bugfix: wrong initialization of ConsensusADMM

[tttamaki]

Fix initialization of $y_1, \ldots, y_m \in R^d$ to have a correct dimension.

The variable y should have a dimension of $m \times d$, where d is the dimension of x0, but it was incorrectly initialized to a dimension of $d$.

Before this change, the test passed incorrectly when $m \le d$ due to the following line;

x = ncp.stack([proxfs[i].prox(x_bar - y[i], tau) for i in range(m)])
# proxfs.shape == m, x_bar.shape == d, y[0],,,y[d-1] are scalar

But this will be an error when $m > d$ because out of index.
The test code in test_solver.py uses only $m=2$ proxfs with $d=$ par["m"] == 10, so I missed this bug.
I found this when I used $m=4$ proxfs and $d=2$ dimension of x0.

Notet that, in the second iteration, the dimension of y become $m \times d$ as expected because of the following line, so that the test have passed after iterations even with the wrong initialization.

y = y + x - x_bar  # y.shape == d, x.shape == (m, d), x_bar.shape == d  --> y.shape == (m, d)

[Copilot]

Pull request overview

This PR fixes a critical bug in the initialization of the dual variable y in the ConsensusADMM function. The variable was incorrectly initialized with dimension d instead of m × d, where m is the number of proximal operators and d is the dimension of the problem. This bug caused index out-of-bounds errors when m > d, though it went undetected in tests where m ≤ d due to numpy's broadcasting behavior in subsequent iterations that inadvertently corrected the shape.

Changes:

• Fixed initialization of dual variable y from ncp.zeros_like(x0) to ncp.zeros((m, x0.shape[0])) to correctly allocate space for m dual variables, each of dimension d

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[mrava87]

@tttamaki good catch. I just have a suggestion to ensure the dtype of y is still correctly setup 😄