[BIFT] restrict the q-range for speed and less noisy data

[kif]

Mark suggests to start at the begining of the guinier region: reasonable

But where to stop ? dq*Dmax<2pi ?

[kif]

Mark wants the full noise: great :)

[kif]

Bad idea ... High intensity region is mandatory for proper representation of the data.

[maaeli]

Hi Jérôme,

I am not surprised as the "Dmax" information mostly comes for the small angles. Do you have test data with low-q artifacts? Because there dropping low q might be the game changer. Looking here https://www.sasbdb.org/browse having no artifact seems to be the exception...

Cheers, Martha

[kif]

Hi Martha,

Nice to read you. How is your new life?

I am also disappointed by those poor results. Here are a couple of images of the same BSA dataset, reconstructed from the (B)IFT transformed data whether the region before Guinier's linerar region was discarded or not. Of course, it is always possible that there was an error in my code.

It is not really surprising when considering the Dmax is completely wrong after BIFT transformation:

[maaeli]

Hi Jérôme,

how do you obtain those plots? Because for me, the bift.py do look like superimposed by a zig-zag...

bsa_005_sub.out.zip

Anyways, the second data above looks a lot like smeared out sphere?

Best, Martha

[kif]

I agree with you, the second solution looks like a sense sphere. This can be understood as the algorithm starts with a sphere and an alpha_max parameter which would make this acceptable. Then starts the optimization process. Different scenarii have been investigated. The one currently implemented is:

• One log scan on alpha (between 1/alpha_max and alpha_max) with Dmax=3Rg
• One linear scan on Dmax from 2 to 4 Rg Of course the Guinier region has better to be acceptable from beginning :( The 2D grid scan has been dropped as it was too time consuming. Then a steepest descent (Powell) to find the best alpha/Dmax. The monte-carlo is at the end, it is just to get some statistics around the best solution.

Are you using the provided test-data ? because I got this: (this is the result of freesas testdata/bsa_005_sub.dat)

[maaeli]

Yes, I am using the provided test data... Btw, how do you get the reciprocal space fits?

[kif]

IFT is an ill-posed inverse problem. This means the opposite is trivial, trivial like a dot product :) I wrote plenty of test based on synthetic data, have a look at the test_autorg there is a test_synthetic. https://github.com/kif/freesas/blob/fit_autorg/freesas/test/test_autorg.py#L68 For example to simulate the SAXS curve for spherical particles:

        R0 = 4
        npt = 1000
        I0 = 1e2
        Dmax = 2 * R0
        size = 5000
        r = numpy.linspace(0, Dmax, npt + 1)
        p = distribution_sphere(I0, Dmax, npt)
        q = numpy.linspace(0, 10, size)
        qr = numpy.outer(q, r / pi)
        T = (4 * pi * (r[-1] - r[0]) / npt) * numpy.sinc(qr)
        I = T.dot(p)
        err = numpy.sqrt(I)
        data = numpy.vstack((q, I, err)).T

[maaeli]

Hi Jérôme,

for me BIFT works on the BSA test set, even if I crop the pre-Guinier data: Fit for complete data: Fit for cropped data:

How exactly did you remove the pre-Guinier region?

[kif]

You are right, it looks OK... I don't know what I did implement