[Ifeffit] Question about PCA

Thu Apr 25 09:48:45 CDT 2019

Hi Joselaine,

I'd like to update my answer, especially in regard to what Athena is doing.
As it turns out, I can very closely reproduce Athena's results with a
simple, straightforward implementation of PCA, and I think it might help
explain the differences you are seeing too.

In Python, for data that is arranged as  `data` with shape (Nspectra,
Nenergy), Athena appears to do

import numpy as np
def pca(data):
    data = (data - data.mean(axis=0)) / data.std(axis=0)
    cor = np.dot(data.T, data) / data.shape[0]
    evals, var = np.linalg.eigh(cor)
    iorder = np.argsort(evals)[::-1]
    evals = evals[iorder]
    evec = np.dot(data, var)[:, iorder]
    return evec, evals

which is to say: Athena *is* normalizing the data, but it is normalizing
along the first axis, not removing the mean spectra.   This gives
eigenvalues that sum to Nspectra and the same component eigenvectors as
Athena shows.  The first component is +/- 1 * mean_spectra.

That appears to be different from what scikit-learn is doing, which removes
the mean spectra (along the other axis).  The eigenvectors the scikit-learn
reports are off by one from the ones reported by Athena.

I suspect the difference you see might also be due to how the data matrix
is ordered, but I don't know why your eigenvalues sum to 160*11 -- I might
have suspected they add to 160 or 11....

I'm not sure what the right approach should be, but I'd like to figure this
out soon.  I think adding Malinowski's IND and perhaps F would be fine
statistics, but I only see how to do that if `s=c` (that is the total
number of spectra).

--Matt

On Wed, Apr 24, 2019 at 10:59 PM Matt Newville <newville at cars.uchicago.edu>
wrote:

> Hi Joselaine,
>
>
> On Wed, Apr 24, 2019 at 8:40 PM Joselaine Cáceres gonzalez <
> joselainecaceres at gmail.com> wrote:
>
>> Hi Matt, thank you for your answer!. The references I have about
>> Malinowski´s work and some applications are:
>>
>> Malinowski, E.R., *Theory of error in factor analysis.* Analytical
>> Chemistry, 1977. *49*(4): p. 606-612.
>>
>>  Malinowski, E.R., *Theory of the distribution of error eigenvalues
>> resulting from principal component analysis with applications to
>> spectroscopic data.* Journal of Chemometrics, 1987. *1*(1): p. 33-40.
>>
>> Malinowski, E.R., *Statistical F-tests for abstract factor analysis and
>> target testing.* Journal of Chemometrics, 1989. *3*(1): p. 49-60.
>>
>> Malinowski, E.R., *Adaptation of the Vogt–Mizaikoff F-test to determine
>> the number of principal factors responsible for a data matrix and
>> comparison with other popular methods.* Journal of Chemometrics, 2004.
>> *18*(9): p. 387-392.
>>
>> McCue, M. and E.R. Malinowski, *Target Factor Analysis of the
>> Ultraviolet Spectra of Unresolved Liquid Chromatographic Fractions.*
>> Applied Spectroscopy, 1983. *37*(5): p. 463-469.
>>
>> Beauchemin, S., D. Hesterberg, and M. Beauchemin, *Principal Component
>> Analysis Approach for Modeling Sulfur K-XANES Spectra of Humic Acids.*
>> Soils Science Society of America Journal, 2002. *66*: p. 83-91.
>>
>> Wasserman, S.R., et al., *EXAFS and principal component analysis: a new
>> shell game.* Journal of Synchrotron Radiation, 1999. *6*: p. 284-286.
>>
>
> OK, thanks, but what I was more interested in was what is the math for
> these tests that you are doing?
>
> I think I don't understand what you mean by "eigenvalues explain exactly
>> the amount of variance ... but they are different".  Can you clarify?
>> Giving an actual example might help.
>>
>> Here are results obtained by ATHENA, the factional variance explained by
>> each eigenvalue is calulated by dividing the eigenvalue between the sum of
>> them all, right?:
>>
>
> Yes, that is my understanding too.
>
>>
>> #  Eignevalues    Variance    Cumulative variance
>> 1 8,864394 0,80585 0,805854
>> 2 1,227578 0,11160 0,917452
>> 3 0,708334 0,06439 0,981846
>> 4 0,129478 0,01177 0,993617
>> 5 0,045127 0,00410 0,997719
>> 6 0,012229 0,00111 0,998831
>> 7 0,009464 0,00086 0,999692
>> 8 0,001489 0,00014 0,999827
>> 9 0,000989 0,00009 0,999917
>> 10 0,000617 0,00006 0,999973
>> 11 0,000298 0,00003 1
>>
>> Here are results obtained with matrix calculator for the same data:
>>
>
> Yes.  Note that for Athena, Sum Eigenvalues = 11, the number of spectra.
>
> What is "matrix calculator"?
>
>> Eigenvalue Explained Variance Cumulative variance
>> 1418,3057 0,80586 0,80586
>> 196,4118 0,11160 0,917453
>> 113,3331 0,06439 0,981846
>> 20,7162 0,01177 0,993617
>> 7,2205 0,00410 0,997719
>> 1,9566 0,00111 0,998831
>> 1,5144 0,00086 0,999692
>> 0,2381 0,00014 0,999827
>> 0,1583 0,00009 0,999917
>> 0,0987 0,00006 0,999973
>> 0,0477 0,00003 1,000000
>>
>
>
> Here, Sum Eigenvalues = 1760.  = 160*11.
>
> The explained variance looks essentially identical.  The eigenvalues
> differ by a constant scale factor that happens to be very close to 160.
> Without seeing your data, I might ask a) are there 160 energy points in the
> matrix you are using? and b) is the integral or the sum of the mean specta
> = 160?
>
> FWIW, when I compare scikit-learn PCA (as used in Larch) with Athena, I
> get answers that are significantly different, and by much more than just a
> scale factor.  I am pretty sure that this is because scikit-learch
> For a case where there are 9 input spectra, Athen gives eigenvalues that
> sum to 9, whereas scikit-learns' eigenvalues sum to 2.3....  I don't know
> where that difference comes from.
>
>
>> The eigenvalues are used then to evaluate the function IND and F test,
>> and depending on the values of eigenvalues, function IND reach a minimum
>> value when the set of primary components are separated from the secondary
>> ones that just explained experimental errors (in the equations lambda are
>> the eigenvalues, r the numbers of rows, c columns, n the number of primary
>> components):
>> [image: image.png]
>>
>>
>> The results obtained with the two sets of eigenvalues are diferent but
>> they reach the minimum in the same n. The F test also gives me similar
>> levels of significance for the two sets, but I do not undestand why I´m not
>> hable to find the same eigenvalues that ATHENA does.
>>
>
> Yeah, I sort of doubt that any of those tests will give a different value
> for 'n' (significant components) based on the scale of the eigenvalues
> themselves.  That is, I think you can use "fractional explained variance"
> (or "explained variance ratio").
>
> I don't see an actual write-up where the formula for IND comes from.  But
> I do not quite get what `s` is -- is that `c-n` ?  I'd also be happy to try
> to add Malinowski's F test to the reporting for Larch's PCA analysis, but I
> don't quite understand it.  The sum is from n+1 to s or n+1 to c (as with
> RE)?  What is `j` in the demoninator (isn't that outside the sum?).
>  Similarly, I'd be happy to report SPOIL if I knew a workable definition...
>
>
> By the way, I tested the possibility you told to not divide the data by
>> the standard deviation and still couldn´t find the same
>>
> eigenvalues.
>>
>
> I am almost certain that Athena is not removing the mean -- that might
> help explain the issue too.
>
> Hope that helps.  I'd like to understand these statistics well enough to
> report them.
>
> --Matt
>
>

-- 
--Matt Newville <newville at cars.uchicago.edu> 630-252-0431
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://millenia.cars.aps.anl.gov/pipermail/ifeffit/attachments/20190425/5306582a/attachment-0001.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: image.png
Type: image/png
Size: 44705 bytes
Desc: not available
URL: <http://millenia.cars.aps.anl.gov/pipermail/ifeffit/attachments/20190425/5306582a/attachment-0001.png>