Very small numbers print as Inf.e-324

[ccsarapas]

Reprex:

library(pillar)

pillar(4e-324)
#> <pillar>
#>     <dbl>
#> Inf.e-324

This looks similar to tidyverse/tibble#377, which was previously closed at d3e1384.

[bbolker]

I dug in a little bit here: https://stackoverflow.com/a/74131254/190277 but didn't make it all the way to the bottom ...

[rikivillalba]

Problem appears to be in the function safe_divide_10_to <- function(x, y) { given in R 4.940656e-324 /(10^-324) gives Inf.

[bbolker]

It's pretty funny that the source code is

safe_divide_10_to <- function(x, y) {
  # Computes x / 10^y in a robust way

  x / (10^y)
}

Did someone mean to implement a robust division by 10^y and forget? Should this be exp(log(x) - y*log(10)) ? Is this called often enough that the computational cost is a problem? (M_LN10 is defined in C/C++ header files ...)

In this particular case the solution to tidyverse/tibble#377 isn't useful, as x*10^(-y) and x/10^y are both Inf ...

[krlmlr]

Thanks, good catch.

Would 10 ^ (log10(x) - y) also work? Or is 10 ^ . just a variant of exp() under the hood?

Is this a case of comment rot? I don't remember what happened here off the top of my head.

[bbolker]

Yes, that will work. Speed is hardly different (note units are in nanoseconds ...)

x <- 4.94e-324
y <- -324
library(microbenchmark)
l10 <- log(10)  ## precompute
f1 <- function(x,y) exp(log(x)-y*l10)
f2 <- function(x,y) 10^(log10(x)-y)
all.equal(f1(x,y), f2(x,y))
mb <- microbenchmark(f1(x,y), f2(x,y), times = 10000L)

Unit: nanoseconds
     expr min  lq     mean median  uq  max neval cld
 f1(x, y) 641 671 712.3680    691 721 6222 10000  a 
 f2(x, y) 661 691 740.3547    711 742 9337 10000   b

[rikivillalba]

This also worked for me https://stackoverflow.com/a/74132326/6912817 in Windows / Intel(R) Core(TM) i5-8350U CPU @ 1.70GHz

library(tibble)

print(tibble(x = c(4e-300, 4e-324, 4e324, 10, -10, 4e-350)))
#> # A tibble: 6 × 1
#>           x
#>       <dbl>
#> 1   4 e-300
#> 2 Inf.e-324
#> 3 Inf      
#> 4   1 e+  1
#> 5  -1 e+  1
#> 6   0

assignInNamespace("safe_divide_10_to", 
                  function(x,y) 10^(log10(x) - y), "pillar")

print(tibble(x = c(4e-300, 4e-324, 4e324, 10, -10, 4e-350)))
#> # A tibble: 6 × 1
#>             x
#>         <dbl>
#> 1   4   e-300
#> 2   4.94e-324
#> 3 Inf        
#> 4   1   e+  1
#> 5  -1   e+  1
#> 6   0