Petar Mlinarić

> Don't you think it is a bit weird that most (all?) of the sysmor reductors internally instantiate another reductor to perform the projection (and nothing else). That your reductors...

I guess ERA could fall under Hankel matrix approaches. I've added QuadBT to the list.

> When envisioning the integration of these data-driven MOR methods into pymor, I do not think that they fit nicely into the current ecosystem. Since they are MOR methods they...

Ok, it looks like the issue was not that SciPy got better when installing from source, but that NumPy got worse :confused: I guess there is an issue with my...

[numpy.linalg.solve](https://github.com/numpy/numpy/blob/67c7066353534eeac8054f0deb75dc5923c6dc20/numpy/linalg/umath_linalg.c.src#L1559) uses [`_gesv`](http://www.netlib.org/lapack/explore-html/d7/d3b/group__double_g_esolve_ga5ee879032a8365897c3ba91e3dc8d512.html#ga5ee879032a8365897c3ba91e3dc8d512), while [scipy.linalg.solve](https://github.com/scipy/scipy/blob/7dd0cd0365b2eda4484efed798000309e7649c77/scipy/linalg/basic.py#L209-L218) directly uses `_getrf` and `_getrs` for LU decomposition and triangular solves, and additionally `_gecon` to estimate the condition number from the LU decomposition. So, as...

I ran it again, but without the `relative_to` part (and on a different computer): - wheels (`pip install numpy scipy`):  - source (`pip install --no-binary :all: numpy scipy`): ...

Here are the results with OPENMP turned off (i.e., running `OMP_NUM_THREADS=1 python script.py`): - wheels:  - source:  The curves are smoother, but there is still the same issue...

Hi @ilayn, thank you for the comment. It's good to see you here. > But when and if [scipy/scipy#12824](https://github.com/scipy/scipy/pull/12824) is done it will hopefully be faster than NumPy conclusively and...

From what I understand: - finding `S_*` is a combinatorial problem (might be NP hard) - computing `S` by Businger-Golub QR with column pivoting (QRP) can be thought of as...

Did we decide how to deal with computing l2 norms of all the dofs? Or instead to go for implementing randomized Q-DEIM?