Żmij for R: Floating-point to string conversion with round-trip guarantees

Why bother?

I like storing data as plain text when possible: humans can read it, diff it, version control it, and edit it with any text editor. However, floating-point numbers make this harder than it should be. By default, R converts them to strings using 15 significant digits and quietly loses information:

x <- c(pi, 0.1, 1 / 3)
x |> as.character()
#> [1] "3.14159265358979" 
#> [2] "0.1"              
#> [3] "0.333333333333333"
x |> as.character() |> as.numeric() |> identical(x)
#> [1] FALSE

You can do a round-trip conversion reliably by asking for 17 significant digits, but then you get this:

x |> sprintf(fmt = "%.17g")
#> [1] "3.1415926535897931" 
#> [2] "0.10000000000000001"
#> [3] "0.33333333333333331"
x |> sprintf(fmt = "%.17g") |> as.numeric() |> identical(x)
#> [1] TRUE

So the traditional choice was between losing precision, bloating your files with digit noise, or giving up on plaintext workflow entirely and using a binary format like .rds, .parquet, or .safetensors.

The dtoa (double to ASCII) algorithms eliminate this trade-off. For every floating-point number, it finds the shortest decimal string that parses back to exactly the same value: 0.1 stays "0.1", while pi keeps all 16 digits it genuinely needs. At the moment, zmij also happens to be the fastest known algorithm to do so with proven correctness: Victor’s blog post has the benchmarks and the tricks behind them. From R, crossing the language boundary will absorb most of those nanoseconds, which I can live with. The round-trip guarantee is what I came for.

Usage

zmij provides two functions. format_double() turns doubles into strings, and parse_double() reverses it:

library(zmij)

x <- c(pi, 0.1, 1 / 3, .Machine$double.xmax)
x |> format_double()
#> [1] "3.141592653589793"      
#> [2] "0.1"                    
#> [3] "0.3333333333333333"     
#> [4] "1.7976931348623157e+308"
x |> format_double() |> parse_double() |> identical(x)
#> [1] TRUE

That TRUE holds for every finite double, meaning your numbers can survive a round-trip through any text file untouched:

path <- tempfile()
x |> format_double() |> writeLines(path)
path |> readLines() |> parse_double() |> identical(x)
#> [1] TRUE

Both functions are vectorized and preserve names and dimensions, so matrices and arrays pass through with their shape intact. Missing values and non-finite values also get explicit representations:

c(NA_real_, NaN, Inf, -Inf) |> format_double()
#> [1] NA     "NaN"  "inf"  "-inf"

parse_double() is useful in its own right beyond undoing format_double(): it uses Rust’s correctly rounded parser, which behaves consistently across platforms even at the far edges of the double range:

c("5e-324", "1.7976931348623157e+308") |> parse_double()
#> [1] 4.940656e-324 1.797693e+308

Development notes

zmij is my first package built with extendr. extendr scaffolds a Rust-powered R package and handles the plumbing between the two languages. The experience was pleasantly boring: most of my time went into building the Rust wrapper logic, and almost none into the bridge or packaging.

One piece of developer tooling deserves a mention. My r-rust-pkgs project tracks every CRAN package that uses Rust, and it now comes with a script that turns that list into something directly useful:

git clone https://github.com/nanxstats/r-rust-pkgs.git
cd r-rust-pkgs
Rscript scripts/monorepo.R

This downloads the source of all those packages into a single “synthetic monorepo” under references/. Move that directory into your own package (remember to add it to .gitignore and .Rbuildignore) and your AI coding agents have a greppable record of how the entire Rust-in-R community has solved the problems you might hit. When zmij failed to build on win-builder, a quick grep showed other packages had already patched their Makevars templates for the same issue, and the fix (nanxstats/zmij#6) only took a few minutes to implement.

Thanks to Victor Zverovich for the algorithm and to David Tolnay for the Rust crate. If you find a number the package mishandles, please create an issue.