window_size 1 and 2 doesn't seem to work about stumpy HOT 2 CLOSED

Thanks for the detailed explanation Sean. I didn't have the time to look deeper in paper's math or code. As a user, I think both options (warning or ValueError) are fine as long as I am given notice of this limitation. From software engineering angle, ValueError may be a better option though. If this wasn't faced during exploratary/development phase, warning may get lost in a complex program utilizing STUMPY where m is computed dynamically. Feel free to close the issue whenever you like.

from stumpy.

@pembeci Thanks for pointing this out and I can confirm this finding.

It looks like the issue comes up when the z-normalization happens. In the case of m=1, z-normalization doesn't make any sense since the standard deviation for an array with only a single element will always be zero and, therefore, you get a divide by zero situation. So, m should never be equal to 1.

In the case of m=2, the mean and standard deviation for any given subsequence are identical and so the z-normalization for any sequence will either be [-1., 1.] or [1., -1.]. Thus, the z-normalized Euclidean distance will be (very likely) zero between any subsequence and its nearest neighbor (assuming that the time series is large enough to contain both scenarios).

In short, this isn't so much a fault of STUMPY but of the ambiguity of what z-normalization means for toy examples where m <= 2. Values of m <= 2 must be avoided. The situation is fine when the m >= 3 or when m and the length of the array gets bigger and bigger.

As a user, would you mind providing some feedback as to how you might expect these scenarios to be handled?

In the simplest case, we can add a check and always enforce that m >= 3 and print a logging warning to let the user know that the value of m must be greater than or equal to 3. On second thought, it probably makes more sense to raise a ValueError. I'd love to hear your thoughts.

from stumpy.