Require positive bounds for log-scale parameters

[dweindl]

Previously, bounds were only checked for non-negativity.

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 76.09%. Comparing base (74ac20d) to head (2f92af9).

Additional details and impacted files

@@           Coverage Diff            @@
##           develop     #259   +/-   ##
========================================
  Coverage    76.09%   76.09%           
========================================
  Files           36       36           
  Lines         3233     3233           
  Branches       786      786           
========================================
  Hits          2460     2460           
  Misses         568      568           
  Partials       205      205           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[dilpath]

Fine for me but it means tools cannot implement model selection via regularization with L1 (admittedly nontrivial on log-scale), because that would contradict the nonzero lower bound.

[dweindl]

Fine for me but it means tools cannot implement model selection via regularization with L1 (admittedly nontrivial on log-scale), because that would contradict the nonzero lower bound.

Thx. No strong opinion. I'm fine with either:

• allowing 0, but making that explicit in the specs and silencing all warnings related to log(0) in the library
• excluding 0

Other opinions?

[dilpath]

Also no strong opinion.

We support nominal values of 0 regardless of parameter scale and bounds, so no issue for PEtab Select.

Since excluding 0 would be a possibly-breaking change, I would be in favor of allowing 0.

[dweindl]

Since excluding 0 would be a possibly-breaking change, I would be in favor of allowing 0.

Okay, I guess we'll leave it as is for PEtab v1.

For PEtab v2 we'll probably allow 0. For further discussion -> PEtab-dev/PEtab#579

[dilpath]

Sampling, e.g. initial start vectors for multi-start optimization, when the lower bound is 0 and the parameter is log10-scaled, would mean sampling in (-inf, upperBound], which isn't sensible.

We could require that, if a user specifies a lowerBound of 0 for a log10-scaled parameter, then they should also specify initializationPriorParameters such that the initial sampling of start vectors can occur on parameterScale within finite bounds.

A PEtab problem where samples are drawn uniformly from non-finite bounds, should be considered invalid.

from discussion with @dweindl

[dweindl]

Closing in favor of #278