title: "Invariants: Comparing behavior with data frames"
output: rmarkdown::html_vignette vignette: > %\VignetteIndexEntry{Invariants: Comparing behavior with data frames} %\VignetteEngine{knitr::rmarkdown} %\VignetteEncoding{UTF-8}
.dftbl { width: 100%; table-layout: fixed; display: inline-table; } .error pre code { color: red; } .warning pre code { color: violet; }This vignette defines invariants for subsetting and subset-assignment for tibbles, and illustrates where their behaviour differs from data frames.
The goal is to define a small set of invariants that consistently define how behaviors interact.
Some behaviors are defined using functions of the vctrs package, e.g. vec_slice()
, vec_recycle()
and vec_as_index()
.
Refer to their documentation for more details about the invariants that they follow.
The subsetting and subassignment operators for data frames and tibbles are particularly tricky, because they support both row and column indexes, both of which are optionally missing.
We resolve this by first defining column access with [[
and $
, then column-wise subsetting with [
, then row-wise subsetting, then the composition of both.
In this article, all behaviors are demonstrated using one example data frame and its tibble equivalent:
library(tibble)
library(vctrs)
new_df <- function() {
df <- data.frame(n = c(1L, NA, 3L, NA))
df$c <- letters[5:8]
df$li <- list(9, 10:11, 12:14, "text")
df
}
new_tbl <- function() {
as_tibble(new_df())
}
Results of the same code for data frames and tibbles are presented side by side:
wzxhzdk:1 wzxhzdk:2If the results are identical (after converting to a data frame if necessary), only the tibble result is shown.
Subsetting operations are read-only. The same objects are reused in all examples:
df <- new_df()
tbl <- new_tbl()
Where needed, we also show examples with hierarchical columns containing a data frame or a matrix:
new_tbl2 <- function() {
tibble(
tb = tbl,
m = diag(4)
)
}
new_df2 <- function() {
df2 <- new_tbl2()
class(df2) <- "data.frame"
class(df2$tb) <- "data.frame"
df2
}
df2 <- new_df2()
tbl2 <- new_tbl2()
wzxhzdk:5
wzxhzdk:6
For subset assignment (subassignment, for short), we need a fresh copy of the data for each test.
The with_*()
functions (omitted here for brevity) allow for a more concise notation.
These functions take an assignment expression, execute it on a fresh copy of the data, and return the data for printing.
The first example prints what's really executed, further examples omit this output.
x[[j]]
x[[j]]
is equal to .subset2(x, j)
.
NB: x[[j]]
always returns an object of size nrow(x)
if the column exists.
j
must be a single number or a string, as enforced by .subset2(x, j)
.
NA
indexes, numeric out-of-bounds (OOB) values, and non-integers throw an error:
Character OOB access is silent because a common package idiom is to check for the absence of a column with is.null(df[[var]])
.
x$name
x$name
and x$"name"
are equal to x[["name"]]
.
Unlike data frames, tibbles do not partially match names.
Because df$x
is rarely used in packages, it can raise a warning:
x[j]
j
is converted to an integer vector by vec_as_index(j, ncol(x), names = names(x))
.
Then x[c(j_1, j_2, ..., j_n)]
is equivalent to tibble(x[[j_1]], x[[j_2]], ..., x[[j_n]])
, keeping the corresponding column names.
This implies that j
must be a numeric or character vector, or a logical vector with length 1 or ncol(x)
.[^subset-extract-commute]
[^subset-extract-commute]: x[j][[jj]]
is equal to x[[ j[[jj]] ]]
, in particular x[j][[1]]
is equal to x[[j]]
for scalar numeric or integer j
.
When subsetting repeated indexes, the resulting column names are undefined, do not rely on them.
wzxhzdk:67 wzxhzdk:68For tibbles with repeated column names, subsetting by name uses the first matching column.
nrow(df[j])
equals nrow(df)
.
Tibbles support indexing by a logical matrix, but only if all values in the returned vector are compatible.
wzxhzdk:71 wzxhzdk:72 wzxhzdk:73 wzxhzdk:74x[, j]
x[, j]
is equal to x[j]
.
Tibbles do not perform column extraction if x[j]
would yield a single column.
x[, j, drop = TRUE]
For backward compatiblity, x[, j, drop = TRUE]
performs column extraction, returning x[j][[1]]
when ncol(x[j])
is 1.
x[i, ]
x[i, ]
is equal to tibble(vec_slice(x[[1]], i), vec_slice(x[[2]], i), ...)
.[^row-subset-efficiency]
[^row-subset-efficiency]: Row subsetting x[i, ]
is not defined in terms of x[[j]][i]
because that definition does not generalise to matrix and data frame columns.
For efficiency and backward compatibility, i
is converted to an integer vector by vec_as_index(i, nrow(x))
first.
This means that i
must be a numeric vector, or a logical vector of length nrow(x)
or 1.
For compatibility, i
can also be a character vector containing positive numbers.
Exception: OOB values generate warnings instead of errors:
wzxhzdk:99 wzxhzdk:100 wzxhzdk:101 wzxhzdk:102Unlike data frames, only logical vectors of length 1 are recycled.
wzxhzdk:103 wzxhzdk:104NB: scalar logicals are recycled, but scalar numerics are not.
That makes the x[NA, ]
and x[NA_integer_, ]
return different results.
x[i, , drop = TRUE]
drop = TRUE
has no effect when not selecting a single row:
x[]
and x[,]
x[]
and x[,]
are equivalent to x
.[^bracket-comma]
[^bracket-comma]: x[,]
is equivalent to x[]
because x[, j]
is equivalent to x[j]
.
x[i, j]
x[i, j]
is equal to x[i, ][j]
.[^bracket-flip]
[^bracket-flip]: A more efficient implementation of x[i, j]
would forward to x[j][i, ]
.
x[[i, j]]
i
must be a numeric vector of length 1.
x[[i, j]]
is equal to x[i, ][[j]]
, or vctrs::vec_slice(x[[j]], i)
.[^bracket2-flip]
[^bracket2-flip]: Cell subsetting x[[i, j]]
is not defined in terms of x[[j]][[i]]
because that definition does not generalise to list, matrix and data frame columns.
A more efficient implementation of x[[i, j]]
would check that j
is a scalar and forward to x[i, j][[1]]
.
df[[1, 1]]
#> [1] 1
df[[1, 3]]
#> [1] 9
This implies that j
must be a numeric or character vector of length 1.
NB: vec_size(x[[i, j]])
always equals 1.
Unlike x[i, ]
, x[[i, ]]
is not valid.
x[[j]] <- a
If a
is a vector then x[[j]] <- a
replaces the j
th column with value a
.
a
is recycled to the same size as x
so must have size nrow(x)
or 1.
(The only exception is when a
is NULL
, as described below.)
Recycling also works for list, data frame, and matrix columns.
j
must be a scalar numeric or a string, and cannot be NA
.
If j
is OOB, a new column is added on the right hand side, with name repair if needed.
df[[j]] <- a
replaces the complete column so can change the type.
[[<-
supports removing a column by assigning NULL
to it.
Removing a nonexistent column is a no-op.
wzxhzdk:184 wzxhzdk:185x$name <- a
x$name <- a
and x$"name" <- a
are equivalent to x[["name"]] <- a
.[^column-assign-symmetry]
[^column-assign-symmetry]: $
behaves almost completely symmetrically to [[
when comparing subsetting and subassignment.
$<-
does not perform partial matching.
x[j] <- a
j
is missing, it's replaced with seq_along(x)
j
is logical vector, it's converted to numeric with seq_along(x)[j]
.a
is a list or data frameIf inherits(a, "list")
or inherits(a, "data.frame")
is TRUE
, then x[j] <- a
is equivalent to x[[j[[1]]] <- a[[1]]
, x[[j[[2]]]] <- a[[2]]
, ...
If length(a)
equals 1, then it is recycled to the same length as j
.
An attempt to update the same column twice gives an error.
wzxhzdk:206 wzxhzdk:207If a
contains NULL
values, the corresponding columns are removed after updating (i.e. position indexes refer to columns before any modifications).
NA
indexes are not supported.
Just like column updates, [<-
supports changing the type of an existing column.
Appending columns at the end (without gaps) is supported. The name of new columns is determined by the LHS, the RHS, or by name repair (in that order of precedence).
wzxhzdk:226 wzxhzdk:227 wzxhzdk:228 wzxhzdk:229 wzxhzdk:230 wzxhzdk:231 wzxhzdk:232 wzxhzdk:233Tibbles support indexing by a logical matrix, but only for a scalar RHS, and if all columns updated are compatible with the value assigned.
wzxhzdk:234 wzxhzdk:235 wzxhzdk:236 wzxhzdk:237 wzxhzdk:238 wzxhzdk:239a
is a matrix or arrayIf is.matrix(a)
, then a
is coerced to a data frame with as.data.frame()
before assigning.
If rows are assigned, the matrix type must be compatible with all columns.
If is.array(a)
and any(dim(a)[-1:-2] != 1)
, an error is thrown.
a
is another type of vectorIf vec_is(a)
, then x[j] <- a
is equivalent to x[j] <- list(a)
.
This is primarily provided for backward compatibility.
Matrices must be wrapped in list()
before assignment to create a matrix column.
a
is NULL
Entire columns can be removed.
Specifying i
is an error.
a
is not a vectorAny other type for a
is an error.
Note that if is.list(a)
is TRUE
, but inherits(a, "list")
is FALSE
, then a
is considered to be a scalar.
See ?vec_is
and ?vec_proxy
for details.
x[i, ] <- list(...)
x[i, ] <- a
is the same as vec_slice(x[[j_1]], i) <- a[[1]]
, vec_slice(x[[j_2]], i) <- a[[2]]
, ... .[^row-assign-symmetry]
[^row-assign-symmetry]: x[i, ]
is symmetrical for subset and subassignment.
Only values of size one can be recycled.
wzxhzdk:292 wzxhzdk:293 wzxhzdk:294 wzxhzdk:295 wzxhzdk:296 wzxhzdk:297 wzxhzdk:298 wzxhzdk:299 wzxhzdk:300 wzxhzdk:301For compatibility, only a warning is issued for indexing beyond the number of rows. Appending rows right at the end of the existing data is supported, without warning.
wzxhzdk:302 wzxhzdk:303 wzxhzdk:304 wzxhzdk:305 wzxhzdk:306 wzxhzdk:307 wzxhzdk:308 wzxhzdk:309 wzxhzdk:310 wzxhzdk:311 wzxhzdk:312 wzxhzdk:313For compatibility, i
can also be a character vector containing positive numbers.
x[i, j] <- a
x[i, j] <- a
is equivalent to x[i, ][j] <- a
.[^bracket-assign-flip]
[^bracket-assign-flip]: x[i, j]
is symmetrical for subsetting and subassignment.
A more efficient implementation of x[i, j] <- a
would forward to x[j][i, ] <- a
.
Subassignment to x[i, j]
is stricter for tibbles than for data frames.
x[i, j] <- a
can't change the data type of existing columns.
A notable exception is the population of a column full of NA
(which is of type logical
), or the use of NA
on the right-hand side of the assignment.
For programming, it is always safer (and faster) to use the correct type of NA
to initialize columns.
For new columns, x[i, j] <- a
fills the unassigned rows with NA
.
Likewise, for new rows, x[i, j] <- a
fills the unassigned columns with NA
.
x[[i, j]] <- a
i
must be a numeric vector of length 1.
x[[i, j]] <- a
is equivalent to x[i, ][[j]] <- a
.[^double-bracket-ij-symmetry]
[^double-bracket-ij-symmetry]: x[[i, j]]
is symmetrical for subsetting and subassignment.
An efficient implementation would check that i
and j
are scalar and forward to x[i, j][[1]] <- a
.
NB: vec_size(a)
must equal 1.
Unlike x[i, ] <-
, x[[i, ]] <-
is not valid.
Any scripts or data that you put into this service are public.
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.