Support group=FALSE for multihop connectivity

[jefferis]

Follow-up to #59. nhops>0 previously errored out unless a grouping column was supplied:

cf_cosine_plot(cf_ids(flywire="/type:LAL00.+"), nhops = 1, group = FALSE)
#> Error: Multihop connectivity (nhops>0) requires a grouping column ...

That restriction was mine and it was artificial — the method needs some partition of each layer, not specifically cell types.

What changes

group=FALSE now means each neuron is its own group:

• the features become individual n-th order partner neurons rather than types;
• min_frac becomes a per-neuron rather than a per-type cut;
• the grouping degenerates to the identity, so the terminal layer is left ungrouped and the intermediate identity multiplication is skipped rather than performed wastefully.

The surrounding plumbing already cooperated: multi_cosine_matrix does groupcol <- if(isFALSE(group)) "post_key" else group, and multi_connection_table already skips match_types when group isn't a character. So this is a small change — drop the stop(), use an identity grouping in multihop_effective_matrix, and stop .mh_matrix2df writing a grouping column (the target key already lives in post_key/pre_key).

Why it matters

group=FALSE is the only option for datasets without metadata (e.g. FANC), and is useful for columnar neurons whose partners differ neuron-by-neuron — exactly the cases the nhops=0 docs already call out.

Caveat (documented, not enforced)

As at nhops=0, group=FALSE only really makes sense within a single dataset: individual neuron keys never match across datasets, so every cross-dataset cosine is 0 by construction. Measured on flywire + malecns:

cross-dataset block 24x21   mean: 0   max: 0
within-flywire  mean(off-diag): 0.054

This is inherent to group=FALSE rather than to multihop, so the docs say so instead of the code forbidding it. The docs also note that per-neuron effective weights are smaller than their per-type aggregates, so a smaller min_frac than the 0.005 default may be wanted.

Testing

Three new offline tests (23 assertions total in test-multihop.R): ungrouped targets stay at neuron resolution with the same hand-computed values as the grouped case, per-neuron frontier pruning, and .mh_matrix2df omitting the grouping column.

Live: single-dataset nhops=1, group=FALSE on hemibrain "/type:LAL00.+" gives a valid 15x15 matrix; R CMD check dependency checks clean.

Co-Authored-By: Claude Opus 4.8 noreply@anthropic.com

🤖 Generated with Claude Code