13:21:54	 From Mirjam Partovi Dilami : Can we agree that we only dived with the sqrt(N), when we need the uncertainty on the mean?
13:26:24	 From Kimi Kreilgaard : In the problem set It wasn’t clear to me whether RMS meant the root mean square (as wikipedia shortens it) or the root mean square deviation RMSE/RMS which I interpret as the standard deviation.  Can you say beforehand what you mean with RMS in general? I tried to look through the material on week 1 where you use both sd, RMS and RMSD to denote root mean square deviation
13:27:10	 From Kimi Kreilgaard : Thank you :)
13:27:27	 From steenbender : What is a clear way of spotting a Poisson distribution when determining the type of distribution in a problem set, specifically compared top the differences in poisson and binomial
13:27:45	 From steenbender : *to
13:29:05	 From Jaime Caballer-Revenga : I have one question related to the solution to problem 1.3 (exam 2016). May I ask verbally?
13:30:29	 From Mirjam Partovi Dilami : Formular 5.24
13:30:46	 From Jaime Caballer-Revenga : thanks
13:31:19	 From Peter Andresen : When trying to fit a peak in a background, one can calculate a local p-value of the peak one identifies. When calculating the global p-value one needs the number of trials, but I am unsure how to estimate the number of trials in this case? 
13:32:39	 From Peter Andresen : Makes sense, thanks :) 
13:35:44	 From Marcus Nørgaard Weng : Kan en jupyter notebook laves om til en pdf, hvor koden er fjernet, men figurerne er med?
13:35:52	 From Mirjam Partovi Dilami : Do you want the code in .py?
13:36:23	 From Daniel : Can we get the feedback of ProblemSet before the exam?
13:37:28	 From Amalie Paulsen : The code in py or pdf?
13:38:13	 From Daniel : oh thanks!
13:39:08	 From Sebastian Øllegaard Utecht : Should we note in the exam if we are using a function-library developed by multiple persons? For plagiat-reasons i mean.
13:39:14	 From Neus : Where do we have to upload the solution? Digital exams or by email?
13:40:17	 From Neus : thank you!
13:40:29	 From Kimi Kreilgaard : Should our exam pdf contain name and ku-id?
13:42:07	 From Qinyi Chen : How many .py files can we upload?
13:42:26	 From steenbender : Would it be possible to get more prep. Exam set?
13:43:03	 From Neus : Will the recorded lectures be available to watch after the course is finished?
13:43:06	 From Eric Planz : In the solution of problem 2.3 of the problem set (DNA problem, slide 19), you had an uncertainty on the mean besides the calculated RMSE. Where did that uncertainty come from?
13:43:08	 From Lasse Bonn : for chauvenets criterion we discard if 0.5<erfc(zscore) but why do we not do 2*erfc because it should be a two sided consideration surely?
13:43:12	 From Rasmus : In the plotting and fitting example from Week 0, does the code do a LLH fit or a Chi-2 fit? (Since you say that we should state which one we are doing).
13:43:38	 From Nikolay Georgiev : “Test if the mean of βinit is constant as a function of energy.” How do you test that? I never managed to figure that out :)
13:44:13	 From Neus : great! thank you
13:44:59	 From Eric Planz : Ah, okay. Tusind tak1
13:45:01	 From Amalie Paulsen : When you do a fisher, you get a separation (Delta). and in the solution you write, that this number corresponds to how many sigma in gaussian separation. But isn't it how many sigma the two means of the histograms are from each other? or is it the same or am I wrong?
13:46:09	 From Lasse Bonn : great thanks
13:49:02	 From Nikolay Georgiev : Yep!
13:49:12	 From Nikolay Georgiev : so just test weighted mean with chi2
13:49:37	 From Liam Joseph Ward : How do you do an unbinned LLH fit to samples from a distribution?
13:50:48	 From Amalie Paulsen : in units of width? of which histogram?
13:51:00	 From Kasper Primdahl Olesen : Please don't take this question the wrong way :)
When going through the problem set last friday you mentioned some approximate point boundaries to be expected between grades 7, 10 and 12 based on previous years. 
Do you have equivalent information for that near-fatal boundary between 0 and 2?
13:51:19	 From Amalie Paulsen : arh thax yea
13:51:20	 From Amalie Paulsen : yes*
13:52:24	 From Liam Joseph Ward : Okay. Thanks!
13:52:27	 From steenbender : When doing the fisher seperation of the data, in the end of your code for the exercise, you add a integer value to the data, why is that? and must the added values have to be identical for the difference data sets?
13:53:59	 From Neus : For the 2016 exam, I couldn't get the solution value for error velocity contribution on the error in 2.2 (I did std/sqrt(N)). I've got it using stats.sem function. But how can I calculate that by myself?
13:54:43	 From Kasper Primdahl Olesen : tops!
13:54:47	 From Kasper Primdahl Olesen : great answer :)
13:55:04	 From Mirjam Partovi Dilami : Is it ok with 50 points in total, or do we need 50 points only in the exam?
13:55:27	 From jakob : in total
13:55:39	 From Marcus Nørgaard Weng : Neus I had the exact same issue
13:55:47	 From steenbender : Okay got it
13:55:52	 From Sebastian Øllegaard Utecht : Neus: np.std has ddof=0 as standard stats.sem() has ddof=1
13:56:06	 From Bia Fonseca : me too actually, re. error on velocity
13:56:56	 From Neus : ok, thank you!!
13:58:18	 From Lasse Bonn : so we should be using sem?
13:58:44	 From Lasse Bonn : ok
13:58:52	 From Mikkel Liisberg : For the week 3 exercise when calculating the unbinned LLH ratio between a fit with only the background and with/without the background - I never get the same very low LLH values that you get. Is this just a matter of how many numbers you include in the distributions?
13:58:53	 From Jakob Riber Rasmussen : In the problemset when doing the calibration in the end, I tried to do a fit the the exponential decaying thing. But I could not the the p-value right,
13:59:07	 From Jakob Riber Rasmussen : But it "looked" nice
13:59:12	 From Daniel : For Exam 2016, Could you explain the plot and the value C in the plot and uncertainty of that?
13:59:47	 From Daniel : Haha thanks!
13:59:48	 From Jakob Riber Rasmussen : What should I do about it? As I suspect it was the last bins with larger errors screwing it up.
13:59:48	 From peter Bagnegaard : When you answered the question “Test if the mean of βinit is constant as a function of energy.”, you said we should bin the data, find uncertainties on the binnings and fit a constant function. Why not just fit the data directly? Don't we get an error by the squared diference directly?
14:00:46	 From Bia Fonseca : fyi, ive just tested the methods to calculate the std on the velocity with the correct ddof and I only get the same value as the solution with stats.sem. np.std(x,ddof=1) is still off.
14:01:08	 From Jaime Caballer-Revenga : thanks
14:01:12	 From Daniel : I mean the plot of 3.1, sorry
14:01:13	 From Jakob Riber Rasmussen : Yeah makes sense, thanks!
14:01:15	 From Nikolay Georgiev : Is that not sort of overfitting? if we calibrate when it doesn’t fit?
14:01:55	 From Nicolai Ree : In problem set 2016 5.1.4 we fitted a sigmoid type function and the results of the fit are shown in Fig 6 as: χ2 / ndf 21.97 / 11
Prob 0.02462 
p0 −0.3594 ± 0.03175. 
But how can the number of degrees of freedom be equal to 11? Don’t you have 41 data points and then 6 parameters, meaning that the n/f should be equal to 35?
14:17:25	 From Jakob Riber Rasmussen : You answered that
14:19:07	 From Peter Andresen : Will you summarize the curriculum today, or will it be questions only :-) ?
14:19:34	 From Jakob Riber Rasmussen : New question regarding Monte Carlo. 
- In real life how would one arrive at the functions we are to generate random numbers from?
14:22:22	 From Jakob Riber Rasmussen : Regarding the sem.function vs np.std(ddof=1). 
stats.Sem uses: np.std "under the hood" 
"s = np.std(a, axis=axis, ddof=ddof) / np.sqrt(n)"
14:23:07	 From Peter Andresen : sounds gofo
14:23:09	 From Peter Andresen : good
14:23:12	 From Jakob Riber Rasmussen : https://github.com/scipy/scipy/blob/v1.6.0/scipy/stats/stats.py#L2365-L2427
14:23:58	 From Jakob Riber Rasmussen : Ahh okay of course!
14:26:04	 From Jakob Riber Rasmussen : It was just answer, @Bia
14:26:32	 From Bia Fonseca : got it, @Jakob. Thank you =)
14:35:42	 From Julie Kiel Holm : Will the solutions for last week be uploaded?
14:36:22	 From Marcus Nørgaard Weng : I don't understand how you got the answer for problem 2.1 from the 2016 exam. 2*(sigma(r)/r) = sigma(L)/L
14:36:43	 From Marcus Nørgaard Weng : From R = L * r^2/pi
14:36:55	 From Nikolay Georgiev : When are we back?
14:59:56	 From eftsia : thanks :)
15:02:24	 From Marcus Nørgaard Weng : Alright, thanks alot!
15:02:38	 From Jens Edelvang-Pejrup : I think you mentioned that one way to test if your data is gaussianly distributed is to do a Kolmogorov-Smirnof test. It takes two histograms as input, so should we simulate gaussianly distributed numbers and bin them as the second input, or how is the comparison done in practise?
15:04:54	 From Marcus Nørgaard Weng : Jens: This is an example from the pi-estimate task: stats.ks_2samp(pi_dist, stats.norm.rvs(np.mean(pi_dist), np.std(pi_dist), size = len(pi_dist)))
15:06:29	 From Elvira Caroline Jarmbæk Jacobsen  To  Troels Christian Petersen(privately) : I am not completely sure how to set confidence limits - is it just like confidence intervals but with a one-tailed Gaussian? 
For example when fitting peaks: If you get a peak with low significance N = 40, sig_N = 30 , you set the confidence limit that the peak has amplitude below 40+1.65*30 at a 95 % confidence level?
15:07:18	 From MagnusJohansen : @Jens If you want to compare values to the normal distrubution, you could do stats.kstest(x, 'norm').
15:07:41	 From Mona Byberg Michelsen : Camilla, there is an example of this in week 0 on GitHub :)
15:08:15	 From Camilla : Ah thanks! :-)
15:08:22	 From Camilla : I'll look into that :-)
15:08:30	 From Jakob Riber Rasmussen : But be aware that the kstest uses a unit gaussian!
15:09:06	 From Jakob Riber Rasmussen : So it won't work if you have a distribution with e.g. µ =24 +/-3
15:09:38	 From Jakob Riber Rasmussen : Generally the Anderson Darling and the Shaipro Wilks test are better for normality tests
15:09:50	 From Jakob Riber Rasmussen : (From my understanding that is)
15:11:58	 From Marcus Nørgaard Weng : I've got a fitting issue, which I think I have to show you to explain, can I screen share?
15:12:05	 From MagnusJohansen : To use a non-unit Gaussian you can provide it with the corresponding cdf from scipy eg. stats.norm.cdf(x,loc, scale)
15:12:41	 From MagnusJohansen : But I agree with Anderson-Darling being better
15:16:38	 From Daniel : Hi Troels,
15:18:59	 From Daniel : I have a question about p-value. If we use smaller bins, we have smaller frequency, and the uncertainty of each frequency value is relatively large compared to frequency value. And in this case, p-value will be bigger if we have smaller bins. I wonder if this is correct.
15:20:28	 From Daniel : Okay thanks!
15:36:17	 From Daniel : After my last question about bins, I understand you explanation and thanks for that! However, I tried it with my code for 2016 exam 4.1.1 question with different bins. When I used bins = 40, p= 0.69 which is close to the solution; when I used bins = 50, p= 0.97; for bins = 200, p = 1.000. I don't understand why... Maybe I made some mistakes but I cannot see where.
15:38:47	 From Bia Fonseca : i get the same as you, Camilla
15:39:16	 From MagnusJohansen : Try using len(B_sick)-1?
15:41:11	 From MagnusJohansen : I also do not get the same values for the sum of the correlation matrices and thus not the same Fisher weights in that excercise. Looking at the result it seems to be a problem with the C values
15:42:29	 From Emma : In exam 2016 you use Chauvenet’s criterion in 5.2.2, how does it work? How do you set p_reject=0.01?
15:47:25	 From Sebastian Øllegaard Utecht : I have a question regarding 1.3 in the problem set from last year.
15:49:40	 From Marcus Nørgaard Weng : I've got another screen-share question
15:49:58	 From Daniel : I think you just skipped my question Troels, here it is again. After my last question about bins, I understand you explanation and thanks for that! However, I tried it with my code for 2016 exam 4.1.1 question with different bins. When I used bins = 40, p= 0.69 which is close to the solution; when I used bins = 50, p= 0.97; for bins = 200, p = 1.000. I don't understand why... Maybe I made some mistakes but I cannot see where.
15:50:30	 From Daniel : But maybe I need to show you my code? perhaps?
15:50:59	 From Marcus Nørgaard Weng : Oh, then you go before me :)
15:52:53	 From Daniel : Thanks Marcus
16:01:37	 From Bia Fonseca : yes =) agreed
16:01:58	 From Liam Joseph Ward : Agreed! Thank you
16:02:21	 From eftsia : Thank you it has been a great help :)
16:04:51	 From Emma : In exam 2016 solution for 5.2.3, you write If anyone tests it with a (high stat.) Kolmogorov-Smirnov-test… how would you make such a test, when you don’t have 2 histograms?
16:11:56	 From Mikkel Liisberg : P value for unbinned LLH : stats.chi2.sf(fitH0.fval - fitH1.fval, 3)
16:16:57	 From Troels Christian Petersen  To  Emma(privately) : Håber det var svar?