fix: Fail if "layout" attribute doesn't match the number of vertices

[schochastics]

Trying to tackle #219.
The overarching issue is that if the layout is saved as a graph attribute (g$layout), independent on how it is done, there is no checks implemented if the layout matrix can actually
serve as a layout (that is, if it has the correct number of rows).
This leads to weird behaviour of the plot function.

This PR could be the starting point for a refactoring. Currently, it only adds a warning if the layout does not match the graph. @szhorvat could this be a good start?

devtools::load_all("~/git/R_packages/rigraph/")
#> ℹ Loading igraph
g <- make_ring(10)
g$layout <- layout_in_circle(g)
plot(induced_subgraph(g,c(1:9)))
#> Warning: The layout has 10 rows, but the graph has 9 vertices. Unintended results may
#> occur.

g <- make_full_graph(10)
g$layout <- layout_in_circle(g)[1:5,]
plot(g)
#> Warning: The layout has 5 rows, but the graph has 10 vertices. Unintended results may
#> occur.

g$layout <- rbind(layout_in_circle(g), cbind(runif(5),runif(5)))
plot(g)
#> Warning: The layout has 15 rows, but the graph has 10 vertices. Unintended results may
#> occur.

^{Created on 2025-06-24 with reprex v2.1.1}

[szhorvat]

I think a complete fix is only possible in the C core. We need support for attributes with vector values. I proposed this here for 1.0:

igraph/igraph#2744

But realistically, @ntamas and I just don't have the capacity to get this done for the 1.0 release, which is due in September.

My suggestion is to not put too much effort into this on the R side until it can be fixed in the C core. Perhaps discourage using $layout and suggest $x, $y instead ...

@ntamas Are there any hooks in the attribute handler that would allow updating such attributes upon vertex deletion/addition? I don't think there are any. Note that $layout is a graph attribute, which is the problem here.

[schochastics]

Thanks. I will augment the warning to suggest using x and y and that should be good enough for now

[ntamas]

Are there any hooks in the attribute handler that would allow updating such attributes upon vertex deletion/addition?

No, not until we implement some kind of a generic event emitter / listener system that would allow third-party code to hook into anything that happens with a graph.

[krlmlr]

An error would be better here. We discussed before that this should be a vertex attribute, AFAIR this isn't something that the C core currently supports.

[krlmlr]

The R core could support it, though, with #1894. What would it take to make the following code produce the right results?

library(igraph, warn.conflicts = FALSE)

g <- make_full_graph(10)
V(g)$layout <- layout_in_circle(g)
plot(g)

^{Created on 2025-06-26 with reprex v2.1.1}

[schochastics]

Cant remember all discussions here: Should I turn the warning into an error message and then merge?

[krlmlr]

My question in #1880 (comment) is still relevant.

[schochastics]

I'll investigate but I think this would be a big change.

1. Matrix attributes are not supported for vertices and 2) The plot code assumes layout either to be a matrix graph attribute or x and y vertex attributes

[krlmlr]

1. The code shows it’s possible to store them? 2. We own the plotting code?

[krlmlr]

Specifically, this works with #1894:

library(igraph, warn.conflicts = FALSE)

g <- make_full_graph(10)
V(g)$layout <- layout_in_circle(g)
V(g)$layout
#>            [,1]          [,2]
#>  [1,]  1.000000  0.000000e+00
#>  [2,]  0.809017  5.877853e-01
#>  [3,]  0.309017  9.510565e-01
#>  [4,] -0.309017  9.510565e-01
#>  [5,] -0.809017  5.877853e-01
#>  [6,] -1.000000  1.224647e-16
#>  [7,] -0.809017 -5.877853e-01
#>  [8,] -0.309017 -9.510565e-01
#>  [9,]  0.309017 -9.510565e-01
#> [10,]  0.809017 -5.877853e-01

^{Created on 2025-07-05 with reprex v2.1.1}

[schochastics]

Then let's make it happen. I could adapt the plot code to also check for a vertex layout attribute

[szhorvat]

There are some issue with this approach. Since vector attributes are not supported in C yet, it will not be possible to export these to the standard formats such as GML.

In the long term it would be good to do this, but I'm wondering if it's a better approach to wait until the C core supports it, or at least wait until upgrading to C/igraph 1.0. I have an issue open for this here: igraph/igraph#2744 Realistically, we probably won't have the capacity to make it happen for 1.0, but I'd like to be optimistic and take another look ...

[szhorvat]

(To be clear, IMO this is absolutely the right way to do it. My concern is making this an R-specific feature, especially before we had the opportunity to take a good look at it in C and determine what interface is feasible.)

[krlmlr]

We also need to check subsetting and subset assignment for such attributes. In the other PR.

[krlmlr]

We could also make V(g)$layout <- … result in setting the x and y vertex attributes. Or perhaps create a layout setting function. I need to review the attribute code to give a better assessment.

[szhorvat]

How I envisioned this working was exactly V(g)$layout <- matrix, and internally, in C, such an attribute would be stored precisely as a matrix. Some of the concerns I have, that I think we need to think through, are:

• Issue with columns-major vs row-major storage in C. @krlmlr, can you confirm that R uses column-major (just like igraph)?
• How well this fits various file formats?
• We had some discussions with @ntamas about proposing a new format, but now that the NASCOL forum is dead (CC @vtraag) this'll be harder to push through.

And of course the biggest issue: Implementing this feature with the very limited capacity we have ...

[szhorvat]

How about going with @krlmlr's design, and setting the layout fixed-length vector attribute, just as you proposed, but also providing a feature (i.e. a conveniene function) to convert between the layout and x/y formats? This would make it easy for people to convert to a format that is easily exportable with the current features.

[schochastics]

Would this be a separate PR?
I feel like the scope of this PR is only to provide a warning/error for layouts as graph attributes. A separate PR would be to allow layout as node attribute + conversion?
Happy to put this all into this PR though if that is fine

[krlmlr]

Let's turn this into an error then, and address the other topics later.

[schochastics]

@krlmlr ok to merge if everything passes?

[krlmlr]

Thanks, merging.