[BUG] spatial_autocorr() returns unbounded Moran's I values (e.g. inf) unlike Squidpy equivalent

[lukasadam]

Describe the bug

The spatial_autocorr() function in rapids-singlecell returns unbounded Moran’s I values, including inf, which is inconsistent with the expected behavior and the reference implementation in squidpy. This behavior occurs for certain samples and breaks downstream interpretation of spatial autocorrelation.

Steps/Code to reproduce bug

# Using squidpy 1.6.5
# Moran's I from squidpy.gr.spatial_autocorr()
Sample: block3, Moran's I min: -0.0014, max: 0.8894
Sample: block7, Moran's I min: -0.0015, max: 0.8727
Sample: block2, Moran's I min: -0.0014, max: 0.8937
Sample: block9, Moran's I min: -0.0013, max: 0.8706

# Using rapids-singlecell 0.12.7
# Moran's I from rapids_singlecell.gr.spatial_autocorr()
Sample: block3, Moran's I min: -0.0014, max: 0.8903
Sample: block7, Moran's I min: -6.80e-06, max: inf
Sample: block2, Moran's I min: -6.43e-17, max: inf
Sample: block9, Moran's I min: -4.17e-32, max: inf

When I look at individual features (for the same sample), it seems to break for features with (very) small I and might indicate some computational instability

Rapids

Gene	I	pval_norm	var_norm	pval_norm_fdr_bh
ASIC1	inf	0.0	3.2141281e-06	0.0
CD40LG	inf	0.0	3.2141281e-06	0.0
GAL	inf	0.0	3.2141281e-06	0.0

Squidpy

Gene	I	pval_norm	var_norm	pval_norm_fdr_bh
ASIC1	0.0465674	0.0	3.2141273e-06	0.0
CD40LG	0.0117390	2.8169134e-11	3.2141273e-06	3.8116098e-11
GAL	0.0154400	0.0	3.2141273e-06	0.0

Expected behavior

Moran’s I values should:

• Be finite (no inf or NaN)
• Fall within the bounded range ([-1, 1])

Be consistent with established implementations like squidpy

Environment details

• Environment location: Bare-metal
• Linux Distro/Architecture: Ubuntu 22.04 amd64
• GPU Model/Driver: A100 and driver 560.35.03
• CUDA: 12.6
• Squidpy Version: '1.6.5'
• Rapids Version: '0.12.7'

[Intron7]

@lukasadam thank you so much for bringing this up. I was not aware of this happening. Do you have a reproducer for this issue? I know that rsc uses 32 bit while squidpy uses 64 bit but i have not seen these differences.