BUG: nanmax/nanmin, 'initial' argument misbehaves #31087

Activity

1. rgbmrc

   added

   00 - Bug

   on Mar 28, 2026
2. rgbmrc

   changed the title [-]BUG: nanmax/nanmin do not ignore initial=nan[/-] [+]BUG: nanmax/nanmin, 'initial' argument misbehaves[/+] on Mar 28, 2026
3. WarrenWeckesser commented on Apr 1, 2026

   

   WarrenWeckesser

   on Apr 1, 2026

   Member

   More actions

   Here's another inconsistency, where the behavior depends on whether the first argument is a Pyhon list or a NumPy array. I expect the first example (pointed out by @rgbmrc in the opening comment) to behave like the second.

   In [12]: np.__version__
   Out[12]: '2.5.0.dev0+git20260330.f91c8a8'
   
   In [13]: np.nanmax([], initial=1.0)
   <ipython-input-13-8564a601b1ff>:1: RuntimeWarning: All-NaN axis encountered
     np.nanmax([], initial=1.0)
   Out[13]: np.float64(nan)
   
   In [14]: np.nanmax(np.array([]), initial=1.0)
   Out[14]: np.float64(1.0)
   
4. Whatsonyourmind commented on Apr 3, 2026

   

   Whatsonyourmind

   on Apr 3, 2026

   More actions

   The issue is that `nanmax` replaces NaN values with `-inf` (the identity for max) before delegating to the underlying `maximum.reduce`. When `initial=nan` is provided, the replacement logic interacts incorrectly with the `initial` parameter.

   The semantics should be:

   1. Replace all NaN in the data with the identity element (`-inf` for max, `+inf` for min)
   2. Then call `reduce` with the user-provided `initial`

   So `nanmax([nan, 1.0], initial=nan)` should compute `max(-inf, 1.0, initial=nan)`. But since `max(anything, nan) = nan` in IEEE 754, the NaN initial poisons the result.

   For the empty-slice case: `nanmax([], initial=1.0)` replaces NaN (none present), then computes `max(initial=1.0)` over an empty array, which should return 1.0. The warning "All-NaN axis encountered" is incorrect here because the slice isn't all-NaN -- it is empty, and the user explicitly provided a fallback via `initial`.

   The fix should distinguish three cases:

   • Non-empty, has non-NaN: return the actual max (current behavior, works)
   • Non-empty, all-NaN with initial: return `initial` if it is not NaN, else NaN (with warning)
   • Empty with initial: return `initial` unconditionally, no warning

   The `initial` parameter should be treated as a sentinel/fallback value, not as another data point subject to NaN replacement.