston. p. 84-85, eqs (3-4),(3-5),(3-6). ISBN 0-03-061703-0.
--Bob K (talk) 02:50, 14 December 2024 (UTC)

Third, in the complex case, $c_{n}={\tfrac {1}{2}}(a_{n}-ib_{n})$ and $c_{-n}={\tfrac {1}{2}}(a_{n}+ib_{n})$. The only thing that changes is that $a_{n}$ and/or $b_{n}$ become complex.

Fourth, there was no clear distinction between the real and complex subsection. That is, the statements in the "real" subsection applied just as well to the complex case.

Fifth, the entire "complex valued functions" section you posted here is basically a very convoluted way of showing there is no difference between the real and complex case, i.e., Eq.7 = Eq.5.

Kind regards, Roffaduft (talk) 14:31, 13 December 2024 (UTC)

which do not overtly specify real-valued s(x)

Well, then they are not references that support your claim.

they do define our Eqs (1), (2), and (5), which I believe do require that assumption

If $s(x)$ is complex then so are (at least) Eq.2 and Eq.5. That is, nowhere does it say that $a_{n}$ and/or $b_{n}$ must be real.

If the same general approach does not hold for the amplitude-phase form, then this is a "special case" which at best merits it's own subsection. It is, however, not part of the general definition.

Kind regards, Roffaduft (talk) 06:13, 14 December 2024 (UTC)

I think I see your point. Let's assume I end up agreeing that Eqs 2 and 5 can be defined from the start with complex s(x). Either way we arrive at the same formulas in the end. Starting with complex s(x) you give up the insight provided by the polar form. Creating a separate subsection is better than nothing, but it is less coherent. Or we could make a separate subsection to show that Eq.2 (and 5) are also valid with complex s(x).
--Bob K (talk) 04:25, 14 December 2024 (UTC)

Strictly speaking, I do not think it's true that every complex signal has an amplitude-phase form (using a real sinusoid ad the basic wavelet). In general, the real and imaginary parts come in different phases. With an amplitude-phase form, one always has I think <math<|c_{-n}|=|c_n|</math>. So it's not exactly true that there is no difference between the two cases. (But I dont disagree that, broadly, the specific passage seems unnecessary.) Tito Omburo (talk) 16:20, 13 December 2024 (UTC)

At best it would seem more appropriate to turn the amplitude-phase form in it’s own subsection rather than be part of the general definition of the Fourier series.

But such a subsection could do with some proper references rather than some lecture notes or online summaries.

Then again, I’m not fully convinced the amplitude-phase form is thst relevant in the definition of the Fourier series.

Kind regards, Roffaduft (talk) 16:45, 13 December 2024 (UTC)

Yes, splitting it out makes the most sense to me. Tito Omburo (talk) 16:58, 13 December 2024 (UTC)

I feel like you're missing the point. The premise of the Fourier transform is the complex case. This has nothing to do with the chronology of the edits but with the fact that it's defined that way, based on a proper source. The real case is just a specific example, so is the amplitude-phase form.

On a different note, could you please stop butchering this talk page please? This is highly inappropriate, as per WP:TPO

• Use the [reply] button at the end of a message
• Don't edit someone else's reply, use Template:Talk_quote_inline and copy/pase the part you want to quoute
• Don't add inline references and subsections

Roffaduft (talk) 06:01, 14 December 2024 (UTC)

Thank you... I will give that template a try. When I started my multi-part reply, there were several [reply] buttons, which I tried to use. I expected each one to insert my reply right below, but they did not do that. Furthermore, I could not get the {cite book} template to work. Hopefully the Template:Talk_quote_inline will work better.
Bob K (talk) 13:49, 14 December 2024 (UTC)

Error in Sine-cosine Transform

Equation (1) and equation (3) contradict each other for n=0. In order for equation (3) to hold, an additional factor 1/2 has to be included in the a₀ term in equation (1). In its current version, inserting equation (1) into equation (3) leads for n=0 to the contradiction a₀ = 2 a₀. RedSiskin (talk) 13:16, 21 December 2024 (UTC)

WP:BEBOLD

Kind regards, Roffaduft (talk) 13:56, 21 December 2024 (UTC)

Well, this will affect the entire article (including the long table of special cases), and I don't have the time right now to go through the enitre artile and check all the definitions of the formulas based on eqs. (1) and (3). So rather than messing around with the definitions and potentially creating a large number of other inconsistencies I mentioned this issue in the hope that people who are currently working on this article can put in in a consistent shape. It can be repaired either in Eq. (1) or (3), but it should be checked first what of the (quite extensive) information following these equations is based on which definition. RedSiskin (talk) 14:27, 21 December 2024 (UTC)

Definition

@Tito Omburo, the recent changes I made to the definition were all supported by inline citations to proper sources.

Unfortunately you reverted them without providing a proper explanation or references. Your latest added inlince citation isn’t of any help either, as Edwards (1982) does not sufficiently cover the old definition to the extent that makes your revision appropriate.

Could you please motivate your recent edits?

Kind regards, Roffaduft (talk) 17:30, 21 December 2024 (UTC)

I'm fine with restricting the definition to $L^{1}$ functions on $[0,P]$. I don't love the idea of putting baroque sufficient conditions for convergence in the definition. The Fourier series is a formal series. It does not have to converge. I'm unclear what the problem is with the sources I provided, and it would be nice if you would more clearly explain the reason. Actually, your version also didn't match your cited source in one particular. Tito Omburo (talk) 17:38, 21 December 2024 (UTC)

I see what you mean. If not the (sufficiency) conditions for conversion, what defines a Fourier series? In other words: what distinguishes a Fourier series from any other trigonometric series?

I tried to refrain from placing too much emphasize on the type of function and kept it as general as possible, i.e., a “complex functuon”. Though a “periodic function” would be appropriate as well. However, when talking about “L1 functions” or “distributions”

I feel it may become too specific.

Also, I have never seen the use of the subscript notation $s_{\infty }$ in the definition before, ever. I think it’s inappropriate.

Kind regards, Roffaduft (talk) 17:50, 21 December 2024 (UTC)

This was a problem with the old version before I added the definition, that a "Fourier series" was the same thing as a trigonometric series. But of course the Fourier series must be gotten by "integrating" something. (There are trigonometric series that are not Fourier series in this sense.) Tito Omburo (talk) 17:53, 21 December 2024 (UTC)

I do agree with the start of the definition now, it’s just more accurate. However, the part regarding convergence is still a bit vague to me (even though I do understand what you try to say).

The series does not need to converge, unless it’s “good” (followed by an example of a “good” s), in which case it does converge in a particular way. But it can also converge in other ways, that are sometimes “convenient”..

It’s unnecessarily woolly and confusing. It may be better to be concise/precise in the definition subsection and elaborate on the various forms of convergence in the appropriate subsection later on Roffaduft (talk) 18:07, 21 December 2024 (UTC)

Ok, I've read up on both volumes (1979 and 1982) of Edwards and I still have some questions.

First, on pages 8-9 of Vol.1. it clearly states that "convergence in the sense of distributions" is suggestive of fruitful generalizations of the concept of Fourier series of such a type that the distinction between Fourier series and trigonometric series largely disappears. It suggests in fact the introduction of so-called distributions or generalized functions

So if you say the Fourier series of a function "(or distribution)" is given by the trigonometric series.... then it makes no sense to say directly afterwards that "The series need not necessarily converge". That is, it are the distinct forms of convergence of functions and distributions that merit mentioning them in the first place.

Talking about (the lack of) convergence makes sense if you compare a function to its Fourier series approximation (which is whatsignal analysis and synthesis is all about). For example, when representing an arbitrary signal $f(x)$by the partial sums, then the partial sums may not converge: $\lim _{N\to \infty }s_{N}(f)(x)\neq s(x)$. Great, but that has nothing to do with the actual definition of the Fourier series itself. Roffaduft (talk) 10:49, 22 December 2024 (UTC)

@Tito Omburo You removed the "existence" claim, which followed directly from the Lion (1986) reference. Why?

Kind regards, Roffaduft (talk) 09:51, 23 December 2024 (UTC)

The previous version said "The existence and (pointwise) convergence of $s(x)$ are guaranteed by the Dirichlet sufficiency conditions." But this is nonsense: $s(x)$ is a function. If it did not exist, it could have no Fourier series. The Fourier series exists because we can write it down, but it might not converge. Lion is focused on convergence of that series. Tito Omburo (talk) 10:01, 23 December 2024 (UTC)

Furthermore, you also state that Folland does not write "s(x) = (series)". This is incorrect, Folland very cleary does on page 19-20. Equation (2.5) is based on (2.2), which is then used in his definition of the Fourier series Roffaduft (talk) 10:04, 23 December 2024 (UTC)

Folland asks if "s(x)=(series)". His definition of Fourier series does not write this. Tito Omburo (talk) 10:12, 23 December 2024 (UTC)

He does. equation (2.2) to be exact. Roffaduft (talk) 10:15, 23 December 2024 (UTC)

Folland writes: "We wish to know if f can be expanded in the form..." His Definition does not write this! Tito Omburo (talk) 10:17, 23 December 2024 (UTC)

Folland page 20:

"If $f$ has a series expansion of the form .. $f(\theta )=\sum c_{n}exp(in\theta )$, and if the series converges decently so that term-by-term integration is permissible, then the coefficients .. $c_{n}$ .. are given by (2.5). But now if $f$ is any Riemann-integrable periodic function, the integrals in (2.5).. make perfectly good sense, and we can use them to define the coefficients ... $c_{n}$"

which perfectly fitted the original statement about existence and convergence. To put it crudely, if these (or any other relevant) sufficiency conditions are not met, then $f(\theta )\neq \sum c_{n}exp(in\theta )$.

You could indeed say that the lhs. of the equation should be removed for the most general and "accurate" definition of the Fourier series, but then you shouldn't replace the lhs. with $s_{\infty }$ either (thereby linking the definition of the Fourier series once more with "some version" of $s$ just because it's "convenient") Roffaduft (talk) 10:36, 23 December 2024 (UTC)

Folland says that if a bunch of things (which he does not specify) are true, then $f(x)$ is equal to its Fourier series. That has no bearing on the definition of the Fourier series, which is valid whether those non-specified conditions hold. Tito Omburo (talk) 10:52, 23 December 2024 (UTC)

As I already said, it is possible to refer to a Fourier series by a symbol other than the symbol of the function $s(x)$. You seem to contend that no such notation is possible, which I find quite mystifying. I would use Zygmund's notation, but the capital S is used elsewhere in the article in a way that conflicts with this, and other parts of the article would also need to be rewritten. Finally, the Fourier series is linked to the function s: its coefficients are determined by integrating s against exponentials. So the notation should have an s in it. Tito Omburo (talk) 10:43, 23 December 2024 (UTC)

You seem to contend that no such notation is possible, which I find quite mystifying

You are the one that is hammering on the exact definition from Folland, just stating the summation. Fine, but then don't use $s_{\infty }$ either.

Which I see you've refrained from in your latest edit, thank you. This makes more sense. Roffaduft (talk) 10:52, 23 December 2024 (UTC)

ps. Folland considers riemann integrable functions, not lebesgue integrable functions Roffaduft (talk) 10:55, 23 December 2024 (UTC)

Folland also requires integrability of the function, so I have restored this change. Also, elsewhere in the article, we appear to use the same symbol $s(x)$ for the function and Fourier series. I have restored the article's old notation, which used $s_{\infty }$ for the Fourier series and $s(x)$ for the function. Tito Omburo (talk) 10:15, 23 December 2024 (UTC)4

I have previously explained why I disagree using $s_{\infty }$, so could we please engage in a discussion first before you start forcing your preferences into the article. Roffaduft (talk) 10:18, 23 December 2024 (UTC)

I didn't see any discussion of this. Writing "s(x)=(series)" is wrong and unsupported by the cited source. Tito Omburo (talk) 10:20, 23 December 2024 (UTC)

It is definitely not wrong and writing "s_\infty = (series)" is also not supported by the cited sources. Roffaduft (talk) 10:22, 23 December 2024 (UTC)

What do you mean "it is definitely not wrong"? We do not have $s(x)=\sum c_{n}e^{2\pi inx}$ in general. The function $s(x)$ is a given function, which may or may not be equal to its Fourier series, which may or may not even converge. A different symbol is required to distinguish the series from the function. I used $s_{\infty }$, as we later use $s_{N}$ to refer to the partial sum. Other common notations are $S[s]$, used in Zygmund for instance. Tito Omburo (talk) 10:29, 23 December 2024 (UTC)

Apart from the fact that I don't really like the Zygmund reference, I'm also fine with defining the Fourier series as an operator $S[\cdot ]$ or something. But then it should be introduced as such, rather than leave the reader guessing where the symbol $S[\cdot ]$ suddenly comes from. Roffaduft (talk) 10:43, 23 December 2024 (UTC)

non-Fourier nature of a series

@Tito Omburo I was looking at the trigonometric series article and came across the series:(Edwards 1979, pp. 158–159. / Edwards 1982, pp. 95–96)

$\sum _{n=2}^{\infty }{\frac {\sin {(nx)}}{\log {n}}}$

as an example of a convergent series that is not a Fourier series.

As this series does not necessarily seem to imply that $s$ is neither periodic nor integrable, it got me wondering about what conditions the Fourier coefficients are subject to.

For example, must $c_{n}$ be Lebesgue integrable, must Riemann-Lebesgue "always" hold, must $c_{n}$ belong to $L^{\infty }$, must $c_{n}$ be of sufficient decay, etc.? And if there is such a condition, shouldn't this be part of the definition?

I'd love your opinion on the matter.

Thank you in advance, Roffaduft (talk) 09:33, 29 December 2024 (UTC)