BUG: `DataFrame.apply(axis='columns')` loses pyarrow-backed dtypes / does not preserve dtype_backend

[mattharrison]

Pandas version checks

• I have checked that this issue has not already been reported.

• I have confirmed this bug exists on the latest version of pandas.

• I have confirmed this bug exists on the main branch of pandas.

Reproducible Example

import pandas as pd

# Create DataFrame and convert to pyarrow-backed dtypes
df = pd.DataFrame({
    'year': [2020, 2021, 2022],
    'make': ['Ford', 'Toyota', 'Honda'],
    'city08': [20, 30, 25],
    'highway08': [30, 40, 35],
}).convert_dtypes(dtype_backend="pyarrow")

print("dtypes:", df.dtypes)
# Row-wise apply (observed problem)
res_apply = df.apply(lambda row: (row['city08'] + row['highway08']) / 2, axis='columns')
print("apply result dtype:", res_apply.dtype, "dtype kind:", type(res_apply.dtype))

# Vectorized / expression using pd.col (expected)
combined = df.assign(combined = pd.col('city08') + pd.col('highway08') / 2).combined
print("assign+pd.col dtype:", combined.dtype, "dtype kind:", type(combined.dtype))

Output:

dtypes: year          int64[pyarrow]
make         string[pyarrow]
city08        int64[pyarrow]
highway08     int64[pyarrow]
dtype: object
apply result dtype: float64 dtype kind: <class 'numpy.dtypes.Float64DType'>
assign+pd.col dtype: double[pyarrow] dtype kind: <class 'pandas.ArrowDtype'>

Issue Description

When a DataFrame is converted with convert_dtypes(dtype_backend='pyarrow'), a row-wise DataFrame.apply(..., axis='columns') that computes a numeric result returns a Series built from Python/numpy scalars and does not preserve a pyarrow-backed extension dtype (e.g., float[pyarrow] or other arrow dtypes). The vectorized expression (e.g., using pd.col or column arithmetic) produces a pyarrow-backed result as expected.

Expected Behavior

Preserve dtype when using apply

Installed Versions

INSTALLED VERSIONS

commit : 1a3230d
python : 3.11.13
python-bits : 64
OS : Darwin
OS-release : 24.6.0
Version : Darwin Kernel Version 24.6.0: Mon Jul 14 11:30:40 PDT 2025; root:xnu-11417.140.69~1/RELEASE_ARM64_T6041
machine : arm64
processor : arm
byteorder : little
LC_ALL : None
LANG : None
LOCALE : en_US.UTF-8

pandas : 3.0.0rc0
numpy : 2.4.0rc1
dateutil : 2.9.0.post0
pip : 25.3
Cython : 3.2.2
sphinx : None
IPython : 9.8.0
adbc-driver-postgresql: None
adbc-driver-sqlite : None
bs4 : 4.14.3
bottleneck : None
fastparquet : None
fsspec : None
html5lib : None
hypothesis : None
gcsfs : None
jinja2 : 3.1.6
lxml.etree : 6.0.2
matplotlib : 3.10.7
numba : None
numexpr : None
odfpy : None
openpyxl : None
psycopg2 : None
pymysql : None
pyarrow : 22.0.0
pyiceberg : None
pyreadstat : None
pytest : None
python-calamine : None
pytz : 2025.2
pyxlsb : None
s3fs : None
scipy : 1.16.3
sqlalchemy : None
tables : None
tabulate : None
xarray : None
xlrd : None
xlsxwriter : None
zstandard : None
qtpy : None
pyqt5 : None

[mattharrison]

Note that this preserves type info:

(df
 .loc[:, ['city08', 'highway08']]
.apply('mean', axis='columns')
)

[rhshadrach]

Thanks for the report! Proposals are more than welcome, but I do wonder if just documenting the current behavior would be sufficient. The issue is that pandas lacks the data to determine when to wrap the scalar results back into a PyArrow array or not. It seems a clear cut case we may try to do so is when the input DataFrame has all pyarrow columns, but we may not be able to an then need to return something like object dtype. I fear this kind of logic can find ways to rise to unexpected behavior in edge cases.

Note that [using 'mean'] preserves type info

I don't think this is relevant; by providing mean, pandas knows the operation being performed. When supplying a UDF, that information is lacking.