fm_rcdt_2d_inla vs. inla.mesh.create differences

[Sean Anderson]

Thank you for all the work extracting mesh construction to the fmesher package on CRAN.

I have noticed a difference in the meshes produced by fmesher::fm_rcdt_2d_inla() compared to INLA::inla.mesh.create() and INLA <= 23.06.29. See a minimal example below using fmesher.evolution to use the old code. The difference seems to be with the outer boundary and triangles between that boundary and the data. Is there an obvious set of arguments I can use to make these fmesher package meshes match the older INLA ones?

set.seed(1)
dat <- matrix(runif(200), ncol = 2L)
INLA::inla.setOption(fmesher.evolution = 1L)

packageVersion("INLA")
#> [1] '23.9.9'

packageVersion("fmesher")
#> [1] ‘0.1.2’

inla_mesh <- INLA::inla.mesh.create(
  dat,
  cutoff = 0.1
)
fmesher_mesh <- fmesher::fm_rcdt_2d_inla(
  dat,
  cutoff = 0.1
)

par(mfrow = c(1, 2))
plot(inla_mesh, asp = 1, main = NA);points(dat)
mtext("INLA::inla.mesh.create() & \nINLA <= INLA_23.06.29 \n fmesher.evolution = 1")
plot(fmesher_mesh);points(dat)
mtext("fmesher::fm_rcdt_2d_inla() | \nINLA >= INLA_23.08.18 \n fmesher.evolution = 2")

Resulting meshes:

https://i.imgur.com/ksc18LC.png

Separately, are there any plans to pull INLA::inla.spde2.matern() into fmesher or some other package hosted on CRAN?

Thanks,

Sean

[Finn Lindgren]

I've taken a quick look at the code&documentation, and the argument
would be extend=list(n=16) to get 16 outer points,

What you're experiencing is a difference in the handling of the case
extend=TRUE, where the fmesher version doesn't set any extension
defaults, causing the defaults to be set by the internal C++ code.
You can use extend = list() to regain the old behaviour until the next
fmesher version (I'll update fmesher on github today).

Interestingly though, both the old and new code have a
documentation/code mismatch in the defaults; the R documentation says
8 but the R code uses 16.
The discrepancy above is because the C++ code default follows the documentation!

Another thing to be aware of is that due to different floating point
behaviour on M1 processors, M1 systems can get different meshes than
on intel and amd. (This is not new behaviour for fmesher, but has
always been true for M1; it was just discovered when testing the
fmesher package for CRAN submission).
So make sure to write code that either doesn't rely on the specific
mesh, and store a specific mesh to disk if/when you need to make sure
it never changes.

Finn

On Tue, 26 Sept 2023 at 08:18, Finn Lindgren <finn.l...@gmail.com> wrote:
>
> I’m not sure why default “n” seems to have changed; I’ll look into it.
> Finn
>
> On 26 Sep 2023, at 08:17, Finn Lindgren <finn.l...@gmail.com> wrote:
>
> Looks like the outer extension setting is different; the new one seems to use n=8 and the old n=16. See the function help text, but it should be an “n” argument.
>
> Inla.spde2.matern itself doesn’t really make sense without inla, but the underlying code is already in fmesher; search for matern and/or precision in the fmesher documentation on github.io.
>
> Finn
>
> On 26 Sep 2023, at 07:33, Sean Anderson <se...@seananderson.ca> wrote:
>
> Thank you for all the work extracting mesh construction to the fmesher package on CRAN.

> --
> You received this message because you are subscribed to the Google Groups "R-inla discussion group" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to r-inla-discussion...@googlegroups.com.
> To view this discussion on the web, visit https://groups.google.com/d/msgid/r-inla-discussion-group/44323df0-a724-4c1c-bd83-ea878f00c45cn%40googlegroups.com.



--
Finn Lindgren
email: finn.l...@gmail.com

[Finn Lindgren]

Looks like the outer extension setting is different; the new one seems to use n=8 and the old n=16. See the function help text, but it should be an “n” argument.

Inla.spde2.matern itself doesn’t really make sense without inla, but the underlying code is already in fmesher; search for matern and/or precision in the fmesher documentation on github.io.

Finn

On 26 Sep 2023, at 07:33, Sean Anderson <se...@seananderson.ca> wrote:

Thank you for all the work extracting mesh construction to the fmesher package on CRAN.

--

[Finn Lindgren]

I’m not sure why default “n” seems to have changed; I’ll look into it.

Finn

On 26 Sep 2023, at 08:17, Finn Lindgren <finn.l...@gmail.com> wrote:

Looks like the outer extension setting is different; the new one seems to use n=8 and the old n=16. See the function help text, but it should be an “n” argument.