PseudoInverse Matrix Glitch

[GerdW]

Hi,

until Christian brings the "ultimate" improvement I just "clean up" a little bit:

[altenbach]

Yes, I actually did the same thing. Not sure why we need option "full" for the SVD. Same result (I did not check if the algorithm is sound ;)):

Personally, I would probably go over all diagonal elements and handle values close to zero differently.

[altenbach]

---

@altenbach wrote:

Personally, I would probably go over all diagonal elements and handle values close to zero differently.

---

Here's how that could look like (I think the elements of S are always nonnegative, but please verify).

The tolerance can be computed based on the elements and machine epsilon, e.g. as described here)

[altenbach]

On a side note, you can of course operate on the vector S instead of matrix S. Maybe more efficient. (not tested).

I am sure things could be further optimized....

[James_McN]

For what its worth I can replicate the original problem on a cRIO (VM) in LabVIEW 2020. Certainly looks like a bug in their RT Linux implementation. Interestingly when I tried it on a Linux desktop it didn't replicate - but I only had LabVIEW 2019 available.

It would be worth sharing your original code with NI through a service request at http://ni.com/ask so they can record it as a LabVIEW bug and hopefully fix it going forward.