Copyright (C) 2016, 2017, H. Conrad Cunningham

Acknowledgements: In Summer 2016, I adapted and revised this chapter from chapter 6 of my Notes on Functional Programming with Haskell and from my notes on Functional Data Structures from the Spring 2016 Scala variant of CSci 555. The latter was based, in part, on chapter 3 of the book Functional Programming in Scala by Paul Chiusano and Runar Bjarnason (Manning, 2015).

I also expanded the discussion of of combinators and functional composition and added discussion of seq, $, and $!. These drew on ideas from chapters 4 and 7 of Richard Bird's Thinking Functionally with Haskell (Cambridge University Press, 2015) and chapter 11 of Simon Thompson's Haskell: The Craft of Programming, Third Edition (Pearson 2011).

In 2017, I continue to develop this chapter.

I maintain these notes as text in Pandoc's dialect of Markdown using embedded LaTeX markup for the mathematical formulas and then translate the notes to HTML, PDF, and other forms as needed.

Advisory: The HTML version of this document may require use of a browser that supports the display of MathML. A good choice as of October 2017 is a recent version of Firefox from Mozilla.

TODO:

5 Higher-Order Functions

5.1 Chapter Introduction

This chapter introduces higher-order functions and constructs a library of useful higher-order functions to process lists.

The goals of the chapter are for the students to be able to:

Upon successful completion of this chapter, students should be able to:

  1. describe the syntax and semantics of Haskell function concepts and operators such as strictness, currying, partial application, operator sections, combinators, lambda expressons, and strict function application

  2. describe the syntax and semantics of first-class and higher-order Haskell functions

  3. develop Haskell programs using using Haskell function concepts and operators such as strictness, currying, partial application, operator sections, combinators, lambda expressons, and strict function application

  4. use a library of higher-order Haskell functions to develop Haskell programs

  5. generalize and develop higher-order Haskell functions for reuse

  6. appreciate the mathematical characteristics of Haskell functions (e.g., to state properties of functions and operators)

The Haskell module for this chapter is in Mod05HigherOrder.hs.

5.2 Higher-Order List Programming

5.2.1 Generalizing procedural abstractions

A function in a programming language is a procedural abstraction. It separates the logical properties of a computation from the details of how the computation is implemented. It abstracts a pattern of behavior and encapsulates it within a program unit.

Suppose we wish to perform the same computation on a set of similar data structures. As we have seen, we can encapsulate the computation in a function having the data structure as an argument. For example, the function length' computes the number of elements in a list of any type.

Suppose instead we wish to perform a similar (but not identical) computation on a set of similar data structures. For example, we want to compute the sum or the product of a list of numbers. In this case, we may can pass the operation itself into the function.

This kind of function is called a higher-order function. A higher-order function is a function that takes functions as arguments or returns functions in a result. Most traditional imperative languages do not fully support higher-order functions.

In most functional programming languages, functions are treated as first class values. That is, functions can be stored in data structures, passed as arguments to functions, and returned as the results of functions. Historically, imperative languages have not treated functions as first-class values. (Recently, many imperative languages, such as Java 8, have added support for functions as first-class values.)

The higher-order functions in Haskell and other functional programming languages enable us to construct regular and powerful abstractions and operations. By taking advantage of a library of higher-order functions that capture common patterns of computation, we can quickly construct concise, yet powerful, programs.

This can increase programmer productivity and program reliability because such programs are shorter, easier to understand, and constructed from well-tested components.

Higher-order functions can also increase the modularity of programs by enabling simple program fragments to be "glued together" readily into more complex programs.

In this section, we examine several common patterns and build a library of useful higher-order functions.

5.2.2 Defining map

Consider the following two functions, noting their type signatures and patterns of recursion.

The first, squareAll, takes a list of integers and returns the corresponding list of squares of the integers.

    squareAll :: [Int] -> [Int]  squareAll :: [Int] -> [Int] 
    squareAll []     = [] 
    squareAll (x:xs) = (x * x) : squareAll xs

The second, lengthAll, takes a list of lists and returns the corresponding list of the lengths of the element lists; it uses the Prelude function length.

    lengthAll :: [[a]] -> [Int] 
    lengthAll []       = [] 
    lengthAll (xs:xss) = (length xs) : lengthAll xss

Although these functions take different kinds of data (a list of integers versus a list of polymorphically typed lists) and apply different operations (squaring versus list length), they exhibit the same pattern of computation. That is, both take a list of some type and apply a given function to each element to generate a resulting list of the same length as the original.

The combination of polymorphic typing and higher-order functions allow us to abstract this pattern of computation into a standard function.

We can abstract the pattern of computation common to squareAll and lengthAll as the (broadly useful) function map, which we define as follows. (In this chapter, we often add a suffix to the base function names to avoid conflicts with the similarly named functions in the Prelude. Here we use map’ instead of map.)

    map' :: (a -> b) -> [a] -> [b]  -- map in Prelude
    map' f []     = [] 
    map' f (x:xs) = f x : map' f xs 

Function map generalizes squareAll, lengthAll, and similar functions by adding a higher-order parameter for the operation applied and making the input and the output lists polymorphic. Specifically, the function takes a function f of type a -> b and a list of type [a], applies function f to each element of the list, and produces a list of type [b].

Thus we can specialize map to give new definitions of squareAll and lengthAll as follows:

    squareAll2 :: [Int] -> [Int] 
    squareAll2 xs = map' sq xs 
                    where sq x = x * x 
    InIn
    lengthAll2 :: [[a]] -> [Int] 
    lengthAll2 xss = map' length xss

Consider the following questions.

5.2.3 Thinking about data transformations

Above we define map as a recursive function that transforms the elements of a list one by one. However, it is often more useful to think of map in one of two ways:

  1. as a powerful list operator that transforms every element of the list. We can combine map with other powerful operators to quickly construct powerful list processing programs.

    We can consider map as operating on every element of the list "simultaneously". In fact, an implementation could use separate processors to transform each element: this is essentially the map operation in Google’s mapReduce distributed "big data" processing framework.

    Referential transparency and immutable data structures make parallelism easier in Haskell than in most imperative languages.

  2. as a operator node in a dataflow network. A stream of data objects flows into the map node. The map node transforms each object by applying the argument function. Then the data object flows out to the next node of the network.

    The lazy evaluation of the Haskell functions enables such an implementation.

Although in the early parts of these notes we give attention to the details of recursion, learning how to think like a functional programmer requires us to think about large-scale transformations of collections of data.

5.2.4 Generalizing function definitions

Whenever we recognize a computational pattern in a set of related functions, we can generalize the function definition as follows:

  1. Do a scope-commonality-variability analysis on the set of related functions.

    That is, identify what is to be included and what not (i.e., the scope), the parts of functions that are the same (the commonalities or frozen spots), and the parts that differ (the variabilities or hot spots)

  2. Leave the commonalities in the generalized function's body.

  3. Move the variabilities into the generalized function's header -- its type signature and parameter list.

  4. Consider other approaches if the generalized function's type type signature and parameter list become too complex.

    For example, we can introduce new data or procedural abstractions for parts of the generalized function. These may be in the same module of the generalized function or in an appropriately defined separate module.

5.2.5 Defining filter

Consider the following two functions.

The first, getEven, takes a list of integers and returns the list of those integers that are even (i.e., are multiples of 2). The function preserves the relative order of the elements in the list.

    getEven :: [Int] -> [Int] 
    getEven []      = [] 
    getEven (x:xs)
        | even x    = x : getEven xs 
        | otherwise = getEven xs

The second, doublePos, takes a list of integers and returns the list of doubles of the positive integers from the input list; it preserves the relative order of the elements.

    doublePos :: [Int] -> [Int] 
    doublePos []          = [] 
    doublePos (x:xs) 
        | 0 < x     = (2 * x) : doublePos xs 
        | otherwise = doublePos xs

Function even is from the Prelude; it returns True if its argument is evenly divisible by 2 and returns False otherwise.

What do these two functions have in common? What differs?

Using the generalization method outlined above, we abstract the pattern of computation common to getEven and doublePos as the (broadly useful) function filter found in the Prelude. (We call the function filter’ below to avoid a name conflict.)

    filter' :: (a -> Bool) -> [a] -> [a]  -- filter in Prelude
    filter' _ []    = [] 
    filter' p (x:xs)
        | p x       = x : xs' 
        | otherwise = xs' 
                    where xs' = filter' p xs

Function filter takes a predicate p of type a -> Bool and a list of type [a] and returns a list containing those elements that satisfy p, in the same order as the input list. Note that the keyword where begins in the same column as the = in the defining equations; thus the scope of the definition of xs’ extends over both legs of the definition.

Function filter does not incorporate the doubling operation from doublePos. We could have included it as another higher-order parameter, but we leave it out to keep the generalized function simple. We can use the already defined map function to achieve this separately.

Therefore, we can specialize filter to give new definitions of getEven and doublePos as follows:

    getEven2 :: [Int] -> [Int] 
    getEven2 xs = filter' even xs 

    doublePos2 :: [Int] -> [Int] 
    doublePos2 xs = map' dbl (filter' pos xs) 
                    where dbl x  = 2 * x
                          pos x = (0 < x)

Note that function doublePos2 exhibits both the filter and the map patterns of computation.

The standard higher-order functions map and filter allow us to restate the three-leg definitions of getEven and doublePos in just one leg each, except that doublePos requires two lines of local definitions. In subsequent subsections, we see how to eliminate these simple local definitions as well.

Consider the following questions.

5.2.6 Defining fold right (foldr)

Consider the sum and product functions we defined in a previous chapter, ignoring the short-cut handling of the zero element in product.

    sum' :: [Int] -> Int             -- sum in Prelude
    sum' []     = 0
    sum' (x:xs) = x + sum' xs 
        
    product' :: [Integer] -> Integer -- product in Prelude
    product' []     = 1 
    product' (x:xs) = x * product' xs 

Both sum' and product' apply arithmetic operations to integers. What about other operations with similar pattern of computation?

Also consider a function concat that concatenates a list of lists of some type into a list of that type with the order of the input lists and their elements preserved.

    concat' :: [[a]] -> [a]  -- concat in Prelude
    concat' []       =  []
    concat' (xs:xss) =  xs ++ concat' xss

For example,

    sum' [1,2,3]          = (1 + (2 + (3 + 0)))
    product' [1,2,3]      = (1 * (2 * (3 * 1)))
    concat' ["1","2","3"] = ("1" ++ ("2" ++ ("3" ++ "")))

What do sum', product', and concat' have in common? What differs?

All exhibit the same pattern of computation.

This group of functions inserts operations of type a -> a -> a between elements a list of type [a].

But these are special cases of more general operations of type a -> b -> b. In this case, the value returned must be of type b in the case of both empty and nonempty lists.

We can abstract the pattern of computation common to sum', product', and concat’ as the function foldr (pronounced "fold right") found in the Prelude. (Here we use foldrX to avoid the name conflict.)

    foldrX :: (a -> b -> b) -> b -> [a] -> b  -- foldr in Prelude
    foldrX f z []     = z
    foldrX f z (x:xs) = f x (foldrX f z xs) 

Function foldr:

The foldr function "folds" the list elements (of type a) into a value (of type b) by "inserting" operation f between the elements, with value z "appended" as the rightmost element.

Often the seed value z is the right identity element for the operation, but foldr may be useful in some circumstances where it is not (or perhaps even if there is no right identity).

For example, foldr f z [1,2,3] expands to f 1 (f 2 (f 3 z)), or, using an infix style:

    1 `f` (2 `f` (3 `f` z)) 

Function foldr does not depend upon f being associative or having either a right or left identity.

Function foldr is backward recursive. If the function application is fully evaluated, it needs a new stack frame for each element of the input list. If its list argument is long or the folding function itself is expensive, then the function can terminate with a stack overflow error.

In Haskell, foldr is called a fold operation. Other languages sometimes call this a reduce or insert operation.

We can specialize foldr to restate the definitions for sum', product', and concat’.

    sum2 :: [Int] -> Int             -- sum
    sum2 xs = foldrX (+) 0 xs

    product2 :: [Int] -> Int         -- product
    product2 xs = foldrX (*) 1 xs

    concat2:: [[a]] -> [a]           -- concat
    concat2 xss = foldrX (++) [] xss

As further examples, consider the folding of the Boolean operators && ("and") and || ("or") over lists of Boolean values as Prelude functions and and or (shown as and’ and or’ below to avoid name conflicts):

    and', or' :: [Bool] -> Bool  -- and, or in Prelude
    and' xs = foldrX (&&) True xs 
    or'  xs = foldrX (||) False xs

Although their definitions look different, and’ and or’ are actually identical to functions and and or in the Prelude.

Consider the following questions.

5.2.7 Using foldr

The fold functions are very powerful. By choosing an appropriate folding function argument, many different list functions can be implemented in terms of foldr.

For example, we can implement map using foldr as follows:

    map2 :: (a -> b) -> [a] -> [b] -- map
    map2 f xs = foldr mf [] xs 
        where mf y ys = (f y) : ys

The folding function mf y ys = (f y):ys applies the mapping function f to the next element of the list (moving right to left) and attaches the result on the front of the processed tail. This is a case where the folding function mf does not have a right identity, but where foldr is quite useful.

We can also implement filter in terms of foldr as follows:

    filter2 :: (a -> Bool) -> [a] -> [a]  -- filter
    filter2 p xs = foldr ff [] xs
        where ff y ys = if p y then (y:ys) else ys

The folding function ff y ys = if p x then (y:ys) else ys applies the filter predicate p to the next element of the list (moving right to left). If the predicate evaluates to True, the folding function attaches that element on the front of the processed tail; otherwise, it omits the element from the result.

We can also use foldr to compute the length of a polymorphic list.

    length2 :: [a] -> Int  -- length
    length2 xs  = foldr len  0  xs
        where len _ acc = acc + 1

This uses the z parameter of foldr to initialize the count to 0. Higher order argument f of foldr is a function that takes an element of the list as its left argument and the previous accumulator as its right argument and returns the accumulator incremented by 1. In this application, z is not the identity element for f but is a convenient beginning value for the counter.

We can construct an "append" function that uses foldr as follows:

    append2 :: [a] -> [a] -> [a]  -- ++
    append2 xs ys = foldr (:) ys xs

Here the the list that foldr operates on the first argument of the append. The z parameter is the entire second argument and the folding function is just (:). So the effect is to replace the [] at the end of the first list by the entire second list.

Function foldr 1s a backward recursive function that processes the elements of a list one by one. However, as we have seen, it is often more useful to think of foldr as a powerful list operator that reduces the element of the list into a single value. We can combine foldr with other operators to conveniently construct list processing programs.

5.2.8 Defining fold left (foldl)

We designed function foldr as a backward linear recursive function with the signature:

    foldr :: (a -> b -> b) -> b -> [a] -> b

As noted:

    foldr f z [1,2,3] ==  f 1 (f 2 (f 3 z))
                      ==  1 `f` (2 `f` (3 `f` z))

Consider a function foldl (pronounced "fold left") such that:

    foldl f z [1,2,3]  ==  f (f (f z 1) 2) 3
                       ==  ((z `f` 1) `f` 2) `f` 3`

This function folds from the left. It offers us the opportunity to use parameter z as an accumulating parameter in a tail recursive implementation. This is shown below as foldlX, which is similar to foldl in the Prelude.

    foldlX :: (a -> b -> a) -> a -> [b] -> a  -- foldl in Prelude
    foldlX f z []     = z  
    foldlX f z (x:xs) = foldlX f (f z x) xs

Note how the second leg of foldlX implements the left binding of the operation. In the recursive call of foldlX the "seed value" argument is used as an accumulating parameter.

Also note how the types of foldr and foldl differ.

Often the beginning value of z is the left identity of the operation f, but foldl (like foldr) can be a quite useful function in circumstances when it is not (or when f has no left identity).

5.2.9 Using foldl

If $\oplus$ is an associative binary operation of type t -> t -> t with identity element z (i.e., $\oplus$ and t form the algebraic structure know as a monoid), then, for any xs,

foldr ($\oplus$) z xs = foldl ($\oplus$) z xs

The classic Bird and Wadler textbook [Bird-Wadler, 1998] calls this property the first duality theorem.

Because +, *, and ++ are all associative operations with identity elements, sum, product, and concat can all be implemented with either foldr or foldl.

Which is better?

Depending upon the nature of the operation, an implementation using foldr may be more efficient than foldl or vice versa.

We defer a more complete discussion of the efficiency until we study evaluation strategies further in a later chapter.

As a rule of thumb, however, if the operation $\oplus$ is nonstrict in either argument, then it is usually better to use foldr. That form takes better advantage of lazy evaluation.

If the operation $\oplus$ is strict in both arguments, then it is often better (i.e., more efficient) to use the optimized version of foldl called foldl' from the standard Haskell module Data.List.

The append operation ++ is nonstrict in its second argument, so it is better to use foldr to implement concat.

Addition and multiplication are strict in both arguments, so we can implement sum and product functions efficiently with foldl', as follows:

    import Data.List   -- to make foldl' available
    sum3, product3 :: Num a =>  [a] -> a -- sum, product
    sum3 xs     = foldl' (+) 0 xs
    product3 xs = foldl' (*) 1 xs

Note that we generalize these functions to operate on polymorphic lists with a base type in class Num. Class Num includes all numeric types.

Function length3 uses foldl. It is like length2 except that the arguments of function len are reversed.

    length3 :: [a] -> Int  -- length
    length3 xs  = foldl len 0  xs
        where len acc _ = acc + 1

However, it is usually better to use the foldr version length2 because the folding function len is nonstrict in the argument corresponding to the list.

We can also implement list reversal using foldl as follows:

    reverse2 :: [a] -> [a]  -- reverse
    reverse2 xs = foldl rev [] xs
        where rev acc x = (x:acc) 

This gives a solution similar to the tail recursive reverse function from a previous chapter. Function foldl's z parameter is initially an empty list; foldl's folding function parameter f uses (:) to "attach" each element of the list as the new head of the accumulator, incrementally building the list in reverse order.

Although cons is nonstrict in its right operand, reverse2 builds up that argument from [], so reverse2 cannot take advantage of lazy evaluation by using foldr instead.

To avoid a stack overflow situation with foldr, we can first apply reverse to the list argument and then apply foldl as follows:

    foldr2 :: (a -> b -> b) -> b -> [a] -> b  -- foldr
    foldr2 f z xs = foldl flipf z (reverse xs)
        where flipf y x = f x y

The combining function in the call to foldl is the same as the one passed to foldr except that its arguments are reversed.

5.2.10 Defining concatMap (flatmap)

The higher-order function map applies its function argument f to every element of a list and returns the list of results. If the function argument f returns a list, then the result is a list of lists. Often we wish to flatten this into a single list, that is, apply a function like concat defined in a previous section.

This computation is sufficiently common that we give it the name concatMap. We can define it in terms of map and concat as

    concatMap' :: (a -> [b]) -> [a] -> [b] 
    concatMap' f xs = concat (map f xs)

or by combining map and concat into one foldr as:

    concatMap2 :: (a -> [b]) -> [a] -> [b] 
    concatMap2 f xs = foldr fmf [] xs 
        where fmf x ys = f x ++ ys

Above, the function argument to foldr applies the concatMap function argument f to each element of the list argument and then appends the resulting list in front of the result from processing the elements to the right.

We can also define filter in terms of concatMap as follows:

    filter3 :: (a -> Bool) -> [a] -> [a]
    filter3 p xs = concatMap' fmf xs
        where fmf x = if p x then [x] else []

The function argument to concatMap generates a one-element list if the filter predicate p is true and an empty list if it is false.

Some other languages (e.g., Scala) call the concatMap function by the name flatmap.

5.3 Haskell Function Concepts and Notation

5.3.1 Strictness

Some expressions cannot be reduced to a simple value, for example, div 1 0. The attempted evaluation of such expressions either return an error immediately or cause the interpreter to go into an "infinite loop".

In our discussions of functions, it is often convenient to assign the symbol $\bot$ (pronounced "bottom") as the value of expressions like div 1 0. We use $\bot$ is a polymorphic symbo--as a value of every type.

The symbol $\bot$ is not in the Haskell syntax and the interpreter cannot actually generate the value $\bot$. It is merely a name for the value of an expression in situations where the expression cannot really be evaluated. It’s use is somewhat analogous to use of symbols such as $\infty$ in mathematics.

Although we cannot actually produce the value $\bot$, we can, conceptually at least, apply any function to $\bot$.

If f $\bot = \bot$, then we say that the function is strict; otherwise, it is nonstrict (sometimes called lenient).

That is, a strict argument of a function must be evaluated before the final result can be computed. A nonstrict argument of a function may not need to be evaluated to compute the final result.

Assume that lazy evaluation is being used and consider the function two that takes an argument of any type and returns the integer value two.

    two :: a -> Int 
    two x = 2

The function two is nonstrict. The argument expression is not evaluated to compute the final result. Hence, two $\bot = 2$.

Consider the following examples.

5.3.2 Currying and partial application

Consider the following two functions:

    add :: (Int,Int) -> Int 
    add (x,y) = x + y 

    add' :: Int -> (Int -> Int) 
    add' x y  = x + y 

These functions are closely related, but they are not identical.

For all integers x and y, add (x,y) = add’ x y. But functions add and add’ have different types.

Function add takes a 2-tuple (Int,Int) and returns an Int. Function add’ takes an Int and returns a function of type Int -> Int.

What is the result of the application add 3? An error.

What is the result of the application add’ 3? The result is a function that "adds 3 to its argument".

What is the result of the application (add’ 3) 4? The result is the integer value .

By convention, function application (denoted by the juxtaposition of a function and its argument) binds to the left. That is, add’ x y = ((add’ x) y).

Hence, the higher-order functions in Haskell allow us to replace any function that takes a tuple argument by an equivalent function that takes a sequence of simple arguments corresponding to the components of the tuple. This process is called currying. It is named after American logician Haskell B. Curry, who first exploited the technique.

Function add’ above is similar to the function (+) from the Prelude (i.e., the addition operator).

We sometimes speak of the function (+) as being partially applied in the expression ((+) 3). In this expression, the first argument of the function is "frozen in" and the resulting function can be passed as an argument, returned as a result, or applied to another argument.

Partially applied functions are very useful in conjunction with other higher-order functions.

For example, consider the partial applications of the relational comparison operator (<) and multiplication operator (*) in the function doublePos3. This function, which is equivalent to the function doublePos discussed in an earlier subsection, doubles the positive integers in a list:

    doublePos3 :: [Int] -> [Int] 
    doublePos3 xs = map' ((*) 2) (filter' ((<) 0) xs) 

Related to the concept of currying is the property of extensionality. Two functions f and g are extensionally equal if f x = g x for all x.

Thus instead of writing the definition of g as

    f, g :: a -> a
    f x = some_expression

    g x = f x

we can write the definition of g as simply:

    g = f

5.3.3 Operator sections

Expressions such as ((*) 2) and ((<) 0), used in the definition of doublePos3 in the previous subsection, can be a bit confusing because we normally use these operators in infix form. (In particular, it is difficult to remember that ((<) 0) returns True for positive integers.)

Also, it would be helpful to be able to use the division operator to express a function that halves (i.e., divides by two) its operand. The function ((/) 2) does not do it; it divides 2 by its operand.

We can use the function flip from the Prelude to state the halving operation. Function flip takes a function and two additional arguments and applies the argument function to the two arguments with their order reversed.

    flip' :: (a -> b -> c) -> b -> a -> c  -- flip in Prelude
    flip' f x y = f y x

Thus we can express the halving operator with the expression (flip (/) 2).

Because expressions such as ((<) 0) and (flip (/) 2) are quite common in programs, Haskell provides a special, more compact and less confusing, syntax.

For some infix operator $\oplus$ and arbitrary expression e, expressions of the form (e$\oplus$) and ($\oplus$e) represent (($\oplus$) e) and (flip ($\oplus$) e), respectively. Expressions of this form are called operator sections.

Examples of operator sections include:

(1+) is the successor function, which returns the value of its argument plus 1.

(0<) is a test for a positive integer.

(/2) is the halving function.

(1.0/) is the reciprocal function.

(:[]) is the function that returns the singleton list containing the argument.

Suppose we want to sum the cubes of list of integers. We can express the function in the following way:

    sumCubes :: [Int] -> Int 
    sumCubes xs = sum' (map' (^3) xs)

Above ^ is the exponentiation operator and sum is the list summation function defined in the Prelude as:

    sum' = foldl' (+) 0  -- sum

5.3.4 Combinators

The function flip in the previous subsection is an example of a useful type of function called a combinator.

A combinator is a function without any free variables. That is, on right side of a defining equation there are no variables or operator symbols that are not bound on the left side of the equation.

For historical reasons, flip is sometimes called the C combinator.

There are several other useful combinators in the Prelude.

The combinator const (shown below as const’) is the constant function constructor; it is a two-argument function that returns its first argument. For historical reasons, this combinator is sometimes called the K combinator.

    const' :: a -> b -> a  -- const in Prelude
    const' k x = k

Example: (const 1) takes any argument and returns the value 1.

Question: What does sum (map (const 1) xs) do?

Function id (shown below as id’) is the identity function; it is a one-argument function that returns its argument unmodified. For historical reasons, this function is sometimes called the I combinator.

    id' :: a -> a  -- id in Prelude
    id' x = x

Combinators fst and snd (shown below as fst’ and snd’) extract the first and second components, respectively, of 2-tuples.

    fst' :: (a,b) -> a  -- fst in Prelude
    fst' (x,_) = x 

    snd' :: (a,b) -> b  -- snd in Prelude
    snd' (_,y) = y

Similarly, fst3, snd3, and thd3 extract the first, second, and third components, respectively, of 3-tuples.

An interesting example that uses a combinator is the function reverse as defined in the Prelude (shown below as reverse’):

    reverse' :: [a] -> [a]          -- reverse in Prelude
    reverse' = foldlX (flip' (:)) []

Function flip (:) takes a list on the left and an element on the right. As this operation is folded through the list from the left it attaches each element as the new head of the list.

We can also define combinators that convert an uncurried function into a curried function and vice versa. The functions curry' and uncurry' defined below are similar to the Prelude functions.

    curry' :: ((a, b) -> c) -> a -> b -> c -- curry in Prelude 
    curry' f x y =  f (x, y) 

    uncurry' :: (a -> b -> c) -> ((a, b) -> c) -- uncurry in Prelude 
    uncurry' f p =  f (fst p) (snd p) 

Two other useful combinators are fork and cross. Combinator fork applies each component of a pair of functions to a value to create a pair of results. Combinator cross applies each component of a pair of functions to the corresponding components of a pair of values to create a pair of results. We can define these as follows:

    fork :: (a -> b, a -> c) -> a -> (b,c)
    fork (f,g) x = (f x, g x)

    cross :: (a -> b, c -> d) -> (a,c) -> (b,d)
    cross (f,g) (x,y) = (f x, g y)

5.3.5 Functional composition

The functional composition operator allows several "smaller" functions to be combined to form “larger” functions. In Haskell, this combinator is denoted by the period (.) symbol and is defined in the Prelude as follows:

    infixr 9 . 
    (.) :: (b -> c) -> (a -> b) -> (a -> c) 
    (f . g) x = f (g x) 

Composition’s default binding is from the right and its precedence is higher than all the operators we have discussed so far except function application itself.

Functional composition is an associative binary operation with the identity function id as its identity element:

     f . (g . h) = (f . g) . h 
     id . f  = f . id

An advantage of the use of functional composition is that some expressions can be written more concisely. For example, the function

    doit x = f1 (f2 (f3 (f4 x))) 

can be written more concisely as:

    doit = f1 . f2 . f3 . f4 

This latter definition is in what is called the point-free style. It defines function doit as being equal to the composition of the other functions. It leaves the parameters (i.e., the points) of doit as implicit; doit has the same parameters as the composition.

The former definition (that gives all the explicit parameters) is in what is called pointful style.

If doit is defined in point-free style as above, then we can use the function in expressions such as map (doit) xs. Of course, if this is the only use of the doit function, we can eliminate it completely and use the composition directly, e.g., map (f1 . f2 . f3 . f4) xs.

As an example, consider the function count that takes two arguments, an integer n and a list of lists, and returns the number of the lists from the second argument that are of length n. Note that all functions composed below are single-argument functions: length, (filter (== n)), (map length).

    count :: Int -> [[a]] -> Int 
    count n 
        | n >= 0    = length' . filter' (== n) . map' length' 
        | otherwise = const' 0   -- discard 2nd arg, return 0 

We can think of the point-free expression length' . filter' (== n) . map' length' as defining a function pipeline through which data flows from right to left.

  1. The pipeline takes a polymorphic list of lists as input.

  2. The map' length' component of the pipeline replaces each inner list by its length.

  3. The filter' (== n) component takes the list created by the previous step and removes all elements not equal to n.

  4. The length' component takes the list created by the previous step and determines how many elements are remaining.

  5. The pipeline outputs the value computed by the previous component. The number of lists within the input list of lists that are of length n.

Thus composition is a powerful form of "glue" that can be used to "stick" simpler functions together to build more powerful functions. The simpler functions in this case include partial applications of higher order functions from the library we have developed.

As we see above in the definition of count, partial applications (e.g., filter (== n)), operator sections (e.g., (== n)), and combinators (e.g., const) are useful as plumbing the function pipeline.

Remember the function doublePos that we discussed in earlier subsections.

    doublePos3 xs = map' ((*) 2) (filter' ((<) 0) xs) 

Using composition, partial application, and operator sections we can restate its definition in point-free style as follows:

    doublePos4 :: [Int] -> [Int] 
    doublePos4 = map' (2*) . filter' (0<)

Consider a function last to return the last element in a non-nil list and a function init to return the initial segment of a non-nil list (i.e., everything except the last element). These could quickly and concisely be written as follows:

    last' = head . reverse'            -- last in Prelude
    init' = reverse' . tail . reverse' -- init in Prelude

However, since these definitions are not very efficient, the Prelude implements functions last and init in a more direct and efficient way similar to the following:

    last2 :: [a] -> a    -- last in Prelude
    last2 [x]    = x 
    last2 (_:xs) = last2 xs 

    init2 :: [a] -> [a]  -- init in Prelude
    init2 [x]    = [] 
    init2 (x:xs) = x : init2 xs 

The definitions for Prelude functions any and all are similar to the definitions show below; they take a predicate and a list and apply the predicate to each element of the list, returning True when any and all, respectively, of the individual tests evaluate to True.

    any', all' :: (a -> Bool) -> [a] -> Bool 
    any' p = or' . map' p   -- any in Prelude
    all' p = and' . map' p  -- all in Prelude

The functions elem and notElem test for an object being an element of a list and not an element, respectively. They are defined in the Prelude similarly to the following:

    elem', notElem' :: Eq a => a -> [a] -> Bool 
    elem'    = any' . (==)  -- elem in Prelude
    notElem' = all' . (/=)  -- notElem in Prelude

These are a bit more difficult to understand since any, all, ==, and /= are two-argument functions. Note that expression elem x xs would be evaluated as follows:

elem’ x xs
$\Longrightarrow$ { expand elem’ }
(any’ . (==)) x xs
$\Longrightarrow$ { expand composition }
any’ ((==) x) xs

The composition operator binds the first argument with (==) to construct the first argument to any’. The second argument of any’ is the second argument of elem’.

5.3.6 Lambda expressions

Remember the function squareAll2 we examined in an earlier subsection on maps:

    squareAll2 :: [Int] -> [Int] 
    squareAll2 xs = map' sq xs 
                    where sq x = x * x 

We introduced the local function definition sq to denote the function to be passed to map. It seems to be a waste of effort to introduce a new symbol for a simple function that is only used in one place in an expression. Would it not be better, somehow, to just give the defining expression itself in the argument position?

Haskell provides a mechanism to do just that, an anonymous function definition. For historical reasons, these nameless functions are called lambda expressions. They begin with a backslash \ and have the syntax:

\ atomicPatterns -> expression

For example, the squaring function (sq) could be replaced by a lambda expression as (\x -> x * x). The pattern x represents the single argument for this anonymous function and the expression x * x is its result.

Thus we can rewrite squareAll2 in point-free style using a lambda expression as follows:

    squareAll3 :: [Int] -> [Int] 
    squareAll3 = map' (\x -> x * x)

A lambda expression to average two numbers can be written (\x y -> (x+y)/2).

An interesting example that uses a lambda expression is the function length as defined in the Prelude--similar to length4 below. (Note that this uses the optimized function foldl' from the standard Haskell Data.List module.)

    length4 :: [a] -> Int   -- length in Prelude
    length4  = foldl' (\n _ -> n+1) 0

The anonymous function (\n _ -> n+1) takes an integer "counter" and a polymorphic value and returns the "counter" incremented by one. As this function is folded through the list from the left, this function counts each element of the second argument.

5.3.7 Application operator $

In Haskell, function application associates to the left and has higher binding power than any infix operator. For example, for some function two-argument function f and values w, x, y, and z

    w + f x y * z

is the same as

    w + (((f x) y) * z)

given the relative binding powers of function application and the numeric operators.

However, sometimes we want to be able to use function application where it associates to the right and binds less tightly than any other operator. Haskell defines the $ operator to enable this style, as follows:

    infixr 0 $
    ($) :: (a -> b) -> a -> b 
    f $ x = f x

Thus, for single argument functions f, g, and h,

    f $ g $ h $ z + 7

is the same as

    (f (g (h (z+7))))

and as:

    (f . g . h) (z+7)

Similarly, for two-argument functions f', g', and h',

    f' w $ g' x $ h' y $ z + 7

is the same as

    ((f' w) ((g' x) ((h' y) (z+7))))

and as:

    (f' w . g' x . h' y) (z+7)

For example, this operator allows us to write

    foldr (+) 0 $ map (2*) $ filter odd $ enumFromTo 1 20

where Prelude function enumFromTo m n generates the sequence of integers from m to n, inclusive.

5.3.8 Eager evaluation using seq and $!

Haskell is a lazily evaluated language. That is, if an argument is nonstrict it may never be evaluated.

Sometimes, using the technique called strictness analysis, the Haskell compiler can detect that an argument's value will always be needed. The compiler can then safely force eager evaluation as an optimization without changing the meaning of the program.

In particular, by selecting the -O option to the Glasgow Haskell Compiler (GHC), we can enable GHC's code optimization processing. GHC will generally create smaller, faster object code at the expense of increased compilation time by taking advantage of strictness analysis and other optimizations.

However, sometimes we may want to force eager evaluation explicitly without invoking a full optimization on all the code (e.g., to make a particular function's evaluation more space efficient). Haskell provides the primitive function seq that enables this. That is,

    seq :: a -> b -> b
    x `seq` y = y

where it just returns the second argument except that, as a side effect, x is evaluated before y is returned. (Technically, x is evaluated to what is called head normal form. It is evaluated until the outer layer of structure such as h:t is revealed, but h and t themselves are not fully evaluated. We examine evaluation in more detail in a later chapter.)

Function foldl, the "optimized" version of foldl can be defined using seq as follows

    foldlP :: (a -> b -> a) -> a -> [b] -> a  -- foldl' in Data.List 
    foldlP f z []     = z 
    foldlP f z (x:xs) = y `seq` foldl' f y xs 
                        where y = f z x

That is, this evaluates the z argument of the tail recursive application eagerly.

Using seq, Haskell also defines $!, a strict version of the $ operator, as follows:

    infixr 0 $! 
    ($!) :: (a -> b) -> a -> b 
    f $! x = x `seq` f x 

The effect of f $! x is the same as f $ x except that $! eagerly evaluates the argument x before applying function f to it.

We can rewrite foldl' using $! as follows:

    foldlQ :: (a -> b -> a) -> a -> [b] -> a  -- foldl' in Data.List 
    foldlQ f z []     = z 
    foldlQ f z (x:xs) = (foldlQ f $! f z x) xs

We can write a tail recursive function to sum the elements of the list as follows:

    sum4 :: [Integer] -> Integer  -- sum in Prelude
    sum4 xs = sumIter xs 0
        where sumIter []     acc = acc
              sumIter (x:xs) acc = sumIter xs (acc+x)

We can then redefine sum4 to force eager evaluation of the accumulating parameter of sumIter as follows:

    sum5 :: [Integer] -> Integer -- sum in Prelude
    sum5 xs = sumIter xs 0
        where sumIter []     acc = acc
              sumIter (x:xs) acc = sumIter xs $! acc + x

However, we need to be careful in applying seq and $!. They change the semantics of the lazily evaluated language in the case where the argument is nonstrict. They may force a program to terminate abnormally and/or cause it to take unnecessary evaluation steps.

5.4 Additional Higher-Order List Functions

5.4.1 List-breaking operations

In a previous chapter we looked at the list-breaking functions take and drop. The Prelude also includes several higher-order list-breaking functions that take two arguments, a predicate that determines where the list is to be broken and the list to be broken.

Here we look at Prelude functions takeWhile and dropWhile. As you would expect, function takeWhile "takes" elements from the beginning of the list "while" the elements satisfy the predicate and dropWhile "drops" elements from the beginning of the list "while" the elements satisfy the predicate. The Prelude definitions are similar to the following:

    takeWhile':: (a -> Bool) -> [a] -> [a] -- takeWhile in Prelude
    takeWhile'  p []  = [] 
    takeWhile'  p (x:xs) 
        | p x        = x : takeWhile' p xs 
        | otherwise  = [] 

    dropWhile' :: (a -> Bool) -> [a] -> [a] -- dropWhile in Prelude
    dropWhile'  p []  = [] 
    dropWhile'  p xs@(x:xs') 
        | p x        = dropWhile' p xs' 
        | otherwise  = xs

Note the use of the pattern xs@(x:xs’) in dropWhile'. This pattern matches a non-nil list with x and xs’ binding to the head and tail, respectively, as usual. Variable xs binds to the entire list.

As an example, suppose we want to remove the leading blanks from a string. We can do that with the expression:

    dropWhile ((==) ' ')

As with take and drop, the above functions can also be related by a "law". For all finite lists xs and predicates p on the same type:

    takeWhile p xs ++ dropWhile p xs = xs

Prelude function span combines the functionality of takeWhile and dropWhile into one function. It takes a predicate p and a list xs and returns a tuple where the first element is the longest prefix (possibly empty) of xs that satisfies p and the second element is the remainder of the list.

    span' :: (a -> Bool) -> [a] -> ([a],[a]) -- span in Prelude
    span' _ xs@[]      =  (xs, xs)
    span' p xs@(x:xs')
        | p x          =  let (ys,zs) = span' p xs' in (x:ys,zs)
        | otherwise    =  ([],xs)

Thus the following "law" holds for all finite lists xs and predicates p on same type:

    span p xs == (takeWhile p xs, dropWhile p xs)

The Prelude also includes the function break, defined as follows:

    break' :: (a -> Bool) -> [a] -> ([a],[a]) -- break in Prelude
    break' p =  span (not . p)

5.4.2 List-combining operations

In a previous chapter, we also looked at the function zip, which takes two lists and returns a list of pairs of the corresponding elements. Function zip applies an operation, in this case tuple-construction, to the corresponding elements of two lists.

We can generalize this pattern of computation with the function zipWith in which the operation is an argument to the function.

    zipWith' :: (a->b->c) -> [a]->[b]->[c] -- zipWith in Prelude
    zipWith' z  (x:xs)  (y:ys) = z x y : zipWith' z xs ys 
    zipWith' _  _     _        = []

Using a lambda expression to state the tuple-forming operation, the Prelude defines zip in terms of zipWith:

    zip'' :: [a] -> [b] -> [(a,b)]  -- zip
    zip'' = zipWith' (\x y -> (x,y)) 

Or can be written more simply as:

    zip''' :: [a] -> [b] -> [(a,b)]  -- zip
    zip''' = zipWith' (,) 

The zipWith function also enables us to define operations such as the scalar product of two vectors in a concise way.

    sp :: Num a => [a] -> [a] -> a 
    sp xs ys = sum' (zipWith' (*) xs ys) 

The Prelude includes versions of zipWith that take up to seven input lists: zipWith3 $\cdots$ zipWith7.

5.5 Rational Arithmetic Revisited

Remember the rational number arithmetic package developed in an earlier chapter. In that package's Rational module, we defined a function eqRat to compare two rational numbers for equality using the appropriate set of integer comparisons.

    eqRat :: Rat -> Rat -> Bool 
    eqRat x y = (numer x) * (denom y) == (numer y) * (denom x)

We could have implemented the other comparison operations similarly.

Because the comparison operations are similar, they are good candidates for us to use a higher-order function. We can abstract out the common pattern of comparisons into a function that takes the corresponding integer comparison as an argument.

To compare two rational numbers, we can express their values in terms of a common denominator (e.g., denom x * denom y) and then compare the numerators using the integer comparisons. We can thus abstract the comparison into a higher-order function compareRat that takes an appropriate integer relational operator and the two rational numbers.

    compareRat :: (Int -> Int -> Bool) -> Rat -> Rat -> Bool 
    compareRat r x y = r (numer x * denom y) (denom x * numer y) 

Then we can define the rational number comparisons in terms of compareRat. (Note that we redefine function eqRat from the package in the earlier chapter.)

    eqRat,neqRat,ltRat,leqRat,gtRat,geqRat :: Rat -> Rat -> Bool 
    eqRat   = compareRat (==) 
    neqRat  = compareRat (/=) 
    ltRat   = compareRat (<)
    leqRat  = compareRat (<=) 
    gtRat   = compareRat (>)
    geqRat  = compareRat (>=) 

5.6 Merge Sort

We defined the insertion sort in a previous chapter. It has an average-case time complexity of O($n^{2}$) where $n$ is the length of the input list.

We now consider a more efficient function to sort the elements of a list into ascending order: merge sort. Merge sort works as follows:

We define function msort to be a polymorphic, higher-order function that has two parameters. The first (less) is the comparison operator and the second (xs) is the list to be sorted. Function less must be defined for every element that appears in the list to be sorted.

    msort :: Ord a => (a -> a -> Bool) -> [a] -> [a]
    msort _    [] = []
    msort less xs = merge (msort less ls) (msort less rs)
        where  n       = (length xs) `div` 2 
               (ls,rs) = splitAt n xs
               merge [] ys     = ys
               merge xs []     = x
               merge ls@(x:xs) rs@(y:ys) 
                   | less x y  = x : merge xs rs
                   | otherwise = y : merge ls ys

By nesting the definition of merge, we enabled it to directly access the the parameters of msort. In particular, we did not need to pass the comparison function to merge.

Assuming that less evaluates in constant time, the time complexity of msort is O($n \log(n)$), where $n$ is the length of the input list.

So the total cost is O($n \log(n)$). The cost is independent of distribution of elements in the original list.

We can apply msort as follows:

    msort (<) [5, 7, 1, 3]

Function msort is defined in curried form with the comparison function first. This enables us to conveniently specialize msort with a specific comparison function. For example,

    descendSort :: Ord a => [a] -> [a]
    descendSort = msort (\ x y -> x > y)

5.7 Exercises

  1. Suppose you need a Haskell function times that takes a list of integers (type Integer) and returns the product of the elements (e.g, times [2,3,4] returns 24). Define the following Haskell functions.

    1. Function times1 that uses the Prelude function foldr(or foldr' from this chapter).

    2. Function times2 that uses backward recursion to compute the product. (Use recursion directly. Do not use the list-folding Prelude functions such as foldr or product.)

    3. Function times3 that uses forward recursion to compute the product. (Hint: use a tail-recursive auxiliary function with an accumulating parameter.)

    4. Function times4 that uses function foldl' from the Haskell library Data.List.

  2. For each of the following specifications, define a Haskell function that has the given arguments and result. Use the higher order library functions (from this chapter) such as map, filter, foldr, and foldl as appropriate.

    1. Function numof takes a value and a list and returns the number of occurrences of the value in the list.

    2. Function ellen takes a list of character strings and returns a list of the lengths of the corresponding strings.

    3. Function ssp takes a list of integers and returns the sum of the squares of the positive elements of the list.

  3. Suppose you need a Haskell function sumSqNeg that takes a list of integers (type Integer) and returns the sum of the squares of the negative values in the list.

    Define the following Haskell functions. Use the higher order library functions (from this chapter) such as map, filter, foldr, and foldl as appropriate.

    1. Function sumSqNeg1 that is backward recursive. (Use recursion directly. Do not use the list-folding Prelude functions such as foldr or sum.)

    2. Function sumSqNeg2 that is tail recursive. (Use recursion directly. Do not use the list-folding Prelude functions such as foldr or sum.)

    3. Function sumSqNeg3 that uses standard prelude functions such as map, filter, foldr, and foldl.

    4. Function sumSqNeg4 that uses list comprehensions (Chapter 7).

  4. Define a Haskell function

        total :: (Integer -> Integer) -> Integer -> Integer

    so that total f n gives f 0 + f 1 + f 2 + ... + f n.

  5. Define a Haskell function

        removeFirst :: (a -> Bool) -> [a] -> [a} 

    so that removeFirst p xs removes the first element of xs that does has the property p.

  6. Define a Haskell function

        removeLast :: (a -> Bool) -> [a] -> [a]

    so that removeLast p xs removes the last occurrence of element of xs that has the property p.

    How could you define it using removeFirst?

  7. Define a Haskell function map2 that takes a list of functions and a list of values and returns the list of results of applying each function in the first list to the corresponding value in the second list.

  8. Define a Haskell function fmap that takes a value and a list of functions and returns the list of results from applying each function to the argument value. (For example, fmap 3 [((*) 2), ((+) 2)] yields [6,5].)

  9. Define a Haskell function composeList that takes a list of functions and composes them into a single function. (Be sure to give the type signature.)

  10. A list s is a prefix of a list t if there is some list u (perhaps nil) such that s ++ u == t. For example, the prefixes of string abc are `,a,ab,abc`.

    A list s is a suffix of a list t if there is some list u (perhaps nil) such that u ++ s == t. For example, the suffixes of abc are abc, bc, c, and ``.

    A list s is a segment of a list t if there are some (perhaps nil) lists u and v such that . For example, the segments of string abc consist of the prefixes and the suffixes plus b.

    Define the following Haskell functions. You may use functions appearing early in the list to implement later ones.

    1. Define a function prefix such that prefix xs ys returns True if xs is a prefix of ys and returns False otherwise.

    2. Define a function suffixes such that suffixes xs returns the list of all suffixes of list xs. (Hint: Generate them in the order given in the example of abc above.)

    3. Define a function indexes such that indexes xs ys returns a list of all the positions at which list xs appears in list ys. Consider the first character of ys as being at position 0. For example, indexes ab abaabbab returns [1,4,7]. (Hint: Remember functions like map, filter, zip, and the functions you just defined?)

    4. Define a function sublist such that sublist xs ys returns True if list xs appears as a segment of list ys and returns False otherwise.

  11. Assume that the following Haskell type synonyms have been defined:

        type Word = String -- word, characters left-to-right
        type Line = [Word] -- line, words left-to-right
        type Page = [Line] -- page, lines top-to-bottom
        type Doc  = [Page] -- document, pages front-to-back

    Further assume that values of type Word do not contain any space characters. Implement the following Haskell text-handling functions:

    1. npages that takes a Doc and returns the number of Pages in the document.

    2. nlines that takes a Doc and returns the number of Lines in the document.

    3. nwords that takes a Doc and returns the number of Words in the document.

    4. nchars that takes a Doc and returns the number of Chars in the document (not including spaces of course).

    5. deblank that takes a Doc and returns the Doc with all blank lines removed. A blank line is a line that contains no words.

    6. linetext that takes a Line and returns the line as a String with the words appended together in left-to-right order separated by space characters and with a newline character ’\n’ appended to the right end of the line. (For example, linetext [Robert, Khayat] yields "Robert Khayat\n".)

    7. pagetext that takes a Page and returns the page as a String--applies linetext to its component lines and appends the result in a top-to-bottom order.

    8. doctext that takes a Doc and returns the document as a String--applies pagetext to its component lines and appends the result in a top-to-bottom order.

    9. wordeq that takes a two Docs and returns True if the two documents are word equivalent and False otherwise. Two documents are word equivalent if they contain exactly the same words in exactly the same order regardless of page and line structure. For example, [[["Robert"],["Khayat"]]] is word equivalent to [[["Robert","Khayat"]]].

5.8 References

[Abelson-Sussman 1996]
Harold Abelson and Gerald Jay Sussman. Structure and Interpretation of Computer Programs (SICP), Second Edition, MIT Press, 1996.
[Bird 2015]
Richard Bird. Thinking Functionally with Haskell Cambridge University Press, 2015.
[Bird-Wadler 1998]
Richard Bird and Philip Wadler. Introduction to Functional Programming, Second Edition, Addison Wesley, 1998. [First Edition, 1988]
[Chiusano-Bjarnason 2015]
Paul Chiusano and Runar Bjarnason, Functional Programming in Scala, Manning, 2015.
[Cunningham 2014]
H. Conrad Cunningham. Notes on Functional Programming with Haskell, 1994-2014.
[Thompson 2011]
Simon Thompon. Haskell: The Craft of Programming, Third Edition, Pearson, 2011.
[Wikipedia 2017]
Wikipedia articles on "Polymorphism", "Ad Hoc Polymorphism", "Parametric Polymorphism", "Subtyping", and "Function Overloading", 2017.

5.9 Terms and Concepts

Procedural abstraction, functions (first-class, higher-order), modularity, interface, function generalization and specialization, scope-commonality-variability (SCV) analysis, hot and frozen spots, think like a functional programmer, common functional programming patterns (map, filter, fold, concatMap), duality theorem, strict and nonstrict functions, bottom, strictness analysis, currying, partial application, operator sections, combinators, functional composition, property of extensionality, pointful and point-free styles, plumbing, function pipeline, lambda expression, application operator $, eager evaluation operators seq and $!, head-normal form, merge sort