NaN output only when using three ouput for QR

I'm trying to use the canoncorr function. I removed all features with NaN values from my two input matrices. Still, I get the error

Error using canoncorr (line 86)
X must contain at least one non-constant column.

This is because the function qr in

[Q1,T11,perm1] = qr(X,0);

returns all-NaN Q1 and T11, therefore at the next line rankX is 0.

Nonetheless, I noted that if I run only [Q1,T11] = qr(X,0); or [T11] = qr(X,0); all looks ok.

Any guess on what is going on?

Thank you

[Chuguang Pan]

I guess the matrix X contains NaN

Answer 1

[Christine Tobler]

The 3-output syntax and the 2-output syntax of QR use different implementations, so that's likely the reason you're seeing different outcomes.

I would usually expect no NaNs in Q and T11 if X has no NaN or Inf entries - but this is possible in extreme cases, for example when the values are close to realmax, where a sum will overflow to Inf.

If you can attach a .mat file with the matrix X, I can take a closer look.