Warning using algorithm 2

[jbruxaux]

Hi!

I am currently using PCAone on a fairly big dataset (195 samples, 4,2 million SNPs), and I get this warning:
WARNING: PCAone converged but the window size did not expand the whole data.
What would you recommend? Using a different algorithm?
I am using the last version of PCAone (v0.3.4) on a linux cluster.

Thanks for your help!

[Zilong-Li]

Hi!

Thanks for trying out PCAone.

Mathmatically, without the warning, we would get A ≈ USV where U matrix is the PCs in pcaone.eigvecs file, S in pcaone.eigvals and V matrix is in the SNP loadings file pcaone.loadings. This warning literately means that PCAone converged in few epochs before the algorithm expands to the whole data, which suggests only U is correct while S and V are not meaningful as what it stands for in the math.

If you care about S and V matrix other than the U , you can force PCAone algorithm2 to run exact 7 epochs where it starts expanding to the whole data A assuming --batches=64 in default

PCAone --bfile plink -m 2 -n 20 -k 10  --svd 2 --tol-rsvd 0 --maxp 7

In fact, you data is quite small :)

Best,
Zilong

[jbruxaux]

Hi Zilong,

I just re-ran the analysis as you suggested, and it worked without any issue.
Thanks for your help!

[Zilong-Li]

Sorry for the confusion. I'll update PCAone in the next release to make algorithm2 running at least 7 epochs although it converges in less epochs.

[Zilong-Li]

this is fixed in the latest release v0.3.5