Test-0.4.0 #57
Test-0.4.0 #57
Conversation
|
Check out this pull request on Review Jupyter notebook visual diffs & provide feedback on notebooks. Powered by ReviewNB |
| @@ -0,0 +1,648 @@ | |||
| { | |||
There was a problem hiding this comment.
What is the purpose of splitting the simulation into interleaving time series and computing the cross-correlation here?
Reply via ReviewNB
| @@ -0,0 +1,648 @@ | |||
| { | |||
There was a problem hiding this comment.
These numbers don't seem to match the numbers appearing below in the keys of the read-in data?
Reply via ReviewNB
| @@ -0,0 +1,648 @@ | |||
| { | |||
There was a problem hiding this comment.
| @@ -0,0 +1,648 @@ | |||
| { | |||
There was a problem hiding this comment.
Maybe I'm looking too hard, but it seems that in each of the panels, there are short sections of realization number where all four standard deviations have estimates that are above/below the truth (eg. at about N=110 in the left-hand plot). Of course, all things statistical are possible, but is this considered highly unlikely (it looks like it does this for ~5-6 realizations in a row). Are you sure there are no correlations between runs?
Reply via ReviewNB
| @@ -0,0 +1,648 @@ | |||
| { | |||
There was a problem hiding this comment.
Again, it seems like the different sigma are correlated -- do they all maybe use the same initial seed? Is this intended? Is there a good reason that ((0,11),(0,11)) should have significantly higher fractional difference?
Reply via ReviewNB
| @@ -0,0 +1,648 @@ | |||
| { | |||
There was a problem hiding this comment.
To me, after about 400 realizations, it looks like the residuals are roughly flat, rather than converging (maybe slightly). Is there an argument for how fast we should expect it to converge? Also, the plot legend seems to be discrepant with the caption (sigma=1mK vs. sigma = 0.5).
Reply via ReviewNB
There was a problem hiding this comment.
Thanks @taylordb, this is great work!
Can you confirm whether the notebook tests a case of P_N < P(k) and also P_N > P(k), as suggested in the corresponding issue #27? I think the main improvement to this would be if you were able to also plot some kind of analytic expectation of the decrease in noise over realizations.
Also, an idea (maybe not a good one) might be to plot the whole distribution of powers (for times and delays) at each realization (with some kind of transparency for each dot).
This notebook validates the hera_pspec pipeline's ability to recover the mean power from a flat-k Gaussian delay spectrum from an ensemble of simulations with random Gaussian noise added to the visibilities. We extract the mean power estimate from each simulation using the hera_pspec pipeline and perform an ensemble average on the resulting power and compare to the mean from a flat-k Gaussian delay spectrum without added noise. The percent difference averaged over 500 simulations is found to be less than 0.04%.
Fixes #27