Talk:Matching pursuit

what is this in the formula?
sorry, what's the _0? Is this the p-norm with p = 0? this is from

$$ \lambda ||s||_0 $$

thanks. --24.130.24.6 (talk) 19:58, 15 September 2008 (UTC)

reference vector of functions
I don't want to make any errors, but I think "reference vector A_j (the dictionary)" could be changed to "reference vector of functions A_j", as there doesn't seem to be much dictionary structure (lookup from j to A_j?). Also, perhaps it could be considered making A_j normalized, and getting rid of the ||A_j|| below. --gatoatigrado 20:13, 15 September 2008 (UTC)

algorithm looks off...
The algorithm looks a bit weird, because it never involves the original signal x again. All of the parameters for refinement are scalars, not functions. This contrasts with the link below. I will look into it further. Apologies if this is a false alarm. http://www.scholarpedia.org/article/Matching_pursuit#Matching_pursuit_algorithm.

cheers, gatoatigrado 20:39, 15 September 2008 (UTC)


 * The algorithm is correct. Rf_n is not a scalar, it is the residual vector. 76.17.115.184 (talk) 20:19, 15 August 2012 (UTC)

Main disadvantage missing
The article says that the main disadvantage of matching pursuit is the encoder complexity. This may be true, but what I think the main disadvantage is, is the fact that the vectors g (i.e. the columns of D) have to be orthogonal w.r.t. each other, to constitute a valid basis. If this is verified the algorithm converges, otherwise unexpected results are obtained. I'd like to receive a feedback before modifying the page. Squalho (talk) 16:26, 5 June 2009 (UTC)


 * That does not sound correct. Matching pursuit is specifically designed for situations in which the dictionary is overcomplete, as explained in the article and its references. --Zvika (talk) 18:12, 5 June 2009 (UTC)


 * The case of the dictionary being an orthogonal basis, or even an under-complete orthogonal set, is indeed pretty trivial, since you can just compute all the coefficients and pick and the largest ones trivially. Zvika is right that the matching pursuit scheme is designed primarily for over-complete bases, though it is sometimes used for non-orthogonal under-complete bases as well.  I'm not sure what you mean by "unexpected results are obtained."  Dicklyon (talk) 01:58, 6 June 2009 (UTC)

orthogonal matching pursuit
The following text appears in the article on least squares spectral analysis.

''matching pursuit with post-backfitting[12] or orthogonal matching pursuit.[13]

12.^ a b Pascal Vincent and Yoshua Bengio (2002). "Kernel Matching Pursuit". Machine Learning 48: 165–187. doi:10.1023/A:1013955821559. http://www.iro.umontreal.ca/~vincentp/Publications/kmp_mlj.pdf.

13.^ Y. C. Pati, R. Rezaiifar, and P. S. Krishnaprasad, "Orthogonal matching pursuit: Recursive function approximation with applications to wavelet decomposition," in Proc. 27th Asilomar Conference on Signals, Systems and Computers, A. Singh, ed., Los Alamitos, CA, USA, IEEE Computer Society Press, 1993.''

It would be helpful if someone who understands this could incorporate this additional information in the article about matching pursuit.

—Preceding unsigned comment added by Skysong263 (talk • contribs) 19:44, 23 April 2010 (UTC)

citation needed
There is no citation under "Orthogonal Matching Pursuit" for the statement

In fact, this algorithm solves the sparse problem : (problem), with the L0 pseudo-norm equal to the number of nonzero elements of x.  — Preceding unsigned comment added by 76.17.115.184 (talk) 03:40, 19 November 2011 (UTC)

The claim is in fact false. The minimization problem posed is known to be NP-Hard. OMP only produces an approximation to the solution. — Preceding unsigned comment added by 132.250.22.10 (talk) 19:29, 1 August 2012 (UTC)

Normalization of a_n
It seems to me that "The Algorithm" assumes the dictionary entries are unity-norm. I think a_n should be defined as

$$ a_n\,\leftarrow\,{\langle R_n, g_{\gamma_n}\rangle \over \| g_{\gamma_n}\|^2}  $$;

... but I lack the confidence to make the edit without an "amen". --Mborg (talk) 01:51, 18 February 2013 (UTC)


 * When I first read the article, I got the impression the functions $$ g_{\gamma_n} $$ were intended to be drawn from a normalized set - equivalent to setting $$ \| g_{\gamma_n}\|^2 = 1$$, but I cannot find an explicit reference in the current version. (I don't know if there has been an edit since I first saw it, or if I got the impression from other, linked pages, or if it's something I picked up from previous experience.) Maximizing the inner product suggests that all the $$ g_{\gamma_n} $$ should be normalized. Also, since the inner product is linear in $$ g_{\gamma_n} $$, either the functions should be normalized, or $$ a_n\,\leftarrow\,{\langle R_n, g_{\gamma_n}\rangle \over \| g_{\gamma_n}\|}  $$. Note I differ from Mborg on the power in the denominator. I do not have a handy reference. Also, there seems to be some confusion in notation. The functions are initially described as $$ \gamma_n $$, but in the article they are referred to as "coefficients" $$ g_{\gamma_n} $$. I suggest removing the "coefficients" $$ g_{\gamma_n} $$ in favor of "functions" $$ \gamma_n $$ and specifying that the latter be normalized. SMesser (talk) 15:01, 28 February 2014 (UTC)


 * You need to divide by $$\|g_{\gamma_n}\|^2$$ because you end up multiplying by $$g_{\gamma_n}$$ twice: once in the inner product and again in the reconstitution on the next line. If you change the scale of $$g_{\gamma_n}$$, you need to divide by the norm squared to get the same approximation and residual.
 * It looks like the other questions on this page have mostly been addressed. I think that clears things up well enough to remove the "in need of expert attention" tag. Motmahp (talk) —Preceding undated comment added 23:25, 20 June 2016 (UTC)

PCA
"In the case that the vectors in $$D$$ are orthonormal instead of redundant, then MP is a form of principal component analysis"

This is not true. PCA is a method for finding an optimal matrix $$D$$. Projecting a vector onto an arbitrary orthonormal $$D$$ has nothing to do with PCA. Keleti (talk) 10:46, 25 January 2018 (UTC)

Original Algorithm from Accelerator Physics
There is an algorithm called MICADO that was developped at CERN in the 70ies which does imho the same thing as the OMP. It can be found here: http://inspirehep.net/record/84507?ln=en

Should it be mentioned? — Preceding unsigned comment added by TshayDee (talk • contribs) 11:12, 7 January 2020 (UTC)