BUG: optimize.minimize backwards compatability in scipy 1.9

Describe your issue.

Hi, we have some code that works with scipy 1.8 but crashes 1.9. An example to reproduce is below. The issue arises in the code than transforms between the new/old bounds definition, but is actually related to the way x0 is defined.

You can ‘fix’ this by changing start to start = np.array([0.0]). We can actually do this in our code too, but I wanted to check if this behavior is intentional or whether this lack of backward compatibility constitutes a bug?

Reproducing Code Example

'''
example which will run on scipy 1.8 but not scipy 1.9
'''
from scipy.optimize import minimize
import numpy as np

def objective(x):
    return(x+10)**2

bounds = np.array([[-20, 20]])
start = np.array([0.0]).reshape(-1, 1)
res = minimize(objective, start, bounds=bounds, method="L-BFGS-B")

Error message

TypeError: object of type 'Bounds' has no len()

SciPy/NumPy/Python version information

1.8/1.9

[andyfaff]

At the start of minimize this code is called:

 x0 = np.atleast_1d(np.asarray(x0))
    if x0.ndim != 1:
        message = ('Use of `minimize` with `x0.ndim != 1` is deprecated. '
                   'Currently, singleton dimensions will be removed from '
                   '`x0`, but an error will be raised in SciPy 1.11.0.')
        warn(message, DeprecationWarning, stacklevel=2)
        x0 = np.squeeze(x0)

Because there is more than one dimension in start the array is squeezed. We do this squeeze because minimize is only designed to work with 1D arrays (problems are experienced with ndim > 1). You should’ve seen this DeprecationWarning when you ran your code. When the code squeezes np.array([[0.0]]) you end up with np.array(0.0), which has shape (). When x0.shape[0] is requested an IndexError is raised, IndexError: tuple index out of range. I get a different error than you see.

We could fix scipy to make it back compatible by doing x0 = np.atleast_1d(np.squeeze(x0)). However, as you can see from the code snippet above we’re going to raise an Error if x0.ndim > 1 in scipy 1.11. This probably makes it less important to make a ‘fix’ in scipy; can you fix the downstream code here?

[mdhaber]

Hmm I do think this was an unintentional mistake on our part to remove all the dimensions, so from that perspective, yes. We might be able to get it in 1.9.1. I can’t push the buttons, but I can review quickly if it’s the one-liner you described, @andyfaff.