Should a user supply data in its original or log-transformed form

[pgkirsch]

GPfit has always taken log-transformed data as its input, which adds a couple lines to a user's script. Is there a good reason for this or should it take untransformed data as the input given that is ultimately what an end user cares about? @whoburg particularly curious if there was a strong reason for wanting to do it this way originally?

[whoburg]

@pgkirsch here's my argument for taking in log-transformed data

The fit will only be successful if the data are "approximately log-convex". Allowing users to supply untransformed data might give the wrong impression, as if GPfit can handle any data one might throw at it. Taking in log-transformed data also eliminates the possibility that the data can't even be log-transformed (negative values), and maybe encourages users to at least look at the log-transformed version of their data.

That's my pitch. willing to hear other views.