Data Weightings in EMTF

[kkappler]

There are several data weighting schemes floating around: Edfwts, huber weights, segment weights and channel weights.

In the matlab codes these are variously, and inconsitently referred to as W, wt, and wtt and it is seldom clear which weights are being applied.

We need

1. Definitions for each of these weights types
2. A flow diagram that shows where in the process each weight is computed,
3. Nomenclature to use while coding that makes it clear which weights are being used where
4. Implementation of these weights (especially those used in TTFestBand.m)

[kkappler]

Regarding 1 EDF weights are implemented to pre-weight the data right before input to regression. There is a draft explanation for these weights in the appendix of the Phase1 FTR. Huber weights: we will reference Gary's papers,but a paragraph explaining that this is warranted. Here is a reference http://users.stat.umn.edu/~sandy/courses/8053/handouts/robust.pdf Segment weights: This is a 1-dimensional array of the same length as the time series' of Fourier coefficients. It can be used to downweight bad segments across all channels. It is not 100% clear how this works w.r.t. RR processing, i.e. is it by station, or arraywide? @egbert-g let's discuss Channel Weights: Like segment weights but multidimensional. Allows individual segments from each channel to be weighted independently

[kkappler]

Before closing this, the following should be addressed

• [ ] Add a test that compares python edfwts vs an octave call to the same method in Gary's matlab code. Test use cases:

  • [ ] With / Without Nan
  • [ ] With / Without RR Here is a link to the original Edfwts.m file The input to the compute_weights() function X is a numpy array, shape (2, n_observations), add to doc when working on this issue

• [ ] The edfweights parameters should be added to the config

• [ ] A flow diagram showing the weights should be added to the documentation.