2 August 2018
Browser Advisory: The HTML version of this textbook requires a browser that supports the display of MathML. A good choice as of August 2018 is a recent version of Firefox from Mozilla.
The goal of this chapter is to show how to state and prove Haskell “laws”.
This chapter depends upon the reader understanding Haskell’s polymorphic, higher-order list programming concepts (e.g. from Chapters 4-5, 8-9, and 13-17), but it is otherwise independent of other preceding chapters.
The chapter provides useful tools that can be used in stating and formally proving function and module contracts (Chapters 6 and 7) and type class laws (Chapter 22). It supports reasoning about program generalization (Chapter 19) and type inference (Chapter 24).
The following two chapters on program synthesis (Chapters 26 and 27) build on the concepts and techniques introduced by this chapter.
Referential transparency is probably the most important property of purely functional programming languages like Haskell.
Chapter 2 defines referential transparency to mean that, within some well-defined context, a variable (or other symbol) always represents the same value. This allows one expression to be replaced by an equivalent expression or, more informally, “equals to be replaced by equals”.
Chapter 8 shows how referential transparency underpins the evaluation (i.e. substitution or reduction) model for Haskell and similar functional languages.
In this chapter, we see that referential transparency allows us to state and prove various “laws” or identities that hold for functions and to use these “laws” to transform programs into equivalent ones. Referential transparency underlies how we reason about Haskell programs.
As a purely functional programming language, Haskell supports mathematical reasoning mostly within the programming language itself. We can state properties of functions and prove them using a primarily equational, or calculational, style of proof. The proof style is similar to that of high school trigonometric identities.
++
associativity and identity elementWe have already seen a number of these laws. Again consider the append operator (++
) for finite lists from Chapter 14.
infixr 5 ++
(++) :: [a] -> [a] -> [a]
[] ++ xs = xs -- append.1
(x:xs) ++ ys = x:(xs ++ ys) -- append.2
The append operator ++
has two useful properties that we have already seen.
xs
, ys
, and zs
,
xs ++ (ys ++ zs) = (xs ++ ys) ++ zs
.
xs
,
[] ++ xs = xs = xs ++ []
.
Note: The above means that the append operator ++
and the set of finite lists form the algebraic structure called a monoid.
How do we prove these properties?
The answer is, of course, induction. But we need a type of induction that allows us to prove theorems over the set of all finite lists. In fact, we have already been using this form of induction in the informal arguments that the list-processing functions terminate.
Induction over the natural numbers is a special case of a more general form of induction called structural induction. This type of induction is over the syntactic structure of recursively (inductively) defined objects. Such objects can be partially ordered by a complexity ordering from the most simple (minimal) to the more complex.
If we think about the usual axiomization of the natural numbers (i.e. Peano’s postulates), then we see that 0
is the only simple (minimal) object and that the successor function ((+) 1)
is the only constructor.
In the case of finite lists, the only simple object is the nil list []
and the only constructor is the cons operator (:)
.
To prove a proposition P(x)
holds for any finite object x
, one must prove the following cases.
P(e)
holds for each simple (minimal) object e
.
C
, if P(x)
holds for some arbitrary object(s) x
, then P(C(x))
also holds.
That is, we can assume P(x)
holds, then prove that P(C(x))
holds. This shows that the constructors preserve proposition P
.
To prove a proposition P(xs)
holds for any finite list xs
, the above reduces to the following cases.
xs = []
:P([])
holds.xs = (a:as)
.P(as)
holds, then P(a:as)
also holds.
One, often useful, strategy for discovering proofs of laws is the following:
Determine whether induction is needed to prove the law. Some laws can be proved directly from the definitions and other previously proved laws.
Carefully choose the induction variable (or variables).
Identify the base and inductive cases.
For each case, use simplification independently on each side of the equation. Often, it is best to start with the side that is the most complex.
Simplification means to substitute the right-hand side of a definition or the induction hypothesis for some expression matching the left-hand side.
Continue simplifying each expression as long as possible.
Often we can show that the two sides of an equation are the same or that simple manipulations (perhaps using previously proved laws) will show that they are the same.
If necessary, identify subcases and prove each subcase independently.
A formal proof of a case should, in general, be shown as a calculation that transforms one side of the equation into the other by substitution of equals for equals.
This formal proof can be constructed from the calculation suggested in the above
++
Now that we have the mathematical machinery we need, let’s prove that ++
is associative for all finite lists. The following proofs assume that all arguments of the functions are defined.
xs
, ys
, and zs
,
xs ++ (ys ++ zs) = (xs ++ ys) ++ zs
.
Proof:
There does not seem to be a non-inductive proof, thus we proceed by structural induction over the finite lists. But on which variable(s)?
By examining the definition of ++
, we see that it has two legs differentiated by the value of the left operand. The right operand is not decomposed. To use this definition in the proof, we need to consider the left operands of the ++
in the associative law.
Thus we choose to do the induction on xs
, the leftmost operand, and consider two cases.
Base case xs = []
:
First, we simplify the left-hand side.
[] ++ (ys ++ zs)
{ append.1
(left to right), omit outer parentheses }
ys ++ zs
We do not know anything about ys
and zs
, so we cannot simplify further.
Next, we simplify the right-hand side.
([] ++ ys) ++ zs
{ append.1
(left to right), omit parentheses around ys
}
ys ++ zs
Thus we have simplified the two sides to the same expression.
Of course, a formal proof can be written more elegantly as:
[] ++ (ys ++ zs)
{ append.1
(left to right) }
ys ++ zs
{ append.1
(right to left, applied to left operand) }
([] ++ ys) ++ zs
Thus the base case is established.
Note the equational style of reasoning. We proved that one expression was equal to another by beginning with one of the expressions and repeatedly substituting “equals for equals” until we got the other expression.
Each transformational step was justified by a definition, a known property, or (as we see later) the induction hypothesis. We normally do not state justifications like “omit parentheses” or “insert parentheses”.
Inductive case xs = (a:as)
:
Assume as ++ (ys ++ zs) = (as ++ ys) ++ zs
;
prove (a:as) ++ (ys ++ zs) = ((a:as) ++ ys) ++ zs
.
First, we simplify the left-hand side.
(a:as) ++ (ys ++ zs)
{ append.2
(left to right) }
a:(as ++ (ys ++ zs))
{ induction hypothesis }
a:((as ++ ys) ++ zs)
We do not know anything further about as
, ys
, and zs
, so we cannot simplify further.
Next, we simplify the right-hand side.
((a:as) ++ ys) ++ zs
{ append.2
(left to right, on inner ++
) }
(a:(as ++ ys)) ++ zs
{ append.2
(left to right, on outer ++
) }
a:((as ++ ys) ++ zs)
Thus we have simplified the two sides to the same expression.
Again, a formal proof can be written more elegantly as follows.
(a:as) ++ (ys ++ zs)
{ append.2
(left to right) }
a:(as ++ (ys ++ zs))
{ induction hypothesis }
a:((as ++ ys) ++ zs)
{ append.2
(right to left, on outer ++
) }
(a:(as ++ ys)) ++ zs
{ append.2
(right to left, on inner ++
) }
((a:as) ++ ys) ++ zs
Thus the inductive case is established.
Therefore, we have proven the ++
associativity property. Q.E.D.
Note: The above proof and the ones that follow assume that the arguments of the functions are all defined (i.e. not equal to ).
You should practice writing proofs in the “more elegant” form given above. This end-to-end calculational style is more useful for synthesis of programs.
Reviewing what we have done, we can identify the following guidelines:
Determine whether induction is really needed.
Choose the induction variable carefully.
Be careful with parentheses.
Substitutions, comparisons, and pattern matches must done with the fully parenthesized forms of definitions, laws, and expressions in mind, that is, with parentheses around all binary operations, simple objects, and the entire expression. We often omit “unneeded” parentheses to make the expression more readable.
Start with the more complex side of the equation.
That gives us more information with which to work.
++
Now let’s prove the identity property.
xs
,
[] ++ xs = xs = xs ++ []
.
Proof:
The equation [] ++ xs = xs
follows directly from append.1
. Thus we consider the equation xs ++ [] = xs
, which we prove by structural induction on xs
.
Base case xs = []
:
[] ++ []
{ append.1
(left to right) }
[]
This establishes the base case.
Inductive case xs = (a:as)
:
Assume as ++ [] = as
; prove (a:as) ++ [] = (a:as)
.
(a:as) ++ []
{ append.2
(left to right) }
a:(as ++ [])
{ induction hypothesis }
a:as
This establishes the inductive case.
Therefore, we have proved that []
is the identity element for ++
. Q.E.D.
length
and ++
Suppose that the list length
function is defined as follows (from Chapter 13).
xs
and ys
:
length (xs++ys) = length xs + length ys
.
Proof:
Because of the way ++
is defined, we choose xs
as the induction variable.
Base case xs = []
:
length [] + length ys
{ length.1
(left to right) }
0 + length ys
{ 0 is identity for addition }
length ys
{ append.1
(right to left) }
length ([] ++ ys)
This establishes the base case.
Inductive case xs = (a:as)
:
Assume length (as ++ ys) = length as + length ys
;
prove length ((a:as) ++ ys) = length (a:as) + length ys
.
length ((a:as) ++ ys)
{ append.2
(left to right) }
length (a:(as ++ ys)) }
{ length.2
(left to right) }
1 + length (as ++ ys)
{ induction hypothesis }
1 + (length as + length ys)
{ associativity of addition }
(1 + length as) + length ys
{ length.2
(right to left, value of a
arbitrary) }
length (a:as) + length ys
This establishes the inductive case.
Therefore, length (xs ++ ys) = length xs + length ys
. Q.E.D.
Note: The proof uses the associativity and identity properties of integer addition.
take
and drop
Remember the definitions for the list functions take
and drop
from Chapter 12.
take :: Int -> [a] -> [a]
take n _ | n <= 0 = [] -- take.1
take _ [] = [] -- take.2
take n (x:xs) = x : take (n-1) xs -- take.3
drop :: Int -> [a] -> [a]
drop n xs | n <= 0 = xs -- drop.1
drop _ [] = [] -- drop.2
drop n (_:xs) = drop (n-1) xs -- drop.3
n
and finite lists xs
,
take n xs ++ drop n xs = xs
.
Proof:
Note that both take
and drop
use both arguments to distinguish the cases. Thus we must do an induction over all natural numbers n
and all finite lists xs
.
We would expect four cases to consider, the combinations from n
being zero and nonzero and xs
being nil and non-nil. But an examination of the definitions for the functions reveal that the cases for n = 0
collapse into a single case.
Base case n = 0
:
take 0 xs ++ drop 0 xs
{ take.1, drop.1
(both left to right) }
[] ++ xs
{ ++
identity xs
}
xs
This establishes the case.
Base case n = m+1
, xs = []
:
take (m+1) [] ++ drop (m+1) []
{ take.2, drop.2
(both left to right) }
[] ++ []
{ ++
identity }
[]
This establishes the case.
Inductive case n = m+1
, xs = (a:as)
:
Assume take m as ++ drop m as = as
;
prove take (m+1) (a:as) ++ drop (m+1) (a:as) = (a:as)
.
take (m+1) (a:as) ++ drop (m+1) (a:as)
{ take.3, drop.3
(both left to right) }
(a:(take m as)) ++ drop m as
{ append.2
(left to right) }
a:(take m as ++ drop m as)
{ induction hypothesis }
(a:as)
This establishes the case.
Therefore, the property is proved. Q.E.D.
What do we mean when we say two functions are equivalent?
Usually, we mean that the “same inputs” yield the “same outputs”. For example, single argument functions f
and g
are equivalent if f x
g x
for all x
.
In Chapter 14 we defined two versions of a function to reverse the elements of a list. Function rev
uses backward recursion and function reverse
(called reverse'
in Chapter 14 uses a forward recursive auxiliary function rev'
.
rev :: [a] -> [a]
rev [] = [] -- rev.1
rev (x:xs) = rev xs ++ [x] -- rev.2
reverse :: [a] -> [a]
reverse xs = rev' xs [] -- reverse.1
where rev' [] ys = ys -- reverse.2
rev' (x:xs) ys = rev' xs (x:ys) -- reverse.3
To show rev
and reverse
are equivalent, we must prove that, for all finite lists xs
:
If we unfold (i.e. simplify) reverse
one step, we see that we need to prove:
rev xs = rev' xs []
Thus let’s try to prove this by structural induction on xs
.
Base case xs = []
:
rev []
{ rev.1
(left to right) }
[]
{ reverse.2
(right to left) }
rev' [] []
This establishes the base case.
Inductive case xs = (a:as)
:
Assume rev as = rev' as []
; prove rev (a:as) = rev' (a:as) []
.
First, we simplify the left side.
rev (a:as)
{ rev.2
(left to right) }
rev as ++ [a]
Then, we simplify the right side.
rev' (a:as) []
{ reverse.3
(left to right) }
rev' as [a]
Thus we need to show that rev as ++ [a] = rev' as [a]
. But we do not know how to proceed from this point.
Maybe another induction. But that would probably just bring us back to a point like this again. We are stuck!
Let’s look back at rev xs = rev' xs []
. This is difficult to prove directly. Note the asymmetry, one argument for rev
versus two for rev'
.
Thus let’s look for a new, more symmetrical, problem that might be easier to solve. Often it is easier to find a solution to a problem that is symmetrical than one which is not.
Note the place we got stuck above (proving rev as ++ [a] = rev' as [a]
) and also note the equation reverse.3
. Taking advantage of the identity element for ++
, we can restate our property in a more symmetrical way as follows:
rev xs ++ [] = rev' xs []
Note that the constant []
appears on both sides of the above equation. We can now apply the following generalization heuristic. (That is, we try to solve a “harder” problem.)
That is, generalize by replacing a constant (or any subexpression) by a new variable.
Thus we try to prove the more general proposition:
The case ys = []
gives us what we really want to hold. Intuitively, this new proposition seems to hold. Now let’s prove it formally. Again we try structural induction on xs
.
Base case xs = []
:
rev [] ++ ys
{ rev.1
(left to right) }
[] ++ ys
{ append.1
(left to right) }
ys
{ reverse.2
(right to left) }
rev' [] ys
This establishes the base case.
Inductive case xs = (a:as)
:
Assume rev as ++ ys = rev' as ys
for any finite list ys
;
prove rev (a:as) ++ ys = rev' (a:as) ys
.
rev (a:as) ++ ys
{ rev.2
(left to right) }
(rev as ++ [a]) ++ ys
{ ++
associativity, Note 1 }
rev as ++ ([a] ++ ys)
{ singleton law, Note 2 }
rev as ++ (a:ys)
{ induction hypothesis }
rev' as (a:ys)
{ reverse.3
(right to left) }
rev' (a:as) ys
This establishes the inductive case.
Notes:
We could apply the induction hypothesis here, but it does not seem profitable. Keeping the expressions in terms of rev
and ++
as long as possible seems better; we know more about those expressions.
The singleton law is [x] ++ xs = x:xs
for any element x
and finite list xs
of the same type. Proof of this is left as an exercise for the reader.
Therefore, we have proved rev xs ++ ys = rev' xs ys
and, hence:
The key to the performance improvement here is the solution of a “harder” problem: function rev'
does both the reversing and appending of a list while rev
separates the two actions.
This chapter illustrated how to state and prove Haskell “laws” about already defined functions.
The next two chapters on program synthesis illustrate how to use similar reasoning methods to synthesize (i.e. derive or calculate) function definitions from their specifications.
This set of exercises uses functions defined in this and previous chapters.
Functions map
, filter
, foldr
, foldl
, and concatMap
are defined in Chapter 15.
Functional composition, identity combinator id
, and function all
are defined in Chapter 16.
Functions takeWhile
and dropWhile
are defined in Chapter 17.
Prove the following properties using the proof methods illustrated in this chapter.
Prove for all x
of some type and finite lists xs
of the same type (i.e. the singleton law):
Consider the definition for length
given in the text of this chapter and the following definition for len
:
Prove for any finite list xs
: len 0 xs = length xs
.
Prove for all finite lists xs
and ys
of the same type:
Hint: The function reverse
(called reverse'
in Chapter 14) uses forward recursion. Backward recursive definitions are generally easier to use in inductive proofs. In Chapter 14, we also defined a backward recursive function rev
and proved that rev xs = reverse xs
for all finite lists xs
. Thus, you may find it easier to substitute rev
for reverse
and instead prove:
Prove for all finite lists xs
of some type:
Prove for all natural numbers m
and n
and all finite lists xs
:
Consider the rational number package from Chapter 7. Prove for any Rat
value r
that satisfied the interface invariant for the abstract module RationalRep
:
Consider the two definitions for the Fibonacci function in Chapter 9. Prove for any natural number n
:
Hint: First prove, for n
2:
Prove that the id
function is the identity element for functional composition. That is, for any function f :: a -> b
, prove:
Prove that functional composition is associative. That is, for any function f :: a -> a
, g :: a -> a
, and h :: a -> a
, prove:
Prove for all finite lists xs
and ys
of the same type and function f
on that type:
Prove for all finite lists xs
and ys
of the same type and predicate p
on that type:
Prove for all finite lists xs
and ys
of the same type and all predicates p
on that type:
The definition for &&
is as follows:
Prove for all finite lists xs
of some type and predicates p
and q
on that type:
Prove for all finite lists xs
and ys
of the same type and for all functions f
and values a
of compatible types:
Prove for all finite lists xs
of some type and all functions f
and g
of conforming types:
Prove for all finite lists of finite lists xss
of some b>ase type and function f
on that type:
Prove for all finite lists xs
of some type and functions f
on that type:
Prove for all lists xs
and predicates p
on the same type:
Prove that, if ***
is an associative binary operation of type t -> t
with identity element z
(i.e. a monoid), then:
Consider the Haskell type for the natural numbers given in an exercise in Chapter 21.
For the functions defined in that exercise, prove the following:
Prove that intToNat
and natToInt
are inverses of each other.
Prove that Zero
is the (right and left) identity element for addNat
.
Prove for any Nat
s x
and y
:
Prove associativity of addition on Nat
’s. That is, for any Nat
s x
, y
, and z
:
Prove commutativity of addition on Nat
’s. That is, for any Nat
s x
and y
:
In Summer 2018, I adapted and revised this chapter from:
These previous notes drew on the presentationsw in the first edition of the Bird and Wadler textbook [Bird 1988] and other sources. ([Bird 1998] and [Bird 2015] are updates of [Bird 1988]).
I incorporated this work as new Chapter 25, Proving Haskell Laws, in the 2018 version of the textbook Exploring Languages with Interpreters and Functional Programming and continue to revise it.
I maintain this chapter as text in Pandoc’s dialect of Markdown using embedded LaTeX markup for the mathematical formulas and then translate the document to HTML, PDF, and other forms as needed.
Referential transparency, equational reasoning, laws, definition, simplification, calculation, associativity, identity, monoid, singleton law, equivalence of functions.