1 **Don't try to read this yet!!! Many substantial edits are still in process.
6 I really don't know much category theory. Just enough to put this
7 together. Also, this really is "put together." I haven't yet found an
8 authoritative source (that's accessible to a category theory beginner like
9 myself) that discusses the correspondence between the category-theoretic and
10 functional programming uses of these notions in enough detail to be sure that
11 none of the pieces here is misguided. In particular, it wasn't completely
12 obvious how to map the polymorphism on the programming theory side into the
13 category theory. And I'm bothered by the fact that our `<=<` operation is only
14 partly defined on our domain of natural transformations. But this does seem to
15 me to be the reasonable way to put the pieces together. We very much welcome
16 feedback from anyone who understands these issues better, and will make
22 A **monoid** is a structure <code>(S,⋆,z)</code> consisting of an associative binary operation <code>⋆</code> over some set `S`, which is closed under <code>⋆</code>, and which contains an identity element `z` for <code>⋆</code>. That is:
26 for all s1, s2, s3 in S:
27 (i) s1⋆s2 etc are also in S
28 (ii) (s1⋆s2)⋆s3 = s1⋆(s2⋆s3)
29 (iii) z⋆s1 = s1 = s1⋆z
32 Some examples of monoids are:
34 * finite strings of an alphabet `A`, with <code>⋆</code> being concatenation and `z` being the empty string
35 * all functions <code>X→X</code> over a set `X`, with <code>⋆</code> being composition and `z` being the identity function over `X`
36 * the natural numbers with <code>⋆</code> being plus and `z` being `0` (in particular, this is a **commutative monoid**). If we use the integers, or the naturals mod n, instead of the naturals, then every element will have an inverse and so we have not merely a monoid but a **group**.)
37 * if we let <code>⋆</code> be multiplication and `z` be `1`, we get different monoids over the same sets as in the previous item.
41 A **category** is a generalization of a monoid. A category consists of a class of **elements**, and a class of **morphisms** between those elements. Morphisms are sometimes also called maps or arrows. They are something like functions (and as we'll see below, given a set of functions they'll determine a category). However, a single morphism only maps between a single source element and a single target element. Also, there can be multiple distinct morphisms between the same source and target, so the identity of a morphism goes beyond its "extension."
43 When a morphism `f` in category <b>C</b> has source `C1` and target `C2`, we'll write <code>f:C1→C2</code>.
45 To have a category, the elements and morphisms have to satisfy some constraints:
48 (i) the class of morphisms has to be closed under composition:
49 where f:C1→C2 and g:C2→C3, g ∘ f is also a
50 morphism of the category, which maps C1→C3.
52 (ii) composition of morphisms has to be associative
54 (iii) every element E of the category has to have an identity
55 morphism 1<sub>E</sub>, which is such that for every morphism f:C1→C2:
56 1<sub>C2</sub> ∘ f = f = f ∘ 1<sub>C1</sub>
59 These parallel the constraints for monoids. Note that there can be multiple distinct morphisms between an element `E` and itself; they need not all be identity morphisms. Indeed from (iii) it follows that each element can have only a single identity morphism.
61 A good intuitive picture of a category is as a generalized directed graph, where the category elements are the graph's nodes, and there can be multiple directed edges between a given pair of nodes, and nodes can also have multiple directed edges to themselves. (Every node must have at least one such, which is that node's identity morphism.)
64 Some examples of categories are:
66 * Categories whose elements are sets and whose morphisms are functions between those sets. Here the source and target of a function are its domain and range, so distinct functions sharing a domain and range (e.g., `sin` and `cos`) are distinct morphisms between the same source and target elements. The identity morphism for any element/set is just the identity function for that set.
68 * any monoid <code>(S,⋆,z)</code> generates a category with a single element `x`; this `x` need not have any relation to `S`. The members of `S` play the role of *morphisms* of this category, rather than its elements. All of these morphisms are understood to map `x` to itself. The result of composing the morphism consisting of `s1` with the morphism `s2` is the morphism `s3`, where <code>s3=s1⋆s2</code>. The identity morphism for the (single) category element `x` is the monoid's identity `z`.
70 * a **preorder** is a structure <code>(S, ≤)</code> consisting of a reflexive, transitive, binary relation on a set `S`. It need not be connected (that is, there may be members `x`,`y` of `S` such that neither <code>x≤y</code> nor <code>y≤x</code>). It need not be anti-symmetric (that is, there may be members `s1`,`s2` of `S` such that <code>s1≤s2</code> and <code>s2≤s1</code> but `s1` and `s2` are not identical). Some examples:
72 * sentences ordered by logical implication ("p and p" implies and is implied by "p", but these sentences are not identical; so this illustrates a pre-order without anti-symmetry)
73 * sets ordered by size (this illustrates it too)
75 Any pre-order <code>(S,≤)</code> generates a category whose elements are the members of `S` and which has only a single morphism between any two elements `s1` and `s2`, iff <code>s1≤s2</code>.
80 A **functor** is a "homomorphism", that is, a structure-preserving mapping, between categories. In particular, a functor `F` from category <b>C</b> to category <b>D</b> must:
83 (i) associate with every element C1 of <b>C</b> an element F(C1) of <b>D</b>
85 (ii) associate with every morphism f:C1→C2 of <b>C</b> a morphism
86 F(f):F(C1)→F(C2) of <b>D</b>
88 (iii) "preserve identity", that is, for every element C1 of <b>C</b>:
89 F of C1's identity morphism in <b>C</b> must be the identity morphism
90 of F(C1) in <b>D</b>: F(1<sub>C1</sub>) = 1<sub>F(C1)</sub>.
92 (iv) "distribute over composition", that is for any morphisms f and g in <b>C</b>:
93 F(g ∘ f) = F(g) ∘ F(f)
96 A functor that maps a category to itself is called an **endofunctor**. The (endo)functor that maps every element and morphism of <b>C</b> to itself is denoted `1C`.
98 How functors compose: If `G` is a functor from category <b>C</b> to category <b>D</b>, and `K` is a functor from category <b>D</b> to category <b>E</b>, then `KG` is a functor which maps every element `C1` of <b>C</b> to element `K(G(C1))` of <b>E</b>, and maps every morphism `f` of <b>C</b> to morphism `K(G(f))` of <b>E</b>.
100 I'll assert without proving that functor composition is associative.
104 Natural Transformation
105 ----------------------
106 So categories include elements and morphisms. Functors consist of mappings from the elements and morphisms of one category to those of another (or the same) category. **Natural transformations** are a third level of mappings, from one functor to another.
108 Where `G` and `H` are functors from category <b>C</b> to category <b>D</b>, a natural transformation η between `G` and `H` is a family of morphisms η[C1]:G(C1)→H(C1)` in <b>D</b> for each element `C1` of <b>C</b>. That is, η[C1]` has as source `C1`'s image under `G` in <b>D</b>, and as target `C1`'s image under `H` in <b>D</b>. The morphisms in this family must also satisfy the constraint:
110 for every morphism f:C1→C2 in <b>C</b>: η[C2] ∘ G(f) = H(f) ∘ η[C1]
112 That is, the morphism via `G(f)` from `G(C1)` to `G(C2)`, and then via η[C2]` to `H(C2)`, is identical to the morphism from `G(C1)` via η[C1]` to `H(C1)`, and then via `H(f)` from `H(C1)` to `H(C2)`.
115 How natural transformations compose:
117 Consider four categories <b>B</b>, <b>C</b>, <b>D</b>, and <b>E</b>. Let `F` be a functor from <b>B</b> to <b>C</b>; `G`, `H`, and `J` be functors from <b>C</b> to <b>D</b>; and `K` and `L` be functors from <b>D</b> to <b>E</b>. Let η be a natural transformation from `G` to `H`; φ be a natural transformation from `H` to `J`; and ψ be a natural transformation from `K` to `L`. Pictorally:
119 - <b>B</b> -+ +--- <b>C</b> --+ +---- <b>D</b> -----+ +-- <b>E</b> --
121 F: -----→ G: -----→ K: -----→
122 | | | | | η | | | ψ
124 | | H: -----→ L: -----→
127 | | J: -----→ | |
128 -----+ +--------+ +------------+ +-------
130 Then `(η F)` is a natural transformation from the (composite) functor `GF` to the composite functor `HF`, such that where `b1` is an element of category <b>B</b>, `(η F)[b1] = η[F(b1)]`---that is, the morphism in <b>D</b> that η assigns to the element `F(b1)` of <b>C</b>.
132 And `(K η)` is a natural transformation from the (composite) functor `KG` to the (composite) functor `KH`, such that where `C1` is an element of category <b>C</b>, `(K η)[C1] = K(η[C1])`---that is, the morphism in <b>E</b> that `K` assigns to the morphism η[C1]` of <b>D</b>.
135 `(φ -v- η)` is a natural transformation from `G` to `J`; this is known as a "vertical composition". We will rely later on this, where `f:C1→C2`:
137 φ[C2] ∘ H(f) ∘ η[C1] = φ[C2] ∘ H(f) ∘ η[C1]
139 by naturalness of φ, is:
141 φ[C2] ∘ H(f) ∘ η[C1] = J(f) ∘ φ[C1] ∘ η[C1]
143 by naturalness of η, is:
145 φ[C2] ∘ η[C2] ∘ G(f) = J(f) ∘ φ[C1] ∘ η[C1]
147 Hence, we can define `(φ -v- η)[x]` as: φ[x] ∘ η[x]` and rely on it to satisfy the constraints for a natural transformation from `G` to `J`:
149 (φ -v- η)[C2] ∘ G(f) = J(f) ∘ (φ -v- η)[C1]
151 An observation we'll rely on later: given the definitions of vertical composition and of how natural transformations compose with functors, it follows that:
153 ((φ -v- η) F) = ((φ F) -v- (η F))
155 I'll assert without proving that vertical composition is associative and has an identity, which we'll call "the identity transformation."
158 `(ψ -h- η)` is natural transformation from the (composite) functor `KG` to the (composite) functor `LH`; this is known as a "horizontal composition." It's trickier to define, but we won't be using it here. For reference:
160 (φ -h- η)[C1] = L(η[C1]) ∘ ψ[G(C1)]
161 = ψ[H(C1)] ∘ K(η[C1])
163 Horizontal composition is also associative, and has the same identity as vertical composition.
169 In earlier days, these were also called "triples."
171 A **monad** is a structure consisting of an (endo)functor `M` from some category <b>C</b> to itself, along with some natural transformations, which we'll specify in a moment.
173 Let `T` be a set of natural transformations `p`, each being between some (variable) functor `P` and another functor which is the composite `MP'` of `M` and a (variable) functor `P'`. That is, for each element `C1` in <b>C</b>, `p` assigns `C1` a morphism from element `P(C1)` to element `MP'(C1)`, satisfying the constraints detailed in the previous section. For different members of `T`, the relevant functors may differ; that is, `p` is a transformation from functor `P` to `MP'`, `q` is a transformation from functor `Q` to `MQ'`, and none of `P`, `P'`, `Q`, `Q'` need be the same.
175 One of the members of `T` will be designated the "unit" transformation for `M`, and it will be a transformation from the identity functor `1C` for <b>C</b> to `M(1C)`. So it will assign to `C1` a morphism from `C1` to `M(C1)`.
177 We also need to designate for `M` a "join" transformation, which is a natural transformation from the (composite) functor `MM` to `M`.
179 These two natural transformations have to satisfy some constraints ("the monad laws") which are most easily stated if we can introduce a defined notion.
181 Let `p` and `q` be members of `T`, that is they are natural transformations from `P` to `MP'` and from `Q` to `MQ'`, respectively. Let them be such that `P' = Q`. Now `(M q)` will also be a natural transformation, formed by composing the functor `M` with the natural transformation `q`. Similarly, `(join Q')` will be a natural transformation, formed by composing the natural transformation `join` with the functor `Q'`; it will transform the functor `MMQ'` to the functor `MQ'`. Now take the vertical composition of the three natural transformations `(join Q')`, `(M q)`, and `p`, and abbreviate it as follows:
183 q <=< p =def. ((join Q') -v- (M q) -v- p)
185 Since composition is associative I don't specify the order of composition on the rhs.
187 In other words, `<=<` is a binary operator that takes us from two members `p` and `q` of `T` to a composite natural transformation. (In functional programming, at least, this is called the "Kleisli composition operator". Sometimes its written `p >=> q` where that's the same as `q <=< p`.)
189 `p` is a transformation from `P` to `MP'` which = `MQ`; `(M q)` is a transformation from `MQ` to `MMQ'`; and `(join Q')` is a transformation from `MMQ'` to `MQ'`. So the composite `q <=< p` will be a transformation from `P` to `MQ'`, and so also eligible to be a member of `T`.
191 Now we can specify the "monad laws" governing a monad as follows:
193 (T, <=<, unit) constitute a monoid
195 That's it. (Well, perhaps we're cheating a bit, because `q <=< p` isn't fully defined on `T`, but only when `P` is a functor to `MP'` and `Q` is a functor from `P'`. But wherever `<=<` is defined, the monoid laws are satisfied:
197 (i) q <=< p is also in T
198 (ii) (r <=< q) <=< p = r <=< (q <=< p)
199 (iii.1) unit <=< p = p (here p has to be a natural transformation to M(1C))
200 (iii.2) p = p <=< unit (here p has to be a natural transformation from 1C)
202 If `p` is a natural transformation from `P` to `M(1C)` and `q` is `(p Q')`, that is, a natural transformation from `PQ` to `MQ`, then we can extend (iii.1) as follows:
206 = ((join -v- (M unit) -v- p) Q')
207 = (join Q') -v- ((M unit) Q') -v- (p Q')
208 = (join Q') -v- (M (unit Q')) -v- q
212 where as we said `q` is a natural transformation from some `PQ'` to `MQ'`.
214 Similarly, if `p` is a natural transformation from `1C` to `MP'`, and `q` is `(p Q)`, that is, a natural transformation from `Q` to `MP'Q`, then we can extend (iii.2) as follows:
218 = (((join P') -v- (M p) -v- unit) Q)
219 = ((join P'Q) -v- ((M p) Q) -v- (unit Q))
220 = ((join P'Q) -v- (M (p Q)) -v- (unit Q))
224 where as we said `q` is a natural transformation from `Q` to some `MP'Q`.
229 The standard category-theory presentation of the monad laws
230 -----------------------------------------------------------
231 In category theory, the monad laws are usually stated in terms of `unit` and `join` instead of `unit` and `<=<`.
234 P2. every element C1 of a category <b>C</b> has an identity morphism 1<sub>C1</sub> such that for every morphism f:C1→C2 in <b>C</b>: 1<sub>C2</sub> ∘ f = f = f ∘ 1<sub>C1</sub>.
235 P3. functors "preserve identity", that is for every element C1 in F's source category: F(1<sub>C1</sub>) = 1<sub>F(C1)</sub>.
238 Let's remind ourselves of some principles:
239 * composition of morphisms, functors, and natural compositions is associative
240 * functors "distribute over composition", that is for any morphisms f and g in F's source category: F(g ∘ f) = F(g) ∘ F(f)
241 * if η is a natural transformation from F to G, then for every f:C1→C2 in F and G's source category <b>C</b>: η[C2] ∘ F(f) = G(f) ∘ η[C1].
244 Let's use the definitions of naturalness, and of composition of natural transformations, to establish two lemmas.
247 Recall that join is a natural transformation from the (composite) functor MM to M. So for elements C1 in <b>C</b>, join[C1] will be a morphism from MM(C1) to M(C1). And for any morphism f:a→b in <b>C</b>:
249 (1) join[b] ∘ MM(f) = M(f) ∘ join[a]
251 Next, consider the composite transformation ((join MQ') -v- (MM q)).
252 q is a transformation from Q to MQ', and assigns elements C1 in <b>C</b> a morphism q*: Q(C1) → MQ'(C1). (MM q) is a transformation that instead assigns C1 the morphism MM(q*).
253 (join MQ') is a transformation from MMMQ' to MMQ' that assigns C1 the morphism join[MQ'(C1)].
255 (2) ((join MQ') -v- (MM q)) assigns to C1 the morphism join[MQ'(C1)] ∘ MM(q*).
257 Next, consider the composite transformation ((M q) -v- (join Q)).
258 (3) This assigns to C1 the morphism M(q*) ∘ join[Q(C1)].
260 So for every element C1 of <b>C</b>:
261 ((join MQ') -v- (MM q))[C1], by (2) is:
262 join[MQ'(C1)] ∘ MM(q*), which by (1), with f=q*: Q(C1)→MQ'(C1) is:
263 M(q*) ∘ join[Q(C1)], which by 3 is:
264 ((M q) -v- (join Q))[C1]
266 So our (lemma 1) is: ((join MQ') -v- (MM q)) = ((M q) -v- (join Q)), where q is a transformation from Q to MQ'.
269 Next recall that unit is a natural transformation from 1C to M. So for elements C1 in <b>C</b>, unit[C1] will be a morphism from C1 to M(C1). And for any morphism f:a→b in <b>C</b>:
270 (4) unit[b] ∘ f = M(f) ∘ unit[a]
272 Next consider the composite transformation ((M q) -v- (unit Q)). (5) This assigns to C1 the morphism M(q*) ∘ unit[Q(C1)].
274 Next consider the composite transformation ((unit MQ') -v- q). (6) This assigns to C1 the morphism unit[MQ'(C1)] ∘ q*.
276 So for every element C1 of <b>C</b>:
277 ((M q) -v- (unit Q))[C1], by (5) =
278 M(q*) ∘ unit[Q(C1)], which by (4), with f=q*: Q(C1)→MQ'(C1) is:
279 unit[MQ'(C1)] ∘ q*, which by (6) =
280 ((unit MQ') -v- q)[C1]
282 So our lemma (2) is: (((M q) -v- (unit Q)) = ((unit MQ') -v- q)), where q is a transformation from Q to MQ'.
285 Finally, we substitute ((join Q') -v- (M q) -v- p) for q <=< p in the monad laws. For simplicity, I'll omit the "-v-".
287 for all p,q,r in T, where p is a transformation from P to MP', q is a transformation from Q to MQ', R is a transformation from R to MR', and P'=Q and Q'=R:
289 (i) q <=< p etc are also in T
291 (i') ((join Q') (M q) p) etc are also in T
294 (ii) (r <=< q) <=< p = r <=< (q <=< p)
296 (r <=< q) is a transformation from Q to MR', so:
297 (r <=< q) <=< p becomes: (join R') (M (r <=< q)) p
298 which is: (join R') (M ((join R') (M r) q)) p
299 substituting in (ii), and helping ourselves to associativity on the rhs, we get:
301 ((join R') (M ((join R') (M r) q)) p) = ((join R') (M r) (join Q') (M q) p)
302 ---------------------
303 which by the distributivity of functors over composition, and helping ourselves to associativity on the lhs, yields:
304 ------------------------
305 ((join R') (M join R') (MM r) (M q) p) = ((join R') (M r) (join Q') (M q) p)
307 which by lemma 1, with r a transformation from Q' to MR', yields:
309 ((join R') (M join R') (MM r) (M q) p) = ((join R') (join MR') (MM r) (M q) p)
311 which will be true for all r,q,p just in case:
313 ((join R') (M join R')) = ((join R') (join MR')), for any R'.
315 which will in turn be true just in case:
317 (ii') (join (M join)) = (join (join M))
320 (iii.1) (unit P') <=< p = p
322 (unit P') is a transformation from P' to MP', so:
323 (unit P') <=< p becomes: (join P') (M unit P') p
324 which is: (join P') (M unit P') p
325 substituting in (iii.1), we get:
326 ((join P') (M unit P') p) = p
328 which will be true for all p just in case:
330 ((join P') (M unit P')) = the identity transformation, for any P'
332 which will in turn be true just in case:
334 (iii.1') (join (M unit) = the identity transformation
337 (iii.2) p = p <=< (unit P)
339 p is a transformation from P to MP', so:
340 unit <=< p becomes: (join P') (M p) unit
341 substituting in (iii.2), we get:
342 p = ((join P') (M p) (unit P))
344 which by lemma (2), yields:
346 p = ((join P') ((unit MP') p)
348 which will be true for all p just in case:
350 ((join P') (unit MP')) = the identity transformation, for any P'
352 which will in turn be true just in case:
354 (iii.2') (join (unit M)) = the identity transformation
357 Collecting the results, our monad laws turn out in this format to be:
359 when p a transformation from P to MP', q a transformation from P' to MQ', r a transformation from Q' to MR' all in T:
361 (i') ((join Q') (M q) p) etc also in T
363 (ii') (join (M join)) = (join (join M))
365 (iii.1') (join (M unit)) = the identity transformation
367 (iii.2')(join (unit M)) = the identity transformation
371 7. The functional programming presentation of the monad laws
372 ------------------------------------------------------------
373 In functional programming, unit is usually called "return" and the monad laws are usually stated in terms of return and an operation called "bind" which is interdefinable with <=< or with join.
375 Additionally, whereas in category-theory one works "monomorphically", in functional programming one usually works with "polymorphic" functions.
377 The base category <b>C</b> will have types as elements, and monadic functions as its morphisms. The source and target of a morphism will be the types of its argument and its result. (As always, there can be multiple distinct morphisms from the same source to the same target.)
379 A monad M will consist of a mapping from types C1 to types M(C1), and a mapping from functions f:C1→C2 to functions M(f):M(C1)→M(C2). This is also known as "fmap f" or "liftM f" for M, and is called "function f lifted into the monad M." For example, where M is the list monad, M maps every type X into the type "list of Xs", and maps every function f:x→y into the function that maps [x1,x2...] to [y1,y2,...].
384 A natural transformation t assigns to each type C1 in <b>C</b> a morphism t[C1]: C1→M(C1) such that, for every f:C1→C2:
385 t[C2] ∘ f = M(f) ∘ t[C1]
387 The composite morphisms said here to be identical are morphisms from the type C1 to the type M(C2).
391 In functional programming, instead of working with natural transformations we work with "monadic values" and polymorphic functions "into the monad" in question.
393 For an example of the latter, let p be a function that takes arguments of some (schematic, polymorphic) type C1 and yields results of some (schematic, polymorphic) type M(C2). An example with M being the list monad, and C2 being the tuple type schema int * C1:
395 let p = fun c → [(1,c), (2,c)]
397 p is polymorphic: when you apply it to the int 0 you get a result of type "list of int * int": [(1,0), (2,0)]. When you apply it to the char 'e' you get a result of type "list of int * char": [(1,'e'), (2,'e')].
399 However, to keep things simple, we'll work instead with functions whose type is settled. So instead of the polymorphic p, we'll work with (p : C1 → M(int * C1)). This only accepts arguments of type C1. For generality, I'll talk of functions with the type (p : C1 → M(C1')), where we assume that C1' is a function of C1.
401 A "monadic value" is any member of a type M(C1), for any type C1. For example, a list is a monadic value for the list monad. We can think of these monadic values as the result of applying some function (p : C1 → M(C1')) to an argument of type C1.