sum() over a dimension crashes when a sibling dimension is empty

[FBumann]

I found this edge case while working on lpspec. I think its better handled here than in lpspecs linopy lane

Note

The following content was generated by AI, on behalf of a user of linopy.

Summing an expression over one dimension raises inside the term reshape when another dimension of the expression has size 0 — even though summing over the empty dimension itself works, and the result (an expression with no terms) is well-defined.

import linopy, pandas as pd

m = linopy.Model()
x = m.add_variables(
    coords=[pd.RangeIndex(4, name='t'), pd.Index([], name='cycle'), pd.Index(['a', 'b'], name='line')],
    name='x',
)
x.sum('cycle').sizes   # Frozen({'t': 4, 'line': 2}) — fine
x.sum('line')          # ValueError: cannot reshape array of size 0 into shape (4,0,newaxis)

Traceback ends in expressions.py, BaseExpression._sum (variables.py:1243 → expressions.py:1633 → expressions.py:2141). Reproduced on 0.9.1.post1.dev3+g09c34dd9d (master).

Where it bites: a PyPSA-style model whose cycle basis is empty (no meshed network) — sum(..., 'line') of a term still carrying the empty cycle dimension. Expected: the same result as LinearExpression.from_constant(m, xr.zeros_like(...)) over the kept coordinates, i.e. {'t': 4, 'cycle': 0} with zero terms — which is what x.sum('cycle') already returns for the mirror case. Presumably the stacking reshape needs the term count taken from the summed dimension rather than inferred with -1, which fails once the array has zero elements.

[FBumann]

Current workaround see fluxopt/lpspec#1365