performance - Here we go again: append an element to a list in R

Question

I am not happy with the accepted answer to Append an object to a list in R in amortized constant time?

> list1 <- list("foo", pi)
> bar <- list("A", "B")

How can I append new element bar to list1? Clearly, c() does not work, it flattens bar:

> c(list1, bar)
[[1]]
[1] "foo"

[[2]]
[1] 3.141593

[[3]]
[1] "A"

[[4]]
[1] "B"

Assignment to index works:

> list1[[length(list1)+1]] <- bar
> list1
[[1]]
[1] "foo"

[[2]]
[1] 3.141593

[[3]]
[[3]][[1]]
[1] "A"

[[3]][[2]]
[1] "B"

What is the efficiency of this method? Is there a more elegant way?

Answer 1

Adding elements to a list is very slow when doing it one element at a time. See these two examples:

I'm keeping the Result variable in the global environment to avoid copies to evaluation frames and telling R where to look for it with .GlobalEnv$, to avoid a blind search with <<-:

Result <- list()

AddItemNaive <- function(item)
{
    .GlobalEnv$Result[[length(.GlobalEnv$Result)+1]] <- item
}

system.time(for(i in seq_len(2e4)) AddItemNaive(i))
#   user  system elapsed 
#  15.60    0.00   15.61 

Slow. Now let's try the second approach:

Result <- list()

AddItemNaive2 <- function(item)
{
    .GlobalEnv$Result <- c(.GlobalEnv$Result, item)
}

system.time(for(i in seq_len(2e4)) AddItemNaive2(i))
#   user  system elapsed 
#  13.85    0.00   13.89

Still slow.

Now let's try using an environment, and creating new variables within this environment instead of adding elements to a list. The issue here is that variables must be named, so I'll use the counter as a string to name each item "slot":

Counter <- 0
Result <- new.env()

AddItemEnvir <- function(item)
{
    .GlobalEnv$Counter <- .GlobalEnv$Counter + 1

    .GlobalEnv$Result[[as.character(.GlobalEnv$Counter)]] <- item
}

system.time(for(i in seq_len(2e4)) AddItemEnvir(i))
#   user  system elapsed 
#   0.36    0.00    0.38 

Whoa much faster. :-) It may be a little awkward to work with, but it works.

A final approach uses a list, but instead of augmenting its size one element at a time, it doubles the size each time the list is full. The list size is also kept in a dedicated variable, to avoid any slowdown using length:

Counter <- 0
Result <- list(NULL)
Size <- 1

AddItemDoubling <- function(item)
{
    if( .GlobalEnv$Counter == .GlobalEnv$Size )
    {
        length(.GlobalEnv$Result) <- .GlobalEnv$Size <- .GlobalEnv$Size * 2
    }

    .GlobalEnv$Counter <- .GlobalEnv$Counter + 1

    .GlobalEnv$Result[[.GlobalEnv$Counter]] <- item
}

system.time(for(i in seq_len(2e4)) AddItemDoubling(i))
#   user  system elapsed 
#   0.22    0.00    0.22

It's even faster. And as easy to a work as any list.

Let's try these last two solutions with more iterations:

Counter <- 0
Result <- new.env()

system.time(for(i in seq_len(1e5)) AddItemEnvir(i))
#   user  system elapsed 
#  27.72    0.06   27.83 


Counter <- 0
Result <- list(NULL)
Size <- 1

system.time(for(i in seq_len(1e5)) AddItemDoubling(i))
#   user  system elapsed 
#   9.26    0.00    9.32

Well, the last one is definetely the way to go.