Different likelihoods between numeric and AD versions of function

[andrjohns]

I've got the following reprex which appears to be returning a slightly different likelihood for AD inputs. From some basic testing, the individual components appear consistent but their combination diverges:

set.seed(2024)
L <- chol(matrix(c(1, 0.5, 0.5, 1), nrow = 2))
z <- matrix(rnorm(1000), ncol = 2) %*% L

y <- qexp(pnorm(z, log.p = TRUE), rate = 2, log.p = TRUE)

library(RTMB)

ll_function <- function(pars) {
  q <- qnorm(pexp(y, rate = pars[1]))
  
  Sigma <- matrix(c(1, pars[2], pars[2], 1), nrow = 2)
  sum(RTMB::dmvnorm(q, Sigma = Sigma, log = TRUE))
}

obj <- MakeADFun(
  func = function(pars) ll_function(pars$v),
  parameters = list(v = c(0.1,0.5)),
  silent=TRUE
)

obj$fn(c(0.1,0.5))
#> [1] -2177.436
ll_function(c(0.1,0.5))
#> [1] -2127.248

^{Created on 2025-10-16 with reprex v2.1.1}

─ Session info ───────────────────────────────────────────────────────────────────────────────────────────────────────────────
 setting  value
 version  R version 4.5.1 (2025-06-13 ucrt)
 os       Windows 11 x64 (build 26100)
 system   x86_64, mingw32
 ui       RStudio
 language (EN)
 collate  English_Australia.utf8
 ctype    English_Australia.utf8
 tz       Australia/Perth
 date     2025-10-16
 rstudio  2025.09.1+401 Cucumberleaf Sunflower (desktop)
 pandoc   3.6.3 @ C:/Program Files/RStudio/resources/app/bin/quarto/bin/tools/ (via rmarkdown)

─ Packages ───────────────────────────────────────────────────────────────────────────────────────────────────────────────────
 package   * version   date (UTC) lib source
 lattice     0.22-7    2025-04-02 [2] CRAN (R 4.5.1)
 MASS        7.3-65    2025-02-28 [2] CRAN (R 4.5.1)
 Matrix      1.7-4     2025-08-28 [1] CRAN (R 4.5.1)
 Rcpp        1.1.0     2025-07-02 [1] CRAN (R 4.5.1)
 RcppEigen   0.3.4.0.2 2024-08-24 [1] CRAN (R 4.5.1)
 RTMB        1.8       2025-10-16 [1] Github (kaskr/RTMB@aab2fb7)
 TMB         1.9.18    2025-10-13 [1] CRAN (R 4.5.1)

 [1] C:/Users/Andrew/AppData/Local/R/win-library/4.5
 [2] C:/Program Files/R/R-4.5.1/library

[kaskr]

Thanks for the report.

The problem is in this line

q <- qnorm(pexp(y, rate = pars[1]))

where the AD version loses the dimension attribute of y.
A workaround, that fixes the problem in this case, is then to add the line:

dim(q) <- dim(y)

I conclude there are two bugs in RTMB:

1. Lost attributes.
2. Why did dmvnorm not throw an error?