Repository navigation
Conda-forge version of slycot 0.3.3 generates errors #217
Description
Activity
This error occurs with conda forge's slycot 0.2.0, 0.3.0, and 0.3.3.
How precise is this test? I can imagine minreal being finicky because of floating point operations. How robust is it to different platforms, architectures, and blas/lapack implementations?
Good to know that this is not a new bug. I was pretty sure I had seen the error message before, but couldn't figure out where.
I chased this down a bit. It looks like what is happening is that one of the (random) state space systems that is getting generated in the unit test has a near pole-zero cancellation.
minreal()is not reducing the system (via a call totb01pd). But when the unit test checks to make sureminrealis not changing the system, it converts the state space system to a transfer function and one of systems is now reduced (inside oftb04ad, where it extracts the controllable-observable subsystem). This is probably just a difference in tolerances, as @moorepants indicated. Interestingly, the error doesn't show up on my local system if I use thelapackvariant ofslycotinstead of theopenblasvariant.Here are the systems that are being compared:
System 1 (original): A = [[-2.07272289 0.16120168 0.05029039 -0.17656953] [-1.2172162 -3.0412386 0.87634026 0.45936988] [ 2.34727018 2.1024825 -2.97194391 -1.72217184] [ 2.70301798 2.1897608 -1.33329543 -3.77801591]] B = [[ 0.22934264] [-1.05015731] [-0.40877983] [-1.282876 ]] C = [[-1.22961654 0.9999803 -0. -0. ]] D = [[0.]] -1.332 s^2 - 10.37 s - 13.7 --------------------------------- s^3 + 9.665 s^2 + 26.21 s + 21.53 Zeros = [-6.09568359 -1.68756532] Poles = [-5.76863334 -2.1991448 -1.69699836] System 2 (output of minreal): A = [[-1.44793116 0.00720403 -1.02465799 -0.32451831] [ 1.29258628 -2.23726554 -1.36828447 -1.39979547] [ 2.18253992 -0.07706622 -4.41025255 -2.57508146] [ 1.75507481 -0.03186882 -2.0133189 -3.76847205]] B = [[-1.09979463] [ 0.78245651] [-0.96370783] [-0.46667011]] C = [[ 0. -1.46759237 -0.09533846 0.59076278]] D = [[0.]] -1.332 s^3 - 13.3 s^2 - 36.51 s - 30.14 --------------------------------------------- s^4 + 11.86 s^3 + 47.46 s^2 + 79.16 s + 47.34 Zeros = [-6.09568359 -2.1991448 -1.68756532] Poles = [-5.76863334+0.00000000e+00j -2.1991448+2.29466803e-07j -2.1991448-2.29466803e-07j -1.69699836+0.00000000e+00j]The cancellation is occurring in the first system but not the second. In particular, in the second system there is a zero at -2.1991448 and a pair of poles at -2.1991448 +/- 2.29466803e-07j (one of which gets apparently gets cancelled in the first system).
There are two issues here:
- The criterion in
minrealfor a pole/zero cancellation is different than inss2tf. - The
minrealunit test is not testing (just)minreal, but is also (implicitly) testing thss2tfconversion.
Not sure yet what the right fix is (if any).
- The criterion in
When installing
slycotfrom conda-forge (v0.3.3), I get the following error when runningpython setup.py test:This error does not occur when using pip version of
slycot. Need to understand what is going on before releasing 0.8.0.Note: conda-forge version of
slycotuses OpenBLAS, pip version useslapack(I think).