Reproducing code example:

1. Procure 2x EC2 instances, ideally a t2.micro (x86) and t4g.micro (ARM) running Ubuntu 20.04 LTS (freetier)
2. sudo apt update && sudo apt install python3-pip && pip3 install numpy to install requirements
3. Open a python3 interactive shell to run the following code:

import numpy as np
values = [-20, 2, 3, 4, 5, 4, 3, 2]
values = values[int(non_zero[0]):int(np.non_zero[-1])+1]
_, e = np.frexp(values[::-1])
c = int(-0.5 * (np.max(e) + np.min(e)))
p = values[::-1]
p = np.ldexp(p, c)
A = np.diag(np.full(p.size - 2, p[0]), k=-1)
A[0,:] = -p[1:]
eigenvalues = np.linalg.eigvals(A)
print(eigenvalues)

(All values up until the eigvals function call are equal on both machines)

On the ARM based CPU, you will see the following output:
[-0.14294258 -0.05014252 -0.06678331 -0.06036967 0.01308302 -0.29849543 0.23065051]

On the x86 based CPU, you will see the following output:
[-0.35561591+0.14999412j -0.35561591-0.14999412j -0.10977392+0.35797579j -0.10977392-0.35797579j 0.15723786+0.31047243j 0.15723786-0.31047243j 0.24130395+0.j ]

There is no stack trace or error thrown.

Error:

I believe that the eigenvalues produced on the ARM CPU are incorrect, as they do not include complex conjugate pairs (with real number input). It's been a while since I've studied linear algebra, but I believe this to be the case.

I found this error originally due to an issue I encountered with the numpy_financial package, which calls the np.linalg.eigvals function from the irr function. More information on the issue over there: numpy/numpy-financial#37

NumPy/Python version information:

Both machines running Ubuntu 20.04 LTS, kernel 5.4.0-1047-aws. Python version 3.8.5. Numpy version 1.20.1.