removed valueError for nonzero eigenvalues

[jordandekraker]

I was rtying to run this for hippocampal surfaces with no masked data (i.e. no medial wall) and received an error on this line. I believe it is an unneeded line.

[ReinderVosDeWael]

I'd be more curious as to why this is happening in the first place. The MSM manuscript is not ambiguous:

A set of n observations with a full-rank spatial weights matrix W of size n × n will result in n – 1 orthogonal and uncorrelated (Griffith 2000) eigenvectors Vk associated with eigenvalues λk, while a single eigenvector with zero eigenvalue is dropped.

Unfortunately, this seems like the kind of deep dive that's beyond what I have time for these days. That being said, I'd see if you have duplicate data in your input. My first instinct is that that's what's going on.

[jordandekraker]

Yes there are likely duplicate values, but this is valid in my case. All my vertices were generated on one regular meshgrid (and then some vertices were removed, leading to appropriate face sizes in hippocampal meshes). Thus, some vertices will have identical "n_ring=1" distances (or in other words, many face sizes are identical)

It looks like the next few lines are meant to handle cases with many 0 Eigenvalues anyway, so I really think this line can just be dropped.

Not sure why the CI has broken

[ReinderVosDeWael]

@OualidBenkarim What are your thoughts; I think we could consider just changing it to a warning. This shouldn't happen, and is likely indicative of another issue, but if the end-user believes their use-case is an exception we can allow it.

FYI these are my fairly uninformed thoughts coalesced in the ten minutes I could find to look at this, so don't put too much stock in my opinion on this :P

[OualidBenkarim]

There's the tol argument to control that. Maybe increasing the tolerance might fix the issue

[Copilot]

Pull request overview

This PR updates Moran spectral randomization (compute_mem) to avoid raising an exception when the doubly-centered weight matrix yields no eigenvalues classified as “zero” under the provided tolerance—motivated by running on hippocampal surfaces without masked regions.

Changes:

• Removed the ValueError previously raised when n_zero == 0 (no eigenvalues below tol).

---

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

[Copilot]

With the ValueError removed, spectrum='nonzero' now returns all eigenvectors when n_zero == 0 (i.e., nothing is filtered by mask_zero). That contradicts the function’s contract of removing the (at least one) zero/constant mode from the doubly-centered matrix and can change downstream behavior (e.g., component count becomes n_vertices instead of the expected <= n_vertices-1). Consider treating the smallest-|eigenvalue| entry as zero when n_zero == 0 (e.g., set mask_zero[ev_abs.argmin()] = True) so the constant mode is still removed even when numerical precision/tol prevents exact zeros.

Suggested change

	# If no eigenvalue is below tolerance but spectrum=='nonzero', still
	# remove at least one mode (the one with smallest |eigenvalue|),
	# corresponding to the constant component of the doubly centered matrix.
	if spectrum == 'nonzero' and n_zero == 0:
	idx_min = ev_abs.argmin()
	mask_zero[idx_min] = True
	n_zero = 1

[Copilot]

There’s no regression test covering the scenario that motivated this change (doubly-centered wc producing no eigenvalues below tol, causing n_zero == 0). Adding a test that forces this condition (e.g., very small tol or a matrix construction that yields tiny-but-nonzero numerical eigenvalues) would help ensure compute_mem still drops the constant/near-zero mode and avoids returning n_vertices components.

[zihuaihuai]

Thanks for the report — your hippocampal-surface use case is real and the current behavior (raising even when the caller asked for spectrum='nonzero') is wrong for it.

However, simply deleting the guard introduces a worse silent-data-corruption bug for the spectrum='all' branch. Walk through with n_zero == 0:

• mask_zero = ev_abs < tol → all False
• mask_zero.argmax() returns 0 when nothing is True
• idx_zero = 0
• ev[idx_zero:-1] = ev[idx_zero+1:] → drops ev[0], the largest eigenvalue (not a zero one)
• compute_mem returns garbage

Your hippocampal call almost certainly uses spectrum='nonzero' (the else branch on line 121), where mask_nonzero = ~mask_zero is all-True and the function happens to do the right thing — that's why you didn't hit the bug.

Suggested fix

Make the guard branch-aware:

if spectrum == 'all' and n_zero == 0:
    raise ValueError(
        "Weight matrix has no zero eigenvalue; spectrum='all' requires "
        "at least one. Use spectrum='nonzero' for matrices without a "
        "kernel (e.g., surfaces with no medial wall)."
    )

That keeps the guard where it actually matters and lets your spectrum='nonzero' call through cleanly.

A small regression test (a non-singular weight matrix → compute_mem(w, spectrum='nonzero') succeeds) would also lock this in.

[zihuaihuai]

Superseded by #157. Your bug report was right — kept the guard branch-aware so spectrum='nonzero' now works on kernel-free matrices like yours, while spectrum='all' still raises (with a clearer message). You're credited as Co-author. Feel free to close this PR.

[zihuaihuai]

Closing — your bug report was correct and the fix has already landed in #157, where it's branch-aware:

• spectrum='nonzero' now accepts kernel-free weight matrices (your hippocampal-subfield use case works)
• spectrum='all' still raises, but with a clearer error pointing at spectrum='nonzero'

This PR's diff overlaps with #157, which is why it now shows a conflict. You're credited as Co-author on the squashed commit. Thanks!