This is an introduction to the R language, explaining evaluation, parsing, object oriented programming, computing on the language, and so forth.
The current version of this document is x.x.x (xxxx-xx-xx) DRAFT.
ISBN 3-900051-13-5
R is a system for statistical computation and graphics. It provides, among other things, a programming language, high level graphics, interfaces to other languages and debugging facilities. This manual details and defines the R language.
The R language is a dialect of S which was designed in the 1980s and has been in widespread use in the statistical community since. Its principal designer, John M. Chambers, was awarded the 1998 ACM Software Systems Award for S.
The language syntax has a superficial similarity with C, but the semantics are of the FPL (functional programming language) variety with stronger affinities with Lisp and APL. In particular, it allows “computing on the language”, which in turn makes it possible to write functions that take expressions as input, something that is often useful for statistical modeling and graphics.
It is possible to get quite far using R interactively, executing simple expressions from the command line. Some users may never need to go beyond that level, others will want to write their own functions either in an ad hoc fashion to systematize repetitive work or with the perspective of writing add-on packages for new functionality.
The purpose of this manual is to document the language per se. That is, the objects that it works on, and the details of the expression evaluation process, which are useful to know when programming R functions. Major subsystems for specific tasks, such as graphics, are only briefly described in this manual and will be documented separately.
Although much of the text will equally apply to S, there are also some substantial differences, and in order not to confuse the issue we shall concentrate on describing R.
The design of the language contains a number of fine points and common pitfalls which may surprise the user. Most of these are due to consistency considerations at a deeper level, as we shall explain. There are also a number of useful shortcuts and idioms, which allow the user to express quite complicated operations succinctly. Many of these become natural once one is familiar with the underlying concepts. In some cases, there are multiple ways of performing a task, but some of the techniques will rely on the language implementation, and others work at a higher level of abstraction. In such cases we shall indicate the preferred usage.
Some familiarity with R is assumed. This is not an introduction to R but rather a programmers' reference manual. Other manuals provide complementary information: in particular Preface (R Introduction) provides an introduction to R and System and foreign language interfaces (Writing R Extensions) details how to extend R using compiled code.
In every computer language variables provide a means of accessing the data stored in memory. R does not provide direct access to the computer's memory but rather provides a number of specialized data structures we will refer to as objects. These objects are referred to through symbols or variables. In R, however, the symbols are themselves objects and can be manipulated in the same way as any other object. This is different from many other languages and has wide ranging effects.
In this chapter we provide preliminary descriptions of the various data
structures provided in R. More detailed discussions of many of them
will be found in the subsequent chapters. The R specific function
typeof
returns the type of an R object. Note that in the C code
underlying R, all objects are pointers to a structure with typedef
SEXPREC; the different R data types are represented in C by
SEXPTYPE, which determines how the information in the various
parts of the structure is used.
The following table describes the possible values returned by
typeof and what they are.
I don't think that the user can get a hold of items marked with a `***'; at least not easily; counter examples would be appreciated.
Function mode gives information about the mode of an object
in the sense of Becker, Chambers & Wilks (1988), and is fully compatible
with other implementations of the S language.
Finally, the function storage.mode returns the storage mode
of its argument in the sense of Becker et al. (1988). It is generally
used when calling functions written in another language, such as C or
FORTRAN, to ensure that R objects have the data type expected by the
routine being called. (In the S language, vectors with integer or
real values are both of mode "numeric", so their storage modes
need to be distinguished.)
> x <- 1:3
> typeof(x)
[1] "integer"
> mode(x)
[1] "numeric"
> storage.mode(x)
[1] "integer"
R objects are often coerced to different types during computations. There are also many functions available to perform explicit coercion. When programming in the R language the type of an object generally doesn't affect the computations, however, when dealing with foreign languages or the operating system it is often necessary to ensure that an object is of the correct type.
Vectors can be thought of as contiguous cells containing data. Cells
are accessed through
indexing operations such as
x[5]. More details are given in Indexing.
R has six basic (`atomic') vector types: logical, integer, real, complex, string (or character) and raw. The modes and storage modes for the different vector types are listed in the following table.
typeof mode storage.mode logicallogicallogicalintegernumericintegerdoublenumericdoublecomplexcomplexcomplexcharactercharactercharacterrawrawraw
Single numbers, such as 4.2, and strings, such as "four
point two" are still vectors, of length 1; there are no more basic
types. Vectors with length zero are possible (and useful).
String vectors have mode and storage mode "character". A single
element of a character vector is often referred to as a character
string.
Lists (“generic vectors”) are another kind of data storage. Lists have elements, each of which can contain any type of R object, i.e. the elements of a list do not have to be of the same type. List elements are accessed through three different indexing operations. These are explained in detail in Indexing.
Lists are vectors, and the basic vector types are referred to as atomic vectors where it is necessary to exclude lists.
There are three types of objects that constitute the R language.
They are calls, expressions, and names.
Since R has objects of type "expression" we will try to avoid
the use of the word expression in other contexts. In particular
syntactically correct expressions will be referred to as
statements.
These objects have modes "call", "expression", and
"name", respectively.
They can be created directly from expressions using the quote
mechanism and converted to and from lists by the as.list and
as.call functions.
Components of the
parse tree can be extracted using the standard
indexing operations.
Symbols refer to R
objects. The
name of any R object is usually a
symbol. Symbols can be created through the function quote.
Symbol have mode "name", storage mode "symbol", and type
"symbol". They can be
coerced to and from character strings
using as.character and as.name.
They naturally appear as atoms of parsed expressions, try e.g.
as.list(quote(x + y)).
In R one can have objects of type "expression". An
expression contains one or more statements. A statement is a
syntactically correct collection of
tokens.
Expression objects are special language objects which contain parsed but
unevaluated R statements. The main difference is that an expression
object can contain several such expressions. Another more subtle
difference is that objects of type "expression" are only
evaluated when
explicitly passed to eval, whereas other language objects may get
evaluated in some unexpected cases.
An expression object behaves much like a list and its components should be accessed in the same way as the components of a list.
In R functions are objects and can be manipulated in much the same way as any other object. Functions (or more precisely, function closures) have three basic components: a formal argument list, a body and an environment. The argument list is a comma-separated list of arguments. An argument can be a symbol, or a symbol = default construct, or the special argument .... The second form of argument is used to specify a default value for an argument. This value will be used if the function is called without any value specified for that argument. The ... argument is special and can contain any number of arguments. It is generally used if the number of arguments is unknown or in cases where the arguments will be passed on to another function.
The body is a parsed R statement. It is usually a collection of statements in braces but it can be a single statement, a symbol or even a constant.
A function's environment is the environment that was active at the time that the function was created. Any symbols bound in that environment are captured and available to the function. This combination of the code of the function and the bindings in its environment is called a `function closure', a term from functional programming theory. In this document we generally use the term `function', but use `closure' to emphasize the importance of the attached environment.
It is possible to extract and manipulate the three parts of a closure
object using formals, body, and environment
constructs (all three can also be used on the left hand side of
assignments).
The last of these can be used to remove unwanted environment capture.
When a function is called, a new environment (called the evaluation environment) is created, whose enclosure (see Environment objects) is the environment from the function closure. This new environment is initially populated with the unevaluated arguments to the function; as evaluation proceeds, local variables are created within it.
There is also a facility for converting functions to and from list
structures using as.list and as.function.
These have been included to provide compatibility with S and their
use is discouraged.
There is a special object called NULL. It is used whenever there
is a need to indicate or specify that an object is absent. It should not be
confused with a vector or list of zero length.
The NULL object has no type and no modifiable properties. There
is only one NULL object in R, to which all instances refer. To
test for NULL use is.null. You cannot set attributes on
NULL.
These two kinds of object contain the built-in
functions of R, i.e.,
those that are displayed as .Primitive in code listings. The
difference between the two lies in the argument handling. Built-in
functions have all their arguments evaluated and passed to the primitive
function, in accordance with call-by-value, whereas special
functions pass the unevaluated expressions to the internal function.
From the R language, these objects are just another kind of function,
except that their definition cannot be listed. The typeof
function can distinguish them from interpreted
functions.
Promise objects are part of R's lazy evaluation mechanism. They contain three slots: a value, an expression, and an environment. When a function is called the arguments are matched and then each of the formal arguments is bound to a promise. The expression that was given for that formal argument and a pointer to the environment the function was called from are stored in the promise.
Until that argument is accessed there is no value associated with
the promise. When the argument is accessed, the stored expression is
evaluated in the stored environment, and the result is returned. The
result is also saved by
the promise. The substitute function will extract the content
of the expression slot. This allows the programmer to
access either the value or the expression associated with the promise.
Within the R language, promise objects are almost only seen
implicitly. (In an upcoming release they will never be visible to
R code, as they will always be evaluated when accessed.)
Actual function arguments are of this type.
There is also a delayedAssign function
that will make a promise out of an expression. There is generally no
way in R code to check whether an object is a promise or not, nor is there a
way to use R code to determine the environment of a promise.
The ... object type is stored as a type of list. The components
of ... can be accessed in the usual list manner from C code, but
is not easily accessed as an object in interpreted code. The object can
be captured as a list, so for example in table one sees
args <- list(...)
## ....
for (a in args) {
## ....
If a function has ... as a formal argument then any actual arguments that do not match a formal argument are matched with ....
Environments can be thought of as consisting of two things. A
frame, consisting of a set of symbol-value pairs, and an
enclosure, a pointer to an enclosing environment. When R
looks up the value for a symbol the frame is examined and if a
matching symbol is found its value will be returned. If not, the
enclosing environment is then accessed and the process repeated.
Environments form a tree structure in which the enclosures play the
role of parents. The tree of environments is rooted in an empty
environment, available through emptyenv(), which has no parent.
It is the direct parent of the environment of the base package
(available through the baseenv() function). Formerly
baseenv() had the special value NULL, but as from
version 2.3.0, the use of NULL as an environment is deprecated.
Environments are created implicitly by function calls, as described in
Function objects and Lexical environment. In this case the
environment contains the variables local to the function (including the
arguments), and its enclosure is the environment of the currently called
function. Environments may also be created directly by new.env.
The frame content of an environment can be accessed and manipulated by
use of ls, get and assign as well as eval and
evalq.
The parent.env function may be used to access the enclosure of
an environment.
Unlike other R objects, environments are not copied when passed to functions or used in assignments. Thus, if you assign the same environment to several symbols and change one, the others will change too. In particular, assigning attributes to an environment can lead to surprises.
Pairlist objects are similar to Lisp's dotted-pair lists. They are used
extensively in the internals of R, but are rarely visible in
interpreted code, although they are returned by formals, and can
be created by (e.g.) the pairlist function. A zero-length
pairlist is NULL, as would be expected in Lisp but in contrast to
a zero-length list.
Each such object has three slots, a CAR value, a CDR value and a TAG
value. The TAG value is a text string and CAR and CDR usually
represent, respectively, a list item (head) and the remainder (tail) of
the list with a NULL object as terminator (the CAR/CDR terminology is
traditional Lisp and originally referred to the address and decrement
registers on an early 60's IBM computer).
Pairlists are handled in the R language in exactly the same way as
generic vectors (“lists”). In particular, elements are accessed using
the same [[]] syntax. The use of pairlists is deprecated since
generic vectors are usually more efficient to use. When an internal
pairlist is accessed from R it is generally (including when
subsetted) converted to a generic vector.
In a very few cases pairlists are user-visible: one is .Options.
It is not really possible for an object to be of “Any” type, but it is
nevertheless a valid type value. It gets used in certain (rather rare)
circumstances, e.g. as.vector(x, "any"), indicating that type
coercion should not be done.
All objects except NULL can have one or more attributes attached
to them. Attributes are stored as a list where all elements are named.
The list of attributes can be obtained using attributes and set
by attributes<-,
individual components are accessed using attr and attr<-.
Some attributes have special accessor
functions (e.g. levels<-
for factors) and these should be used when available. In addition to
hiding details of implementation they may perform additional operations.
R attempts to intercept calls to attr<- and to
attributes<- that involve the special attributes and to enforce
the consistency checks.
Matrices and arrays are simply vectors with the attribute dim and
optionally dimnames attached to the vector.
Attributes are used to implement the class structure used in R. If an
object has a class attribute then that attribute will be examined
during
evaluation. The class structure in R is described in detail
in Object-oriented programming.
A names attribute, when present, labels the individual elements of
a vector or list. When an object is printed the names attribute,
when present, is used to label the elements. The names attribute
can also be used for indexing purposes, for example,
quantile(x)["25%"].
One may get and set the names using names and names<-
constructions.
The latter will perform the necessary consistency checks to ensure that
the names attribute has the proper type and length.
Pairlists and one-dimensional arrays are treated specially. For pairlist
objects, a virtual names attribute is used; the names
attribute is really constructed from the tags of the list components.
For one-dimensional arrays the names attribute really accesses
dimnames[[1]].
The dim attribute is used to implement arrays. The content of
the array is stored in a vector in column-major order and the dim
attribute is a vector of integers specifying the respective extents of
the array. R ensures that the length of the vector is the product of
the lengths of the dimensions. The length of one or more dimensions may
be zero.
A vector is not the same as a one-dimensional array since the latter has
a dim attribute of length one, whereas the former has no
dim attribute.
Arrays may name each dimension separately using the dimnames
attribute which is a list of character vectors. The dimnames
list may itself have names which are then used for extent headings when
printing arrays.
R has an elaborate class system, controlled via the class
attribute. This attribute is a character vector containing the list of
classes that an object inherits from. This forms the basis of the
“generic methods” functionality in R.
This attribute can be accessed and manipulated virtually without
restriction by users. There is no checking that an object actually
contains the components that class methods expect. Thus, altering the
class attribute should be done with caution, and when they are
available specific creation and
coercion functions should be preferred.
The tsp attribute is used to hold parameters of time series,
start, end, and frequency. This construction is mainly used to handle
series with periodic substructure such as monthly or quarterly data.
Factors are used to describe items that can have a finite number of
values (gender, social class, etc.). A factor has a levels
attribute and class "factor". Optionally, it may also contain a
contrasts attribute which controls the parametrisation used when
the factor is used in a
modeling functions.
A factor may be purely nominal or may have ordered categories. In the
latter case, it should be defined as such and have a class vector
c("ordered"," factor").
Factors are currently implemented using an integer array to specify the actual levels and a second array of names that are mapped to the integers. Rather unfortunately users often make use of the implementation in order to make some calculations easier. This, however, is an implementation issue and is not guaranteed to hold in all implementations of R.
Data frames are the R structures which most closely mimic the SAS or SPSS data set, i.e. a “cases by variables” matrix of data.
A data frame is a list of vectors, factors, and/or matrices all having
the same length (number of rows in the case of matrices). In addition,
a data frame generally has a names attribute labeling the
variables and a row.names attribute for labeling the cases.
A data frame can contain a list that is the same length as the other components. The list can contain elements of differing lengths thereby providing a data structure for ragged arrays. However, as of this writing such arrays are not generally handled correctly.
When a user types a command at the prompt (or when an expression is read from a file) the first thing that happens to it is that the command is transformed by the parser into an internal representation. The evaluator executes parsed R expressions and returns the value of the expression. All expressions have a value. This is the core of the language.
This chapter describes the basic mechanisms of the evaluator, but avoids discussion of specific functions or groups of functions which are described in separate chapters later on or where the help pages should be sufficient documentation.
Users can construct expressions and invoke the evaluator on them.
Any number typed directly at the prompt is a constant and is evaluated.
> 1
[1] 1
Constants are fairly boring and to do more we need symbols.
When a new variable is created it must have a name so it can be referenced and it usually has a value. The name itself is a symbol. When a symbol is evaluated its value is returned. Later we shall explain in detail how to determine the value associated with a symbol.
In this small example y is a symbol and its value is 4. A symbol
is an R object too, but one rarely needs to deal with symbols
directly, except when doing “programming on the language”
(Computing on the language).
> y <- 4
> y
[1] 4
Most of the computations carried out in R involve the evaluation of functions. We will also refer to this as function invocation. Functions are invoked by name with a list of arguments separated by commas.
> mean(1:10)
[1] 5.5
In this example the function mean was called with one argument,
the vector of integers from 1 to 10.
R contains a huge number of functions with different purposes. Most are used for producing a result which is an R object, but others are used for their side effects, e.g., printing and plotting functions.
Function calls can have tagged (or named arguments), as in
plot(x, y, pch = 3) arguments without tags are known as
positional since the function must distinguish their meaning from
their sequential positions among the arguments of the call, e.g., that
x denotes the abscissa variable and y the ordinate. The
use of tags/names is an obvious convenience for functions with a large
number of optional arguments.
A special type of function calls can appear on the left hand side of the assignment operator as in
> class(x) <- "foo"
What this construction really does is to call the function
class<- with the original object and the right hand side. This
function performs the modification of the object and returns the result
which is then stored back into the original variable. (At least
conceptually, this is what happens. Some additional effort is made to
avoid unnecessary data duplication.)
R allows the use of arithmetic expressions using operators similar to those of the C programming language, for instance
> 1 + 2
[1] 3
Expressions can be grouped using parentheses, mixed with function calls, and assigned to variables in a straightforward manner
> y <- 2 * (a + log(x))
R contains a number of operators. They are listed in the table below.
-Minus, can be unary or binary +Plus, can be unary or binary !Unary not ~Tilde, used for model formulae, can be either unary or binary ?Help :Sequence, binary (in model formulae: interaction) *Multiplication, binary /Division, binary ^Exponentiation, binary %x%Special binary operators, x can be replaced by any valid name %%Modulus, binary %/%Integer divide, binary %*%Matrix product, binary %o%Outer product, binary %x%Kronecker product, binary %in%Matching operator, binary (in model formulae: nesting) <Less than, binary >Greater than, binary ==Equal to, binary >=Greater than or equal to, binary <=Less than or equal to, binary &And, binary, vectorized &&And, binary, not vectorized |Or, binary, vectorized ||Or, binary, not vectorized <-Left assignment, binary ->Right assignment, binary $List subset, binary
Except for the syntax, there is no difference between applying an
operator and calling a function. In fact, x + y can equivalently
be written "+"(x, y). Notice that since + is a
non-standard function name, it needs to be quoted.
R deals with entire vectors of data at a time, and most of the
elementary operators and basic mathematical functions like log
are vectorized (as indicated in the table above). This means that
e.g. adding two vectors of the same length will create a vector
containing the element-wise sums, implicitly looping over the vector
index. This applies also to other operators like -, *,
and / as well as to higher dimensional structures. Notice in
particular that multiplying two matrices does not produce the usual
matrix product (the %*% operator exists for that purpose). Some
finer points relating to vectorized operations will be discussed in
Elementary arithmetic operations.
To access individual elements of a vector, one generally uses the
x[i] construction.
> x <- rnorm(5)
> x
[1] -0.12526937 -0.27961154 -1.03718717 -0.08156527 1.37167090
> x[2]
[1] -0.2796115
List components are more commonly accessed using x$a or
x[[i]].
> x <- options()
> x$prompt
[1] "> "
Indexing constructs can also appear on the right hand side of an assignment.
Like the other operators, indexing is really done by functions, and one
could have used "["(x, 2) instead of x[2].
R's indexing operations contain many advanced features which are further described in Indexing.
Computation in R consists of sequentially evaluating
statements. Statements, such as x<-1:10 or
mean(y), can be separated by either a semi-colon or a new line.
Whenever the
evaluator is presented with a syntactically complete
statement that statement is evaluated and the value returned.
The result of evaluating a statement can be referred to as the value of
the statement1 The value can
always be assigned to a symbol.
Both semicolons and new lines can be used to separate statements. A semicolon always indicates the end of a statement while a new line may indicate the end of a statement. If the current statement is not syntactically complete new lines are simply ignored by the evaluator. If the session is interactive the prompt changes from > to +.
> x <- 0; x + 5
[1] 5
> y <- 1:10
> 1; 2
[1] 1
[1] 2
Statements can be grouped together using braces { and }. A group of statements is sometimes called a block. Single statements are evaluated when a new line is typed at the end of the syntactically complete statement. Blocks are not evaluated until a new line is entered after the closing brace. In the remainder of this section, statement refers to either a single statement or a block.
> { x <- 0
+ x + 5
+ }
[1] 5
The if/else statement conditionally evaluates two
statements. There is a condition which is evaluated and if the
value is TRUE then the first statement is evaluated;
otherwise the second statement will be evaluated. The
if/else statement returns, as its value, the value of the
statement that was selected. The formal syntax is
if ( statement1 )
statement2
else
statement3
First, statement1 is evaluated to yield value1. If
value1 is a logical vector with first element TRUE then
statement2 is evaluated. If the first element of value1 is
FALSE then statement3 is evaluated. If value1 is a
numeric vector then statement3 is evaluated when the first element
of value1 is zero and otherwise statement2 is evaluated.
Only the first element of value1 is used. All other elements are
ignored. If value1 has any type other than a logical or a numeric
vector an error is signalled.
If/else statements can be used to avoid numeric problems such as taking the logarithm of a negative number. Because if/else statements are the same as other statements you can assign the value of them. The two examples below are equivalent.
> if( any(x <= 0) ) y <- log(1+x) else y <- log(x)
> y <- if( any(x <= 0) ) log(1+x) else log(x)
The else clause is optional. The statement if(any(x <= 0))
x <- x[x <= 0] is valid. When the if statement is not in a
block the else, if present, must appear on the same line as
statement1. Otherwise the new line at the end of statement1
yields a syntactically complete statement that is evaluated.
If/else statements can be nested.
if ( statement1 )
statement2
else if ( statement3 )
statement4
else if ( statement5 )
statement6
else
statement8
One of the even numbered statements will be evaluated and the resulting
value returned. If the optional else clause is omitted and all
the odd numbered statement's evaluate to FALSE no statement
will be evaluated and NULL is returned.
The odd numbered statements are evaluated, in order, until one
evaluates to TRUE and then the associated even numbered
statement is evaluated. In this example, statement6 will
only be evaluated if statement1 is FALSE and
statement3 is FALSE and statement5 is TRUE.
There is no limit to the number of else if clauses that are
permitted.
R has three statements that provide explicit
looping.2 They are for, while and
repeat. The two built-in constructs, next and
break, provide additional control over the evaluation. Each of
the three statements returns the value of the last statement that was
evaluated. It is possible, although uncommon, to assign the result of
one of these statements to a symbol. R provides other functions for
implicit looping such as tapply, apply, and lapply.
In addition many operations, especially arithmetic ones, are vectorized
so you may not need to use a loop.
There are two statements that can be used to explicitly control looping.
They are break and next.
The break statement causes an exit from the innermost loop that
is currently being executed. The next statement immediately
causes control to return to the start of the loop. The next iteration
of the loop (if there is one) is then executed. No statement below
next in the current loop is evaluated.
The repeat statement causes repeated evaluation of the body until
a break is specifically requested. This means that you need to be
careful when using repeat because of the danger of an infinite
loop. The syntax of the repeat loop is
repeat statement
When using repeat, statement must be a block statement.
You need to both perform some computation and test whether or not to
break from the loop and usually this requires two statements.
The while statement is very similar to the repeat
statement. The syntax of the while loop is
while ( statement1 ) statement2
where statement1 is evaluated and if its value is TRUE then
statement2 is evaluated. This process continues until
statement1 evaluates to FALSE. If statement2 is
never evaluated then while returns NULL and otherwise it
returns the value of the last evaluation of statement2.
for ( name in vector )
statement1
where vector can be either a vector or a list. For each element in vector the variable name is set to the value of that element and statement1 is evaluated. A side effect is that the variable name still exists after the loop has concluded and it has the value of the last element of vector that the loop was evaluated for.
Technically speaking, switch is just another function, but its
semantics are close to those of control structures of other programming
languages.
The syntax is
switch (statement, list)
where the elements of list may be named. First, statement
is evaluated and the result, value, obtained. If value is a
number between 1 and the length of list then the corresponding
element list is evaluated and the result returned. If value
is too large or too small NULL is returned.
> x <- 3
> switch(x, 2+2, mean(1:10), rnorm(5))
[1] 2.2903605 2.3271663 -0.7060073 1.3622045 -0.2892720
> switch(2, 2+2, mean(1:10), rnorm(5))
[1] 5.5
> switch(6, 2+2, mean(1:10), rnorm(5))
NULL
If value is a character vector then the element of ... with
a name that exactly matches value is evaluated. If there is no
match NULL is returned.
> y <- "fruit"
> switch(y, fruit = "banana", vegetable = "broccoli", meat = "beef")
[1] "banana"
A common use of switch is to branch according to the character
value of one of the arguments to a function.
> centre <- function(x, type) {
+ switch(type,
+ mean = mean(x),
+ median = median(x),
+ trimmed = mean(x, trim = .1))
+ }
> x <- rcauchy(10)
> centre(x, "mean")
[1] 0.8760325
> centre(x, "median")
[1] 0.5360891
> centre(x, "trimmed")
[1] 0.6086504
switch returns either the value of the statement that was
evaluated or NULL if no statement was evaluated.
To choose from a list of alternatives that already exists switch
may not be the best way to select one for evaluation. It is often
better to use eval and the subset operator, [[, directly
via eval(x[[condition]]).
In this section, we discuss the finer points of the rules that apply to basic operation like addition or multiplication of two vectors or matrices.
If one tries to add two structures with a different number of elements,
then the shortest is recycled to length of longest. That is, if for
instance you add c(1, 2, 3) to a six-element vector then you will
really add c(1, 2, 3, 1, 2, 3). If the length of the longer
vector is not a multiple of the shorter one, a warning is given.
As from R 1.4.0, any arithmetic operation involving a zero-length vector has a zero-length result.
One exception is that when adding vectors to matrices, a warning is not given if the lengths are incompatible.
propagation of names (first one wins, I think - also if it has no names?? —– first one *with names* wins, recycling causes shortest to lose names)
(matrix+matrix, dimensions must match. vector+matrix: first recycle, then check if dims fit, error if not)
Missing values in the statistical sense, that is, variables whose value
is not known, have the value NA. This should not be confused with
the missing property for a function argument that has not been
supplied (see Arguments).
As the elements of an atomic vector must be of the same type there are
multiple types of NA values. There is one case where this is
particularly important to the user. The default type of NA is
logical, unless coerced to some other type, so the appearance of
a missing value may trigger logical rather than numeric indexing (see
Indexing for details).
Numeric and logical calculations with NA generally return
NA. In cases where the result of the operation would be the same
for all possible values the NA could take, the operation may
return this value. In particular, FALSE & NA is FALSE,
TRUE | NA is TRUE. NA is not equal to any other
value or to itself; testing for NA is done using is.na.
However, an NA value will match another NA value in
match.
Numeric calculations whose result is undefined, such as 0/0,
produce the value NaN. This exists only in the double
type and for real or imaginary components of the complex type. The
function is.nan is provided to check specifically for
NaN, is.na also returns TRUE for NaN.
Coercing NaN to logical or integer type gives an NA of the
appropriate type, but coercion to character gives the string
"NaN". NaN values are incomparable so tests of equality
or collation involving NaN will result in NA. They are
regarded as matching any NaN value (and no other value, not even
NA) by match.
The NA of character type is as from R 1.5.0 distinct from the
string "NA". Programmers who need to specify an explicit string
NA should use as.character(NA) rather than "NA", or
set elements to NA using is.na<-.
There is no NA value for raw vectors.
R contains several constructs which allow access to individual
elements or subsets through indexing operations. In the case of the
basic vector types one can access the i-th element using x[i],
but there is also indexing of lists, matrices, and multi-dimensional
arrays. There are several forms of indexing in addition to indexing
with a single integer. Indexing can be used both to extract part of an
object and to replace parts of an object (or to add parts).
R has three basic indexing operators, with syntax displayed by the following examples
x[i]
x[i, j]
x[[i]]
x[[i, j]]
x$a
x$"a"
For vectors and matrices the [[ forms are rarely used, although
they have some slight semantic differences from the [ form (e.g.
it drops any names or dimnames attribute, and that partial
matching is used for character indices). When indexing
multi-dimensional structures with a single index, x[[i]] or
x[i] will return the ith sequential element of x.
For lists, one generally uses [[ to select any single element,
whereas [ returns a list of the selected elements.
The [[ form allows only a single element to be selected using
integer or character indices, whereas [ allows indexing by
vectors. Note though that for a list, the index can be a vector and
each element of the vector is applied in turn to the list, the
selected component, the selected component of that component, and so on.
The result is still a single element.
The form using $ applies to recursive objects such as lists and
pairlists. It allows only a literal character string or a symbol as the
index. That is, the index is not computable: for cases where you need
to evaluate an expression to find the index, use x[[expr]]. When
$ is applied to a non-recursive object the result is always
NULL.
R allows some powerful constructions using vectors as indices. We
shall discuss indexing of simple vectors first. For simplicity, assume
that the expression is x[i]. Then the following possibilities
exist according to the type of i.
i must have the same sign. If
they are positive, the elements of x with those index numbers are
selected. If i contains negative elements, all elements except
those indicated are selected.
If i is positive and exceeds length(x) then the
corresponding selection is NA. A negative out of bounds value
for i causes an error.
A special case is the zero index, which has null effects: x[0] is
an empty vector and otherwise including zeros among positive or negative
indices has the same effect as if they were omitted.
i should generally have the same
length as x. If it is shorter, then its elements will be
recycled as discussed in Elementary arithmetic operations. If it
is longer, then x is conceptually extended with NAs. The
selected values of x are those for which i is TRUE.
i are matched against the
names attribute of x and the resulting integers are used. For
[[ and $ partial matching is used if exact matching fails,
so x$aa will match x$aabb if x does not a component
named "aa" and "aabb" is the first name which has prefix
"aa". However, [ requires an exact match. The string
"" is treated specially: it indicates `no name' and matches no
element (not even those without a name). Note that partial matching is
only used when extracting and not when replacing.
x[as.integer(i)].
The factor levels are never used. If so desired, use
x[as.character(i)] or a similar construction.
x[] returns x, but drops
“irrelevant” attributes from the result. Only names and in
multi-dimensional arrays dim and dimnames attributes are
retained.
integer(0).
Indexing with a missing (i.e. NA) value give an NA
result. This rule applies also to the case of logical indexing,
i.e. the elements of x that have an NA selector in
i get included in the result, but their value will be NA.
Notice however, that there are different modes of NA—the
literal constant is of mode "logical", but it is frequently
automatically coerced to other types. One effect of this is that
x[NA] has the length of x, but x[c(1, NA)] has
length 2. That is because the rules for logical indices apply in the
former case, but those for integer indices in the latter.
Indexing with [ will also carry out the relevant subsetting of
any names attributes.
Subsetting multi-dimensional structures generally follows the same rules
as single-dimensional indexing for each index variable, with the
relevant component of dimnames taking the place of names.
A couple of special rules apply, though:
Normally, a structure is accessed using the number of indices
corresponding to its dimension. It is however also possible to use a
single index in which case the dim and dimnames attributes
are disregarded and the result is effectively that of c(m)[i].
Notice that m[1] is usually very different from m[1, ] or
m[, 1].
It is possible to use a matrix of integers as an index. In this case,
the number of columns of the matrix should match the number of
dimensions of the structure, and the result will be a vector with length
as the number of rows of the matrix. The following example shows how
to extract the elements m[1, 1] and m[2, 2] in one
operation.
> m <- matrix(1:4, 2)
> m
[,1] [,2]
[1,] 1 3
[2,] 2 4
> i <- matrix(c(1, 1, 2, 2), 2, byrow = TRUE)
> i
[,1] [,2]
[1,] 1 1
[2,] 2 2
> m[i]
[1] 1 4
Negative indices are not allowed in indexing matrices. NA and
zero values are allowed: rows in an index matrix containing a zero are
ignored, whereas rows containing an NA produce an NA in
the result.
Both in the case of using a single
index and in matrix indexing, a names attribute is used if
present, as had the structure been one-dimensional.
If an indexing operation causes the result to have one of its extents of
length one, as in selecting a single slice of a three-dimensional matrix
with (say) m[2, , ], the corresponding dimension is generally
dropped from the result. If a single-dimensional structure results, a
vector is obtained. This is occasionally undesirable and can be turned
off by adding the drop = FALSE to the indexing operation. Notice
that this is an additional argument to the [ function and doesn't
add to the index count. Hence the correct way of selecting the first
row of a matrix as a 1 by n matrix is m[1, , drop =
FALSE]. Forgetting to disable the dropping feature is a common cause
of failure in general subroutines where an index occasionally, but not
usually has length one. This rule still applies to a one-dimensional
array, where any subsetting will give a vector result unless drop
= FALSE is used.
Notice that vectors are distinct from one-dimensional arrays in that the
latter have dim and dimnames attributes (both of length
one). One-dimensional arrays are not easily obtained from subsetting
operations but they can be constructed explicitly and are returned by
table. This is sometimes useful because the elements of the
dimnames list may themselves be named, which is not the case for
the names attribute.
Some operations such as m[FALSE, ] result in structures in which
a dimension has zero extent. R generally tries to handle these
structures sensibly.
The operator [ is a generic function which allows class methods
to be added, and the $ and [[ operators likewise. Thus,
it is possible to have user-defined indexing operations for any
structure. Such a function, say [.foo is called with a set of
arguments of which the first is the structure being indexed and the rest
are the indices. In the case of $, the index argument is of mode
"symbol" even when using the x$"abc" form. It is
important to be aware that class methods do not necessarily behave in
the same way as the basic methods, for example with respect to partial
matching.
The most important example of a class method for [ is that used
for data frames. It is not be described in detail here (see the help
page for [.data.frame, but in broad terms, if two indices are
supplied (even if one is empty) it creates matrix-like indexing for a
structure that is basically a list of vectors of the same length. If a
single index is supplied, it is interpreted as indexing the list of
columns—in that case the drop argument is ignored, with a
warning.
The basic operators $ and [[ can be applied to
environments. Only character indices are allowed and no partial
matching is done.
Assignment to subsets of a structure is a special case of a general mechanism for complex assignment:
x[3:5] <- 13:15
The result of this commands is as if the following had been executed
`*tmp*` <- x
x <- "[<-"(`*tmp*`,3:5, value=13:15)
The same mechanism can be applied to other functions than [. The
assignment function has the same name with <- pasted on. Its last
argument, which must be called value, is the new value to be
assigned.
names(x) <- c("a","b")
is equivalent to
`*tmp*` <- x
x <- "names<-"(`*tmp*`, value=c("a","b"))
Nesting of complex assignments is evaluated recursively
names(x)[3] <- "Three"
is equivalent to
`*tmp*` <- x
x <- "names<-"(`*tmp*`, value="[<-"(names(`*tmp*`), 3, value="Three"))
Complex assignments in the enclosing environment (using <<-) are
also permitted:
names(x)[3] <<- "Three"
is equivalent to
`*tmp*` <<- get(x, envir=parent.env(), inherits=TRUE)
names(`*tmp*`)[3] <- "Three"
x <<- `*tmp*`
and also to
`*tmp*` <- get(x,envir=parent.env(), inherits=TRUE)
x <<- "names<-"(`*tmp*`, value="[<-"(names(`*tmp*`), 3, value="Three"))
Only the target variable is evaluated in the enclosing environment, so
e<-c(a=1,b=2)
i<-1
local({
e <- c(A=10,B=11)
i <-2
e[i] <<- e[i]+1
})
uses the local value of i on both the LHS and RHS, and the local
value of e on the RHS of the superassignment statement. It sets
e in the outer environment to
a b
1 12
That is, the superassignment is equivalent to the three lines
`*tmp*` <- get(x,envir=parent.env(), inherits=TRUE)
`*tmp*`[i] <- e[i]+1
x <<- `*tmp*`
Similarly
x[is.na(x)] <<- 0
is equivalent to
`*tmp*` <- get(x,envir=parent.env(), inherits=TRUE)
`*tmp*`[is.na(x)] <- 0
x <<- `*tmp*`
and not to
`*tmp*` <- get(x,envir=parent.env(), inherits=TRUE)
`*tmp*`[is.na(`*tmp*`)] <- 0
x <<- `*tmp*`
These two candidate interpretations differ only if there is also a
local variable x. It is a good idea to avoid having a local
variable with the same name as the target variable of a
superassignment. As this case was handled incorrectly in versions
1.9.1 and earlier there must not be a serious need for such code.
Almost every programming language has a set of scoping rules, allowing the same name to be used for different objects. This allows, e.g., a local variable in a function to have the same name a global object.
R uses a lexical scoping model, similar to languages like Pascal. However, R is a functional programming language and allows dynamic creation and manipulation of functions and language objects, and has additional features reflecting this fact.
The global environment is the root of the user workspace. An assignment operation from the command line will cause the relevant object to belong to the global environment. Its enclosing environment is the next environment on the search path, and so on back to the empty environment that is the enclosure of the base environment.
Every call to a function creates a frame which contains the local variables created in the function, and is evaluated in an environment, which in combination creates a new environment.
Notice the terminology: A frame is a set of variables, an environment is a nesting of frames (or equivalently: the innermost frame plus the enclosing environment).
Environments may be assigned to variables or be contained in other objects. However, notice that they are not standard objects—in particular, they are not copied on assignment.
A closure (mode "function") object will contain the environment
in which it is created as part of its definition (By default. The
environment can be manipulated using environment<-). When the
function is subsequently called, its
evaluation environment is created with the closure's environment as
enclosure. Notice that this is not
necessarily the environment of the caller!
Thus, when a variable is requested inside a function, it is first sought in the evaluation environment, then in the enclosure, the enclosure of the enclosure, etc.; once the global environment or the environment of a package is reached, the search continues up the search path to the environment of the base package. If the variable is not found there, the search will proceed next to the empty environment, and will fail.
Every time a function is invoked a new evaluation frame is created. At any point in time during the computation the currently active environments are accessible through the call stack. Each time a function is invoked a special construct called a context is created internally and is placed on a list of contexts. When a function has finished evaluating its context is removed from the call stack.
Making variables defined higher up the call stack available is called dynamic scope. The binding for a variable is then determined by the most recent (in time) definition of the variable. This contradicts the default scoping rules in R, which use the bindings in the environment in which the function was defined (lexical scope). Some functions, particularly those that use and manipulate model formulas, need to simulate dynamic scope by directly accessing the call stack.
Access to the call stack is provided through a family of functions which have names that start with sys.. They are listed briefly below.
sys.callsys.framesys.nframesys.functionsys.parentsys.callssys.framessys.parentssys.on.exitsys.statussys.frames, sys.parents and sys.calls.
parent.frameIn addition to the evaluation environment structure, R has a search path of environments which are searched for variables not found elsewhere. This is used for two things: packages of functions and attached user data.
The first element of the search path is the global environment and the
last is the base package. An Autoloads environment is used for
holding proxy objects that may be loaded on demand. Other environments
are inserted in the path using attach or library.
Packages which have a namespace have a different search path. When a search for an R object is started from an object in such a package, the package itself is searched first, then its imports, then the base namespace and finally the global environment and the rest of the regular search path. The effect is that references to other objects in the same package will be resolved to the package, and objects cannot be masked by objects of the same name in the global environment or in other packages.
While R can be very useful as a data analysis tool most users very quickly find themselves wanting to write their own functions. This is one of the real advantages of R. Users can program it and they can, if they want to, change the system level functions to functions that they find more appropriate.
R also provides facilities that make it easy to document any functions that you have created. See Writing R documentation (Writing R Extensions).
The syntax for writing a function is
function ( arglist ) body
The first component of the function declaration is the keyword
function which indicates to R that you want to create a
function.
An argument list is a comma separated list of formal arguments. A formal argument can be a symbol, a statement of the form symbol = expression, or the special formal argument ....
The body can be any valid R expression. Generally, the body is a group of expressions contained in curly braces ({ and }).
Generally
functions are assigned to symbols but they don't need to be.
The value returned by the call to function is a function. If
this is not given a name it is referred to as an
anonymous
function. Anonymous functions are most frequently used as arguments
other functions such as the apply family or outer.
Here is a simple function: echo <- function(x) print(x). So
echo is a function that takes a single argument and when
echo is invoked it prints its argument.
The formal arguments to the function define the variables whose values will be supplied at the time the function is invoked. The names of these arguments can be used within the function body where they obtain the value supplied at the time of function invocation.
Default values for arguments can be specified using the special form name = expression. In this case, if the user does not specify a value for the argument when the function is invoked the expression will be associated with the corresponding symbol. When a value is needed the expression is evaluated in the evaluation frame of the function.
Default behaviours can also be specified by using the function
missing. When missing is called with the
name of a formal
argument it returns TRUE if the formal argument was not matched
with any actual argument and has not been subsequently modified in the
body of the function. An argument that is missing will thus
have its default value, if any. The missing function does not
force evaluation of the argument.
The special type of argument ... can contain any number of supplied arguments. It is used for a variety of purposes. It allows you to write a function that takes an arbitrary number of arguments. It can be used to absorb some arguments into an intermediate function which can then be extracted by functions called subsequently.
Functions are first class objects in R. They can be used anywhere that an R object is required. In particular they can be passed as arguments to functions and returned as values from functions. See Function objects for the details.
When a function is called or invoked a new evaluation frame is created. In this frame the formal arguments are matched with the supplied arguments according to the rules given in Argument matching. The statements in the body of the function are evaluated sequentially in this environment frame.
The enclosing frame of the evaluation frame is the environment frame
associated with the function being invoked. This may be different from
S. While many functions have .GlobalEnv as their environment
this does not have to be true and functions defined in packages with
namespaces (normally) have the package namespace as their environment.
The first thing that occurs in a function evaluation is the matching of formal to the actual or supplied arguments. This is done by a three-pass process:
f <- function(fumble,
fooey) fbody, then f(f = 1, fo = 2) is illegal, even though the
2nd actual argument only matches fooey. f(f = 1, fooey =
2) is legal though since the second argument matches exactly and
is removed from consideration for partial matching. If the formal
arguments contain ... then partial matching is only applied to
arguments that precede it.
If any arguments remain unmatched an error is declared.
Argument matching is augmented by the functions match.arg,
match.call and match.fun.
Access to the partial matching algorithm used by R is via
pmatch.
One of the most important things to know about the evaluation of arguments to a function is that supplied arguments and default arguments are treated differently. The supplied arguments to a function are evaluated in the evaluation frame of the calling function. The default arguments to a function are evaluated in the evaluation frame of the function.
The semantics of invoking a function in R argument are call-by-value. In general, supplied arguments behave as if they are local variables initialized with the value supplied and the name of the corresponding formal argument. Changing the value of a supplied argument within a function will not affect the value of the variable in the calling frame.
R has a form of lazy evaluation of function arguments. Arguments are
not evaluated until needed. It is important to realize that in some
cases the argument will never be evaluated. Thus, it is bad style to
use arguments to functions to cause side-effects. While in C it is
common to use the form, foo(x = y) to invoke foo with the
value of y and simultaneously to assign the value of y to
x this same style should not be used in R. There is no
guarantee that the argument will ever be evaluated and hence the
assignment may not take place.
It is also worth noting that the effect of foo(x <- y) if the
argument is evaluated is to change the value of x in the calling
environment and not in the
evaluation environment of foo.
It is possible to access the actual (not default) expressions used as
arguments inside the function. The mechanism is implemented via
promises. When a
function is being evaluated the actual expression used as an argument is
stored in the promise together with a pointer to the environment the
function was called from. When (if) the argument is evaluated the
stored expression is evaluated in the environment that the function was
called from. Since only a pointer to the environment is used any
changes made to that environment will be in effect during this
evaluation. The resulting value is then also stored in a separate spot
in the promise. Subsequent evaluations retrieve this stored value (a
second evaluation is not carried out). Access to the unevaluated
expression is also available using substitute.
When a function is called, each formal argument is assigned a promise in the local environment of the call with the expression slot containing the actual argument (if it exists) and the environment slot containing the environment of the caller. If no actual argument for a formal argument is given in the call and there is a default expression, it is similarly assigned to the expression slot of the formal argument, but with the environment set to the local environment.
The process of filling the value slot of a promise by evaluating the contents of the expression slot in the promises environment is called forcing the promise. A promise will only be forced once, the value slot content being used directly later on.
A promise is forced when its value is needed. This usually happens
inside internal
functions, but a promise can also be forced by direct evaluation of the
promise itself. This is occasionally useful when a default expression
depends on the value of another formal argument or other variable in the
local environment. This is seen in the following example where the lone
label ensures that the label is based on the value of x
before it is changed in the next line.
function(x, label = deparse(x)) {
label
x <- x + 1
print(label)
}
The expression slot of a promise can itself involve other promises. This happens whenever an unevaluated argument is passed as an argument to another function. When forcing a promise, other promises in its expression will also be forced recursively as they are evaluated.
Scope or the scoping rules are simply the set of rules used by the evaluator to find a value for a symbol. Every computer language has a set of such rules. In R the rules are fairly simple but there do exist mechanisms for subverting the usual, or default rules.
R adheres to a set of rules that are called lexical scope. This means the variable bindings in effect at the time the expression was created are used to provide values for any unbound symbols in the expression.
Most of the interesting properties of scope are involved with evaluating functions and we concentrate on this issue. A symbol can be either bound or unbound. All of the formal arguments to a function provide bound symbols in the body of the function. Any other symbols in the body of the function are either local variables or unbound variables. A local variable is one that is defined within the function. Because R has no formal definition of variables, they are simply used as needed, it can be difficult to determine whether a variable is local or not. Local variables must first be defined, this is typically done by having them on the left-hand side of an assignment.
During the evaluation process if an unbound symbol is detected then R attempts to find a value for it. The scoping rules determine how this process proceeds. In R the environment of the function is searched first, then its enclosure and so on until the global environment is reached.
The global environment heads a search list of environments that are searched sequentially for a matching symbol. The value of the first match is then used.
When this set of rules is combined with the fact that functions can be returned as values from other functions then some rather nice, but at first glance peculiar, properties obtain.
A simple example,
f <- function(x) {
y <- 10
g <- function(x) x + y
return(g)
}
h <- f()
h(3)
A rather interesting question is what happens when h is
evaluated. To describe this we need a bit more notation. Within a
function body variables can be bound, local or unbound. The bound
variables are those that match the formal arguments to the function.
The local variables are those that were created or defined within the
function body. The unbound variables are those that are neither local
or bound. When a function body is evaluated there is no problem
determining values for local variables or for bound variables. Scoping
rules determine how the language will find values for the unbound
variables.
When h(3) is evaluated we see that its body is that of g.
Within that body x and y are unbound. In a language with
lexical scope x will be associated with the value 3 and y
with the value 10 so h() should return the value 13. In R
this is indeed what happens.
In S, because of the different scoping rules one will get an error
indicating that y is not found, unless there is a variable
y in your workspace in which case its value will be used.
Object-oriented programming is a style of programming that has become popular in recent years. Much of the popularity comes from the fact that it makes it easier to write and maintain complicated systems. It does this through several different mechanisms.
Central to any object-oriented language are the concepts of class and of methods. A class is a definition of an object. Typically a class contains several slots that are used to hold class-specific information. An object in the language must be an instance of some class. Programming is based on objects or instances of classes.
Computations are carried out via methods. Methods are basically functions that are specialized to carry out specific calculations on objects, usually of a specific class. This is what makes the language object oriented. In R, generic functions are used to determine the appropriate method. The generic function is responsible for determining the class of its argument(s) and uses that information to select the appropriate method.
Another feature of most object-oriented languages is the concept of inheritance. In most programming problems there are usually many objects that are related to one another. The programming is considerably simplified if some components can be reused.
If a class inherits from another class then generally it gets all the slots in the parent class and can extend it by adding new slots. On method dispatching (via the generic functions) if a method for the class does not exist then a method for the parent is sought.
In this chapter we discuss how this general strategy has been implemented in R and discuss some of the limitations within the current design. One of the advantages that most object systems impart is greater consistency. This is achieved via the rules that are checked by the compiler or interpretor. Unfortunately because of the way that the object system is incorporated into R this advantage does not obtain. Users are cautioned to use the object system in a straightforward manner. While it is possible to perform some rather interesting feats these tend to lead to obfuscated code and may depend on implementation details that will not be carried forward.
The greatest use of object oriented programming in R is through
print methods, summary methods and plot methods.
These methods allow us to have one generic
function call, plot
say, that dispatches on the type of its argument and calls a plotting
function that is specific to the data supplied.
In order to make the concepts clear we will consider the implementation of a small system designed to teach students about probabi