Unable to concatenate a cp.Variable vector

[antonelloventurino]

Describe the bug
I'm trying to make an LMI constraint like:

[[scalar, variable2D'],
[variable2D, identity2D]] >>0

but is impossible due to the fact that the solver does not concatenate the elements. The entire matrix at the end has a shape = (2,2) and not (3,3) as expected.

To Reproduce

import cvxpy as cp
import numpy as np

# scalar
u = 3
# variable
x = cp.Variable(2)
c_vert = cp.vstack([u, x])

Expected behavior
c_vert must be a vector with 3 elements. First one as known scalar and second ones from x variable vector.

Output

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/*/.local/lib/python3.8/site-packages/cvxpy/atoms/affine/vstack.py", line 29, in vstack
    return Vstack(*arg_list)
  File "/home/*/.local/lib/python3.8/site-packages/cvxpy/atoms/atom.py", line 50, in __init__
    self.validate_arguments()
  File "/home/*/.local/lib/python3.8/site-packages/cvxpy/atoms/affine/vstack.py", line 74, in validate_arguments
    raise ValueError(("All the input dimensions except"
ValueError: All the input dimensions except for axis 0 must match exactly.

Version

• OS: Linux Ubuntu 20.04
• CVXPY Version: 1.4.1

[SteveDiamond]

This is not a bug. The behavior is the same as numpy. Try using the bmat function and make all elements 2 dimensional. For example,
scalar = np.ones((1,1)) * value

[antonelloventurino]

This is not a bug. The behavior is the same as numpy. Try using the bmat function and make all elements 2 dimensional. For example, scalar = np.ones((1,1)) * value

Thanks, it works this way. However, just for information, numpy is a bit smarter.
Example 1 with cvxpy gives an error:

import cvxpy as cp
import numpy as np
U = cp.Variable(shape=(2,1))
u_bounds_max=2
cp.bmat([[ u_bounds_max ** 2, U.T], [U, cp.bmat(np.identity(2))]])

Error:

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/*/.local/lib/python3.8/site-packages/cvxpy/atoms/affine/bmat.py", line 37, in bmat
    row_blocks = [hstack(blocks) for blocks in block_lists]
  File "/home/*/.local/lib/python3.8/site-packages/cvxpy/atoms/affine/bmat.py", line 37, in <listcomp>
    row_blocks = [hstack(blocks) for blocks in block_lists]
  File "/home/*/.local/lib/python3.8/site-packages/cvxpy/atoms/affine/hstack.py", line 38, in hstack
    return Hstack(*arg_list)
  File "/home/*/.local/lib/python3.8/site-packages/cvxpy/atoms/atom.py", line 50, in __init__
    self.validate_arguments()
  File "/home/*/.local/lib/python3.8/site-packages/cvxpy/atoms/affine/hstack.py", line 68, in validate_arguments
    raise error
ValueError: All the input dimensions except for axis 1 must match exactly.

Same example with numpy works fine:

import numpy as np
U = np.ones((2,1))
u_bounds_max=2
np.block([[u_bounds_max ** 2, U.T], [U, np.identity(2)]])

Output correct:

array([[4., 1., 1.],
       [1., 1., 0.],
       [1., 0., 1.]])

[SteveDiamond]

That's good to know. We will make it work in cvxpy then.

[SteveDiamond]

For reference numpy.bmat does not handle scalars gracefully like numpy.block.

[rileyjmurray]

bmat also returns a matrix object and not an ndarray, so will probably be discouraged moving forward. It would be prudent for us to make an analog of NumPy’s block function.

…

On Wed, Feb 14, 2024 at 5:58 PM Steven Diamond ***@***.***> wrote: For reference numpy.bmat does not handle scalars gracefully like numpy.block. — Reply to this email directly, view it on GitHub <#2328 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACRLIFFW5ETYKRYBXQH2Y2DYTU6PXAVCNFSM6AAAAABB6R5SMWVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTSNBUHEZDCMRQGM> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>