Talk:Fourier–Bros–Iagolnitzer transform
![]() | This article is rated B-class on Wikipedia's content assessment scale. It is of interest to the following WikiProjects: | ||||||||||
|
Merge proposal
[edit]- The following discussion is closed. Please do not modify it. Subsequent comments should be made in a new section. A summary of the conclusions reached follows.
- To not merge on the grounds that each page is distinct and distinctly notable (consensus reached). Klbrain (talk) 18:05, 3 August 2016 (UTC)
I propose that Gabor transform be merged into this page. They seem to be the same thing. The Gabor transform article has only one reference, and it's just some lecture notes, whereas FBI transform has many high-quality references, so I think FBI transform is the more common name for the concept. Quietbritishjim (talk) 15:04, 25 January 2013 (UTC)
No they are not the same thing. Please keep separate. There is a lot of signal processing work that is baed on Gabor filters, and they are well known. Thanks. — Preceding unsigned comment added by 108.175.218.4 (talk) 22:37, 19 August 2013 (UTC)
I second the notion above that they should not be merged, if for no other reason than the inclusion of the constant a, as opposed to pi, which plays a very important role in the Bros-Iagolnitzer theorem as an integration variable. In other words, it would be very hard to prove local analyticity with just a Gabor transform, and ideas of local analyticity aren't very relevant to the study of signals and systems either.
I am also opposed to the merger. The signal processing community often uses the term Gabor transform, but I have never seen the term FBI transform in any signal processing journals. GeorgeSienceNut (talk) 07:10, 22 June 2014 (UTC)
Opposing the proposed merger. I'm researching an engineering project involving audio DSP - which leads me quickly to 'Gabor Transform' and 'Discrete Gabor Transform.' These seem to be well-known terms & concepts for a practical applied technique in DSP. The FBI transform article is pure math, and way more general than anything I need or want to learn about. If I arrived there by following 'gabor transform' I'd be (a) lost and (b) not helped at all even if I could grasp the content. I don't doubt that to a mathematician one concept is a subset of the other, but how many people reading about the Gabor Transform are mathematicians? Might as well have the article on Zero say "additive identity of the complex number field, see Field (Mathematics)." — Preceding unsigned comment added by Spike0xff (talk • contribs) 19:12, 19 May 2016 (UTC)
Incorrect Inversion Formula
[edit]More pressing is the issue that the inverse equation for the so-called FBI transform is simply not true. This can be easily verified by the reader familiar with unit analysis. The constant, a, has units [1/x^2] which means the right summand must have units [1/x]. This is a clear sign the math is in error, for it means a shift in variables: a-> v^2 a, x -> x/v, t-> vt, which should a priori leave the expression for f(0) unchanged, fails to do so, and in fact multiplies the second summand by v. On a related note, the suggestion that the proof of the equation "easily follows" is wholly out of place in this context. If the result so easily follows, then why not prove it instead of providing a sketch of the argument? — Preceding unsigned comment added by 128.151.144.125 (talk) 22:55, 13 October 2013 (UTC)
Update: Having now read the original paper by Iagolnitzer I am quite certain the inversion expression is erroneous. The original paper is very clear and even elegant in its composition and does not rely on Euler operators or Laplacians which only cloud the argument (as well as the mistake!). The proper inverse (given in Iagolnitzer, eq. 6) is exactly the same as the Fourier inverse (as it should be, intuitively), but writing the inverse in that way does not make the local analyticity transparent, so Iagolnitzer introduces a surface integral that is equivalent to the previous integral (by a clever construction involving Stoke's theorem) and it is this expression for which the bound clearly implies analyticity. If the inversion formula had to be corrected, the easiest way would be to simply remove the second summand (multiplying a) so that the inverse is identical to the Fourier inverse. However, there seems to be some muddled effort to give the path-deformed inverse offered by Iagolnitzer, and if this expression is to be included then the entire expression must be changed (for example, the integral over dt^n is actually a surface integral that would include an integration over da). I would change it if that pesky bot wasn't all in my space erasing my edits.
Update Update! I have fixed it. — Preceding unsigned comment added by Levi Greenwood (talk • contribs) 23:21, 15 October 2013 (UTC)
FBI vs Gabor transform
[edit]This article and Gabor transform indeed define the same thing (and should have been merged), but there seems to be a disagreement about the definitions of an FBI transform in the literaure: Folland (1989) section 3, Tataru (1999) equation (1), and Hitrik and Sjöstrand (2015) section 1.3 appear to define the same thing, but Antoine and Trapani (2009) section 8.4.1 and its notes, and Matsuzawa (2002) equation (2.2) seem to define a substantially different object (not just a trivial rescaling like the Gabor transform). Indeed AT comment on the difference between Gabor and FBI in the notes. Ashpilkin (talk) 23:28, 20 March 2018 (UTC)
Comparing and Contrasting the Two
[edit]It is worth pointing out that the "different" definitions of the FBI transform are not substantially different. Moreover, the presence of an adjective in front of FBI transform (such as metaplectic) changes the definition, the intention, and the application. Whether the integrand contains a Gaussian (a.k.a. standard normal distribution) function vs a plane wave function () (or both for that matter) is actually not tremendously different. The action of an (the) FBI transform on a tempered distribution (not just any distribution) is the key. Tempered distributions are the elements of the topological dual vector space of the space of Schwartz functions , and are denoted as . The Fourier transform (presence of only the plane wave in the integrand), of a Schwartz function is another Schwartz function because the Fourier transform is an isometric automorphism on . So to each there is an associated such that . But because the Fourier map is invertible as a cause of being an automorphism, it also makes sense to write . Since is the continuous linear dual to , the Fourier map lifts to , and the Fourier transform of a tempered distribution is a tempered distribution. So the inclusion of a plane wave in the integrand is a matter of nomenclature as to where one starts, and where one wishes to end up. To see that this is also an issue of nomenclature for the Gaussian, one must note that the space of Schwartz functions have a basis given by the Hermite polynomials times a Gaussian. The existence of the basis means that for each there exists constant coefficients () such that for fixed (typically 1 or 1/2), the Hermite polynomials, and .
It should now be clear that the Hermite basis times the Gaussian implies that each element of has a Gaussian function implicitly present. Every element of is also an element of through the dual map embedding, and so is a dense proper subset of . There are thus various different characterization theorems for tempered distributions. Consequently depending upon what aspect is to be emphasized, FBI integrand for tempered distributions does not need to incorporate neither the Gaussian, nor the plane wave explicitly.
The Gabor and FBI articles should not be merged because (as was noted above in the original debate) the Gabor transform has a fixed value of in the exponent and this is for purposes of norm comparison of signals, especially relative to a fixed signal. In contrast, the FBI transform must have a free parameter in order to conduct wave-front analysis. Again as it was pointed out originally, obviously pi does not vary. More precisely though, the article on Gabor analysis is geared more toward practical applications for signals and images, whereas the FBI transform is concerning pure mathematics. The connection between the two being extremely fleeting, namely only for when (a set of measure zero) are they related.
Notice that neither the Gabor Transform nor the FBI transform are a Short-time Fourier transform. The three all being distinct. 146.115.76.93 (talk) 23:52, 27 June 2019 (UTC)