26 Deprecated

26.1 Conditions

  1. Q: What does options(error = recover) do? Why might you use it?

    A: In case of options(error = recover) utils::recover() will be called (without arguments) in case of an error. This will print out a list of calls which precede the error and lets the user choose to incorporate browser() directly in any of the regarding environments allowing a practical mode for debugging.

  2. Q: What does options(error = quote(dump.frames(to.file = TRUE))) do? Why might you use it?

    A: This option writes a dump of the evaluation environment where an error occurs into a file ending on .rda. When this option is set, R will continue to run after the first error. To stop R at the first error use quote({dump.frames(to.file=TRUE); q()}). These options are especially useful for debugging non-interactive R scripts afterwards (“post mortem debugging”).

26.2 Expressions (new)

  1. Q: base::alist() is useful for creating pairlists to be used for function arguments:

    What makes alist() special compared to list()?

    A: From ?alist:

    alist handles its arguments as if they described function arguments. So the values are not evaluated, and tagged arguments with no value are allowed whereas list simply ignores them. alist is most often used in conjunction with formals.

26.3 Functionals

26.3.1 My first functional: lapply()

  1. Q: Why are the following two invocations of lapply() equivalent?

    A: In the first statement each element of trims is explicitly supplied to mean()’s second argument. In the latter statement this happens via positional matching, since mean()’s first argument is supplied via name in lapply()’s third argument (...).

  2. Q: The function below scales a vector so it falls in the range [0, 1]. How would you apply it to every column of a data frame? How would you apply it to every numeric column in a data frame?

    A: Since this function needs numeric input, one can check this via an if clause. If one also wants to return non-numeric input columns, these can be supplied to the else argument of the if() “function”:

  3. Q: Use both for loops and lapply() to fit linear models to the mtcars using the formulas stored in this list:

    A: Like in the first exercise, we can create two lapply() versions:

    Note that all versions return the same content, but they won’t be identical, since the values of the “call” element will differ between each version.

  4. Q: Fit the model mpg ~ disp to each of the bootstrap replicates of mtcars in the list below by using a for loop and lapply(). Can you do it without an anonymous function?


  5. Q: For each model in the previous two exercises, extract \(R^2\) using the function below.

    A: For the models in exercise 3:

    And the models in exercise 4:

26.3.2 For loops functionals: friends of lapply():

  1. Q: Use vapply() to:

    1. Compute the standard deviation of every column in a numeric data frame.

    2. Compute the standard deviation of every numeric column in a mixed data frame. (Hint: you’ll need to use vapply() twice.)

    A: As a numeric data.frame we choose cars:

    And as a mixed data.frame we choose iris:

  2. Q: Why is using sapply() to get the class() of each element in a data frame dangerous?

    A: Columns of data.frames might have more than one class, so the class of sapply()’s output may differ from time to time (silently). If …

    • all columns have one class: sapply() returns a character vector
    • one column has more classes than the others: sapply() returns a list
    • all columns have the same number of classes, which is more than one: sapply() returns a matrix

    For example:

    Note that this case often appears, wile working with the POSIXt types, POSIXct and POSIXlt.

  3. Q: The following code simulates the performance of a t-test for non-normal data. Use sapply() and an anonymous function to extract the p-value from every trial.

    Extra challenge: get rid of the anonymous function by using [[ directly.


  4. Q: What does replicate() do? What sort of for loop does it eliminate? Why do its arguments differ from lapply() and friends?

    A: As stated in ?replicate:

    replicate is a wrapper for the common use of sapply for repeated evaluation of an expression (which will usually involve random number generation).

    We can see this clearly in the source code:

    #> function (n, expr, simplify = "array") 
    #> sapply(integer(n), eval.parent(substitute(function(...) expr)), 
    #>     simplify = simplify)
    #> <bytecode: 0x3f03270>
    #> <environment: namespace:base>

    Like sapply() replicate() eliminates a for loop. As explained for Map() in the textbook, also every replicate() could have been written via lapply(). But using replicate() is more concise, and more clearly indicates what you’re trying to do.

  5. Q: Implement a version of lapply() that supplies FUN with both the name and the value of each component.


  6. Q: Implement a combination of Map() and vapply() to create an lapply() variant that iterates in parallel over all of its inputs and stores its outputs in a vector (or a matrix). What arguments should the function take?

    A As we understand this exercise, it is about working with a list of lists, like in the following example:

    So we can get the same result with a more specialized function:

  7. Q: Implement mcsapply(), a multi-core version of sapply(). Can you implement mcvapply(), a parallel version of vapply()? Why or why not?

26.3.3 Manipulating matrices and data frames

  1. Q: How does apply() arrange the output? Read the documentation and perform some experiments.


    apply() arranges its output columns (or list elements) according to the order of the margin. The rows are ordered by the other dimensions, starting with the “last” dimension of the input object. What this means should become clear by looking at the three and four dimensional cases of the following example:

  2. Q: There’s no equivalent to split() + vapply(). Should there be? When would it be useful? Implement one yourself.

    A: We can modify the tapply2() approach from the book, where split() and sapply() were combined:

    tapply() has a SIMPLIFY argument. When you set it to FALSE, tapply() will always return a list. It is easy to create cases where the length and the types/classes of the list elements vary depending on the input. The vapply() version could be useful, if you want to control the structure of the output to get an error according to some logic of a specific usecase or you want typestable output to build up other functions on top of it.

  3. Q: Implement a pure R version of split(). (Hint: use unique() and subsetting.) Can you do it without a for loop?


  4. Q: What other types of input and output are missing? Brainstorm before you look up some answers in the plyr paper.

    A: From the suggested plyr paper, we can extract a lot of possible combinations and list them up on a table. Sean C. Anderson already has done this based on a presentation from Hadley Wickham and provided the following result here.

    object type array data frame list nothing
    array apply . . .
    data frame . aggregate by .
    list sapply . lapply .
    n replicates replicate . replicate .
    function arguments mapply . mapply .

    Note the column nothing, which is specifically for usecases, where sideeffects like plotting or writing data are intended.

26.3.4 Manipulating lists

  1. Q: Why isn’t is.na() a predicate function? What base R function is closest to being a predicate version of is.na()?

    A: Because a predicate function always returns TRUE or FALSE. is.na(NULL) returns logical(0), which excludes it from being a predicate function. The closest in base that we are aware of is anyNA(), if one applies it elementwise.

  2. Q: Use Filter() and vapply() to create a function that applies a summary statistic to every numeric column in a data frame.


  3. Q: What’s the relationship between which() and Position()? What’s the relationship between where() and Filter()?

    A: which() returns all indices of true entries from a logical vector. Position() returns just the first (default) or the last integer index of all true entries that occur by applying a predicate function on a vector. So the default relation is Position(f, x) <=> min(which(f(x))).

    where(), defined in the book as:

    is useful to return a logical vector from a condition asked on elements of a list or a data frame. Filter(f, x) returns all elements of a list or a data frame, where the supplied predicate function returns TRUE. So the relation is Filter(f, x) <=> x[where(f, x)].

  4. Q: Implement Any(), a function that takes a list and a predicate function, and returns TRUE if the predicate function returns TRUE for any of the inputs. Implement All() similarly.

    A: Any():


  5. Q: Implement the span() function from Haskell: given a list x and a predicate function f, span returns the location of the longest sequential run of elements where the predicate is true. (Hint: you might find rle() helpful.)

    A: Our span_r() function returns the first index of the longest sequential run of elements where the predicate is true. In case of more than one longest sequenital, more than one first_index is returned.

26.3.5 List of functions

  1. Q: Implement a summary function that works like base::summary(), but uses a list of functions. Modify the function so it returns a closure, making it possible to use it as a function factory.

  2. Q: Which of the following commands is equivalent to with(x, f(z))?

    1. x$f(x$z).
    2. f(x$z).
    3. x$f(z).
    4. f(z).
    5. It depends.

26.3.6 Mathematical functionals

  1. Q: Implement arg_max(). It should take a function and a vector of inputs, and return the elements of the input where the function returns the highest value. For example, arg_max(-10:5, function(x) x ^ 2) should return -10. arg_max(-5:5, function(x) x ^ 2) should return c(-5, 5). Also implement the matching arg_min() function.

    A: arg_max():


  2. Q: Challenge: read about the fixed point algorithm. Complete the exercises using R.

26.3.7 A family of functions

  1. Q: Implement smaller and larger functions that, given two inputs, return either the smaller or the larger value. Implement na.rm = TRUE: what should the identity be? (Hint: smaller(x, smaller(NA, NA, na.rm = TRUE), na.rm = TRUE) must be x, so smaller(NA, NA, na.rm = TRUE) must be bigger than any other value of x.) Use smaller and larger to implement equivalents of min(), max(), pmin(), pmax(), and new functions row_min() and row_max().

    A: We can do almost everything as shown in the case study in the textbook. First we define the functions smaller_() and larger_(). We use the underscore suffix, to built up non suffixed versions on top, which will include the na.rm parameter. In contrast to the add() example from the book, we change two things at this step. We won’t include errorchecking, since this is done later at the top level and we return NA_integer_ if any of the arguments is NA (this is important, if na.rm is set to FALSE and wasn’t needed by the add() example, since + already returns NA in this case.)

    We can take na.rm() from the book:

    To find the identity value, we can apply the same argument as in the textbook, hence our functions are also associative and the following equation should hold:

    3 = smaller(smaller(3, NA), NA) = smaller(3, smaller(NA, NA)) = 3

    So the identidy has to be greater than 3. When we generalize from 3 to any real number this means that the identity has to be greater than any number, which leads us to infinity. Hence identity has to be Inf for smaller() (and -Inf for larger()), which we implement next:

    Like min() and max() can act on vectors, we can implement this easyly for our new functions. As shown in the book, we also have to set the init parameter to the identity value.

    We can also create vectorised versions as shown in the book. We will just show the smaller() case to become not too verbose.

    Of course, we are also able to copy paste the rest from the textbook, to solve the last part of the exercise:

  2. Q: Create a table that has and, or, add, multiply, smaller, and larger in the columns and binary operator, reducing variant, vectorised variant, and array variants in the rows.

    1. Fill in the cells with the names of base R functions that perform each of the roles.

    2. Compare the names and arguments of the existing R functions. How consistent are they? How could you improve them?

    3. Complete the matrix by implementing any missing functions.

    A In the following table we can see the requested base R functions, that we are aware of:

    and or add multiply smaller larger
    binary && ||
    reducing all any sum prod min max
    vectorised & | + * pmin pmax

    Notice that we were relatively strict about the binary row. Since the vectorised and reducing versions are more general, then the binary versions, we could have used them twice. However, this doesn’t seem to be the intention of this exercise.

    The last part of this exercise can be solved via copy pasting from the book and the last exercise for the binary row and creating combinations of apply() and the reducing versions for the array row. We think the array functions just need a dimension and an rm.na argument. We don’t know how we would name them, but sth. like sum_array(1, na.rm = TRUE) could be ok.

    The second part of the exercise is hard to solve complete. But in our opinion, there are two important parts. The behaviour for special inputs like NA, NaN, NULL and zero length atomics should be consistent and all versions should have a rm.na argument, for which the functions also behave consistent. In the follwing table, we return the output of `f`(x, 1), where f is the function in the first column and x is the special input in the header (the named functions also have an rm.na argument, which is FALSE by default). The order of the arguments is important, because of lazy evaluation.

    NA NaN NULL logical(0) integer(0)
    && NA NA error NA NA
    & NA NA error logical(0) logical(0)
    || TRUE TRUE error TRUE TRUE
    | TRUE TRUE error logical(0) logical(0)
    sum NA NaN 1 1 1
    + NA NaN numeric(0) numeric(0) numeric(0)
    prod NA NaN 1 1 1
    * NA NaN numeric(0) numeric(0) numeric(0)
    min NA NaN 1 1 1
    pmin NA NaN numeric(0) numeric(0) numeric(0)
    max NA NaN 1 1 1
    pmax NA NaN numeric(0) numeric(0) numeric(0)

    We can see, that the vectorised and reduced numerical functions are all consistent. However it is not, that the first three logical functions return NA for NA and NaN, while the 4th till 6th function all return TRUE. Then FALSE would be more consistent for the first three or the return of NA for all and an extra na.rm argument. In seems relatively hard to find an easy rule for all cases and especially the different behaviour for NULL is relatively confusing. Another good opportunity for sorting the functions would be to differentiate between “numerical” and “logical” operators first and then between binary, reduced and vectorised, like below (we left the last colum, which is redundant, because of coercion, as intended):

    `f(x,1)` NA NaN NULL logical(0)
    && NA NA error NA
    || TRUE TRUE error TRUE
    & NA NA error logical(0)
    | TRUE TRUE error logical(0)
    sum NA NaN 1 1
    prod NA NaN 1 1
    min NA NaN 1 1
    max NA NaN 1 1
    + NA NaN numeric(0) numeric(0)
    * NA NaN numeric(0) numeric(0)
    pmin NA NaN numeric(0) numeric(0)
    pmax NA NaN numeric(0) numeric(0)

    The other point are the naming conventions. We think they are clear, but it could be useful to provide the missing binary operators and name them for example ++, **, <>, >< to be consistent.

  3. Q: How does paste() fit into this structure? What is the scalar binary function that underlies paste()? What are the sep and collapse arguments to paste() equivalent to? Are there any paste variants that don’t have existing R implementations?

    A paste() behaves like a mix. If you supply only length one arguments, it will behave like a reducing function, i.e. :

    If you supply at least one element with length greater then one, it behaves like a vectorised function, i.e. :

    We think it should be possible to implement a new paste() starting from

    The sep argument is equivalent to bind sep on every ... input supplied to paste(), but the last and then bind these results together. In relations:

    paste(n1, n2, ...,nm , sep = sep) <=>
    paste0(paste0(n1, sep), paste(n2, n3, ..., nm, sep = sep)) <=>
    paste0(paste0(n1, sep), paste0(n2, sep), ..., paste0(nn, sep), paste0(nm))

    We can check this for scalar and non scalar input

    collapse just binds the outputs for non scalar input together with the collapse input. In relations:

    for input A1, ..., An, where Ai = a1i:ami,
    paste(A1 , A2 , ...,  An, collapse = collapse) 
          paste0(paste(  a11,   a12, ...,   a1n), collapse),
          paste0(paste(  a21,   a22, ...,   a2n), collapse),
          paste0(paste(am-11, am-12, ..., am-1n), collapse),      
                 paste(  am1,   am2, ...,   amn)

    One can see this easily by intuition from examples:

    We think the only paste version that is not implemented in base R is an array version. At least we are not aware of sth. like row_paste or paste_apply etc.

26.4 S3

  1. Q: The most important S3 objects in base R are factors, data frames, difftimes, and date/times (Dates, POSIXct, POSIXlt). You’ve already seen the attributes and base type that factors are built on. What base types and attributes are the others built on?

    A: TODO: Add answer for difftime.

    data frame: Data frames are build up on (named) lists. Together with the row.names attribute and after setting the class to “data.frame”, we get a classical data frame

    date/times (Dates, POSIXct, POSIXlt): Date is just a double with the class attribute set to “Date”

    POSIXct is a class for date/times that inherits from POSIXt and is built on doubles as well. The only attribute is tz (for timezone)

    POSIXlt is another date/time class that inherits from POSIXt. It is built on top of a named list and a tzone attribute. Differences between POSIXct and POSIXlt are described in ?DateTimeClasses.

  2. Q: Draw a Venn diagram illustrating the relationships between functions, generics, and methods.

    A: Funtions don’t have to be generics or methods, but both the latter are functions. It is also possible that a function is both, a method and a generic, at the same time, which seems to be relatively awkward, so that also the author of the textbook doesn’t recommend it, see ?pryr::ftype

    This function figures out whether the input function is a regular/primitive/internal function, a internal/S3/S4 generic, or a S3/S4/RC method. This is function is slightly simplified as it’s possible for a method from one class to be a generic for another class, but that seems like such a bad idea that hopefully no one has done it.

  3. Q: Write a constructor for difftime objects. What base type are they built on? What attributes do they use? You’ll need to consult the documentation, read some code, and perform some experiments.

    A: Our constructor should be named new_class_name, have one argument for its base type and each attribute and check the base types of these arguments as well.

    However, since the following result prints awkward

    we get a little bit more “inspiration” by the original difftime() function and make the regarding changes. Basically we need to implement logic for the units attribute, in case it is set to "auto" and convert the value of the underlying double from seconds to the regarding unit, as commented in the following

26.4.1 Inheritance

  1. Q: The ordered class is a subclass of factor, but it’s implemented in a very ad hoc way in base R. Implement it in a principled way by building a constructor and providing vec_restore() method.

    A: TODO: the olad exercise text ended on “an as_ordered generic”. Check the answer if it needs to be updated.

    ordered is a subclass of factor, so we need to do the following

    • for factors: add a subclass argument to the constructor and helper
    • for ordered: add a constructor
    • write an as_ordered() generic with methods ordered, factor and default

    We use the factor constructor from the textbook and add the subclass argument

    We also use the validator for factors from the textbook

    And we add the subclass argument for the helper from the textbook and the exercises

    A constructor for ordered is already implemented in the sloop package:

    The implementation of the generic and the first two methods is straight forward

    For the factor method of as_ordered() we use the factor helper, since it saves us some typing:

    Finally, our new method preserves all levels:

    For a real scenario, we might want to add an as_factor.ordered() method to the as_factor() generic from the textbook.

26.6 Expressions

26.6.1 Abstract syntax trees

  1. Q: Use ast() and experimentation to figure out the three arguments to an if() call. What would you call them? Which arguments are required and which are optional?

    A: You can write an if() statement in several ways: with or without else, formatted or in one line and also in prefix notation. Here are several versions focussing on the possibility of leaving out curly brackets.

    One possible way of naming the arguments would be: condition (1), conclusion (2), alternative (3).

    The condition is always required. If the condition is TRUE, also the conclusion is required. If the condition is FALSE and if() is called in combination with else(), then also the alternative is required.

  2. Q: What are the arguments to the for() and while() calls?

    A: for() requires an index (called var in the docs), a sequence and an expression, for example

    while() requires a condition and an expression. Again, an example in prefix notation:

    Note that a minimal expression can consist of { only.

  3. Q: Two arithmetic operators can be used in both prefix and infix style. What are they?

    A: I am not sure how this is meant to be. Theoretically every arithmetic operator can be written in prefix notation via backticks. On the other hand, + and - seem to be the only ones, which can be written in infix notation without backticks.

    However, when we look more closely, the call tree is not what we would expect from a prefix function

    So maybe it is meant to look like this…

    Of course also this doesn’t make too much sense, since in ?Syntax one can read, that R clearly differentiates between unary and binary + and - operators and a unary operator is not really what we mean, when we speak about infix operators.

    However, if we don’t differentiate in this way, this is probably the solution, since it’s obviously also an infix function:

26.7 Quasiquotation (new)

  1. Q: Why does as.Date.default() use substitute() and deparse()? Why does pairwise.t.test() use them? Read the source code.

    A: as.Date.default() uses them to convert unexpected input expressions (neither dates, nor NAs) into a character string and return it within an error message.

    pairwise.t.test() uses them to convert the names of its datainputs (response vector x and grouping factor g) into character strings to format these further into a part of the desired output.

  2. Q: pairwise.t.test() assumes that deparse() always returns a length one character vector. Can you construct an input that violates this expectation? What happens?

    A: We can pass an expression to one of pairwise.t.test()’s data input arguments, which exceeds the default cutoff width in deparse(). The expression will be split into a character vector of length greater 1. The deparsed data inputs are directly pasted (read the source code!) with “and” as separator and the result is just used to be displayed in the output. Just the data.name output will change (it will include more than one “and”).

26.8 FO

26.8.1 Behavioural FOs

  1. Q: What does the following function do? What would be a good name for it?

    A: It returns a new version of the inputfunction. That version will always return the result of it’s first run (in case this not NULL), no matter how the input changes. Good names could be first_run() or initial_return().

  2. Q: Modify delay_by() so that instead of delaying by a fixed amount of time, it ensures that a certain amount of time has elapsed since the function was last called. That is, if you called g <- delay_by(1, f); g(); Sys.sleep(2); g() there shouldn’t be an extra delay.

  3. Q: Write wait_until() which delays execution until a specific time.


  4. Q: There are three places we could have added a memoise call: why did we choose the one we did?

    A: The second was chosen. It’s easy to see why, if we eliminate the other two options:

    • The first version only prints a dot at every tenth download() call with a new input. This is because dot_every() is inside of memoise() and the counter created by dot_every() is not “activated” if the input is known.

    • The third version takes one second for every call. Even if we already know the result and don’t download anything again.

  5. Q: Why is the remember() function inefficient? How could you implement it in more efficient way?

  6. Q: Why does the following code, from stackoverflow, not do what you expect?

    How can you modify f so that it works correctly?

    A: You can read in the stackoverflow link that the question arose, because the original return of fs[[1]](3) was 4, which is due to lazy evaluation and could be solved by two users via force():

    However you can see in the result within the question that R’s behaviour was changed in this case and as Jan Kislinger points out on twitter:

    The real question should be: “How did they modify #rstats so that it works correctly?” otherwise it’s a tricky question :D

    Note that the same issue appears in the textbook:

    In the following example, we take a list of functions and delay each one. But when we try to evaluate the mean, we get the sum instead.

    Which (as one can see) is not true anymore…actually it changed in R version 3.2:

    Higher order functions such as the apply functions and Reduce() now force arguments to the functions they apply in order to eliminate undesirable interactions between lazy evaluation and variable capture in closures. This resolves PR#16093.

    For further interested: PR#16093 will lead you to the subject “iterated lapply” within the R-devel Archives. Note that the behaviour in for loops is still as “the old lapply()” behaviour.

26.8.2 Output FOs

  1. Q: Create a negative() FO that flips the sign of the output of the function to which it is applied.


  2. Q: The evaluate package makes it easy to capture all the outputs (results, text, messages, warnings, errors, and plots) from an expression. Create a function like capture_it() that also captures the warnings and errors generated by a function.

    A: One way is just to capture the output of tryCatch() with identity handlers for errors and warnings:

  3. Q: Create a FO that tracks files created or deleted in the working directory (Hint: use dir() and setdiff().) What other global effects of functions might you want to track?


26.8.3 Input FOs

  1. Q: Our previous download() function only downloads a single file. How can you use partial() and lapply() to create a function that downloads multiple files at once? What are the pros and cons of using partial() vs. writing a function by hand?

  2. Q: Read the source code for plyr::colwise(). How does the code work? What are colwise()’s three main tasks? How could you make colwise() simpler by implementing each task as a function operator? (Hint: think about partial().)

    A: We describe how it works by commenting the source code:

    function (.fun, .cols = true, ...) 
      # We check if .cols is not a function, since it is possible to supply a
      # predicate function.
      # if so, the .cols arguments will be "quoted", and filter() will 
      # be a function that checks and evaluates these .cols within its other argument
      if (!is.function(.cols)) {
        .cols <- as.quoted(.cols)
        filter <- function(df) eval.quoted(.cols, df)
      # otherwise, filter will be be Filter(), which applies the function 
      # in .cols to every element of its other argument
      else {
        filter <- function(df) Filter(.cols, df)
      # the ... arguments are caught in the list dots
      dots <- list(...)
      # a function is created, which will also be the return value.
      # it checks if its input is a data frame
      function(df, ...) {
        # if df is split (in "plyr" speaking), this will be taken into account...
        df <- strip_splits(df)
        # now the columns of the data frame are chosen, depending on the input of .cols
        # this can chosen directly, via a predicate function, or all columns (default)
        filtered <- filter(df)
        # if this means, that no columns are selected, an empty data frame will be returned
        if (length(filtered) == 0) 
        # otherwise lapply will be called on all filtered columns, with 
        # the .fun argument, which has to be provided by the user, and some other
        # arguments provided by the user, when calling the function (...) and
        # when defining the function (dots)
        out <- do.call("lapply", c(list(filtered, .fun, ...), 
        # the output will be named and converted from list into a data frame again
        names(out) <- names(filtered)
    <environment: namespace:plyr>
  3. Q: Write FOs that convert a function to return a matrix instead of a data frame, or a data frame instead of a matrix. If you understand S3, call them as.data.frame.function() and as.matrix.function().


  4. Q: You’ve seen five functions that modify a function to change its output from one form to another. What are they? Draw a table of the various combinations of types of outputs: what should go in the rows and what should go in the columns? What function operators might you want to write to fill in the missing cells? Come up with example use cases.

  5. Q: Look at all the examples of using an anonymous function to partially apply a function in this and the previous chapter. Replace the anonymous function with partial(). What do you think of the result? Is it easier or harder to read?

    A: The results are easy to read. Especially the Map() examples profit in readability:

    #> Registered S3 method overwritten by 'pryr':
    #>   method      from
    #>   print.bytes Rcpp
    #> Attaching package: 'pryr'
    #> The following object is masked _by_ '.GlobalEnv':
    #>     f
    ## From Functionals
    # 1
    trims <- c(0, 0.1, 0.2, 0.5)
    x <- rcauchy(1000)
    unlist(lapply(trims, function(trim) mean(x, trim = trim)))
    #> [1] -0.00498  0.05088  0.03304  0.02733
    unlist(lapply(trims, partial(mean, x)))
    #> [1] -0.00498  0.05088  0.03304  0.02733
    # 2
    xs <- replicate(5, runif(10), simplify = FALSE)
    ws <- replicate(5, rpois(10, 5) + 1, simplify = FALSE)
    unlist(Map(function(x, w) weighted.mean(x, w, na.rm = TRUE), xs, ws))
    #> [1] 0.453 0.521 0.500 0.443 0.525
    unlist(Map(partial(weighted.mean, na.rm = TRUE), xs, ws))
    #> [1] 0.453 0.521 0.500 0.443 0.525
    # 3
    add <- function(x, y, na.rm = FALSE) {
      if (na.rm && (is.na(x) || is.na(y))) rm_na(x, y, 0) else x + y
    r_add <- function(xs, na.rm = TRUE) {
      Reduce(function(x, y) add(x, y, na.rm = na.rm), xs)
    r_add_compact <- function(xs, na.rm = TRUE) {
      Reduce(partial(add, na.rm = na.rm), xs)
    #> [1] 10
    #> [1] 10
    # 4
    v_add1 <- function(x, y, na.rm = FALSE) {
      stopifnot(length(x) == length(y), is.numeric(x), is.numeric(y))
      if (length(x) == 0) return(numeric())
        Map(function(x, y) add(x, y, na.rm = na.rm), x, y)
    v_add1_compact <- function(x, y, na.rm = FALSE) {
      stopifnot(length(x) == length(y), is.numeric(x), is.numeric(y))
      if (length(x) == 0) return(numeric())
        Map(partial(add, na.rm = na.rm), x, y)
    v_add1(1:3, 2:4)
    #> [1] 3 5 7
    v_add1_compact(1:3, 2:4)
    #> [1] 3 5 7
    # 5
    c_add <- function(xs, na.rm = FALSE) {
      Reduce(function(x, y) add(x, y, na.rm = na.rm), xs,
             accumulate = TRUE)
    c_add_compact <- function(xs, na.rm = FALSE) {
      Reduce(partial(add, na.rm = na.rm), xs,
             accumulate = TRUE)
    #> [1] 1 3 6
    #> [1] 1 3 6
    ## From Function operators
    # 6
    f <- function(x) x ^ 2
    #> function (...) 
    #> f(...)
    # 7
    # Map(function(x, y) f(x, y, zs), xs, ys)
    # Map(partial(f, zs = zs), xs, yz)
    # 8
    # f <- function(a) g(a, b = 1)
    # f <- partial(g, b = 1)
    # 9
    compact <- function(x) Filter(Negate(is.null), x)
    compact <- partial(Filter, Negate(is.null))
    # 10
    # Map(function(x, y) f(x, y, zs), xs, ys)
    # Map(partial(f, zs = zs), xs, ys)
    # 11
    funs2 <- list(
      sum = function(...) sum(..., na.rm = TRUE),
      mean = function(...) mean(..., na.rm = TRUE),
      median = function(...) median(..., na.rm = TRUE)
    funs2 <- list(
      sum = partial(sum, na.rm = TRUE),
      mean = partial(mean, na.rm = TRUE),
      median = partial(median, na.rm = TRUE)

26.8.4 Combining FOs

  1. Q: Implement your own version of compose() using Reduce and %o%. For bonus points, do it without calling function.

    A: We use the definition from the textbook:

    And then we build two versions. One via an anonymous function and one via partial():

  2. Q: Extend and() and or() to deal with any number of input functions. Can you do it with Reduce()? Can you keep them lazy (e.g., for and(), the function returns once it sees the first FALSE)?

    A: We use and() and or() as defined in the textbook. They are lazy, since they are build up on && and ||. Also their reduced versions stay lazy, as we will show at the end of the code

  3. Q: Implement the xor() binary operator. Implement it using the existing xor() function. Implement it as a combination of and() and or(). What are the advantages and disadvantages of each approach? Also think about what you’ll call the resulting function to avoid a clash with the existing xor() function, and how you might change the names of and(), not(), and or() to keep them consistent.

    A: Both versions are implemented straight forward, as also the reduced versions. However, the parallel versions need a little bit more care:

    xor_fb1 <- function(f1, f2){
      force(f1); force(f2)
        xor(f1(...), f2(...)) 
    xor_fb2 <- function(f1, f2){
      force(f1); force(f2)
        or(f1, f2)(...) && !(and(f1, f2)(...))
    # binary combination
    xor_fb1(is.logical, is.character)("a")
    #> [1] TRUE
    xor_fb2(is.logical, is.character)("a")
    #> [1] TRUE
    # parallel combination (results in an error)
    xor_fb1(c(is.logical, is.character), c(is.logical, is.character))("a")
    #> Error in f1(...): could not find function "f1"
    xor_fb2(c(is.logical, is.character), c(is.logical, is.character))("a")
    #> Error in f1(...): could not find function "f1"
    # reduced combination (results in an error)
    xor_fb1(c(is.logical, is.character, is.logical, is.character))("a")
    #> Error in force(f2): argument "f2" is missing, with no default
    xor_fb2(c(is.logical, is.character, is.logical, is.character))("a")
    #> Error in force(f2): argument "f2" is missing, with no default
    ### Reduced version
    xor_fb1_red <- function(fs){
      Reduce(function(f, g) xor_fb1(f, g), fs)
    xor_fb2_red <- function(fs){
      Reduce(function(f, g) xor_fb2(f, g), fs)
    # should return TRUE
    xor_fb1_red(c(is.logical, is.character, is.logical, is.character))("a")
    #> [1] FALSE
    xor_fb2_red(c(is.logical, is.character, is.logical, is.character))("a")
    #> [1] FALSE
    # should return FALSE
    xor_fb1_red(c(is.logical, is.logical, is.character, is.logical))("a")
    #> [1] TRUE
    xor_fb2_red(c(is.logical, is.logical, is.character, is.logical))("a")
    #> [1] TRUE
    # should return FALSE
    xor_fb1_red(c(is.logical, is.logical, is.character, is.character))("a")
    #> [1] FALSE
    xor_fb2_red(c(is.logical, is.logical, is.character, is.character))("a")
    #> [1] FALSE
  4. Q: Above, we implemented boolean algebra for functions that return a logical function. Implement elementary algebra (plus(), minus(), multiply(), divide(), exponentiate(), log()) for functions that return numeric vectors.


26.9 Expressions (again)

26.9.1 Data structures

  1. Q: How is rlang::maybe_missing() implemented? Why does it work?

    A: Let us take a look at the functions source code to see what’s going on

    First it is checked if the argument is missing. If so, the missing arg is returned, otherwise the argument (x) itsself is returned.

26.9.2 Parsing and deparsing

  1. Q: Why does as.Date.default() use substitute() and deparse()? Why does pairwise.t.test() use them? Read the source code.


26.9.3 R’s grammar

  1. Q: deparse() produces vectors when the input is long. For example, the following call produces a vector of length two:

    What do expr_text(), expr_name(), and expr_label() do with this input?


    • expr_text() pastes the output string into one and inserts \n (new line identifiers) as separators
    • expr_name() recreates the call into the form f(…) and deparses this expression into a string
    • expr_label() does the same as expr_name(), but surrounds the output also with backticks

26.10 Quasiquotation

26.10.1 Old exercises Unquoting

  1. Q: What does the following command return? What information is lost? Why?

    A: When we look at the captured expression, we see that the extra whitespaces and comments are lost. R ignores them when parsing an expression. They do do not need to be represented in the AST, because they do not affect the evaluation of the expression.

    However, it is possible to retrieve the original input through the attributes of the captured expression:

26.10.2 Unquoting

  1. Q: Explain why both !0 + !0 and !1 + !1 return FALSE while !0 + !1 returns TRUE.

    A: To answer this question we look at the AST of the first example:

    As the coercion rules are the same in all examples, we can use the precedence order (right to left) to explain all three examples:

    • !0 + !0:
      So the second zero gets coerced to FALSE and !FALSE becomes TRUE.
      0 + TRUE gets coerced to 1.
      !1 becomes !TRUE which is FALSE
    • !1 + !1:
      So !1 is FALSE.
      1 + FALSE is 1.
      !1 is !TRUE so FALSE.
    • !0 + !1:
      !1 is FALSE.
      0 + FALSE is 0.
      !0 is TRUE.
  2. Q: Base functions match.fun(), page(), and ls() all try to automatically determine whether you want standard or non-standard evaluation. Each uses a different approach. Figure out the essence of each approach by reading the source code, then compare and contrast the techniques.

26.10.3 Case studies

  1. Q: Implement arrange_desc(), a variant of dplyr::arrange() that sorts in descending order by default.

    A: We just have to catch the ... from arrange() as an expression and modify the expression to be wrapped inside desc(). Afterwards we evaluate this new code within a regular arrange() call:

    Let’s try it out

  2. Q: Implement filter_or(), a variant of dplyr::filter() that combines multiple arguments using | instead of &.

    A: This time we just need to collapse the ... arguments with |. Therefore we can use purrr::reduce() and afterwards we just need to evaluate the new code within a regular filter call:

  3. Q:Implement partition_rows() which, like partition_cols(), returns two data frames, one containing the selected rows, and the other containing the rows that weren’t selected.

    A: We just have to decide if we focus on integer subsetting via dplyr::slice() or logical subsetting via dplyr::filter(). The rest is straightforward. Since the implementations of both subsetting styles are completely equivalent we just choose one without any particular reason:

  4. Q:Add error handling to slice(). Give clear error messages if either along or index have invalid values (i.e. not numeric, not length 1, too small, or too big).

26.11 More old evaluation exercises

  1. Q: Run this code in your head and predict what it will print. Confirm or refute your prediction by running the code in R.

  2. Q: What happens if you use expr() instead of enexpr() inside of subset2()?

  3. Q: Improve subset2() to make it more like real base::subset():

    • Drop rows where subset evaluates to NA
    • Give a clear error message if subset doesn’t yield a logical vector
    • What happens if subset yields a vector that’s not the same as the number of rows in data? What do you think should happen?
  4. Q: The third argument in base::subset() allows you to select variables. It treats variable names as if they were positions. This allows you to do things like subset(mtcars, , -cyl) to drop the cylinder variable, or subset(mtcars, , disp:drat) to select all the variables between disp and drat. How does this work? I’ve made this easier to understand by extracting it out into its own function that uses tidy evaluation.

  5. Q: Here’s an alternative implementation of arrange():

    Describe the primary difference in approach compared to the function defined in the text.

    One advantage of this approach is that you could check each element of ... to make sure that input is correct. What property should each element of ords have?

  6. Q: Here’s an alternative implementation of subset2():

    Use intermediate variables to make the function easier to understand, then explain how this approach differs to the approach in the text.

  7. Q: Implement a form of arrange() where you can request a variable to sorted in descending order using named arguments:

    (Hint: The descreasing argument to order() will not help you. Instead, look at the definition of dplyr::desc(), and read the help for xtfrm().)

  8. Q: Why do you not need to worry about ambiguous argument names with ... in arrange()? Why is it a good idea to use the . prefix anyway?

  9. Q: What does transform() do? Read the documentation. How does it work? Read the source code for transform.data.frame(). What does substitute(list(...)) do?

  10. Q: Use tidy evaluation to implement your own version of transform(). Extend it so that a calculation can refer to variables created by transform, i.e. make this work:

  11. Q: What does with() do? How does it work? Read the source code for with.default(). What does within() do? How does it work? Read the source code for within.data.frame(). Why is the code so much more complex than with()?

  12. Q: Implement a version of within.data.frame() that uses tidy evaluation. Read the documentation and make sure that you understand what within() does, then read the source code.

  1. Q: When model building, typically the response and data are relatively constant while you rapidly experiment with different predictors. Write a small wrapper that allows you to reduce duplication in this situation.

  2. Q: Another way to way to write boot_lm() would be to include the boostrapping expression (data[sample(nrow(data), replace = TRUE), , drop = FALSE]) in the data argument. Implement that approach. What are the advantages? What are the disadvantages?

  3. Q: To make these functions somewhat more robust, instead of always using the caller_env() we could capture a quosure, and then use its environment. However, if there are multiple arguments, they might be associated with different environments. Write a function that takes a list of quosures, and returns the common environment, if they have one, or otherwise throws an error.

  4. Q: Write a function that takes a data frame and a list of formulas, fitting a linear model with each formula, generating a useful model call.

  5. Q: Create a formula generation function that allows you to optionally supply a transformation function (e.g. log()) to the response or the predictors.

26.12 Deprecated evaluation basics

  1. Q: The code generated by source2() lacks source references. Read the source code for sys.source() and the help for srcfilecopy(), then modify source2() to preserve source references. You can test your code by sourcing a function that contains a comment. If successful, when you look at the function, you’ll see the comment and not just the source code.


    • the comment is still missing

26.13 Measuring performance (Old exercises)

  1. Q: Instead of using microbenchmark(), you could use the built-in function system.time(). But system.time() is much less precise, so you’ll need to repeat each operation many times with a loop, and then divide to find the average time of each operation, as in the code below.

    How do the estimates from system.time() compare to those from microbenchmark()? Why are they different?

  2. Q: Here are two other ways to compute the square root of a vector. Which do you think will be fastest? Which will be slowest? Use microbenchmarking to test your answers.

    A: The second one looks more complex, but you never know…unless you test it.

  3. Q: Use microbenchmarking to rank the basic arithmetic operators (+, -, *, /, and ^) in terms of their speed. Visualise the results. Compare the speed of arithmetic on integers vs. doubles.

    A: On a Windows system, these short execution times are hard to measure. The following code was executed on Linux and pasted here:

    Let’s compare the results visually.

  4. Q: You can change the units in which the microbenchmark results are expressed with the unit parameter. Use unit = "eps" to show the number of evaluations needed to take 1 second. Repeat the benchmarks above with the eps unit. How does this change your intuition for performance?

  1. Q: scan() has the most arguments (21) of any base function. About how much time does it take to make 21 promises each time scan is called? Given a simple input (e.g., scan(text = "1 2 3", quiet = T)) what proportion of the total run time is due to creating those promises?

    A: According to the textbook every extra argument slows the function down by approximately 20 nanoseconds, which I can’t reproduce on my system:

    However, for now we just assume that 20 nanoseconds are correct and in kind of doubt, we recommend to benchmark this value individually. With this assumption we calculate 21 * 20 = 420 nanoseconds of extra time for each call of scan().

    For a percentage, we first benchmark a simple call of scan():

    This lets us calculate, that ~1.18% of the median run time are caused by the extra arguments.

  2. Q: Read “Evaluating the Design of the R Language”. What other aspects of the R-language slow it down? Construct microbenchmarks to illustrate.

  3. Q: How does the performance of S3 method dispatch change with the length of the class vector? How does performance of S4 method dispatch change with number of superclasses? How about RC?

  4. Q: What is the cost of multiple inheritance and multiple dispatch on S4 method dispatch?

  5. Q: Why is the cost of name lookup less for functions in the base package?

  1. Q: The performance characteristics of squish_ife(), squish_p(), and squish_in_place() vary considerably with the size of x. Explore the differences. Which sizes lead to the biggest and smallest differences?

  2. Q: Compare the performance costs of extracting an element from a list, a column from a matrix, and a column from a data frame. Do the same for rows.