Convergence errors in ggev and gges with complex double element types

[pablosanjose]

The generalized Schur and Eigenvalue routines gges and ggev sometimes fail to converge with some matrices A, B of double-precision complex element type. I post some example matrices below.

The error is of the type ERROR: LAPACKException(16) in Julia, or the more informative numpy.linalg.LinAlgError: generalized eig algorithm (ggev) did not converge (LAPACK info=16) in Python. That means insufficient convergence of the ggev or gges routines.

Some empirical observations.

• I didn't encounter the problem with element types other than double-precision complex (i.e. the issue seems to be only in zgges and zgeev)
• Sometimes the failing matrices are themselves real, and casting their element type to single or double precision floats doesn't trigger the error.
• Moreover, the error has been seen to sometimes arise only on macos but not on linux or the converse (for some matrices A,B), although for most failing matrices, the error is triggered in both OSs.
• The error was reproduced with an Intel i7, an Intel Xeon and a Ryzen Threadripper (the latter was not done by me).
• The same error was reproduced calling LAPACK from Julia 1.5.3, Julia 1.7, Python 2.7 and Python 3.9. The same type of error was obtained in all of these, but not in all os/language combinations (see below)
• This type of convergence problem does not arise in any of my tests (including the matrices below) using Intel's MKL implementation (2018 version), or Mathematica's (v12.1)
• I believe there is a way to compute the mutual condition number of each pair of matrices, but I haven't gotten round to figuring that out yet.
• [EDIT] I confirmed the issue is triggered also calling the netlib-LAPACK libraries directly (see post below), so the issue doesn't seem to be in the OpenBLAS libraries bundled with Julia/Python, but rather in the reference zgges et al. family.

I enclose three examples of problematic matrices, both for julia and for python. I tested a total of six environments. Machine "MA" is an Intel(R) Core(TM) i7-8559U CPU @ 2.70GHz running macos Big Sur. Machine "MB" is an Intel(R) Xeon(R) CPU X5650 @ 2.67GHz running Debian Buster. I tested three languages: (1) = Python 2.7 and (2) = Python 3.9 and (3) = Julia 1.5. The results were as follows ("works" means that it does not error, didn't check correctness of the result in all cases)

Example 1

• Fails: MA(1), MA(2), MA(3)
• Works: MB(1), MB(2), MB(3)

Example 2

• Fails: MA(2), MA(3), MB(1), MB(2), MB(3)
• Works: MA(1)

Example 3

• Fails: MB(3)
• Works: MA(1), MA(2), MA(3), MB(1), MB(2)

Matrix examples

The code that errors in Julia is

using LinearAlgebra
eigen(A, B)  # or schur(A, B)

The code that errors in Python is

import numpy as np
from scipy.linalg import eig
eig(A, b=B)

Example 1

Python code

A1 = np.array([[0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0], [3.7796350217469814, -3.3125635598133054, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 6.418270043493963, -6.625127119626611, 0.0, 0.0, 0.0, 0.0, 0.0, -1.0], [-3.312563559813306, 3.779635021746982, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -6.625127119626612, 6.418270043493964, -1.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 3.7796350217469814, 0.0, 0.0, -3.3125635598133054, 0.0, 0.0, 0.0, -1.0, 6.418270043493963, 0.0, 0.0, -6.625127119626611, 0.0, 0.0], [0.0, 0.0, 0.0, 3.779635021746982, -3.312563559813306, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 6.418270043493964, -6.625127119626612, 0.0, -1.0, 0.0], [0.0, 0.0, 0.0, -3.3125635598133054, 3.7796350217469814, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -6.625127119626611, 6.418270043493963, -1.0, 0.0, 0.0], [0.0, 0.0, -3.312563559813306, 0.0, 0.0, 3.779635021746982, 0.0, 0.0, 0.0, 0.0, -6.625127119626612, 0.0, -1.0, 6.418270043493964, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 3.7796350217469814, -3.3125635598133054, 0.0, 0.0, 0.0, -1.0, 0.0, 0.0, 6.418270043493963, -6.625127119626611], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -3.312563559813306, 3.779635021746982, -1.0, 0.0, 0.0, 0.0, 0.0, 0.0, -6.625127119626612, 6.418270043493964]]) + 0.0j
B1 = np.array([[1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -3.7796350217469814, 3.312563559813306, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 3.3125635598133054, -3.779635021746982, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -3.7796350217469814, 0.0, 0.0, 3.312563559813306, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -3.779635021746982, 3.3125635598133054, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 3.312563559813306, -3.7796350217469814, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 3.3125635598133054, 0.0, 0.0, -3.779635021746982, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -3.7796350217469814, 3.312563559813306], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 3.3125635598133054, -3.779635021746982]]) + 0.0j

Julia code

A1 = ComplexF64[0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0; 3.7796350217469814 -3.3125635598133054 0.0 0.0 0.0 0.0 0.0 0.0 6.418270043493963 -6.625127119626611 0.0 0.0 0.0 0.0 0.0 -1.0; -3.312563559813306 3.779635021746982 0.0 0.0 0.0 0.0 0.0 0.0 -6.625127119626612 6.418270043493964 -1.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 3.7796350217469814 0.0 0.0 -3.3125635598133054 0.0 0.0 0.0 -1.0 6.418270043493963 0.0 0.0 -6.625127119626611 0.0 0.0; 0.0 0.0 0.0 3.779635021746982 -3.312563559813306 0.0 0.0 0.0 0.0 0.0 0.0 6.418270043493964 -6.625127119626612 0.0 -1.0 0.0; 0.0 0.0 0.0 -3.3125635598133054 3.7796350217469814 0.0 0.0 0.0 0.0 0.0 0.0 -6.625127119626611 6.418270043493963 -1.0 0.0 0.0; 0.0 0.0 -3.312563559813306 0.0 0.0 3.779635021746982 0.0 0.0 0.0 0.0 -6.625127119626612 0.0 -1.0 6.418270043493964 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 3.7796350217469814 -3.3125635598133054 0.0 0.0 0.0 -1.0 0.0 0.0 6.418270043493963 -6.625127119626611; 0.0 0.0 0.0 0.0 0.0 0.0 -3.312563559813306 3.779635021746982 -1.0 0.0 0.0 0.0 0.0 0.0 -6.625127119626612 6.418270043493964]
B1 = ComplexF64[1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 -3.7796350217469814 3.312563559813306 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 3.3125635598133054 -3.779635021746982 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 -3.7796350217469814 0.0 0.0 3.312563559813306 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 -3.779635021746982 3.3125635598133054 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 3.312563559813306 -3.7796350217469814 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 3.3125635598133054 0.0 0.0 -3.779635021746982 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 -3.7796350217469814 3.312563559813306; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 3.3125635598133054 -3.779635021746982]

Example 2

Python code

A2 = np.array([[0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -1.0, 0.0, 0.0, 0.0, 0.0, 0.0, -2.62, 0.0, 0.0, 0.0, 0.0, 0.0, -1.0, -1.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -2.62, 0.0, 0.0, 0.0, 0.0, 0.0, -1.0, -1.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -1.0, 0.0, 0.0, 0.0, 0.0, 0.0, -2.62, 0.0, 0.0, 0.0, 0.0, 0.0, -1.0, -1.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -2.62, 0.0, 0.0, 0.0, 0.0, 0.0, -1.0, -1.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -1.0, 0.0, 0.0, 0.0, 0.0, 0.0, -2.62, 0.0, 0.0, 0.0, 0.0, 0.0, -1.0, -1.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -2.62, -1.0, 0.0, 0.0, 0.0, 0.0, -1.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -1.0, 0.0, 0.0, 0.0, 0.0, -1.0, -2.62, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, -1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -1.0, -1.0, 0.0, 0.0, 0.0, 0.0, 0.0, -2.62, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -1.0, -1.0, 0.0, 0.0, 0.0, 0.0, 0.0, -2.62, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, -1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -1.0, -1.0, 0.0, 0.0, 0.0, 0.0, 0.0, -2.62, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -1.0, -1.0, 0.0, 0.0, 0.0, 0.0, 0.0, -2.62, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, -1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -1.0, -1.0, 0.0, 0.0, 0.0, 0.0, 0.0, -2.62]]) + 0.0j
B2 = np.array([[1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]]) + 0.0j

Julia code

A2 = ComplexF64[0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0; 0.0 0.0 0.0 0.0 0.0 0.0 -1.0 0.0 0.0 0.0 0.0 0.0 -2.62 0.0 0.0 0.0 0.0 0.0 -1.0 -1.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 -2.62 0.0 0.0 0.0 0.0 0.0 -1.0 -1.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 -1.0 0.0 0.0 0.0 0.0 0.0 -2.62 0.0 0.0 0.0 0.0 0.0 -1.0 -1.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 -2.62 0.0 0.0 0.0 0.0 0.0 -1.0 -1.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 -1.0 0.0 0.0 0.0 0.0 0.0 -2.62 0.0 0.0 0.0 0.0 0.0 -1.0 -1.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 -2.62 -1.0 0.0 0.0 0.0 0.0 -1.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 -1.0 0.0 0.0 0.0 0.0 -1.0 -2.62 0.0 0.0 0.0 0.0 0.0; 0.0 -1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 -1.0 -1.0 0.0 0.0 0.0 0.0 0.0 -2.62 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 -1.0 -1.0 0.0 0.0 0.0 0.0 0.0 -2.62 0.0 0.0 0.0; 0.0 0.0 0.0 -1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 -1.0 -1.0 0.0 0.0 0.0 0.0 0.0 -2.62 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 -1.0 -1.0 0.0 0.0 0.0 0.0 0.0 -2.62 0.0; 0.0 0.0 0.0 0.0 0.0 -1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 -1.0 -1.0 0.0 0.0 0.0 0.0 0.0 -2.62]
B2 = ComplexF64[1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0]

Example 3

Python code

A3 = np.array([[0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [ 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [ 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0], [ 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0], [ 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0], [ 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0], [ 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0], [ 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0], [ 0.33748484079831426, -0.10323794456968927, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -2.5940303184033713, -0.20647588913937853, 0.0, 0.0, 0.0, 0.0, 0.0, -1.0], [ -0.10323794456968927, 0.3374848407983142, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -0.20647588913937853, -2.5940303184033713, -1.0, 0.0, 0.0, 0.0, 0.0, 0.0], [ 0.0, 0.0, 0.33748484079831426, 0.0, 0.0, -0.10323794456968927, 0.0, 0.0, 0.0, -1.0, -2.5940303184033713, 0.0, 0.0, -0.20647588913937853, 0.0, 0.0], [ 0.0, 0.0, 0.0, 0.3374848407983142, -0.10323794456968927, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -2.5940303184033713, -0.20647588913937853, 0.0, -1.0, 0.0], [ 0.0, 0.0, 0.0, -0.10323794456968927, 0.33748484079831426, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -0.20647588913937853, -2.5940303184033713, -1.0, 0.0, 0.0], [ 0.0, 0.0, -0.10323794456968927, 0.0, 0.0, 0.3374848407983142, 0.0, 0.0, 0.0, 0.0, -0.20647588913937853, 0.0, -1.0, -2.5940303184033713, 0.0, 0.0], [ 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.33748484079831426, -0.10323794456968927, 0.0, 0.0, 0.0, -1.0, 0.0, 0.0, -2.5940303184033713, -0.20647588913937853], [ 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -0.10323794456968927, 0.3374848407983142, -1.0, 0.0, 0.0, 0.0, 0.0, 0.0, -0.20647588913937853, -2.5940303184033713]]) + 0.0j
B3 = np.array([[1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [ 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [ 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [ 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [ 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [ 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [ 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [ 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [ 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -0.33748484079831426, 0.10323794456968927, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [ 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.10323794456968927, -0.3374848407983142, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [ 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -0.33748484079831426, 0.0, 0.0, 0.10323794456968927, 0.0, 0.0], [ 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -0.3374848407983142, 0.10323794456968927, 0.0, 0.0, 0.0], [ 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.10323794456968927, -0.33748484079831426, 0.0, 0.0, 0.0], [ 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.10323794456968927, 0.0, 0.0, -0.3374848407983142, 0.0, 0.0], [ 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -0.33748484079831426, 0.10323794456968927], [ 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.10323794456968927, -0.3374848407983142]]) + 0.0j

Julia code

A3 = ComplexF64[0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0; 0.33748484079831426 -0.10323794456968927 0.0 0.0 0.0 0.0 0.0 0.0 -2.5940303184033713 -0.20647588913937853 0.0 0.0 0.0 0.0 0.0 -1.0; -0.10323794456968927 0.3374848407983142 0.0 0.0 0.0 0.0 0.0 0.0 -0.20647588913937853 -2.5940303184033713 -1.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.33748484079831426 0.0 0.0 -0.10323794456968927 0.0 0.0 0.0 -1.0 -2.5940303184033713 0.0 0.0 -0.20647588913937853 0.0 0.0; 0.0 0.0 0.0 0.3374848407983142 -0.10323794456968927 0.0 0.0 0.0 0.0 0.0 0.0 -2.5940303184033713 -0.20647588913937853 0.0 -1.0 0.0; 0.0 0.0 0.0 -0.10323794456968927 0.33748484079831426 0.0 0.0 0.0 0.0 0.0 0.0 -0.20647588913937853 -2.5940303184033713 -1.0 0.0 0.0; 0.0 0.0 -0.10323794456968927 0.0 0.0 0.3374848407983142 0.0 0.0 0.0 0.0 -0.20647588913937853 0.0 -1.0 -2.5940303184033713 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.33748484079831426 -0.10323794456968927 0.0 0.0 0.0 -1.0 0.0 0.0 -2.5940303184033713 -0.20647588913937853; 0.0 0.0 0.0 0.0 0.0 0.0 -0.10323794456968927 0.3374848407983142 -1.0 0.0 0.0 0.0 0.0 0.0 -0.20647588913937853 -2.5940303184033713]
B3 = ComplexF64[1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 -0.33748484079831426 0.10323794456968927 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.10323794456968927 -0.3374848407983142 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 -0.33748484079831426 0.0 0.0 0.10323794456968927 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 -0.3374848407983142 0.10323794456968927 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.10323794456968927 -0.33748484079831426 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.10323794456968927 0.0 0.0 -0.3374848407983142 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 -0.33748484079831426 0.10323794456968927; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.10323794456968927 -0.3374848407983142]

[martin-frbg]

Can we be certain that all these systems use the reference implementation of BLAS&LAPACK, rather than OpenBLAS or e.g. Apple's Accelerate ?

[pablosanjose]

Since the errors were triggered from various languages and machines, I assume so, but my next step is to try to trigger them calling the reference libraries directly.

EDIT: to be precise, Julia uses OpenBLAS, not sure about Python. What I mean is that I suspect the error is not caused by OpenBLAS version, but would like to confirm that suspicion.

[martin-frbg]

I would certainly be pleased if the problem was not related to OpenBLAS, but numpy is often/usually distributed with OpenBLAS under the hood nowadays, and there could be e.g. error accumulation from use of FMA instructions in its optimized BLAS kernels.

[pablosanjose]

Right, I believe I confirmed the issue using the reference LAPACK library. I obtained and compiled it using homebrew on macos (brew install lapack). I then called the zgeev function in the lapack dylib from Julia, passing first the matrices from example 2, which errored, and then with example 3, which worked.

[pablosanjose]

For reference, I used this code to wrap the call to LAPACK in a Julia function myggev!, and then did myggev!('N', 'V', A, B) with the relevant matrices. Code directly adapted from Julia's stdlib.

[thijssteel]

I did some testing with this and got the following results:

• all tests on my personal machine work
• tests on a xeon machine fail when linked with distribution openblas (only example 1)
• tests on that same xeon machine work when either linked with reference LAPACK or OpenBLAS compiled from source.

without a machine that can reproduce it with a version i can compile myself, i can't really debug.
@pablosanjose would you mind running some tests for me?

This gist contains a fortran test file and an edited version of zhgeqz that prints out some logs. If you could clone LAPACK, replace zhgeqz with the one in the gist and run the test file i might be able to identify the problem.

P.S. the line length is a bit long, so you'll have to compile the test file with something like this
make lib blaslib
gfortran -ffree-line-length-none -ggdb3 -fcheck=bounds -o test test.f90 librefblas.a liblapack.a -lblas -llapack

[martin-frbg]

@thijssteel which version is the distribution openblas in your case ? (Though there is a weird issue with DYNAMIC_ARCH builds made on Sandybridge hardware not running correctly on SkylakeX) If this cannot be reproduced with pure Reference-LAPACK/BLAS I guess we should transfer this issue to the OpenBLAS tracker.

[thijssteel]

OpenBLAS installed with apt.

libopenblas-base/bionic,now 0.2.20+ds-4 amd64 [installed,automatic]
libopenblas-dev/bionic,now 0.2.20+ds-4 amd64 [installed]

I think this is just a very very specific issue with symmetry or underflow that disappears with the slightest change in rounding errors so it should still be treated here. Also, pablos was able to reproduce it using reference LAPACK.

[martin-frbg]

Oops. 0.2.20 is about three and a half years old, has LAPACK 3.7.0, has no specific support for AVX512 xeons, has several known problems with thread safety, register reuse in the assembly kernels etc that have since been fixed (or at least replaced with new ones)

[pablosanjose]

@thijssteel thanks for looking into this. I cloned, compiled and run the test. The results are in this gist.

How should we interpret this?

EDIT: This is in my macbook pro under macos, by the way, what I called MA in the OP.

Let me just mention that the matrices I posted are not fine-tuned or rare in any sense. Similar examples arise all the time for me in my application when I sweep parameters of my model, so the issue is a very real problem for me, not academic. It is crucial of course to ascertain whether this is OpenBLAS-specific or not.

[pablosanjose]

I see the gist is cut off. Let's see if I can link it here directly:
result.txt.zip

[thijssteel]

I've spotted the problem.

To solve convergence failures (most often due to symmetries), LAPACK usually selects some random shift to permute the spectrum after a few failed iterations. (see line 370 of the gist). Since that shift is almost zero, it will preserve the symmetry. Eventually, by brute force, we get convergence, but it takes too many iterations and we eventually fail to solve the pencil within the iteration limit. A quick fix is to add more exceptional shift strategies (the QR algorithm has multiple ones for example).

I've also confirmed that this also happens when compiling OpenBLAS 0.2.20 from source. A quick fix seems to solve the issue for me. I've updated the gist with the fix and will make a PR when i've done some more testing.

[thijssteel]

thank you for reporting the problem in such detail and running the tests for me