Encountering NaN when fitting Vectors with CCIPCA

[hv10]

Hi,

I am encountering the following issue when using CCIPCA:

When I call fit! repeatedly with the same input vector it can happen that the CCIPCA eigenvectors & -values become NaN.

Minimal Example:

a = [0.027667, 0.0428616, 0.57036, 0.382638, 68.4809, 24.4805, 230.786, 32.9694]
o = OnlineStats.CCIPCA(3, 8; l=3)
fit!(o, a)
display(o.U)

fit!(o,a)
display(o.U)

Leading to output:

8×3 Matrix{Float64}:
 0.000113292  0.0  0.0
 0.000175511  0.0  0.0
 0.00233553   0.0  0.0
 0.00156684   0.0  0.0
 0.280418     0.0  0.0
 0.100244     0.0  0.0
 0.945031     0.0  0.0
 0.135004     0.0  0.0

8×3 Matrix{Float64}:
 -0.000113292  NaN  0.0
 -0.000175511  NaN  0.0
 -0.00233553   NaN  0.0
 -0.00156684   NaN  0.0
 -0.280418     NaN  0.0
 -0.100244     NaN  0.0
 -0.945031     NaN  0.0
 -0.135004     NaN  0.0

From my limited experience I can see that the issue occurs when we divide by o.lambda within the fitting procedure.
Either at

OnlineStats.jl/src/stats/pca.jl

Line 103 in b99b6cc

o.U[:, i] = o.xi / o.lambda[i]

or

OnlineStats.jl/src/stats/pca.jl

Line 108 in b99b6cc

o.U[:, i] = o.v/o.lambda[i]

when the current eigenvalue is o.lambda[i]=0.0.

Is this something which is attributable to user error? If so how can I make sure to avoid it in the future?

[joshday]

@robertfeldt