Re: [eigen] Re: SparseQR crashes on rectangular matrix

QR is for matrices with rows>=cols, otherwise the best you could expect is to have R as a upper trapezoidal matrix (as in the dense case), but that's is not really useful and could be forbid en IMO.

Now regarding the case where rows==cols, can you be more specific about what does not work and in which case? This case has been well tested, even with matrices with empty columns.

thanks,

gael

On Wed, Sep 11, 2013 at 3:25 PM, Pavel Holoborodko <pavel@xxxxxxxxxxxxxxx> wrote:

This bug goes much deeper, SparseQR::factorize cannot handle rectangular matrices when rows <= cols (yes, even when rows == cols).
So this is not only coletree, but SparseQR itself.

This is definitely the major issue, since rectangular matrices are the main domain of QR application.

Please advise what is the maturity level of SparseQR code and if there are any plans to improve it?

Probably I can help with better implementation with the help of core devs.

Thank you.
Pavel.

On Mon, Sep 2, 2013 at 4:12 PM, Pavel Holoborodko <pavel@xxxxxxxxxxxxxxx> wrote:

SparseQR crashes in coletree when applied for rectangular matrix with rows < cols.

After quick study - apparently coletree doesn't support rectangular matrices.

Vector firstRowElt is allocated with rows() size (line #73).

However it is indexed as having cols() size in the code (lines #75, #108).

Would appreciate any help in this regard.

Thank you in advance.
Pavel.