R: Global assignment of vector element works only inside a function

问题

I'm working on a project where there are some global assignments, and I ran into something sort of odd. I was hoping someone could help me with it.

I wrote this toy example to demonstrate the problem:

x <-  1:3 ; x <-  c(1, 2, 5) # this works fine
x <-  1:3 ; x[3] <- 5        # this works fine

x <<- 1:3 ; x <<- c(1, 2, 5) # this works fine
x <<- 1:3 ; x[3] <<- 5       # this does not work
# Error in x[3] <<- 5 : object 'x' not found

same.thing.but.in.a.function = function() {
  x <<- 1:3
  x[3] <<- 5
}
same.thing.but.in.a.function(); x
# works just fine

So, it seems it's not possible to change part of a vector using a global assignment -- unless that assignment is contained within a function. Can anyone explain why this is the case?

回答1:

I figured out the problem.

Basically, in this manifestation of <<- (which is more accurately called the "superassignment operator" rather than the "global assignment operator"), it actually skips checking the global environment when trying to access the variable.

On page 19 of R Language Definition, it states the following:

x <<- data.frame(0, 0, 0) # (I added this so the code can be run)
names(x)[3] <<- "Three"

is equivalent to

x <<- data.frame(0, 0, 0) # (I added this so the code can be run)
`*tmp*` <<- get(x, envir=parent.env(), inherits=TRUE)
names(`*tmp*`)[3] <- "Three"
x <<- `*tmp*`
rm(`*tmp*`)

When I tried to run those four lines, it threw an error -- parent.env requires an argument and has no default. I can only assume that the documentation was written at a time when parent.env() contained a default value for its first argument. But I can safely guess that the default would have been environment() which returns the current environment. It then throws an error again -- x needs to be in quotes. So I fixed that too. Now, when I run the first line, it throws the same error message as I encountered originally, but with more detail:

# Error in get("x", envir = parent.env(environment()), inherits = TRUE) :
#   object 'x' not found

This makes sense -- environment() itself returns .GlobalEnv, so parent.env(.GlobalEnv) misses out on the global environment entirely, instead returning the most recently loaded package environment. Then, since inherits is set to TRUE, the get() function keeps going up the levels, searching through each of the loaded package environments before eventually reaching the empty environment, and at that point it has still not found x. Thus the error.

Since parent.env(environment()) will return .GlobalEnv (or another environment below it) as long as you start inside a local environment, this same problem does not occur when the same lines are run from inside a local environment:*

local({
  x <<- data.frame(0, 0, 0) # (I added this so the code can be run)
  `tmp` <<- get("x", envir=parent.env(environment()), inherits=TRUE)
  names(`tmp`)[3] <- "Three"
  x <<- `tmp`
  rm(`tmp`)
})
x
#   X0 X0.1 Three
# 1  0    0     0

# so, it works properly

In contrast, when <<- is used in general, there is no extra subsetting code that occurs behind the scenes, and it first attempts to access the value in the current environment (which might be the global environment), before moving upwards. So in that situation, it doesn't run into the problem where it skips the global environment.

_{* I had to change the variable from *tmp* to tmp because one of the behind-the-scenes operations in the code uses the *tmp* variable and then removes it, so *tmp* disappears in the middle of line 3 and so it throws an error when I then try to access it.}