Getting `TypeError: Dask sparse arrays do not support zero-centering (yet)` when `zero_center=False`

[mumichae]

Please make sure these conditions are met

• I have checked that this issue has not already been reported.
• I have confirmed this bug exists on the latest version of scanpy.
• (optional) I have confirmed this bug exists on the main branch of scanpy.

What happened?

When using a dask array in .X, zero_center does not seem to get handled correctly in sc.pp.pca. When zero_center=False there is an error, while for zero_center=True the function computes without error.

Minimal code sample

import scanpy as sc
import dask.array as da
import anndata as ad
from scipy.sparse import csr_matrix

X = da.random.random((1000, 500)).map_blocks(csr_matrix)
adata = ad.AnnData(X=X)

sc.pp.pca(adata, zero_center=True) # does not produce error
sc.pp.pca(adata, zero_center=False) # produces error
sc.pp.pca(adata, zero_center=None) # produces error

Error output

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
Cell In[23], line 9
      6 X = X.map_blocks(scipy.sparse.csr_matrix)
      7 adata = ad.AnnData(X=X)
----> 9 sc.pp.pca(adata, zero_center=False)

File [...]/envs/scanpy/lib/python3.10/site-packages/scanpy/preprocessing/_pca/__init__.py:352, in pca(***failed resolving arguments***)
    350 if issparse(X._meta):
    351     msg = "Dask sparse arrays do not support zero-centering (yet)"
--> 352     raise TypeError(msg)
    353 from dask_ml.decomposition import TruncatedSVD
    355 svd_solver = _handle_dask_ml_args(svd_solver, TruncatedSVD)

TypeError: Dask sparse arrays do not support zero-centering (yet)

Versions

dask    2024.7.1
----    ----
pillow  11.1.0
pyarrow 18.1.0
importlib_metadata      8.5.0
PyYAML  6.0.2
pickleshare     0.7.5
pytz    2024.1
zarr    2.18.3
backports.tarfile       1.2.0
sparse  0.17.0
texttable       1.7.0
tornado 6.4.2
psutil  6.1.1
jedi    0.19.2
six     1.17.0
llvmlite        0.43.0
jaraco.context  5.3.0
leidenalg       0.10.2
zipp    3.21.0
zict    3.0.0
ipython 8.31.0
asciitree       0.3.3
lz4     4.3.3
parso   0.8.4
pure_eval       0.2.3
joblib  1.4.2
asttokens       3.0.0
cytoolz 1.0.1
MarkupSafe      3.0.2
decorator       5.1.1
stack_data      0.6.3
toolz   1.0.0 (toolz: 1.0.0, tlz: 1.0.1)
wcwidth 0.2.13
setuptools      75.6.0
igraph  0.11.8
kiwisolver      1.4.7
louvain 0.8.2
cycler  0.12.1
sortedcontainers        2.4.0
jaraco.collections      5.1.0
tblib   3.0.0
executing       2.1.0
tqdm    4.67.1
prompt_toolkit  3.0.48
msgpack 1.1.0
distributed     2024.7.1
locket  1.0.0
numba   0.60.0
h5py    3.10.0
jaraco.text     3.12.1
numcodecs       0.13.1
jaraco.functools        4.0.1
natsort 8.4.0
python-dateutil 2.9.0.post0
zstandard       0.23.0
dask-expr       1.1.9
----    ----
Python  3.10.17 | packaged by conda-forge | (main, Apr 10 2025, 22:19:12) [GCC 13.3.0]
OS      Linux-4.18.0-553.33.1.el8_10.x86_64-x86_64-with-glibc2.28
CPU     128 logical CPU cores, x86_64
GPU     No GPU found
Updated 2025-05-26 13:34

[flying-sheep]

In your versions list, I don’t see scanpy at all. Which scanpy version is this?

[mumichae]

Strange, I didn't notice that it didn't show in the logging command. The scanpy version is 1.11.1 from conda, and anndata is 0.11.4

[flying-sheep]

OK, so this is just a problem with the error message, it should read

zero_center=False is not supported for sparse Dask arrays (yet)

See warning box in https://ml.dask.org/modules/generated/dask_ml.decomposition.TruncatedSVD.html

Issue: dask/dask-ml#123

[mumichae]

Thanks, I think I got confused which mode was not supported for sparse matrices.

Might be good to also add update the documentation of the zero_center parameter. Currently the following suggests that using zero_center=False should be preferred for sparse data, but in reality isn't supported at the moment

scanpy/src/scanpy/preprocessing/_pca/__init__.py

Lines 104 to 107 in 5bfa840

	If `False`, omit zero-centering variables
	(uses *scikit-learn* :class:`~sklearn.decomposition.TruncatedSVD` or
	*dask-ml* :class:`~dask_ml.decomposition.TruncatedSVD`),
	which allows to handle sparse input efficiently.

[flying-sheep]

Great point. Our PCA/TruncatedSVD implementation is pretty complex at this point.

Ignoring chunks and backwards compatibility, the following is basically true, except that only some svd_solvers handle sparse data:

	zero_center=True	zero_center=False
np.ndarray/ scipy.sparse.sp{array,matrix}	sklearn.decomposition.PCA	sklearn.decomposition.TruncatedSVD
dense dask.Array	dask_ml.decomposition.PCA	dask_ml.decomposition.TruncatedSVD
sparse dask.Array	scanpy’s custom PCAEighDask	🆘