vn-verdnaturachat/ios/Pods/boost-for-react-native/boost/hana/fwd/concept/monad.hpp

199 lines
9.3 KiB
C++
Raw Normal View History

/*!
@file
Forward declares `boost::hana::Monad`.
@copyright Louis Dionne 2013-2016
Distributed under the Boost Software License, Version 1.0.
(See accompanying file LICENSE.md or copy at http://boost.org/LICENSE_1_0.txt)
*/
#ifndef BOOST_HANA_FWD_CONCEPT_MONAD_HPP
#define BOOST_HANA_FWD_CONCEPT_MONAD_HPP
#include <boost/hana/config.hpp>
BOOST_HANA_NAMESPACE_BEGIN
//! @ingroup group-concepts
//! @defgroup group-Monad Monad
//! The `Monad` concept represents `Applicative`s with the ability to
//! flatten nested levels of structure.
//!
//! Historically, Monads are a construction coming from category theory,
//! an abstract branch of mathematics. The functional programming
//! community eventually discovered how Monads could be used to
//! formalize several useful things like side effects, which led
//! to the wide adoption of Monads in that community. However, even
//! in a multi-paradigm language like C++, there are several constructs
//! which turn out to be Monads, like `std::optional`, `std::vector` and
//! others.
//!
//! Everybody tries to introduce `Monad`s with a different analogy, and
//! most people fail. This is called the [Monad tutorial fallacy][1]. We
//! will try to avoid this trap by not presenting a specific intuition,
//! and we will instead present what monads are mathematically.
//! For specific intuitions, we will let readers who are new to this
//! concept read one of the many excellent tutorials available online.
//! Understanding Monads might take time at first, but once you get it,
//! a lot of patterns will become obvious Monads; this enlightening will
//! be your reward for the hard work.
//!
//! There are different ways of defining a Monad; Haskell uses a function
//! called `bind` (`>>=`) and another one called `return` (it has nothing
//! to do with C++'s `return` statement). They then introduce relationships
//! that must be satisfied for a type to be a Monad with those functions.
//! Mathematicians sometimes use a function called `join` and another one
//! called `unit`, or they also sometimes use other category theoretic
//! constructions like functor adjunctions and the Kleisli category.
//!
//! This library uses a composite approach. First, we use the `flatten`
//! function (equivalent to `join`) along with the `lift` function from
//! `Applicative` (equivalent to `unit`) to introduce the notion of
//! monadic function composition. We then write the properties that must
//! be satisfied by a Monad using this monadic composition operator,
//! because we feel it shows the link between Monads and Monoids more
//! clearly than other approaches.
//!
//! Roughly speaking, we will say that a `Monad` is an `Applicative` which
//! also defines a way to compose functions returning a monadic result,
//! as opposed to only being able to compose functions returning a normal
//! result. We will then ask for this composition to be associative and to
//! have a neutral element, just like normal function composition. For
//! usual composition, the neutral element is the identity function `id`.
//! For monadic composition, the neutral element is the `lift` function
//! defined by `Applicative`. This construction is made clearer in the
//! laws below.
//!
//! @note
//! Monads are known to be a big chunk to swallow. However, it is out of
//! the scope of this documentation to provide a full-blown explanation
//! of the concept. The [Typeclassopedia][2] is a nice Haskell-oriented
//! resource where more information about Monads can be found.
//!
//!
//! Minimal complete definitions
//! ----------------------------
//! First, a `Monad` must be both a `Functor` and an `Applicative`.
//! Also, an implementation of `flatten` or `chain` satisfying the
//! laws below for monadic composition must be provided.
//!
//! @note
//! The `ap` method for `Applicatives` may be derived from the minimal
//! complete definition of `Monad` and `Functor`; see below for more
//! information.
//!
//!
//! Laws
//! ----
//! To simplify writing the laws, we use the comparison between functions.
//! For two functions `f` and `g`, we define
//! @code
//! f == g if and only if f(x) == g(x) for all x
//! @endcode
//!
//! With the usual composition of functions, we are given two functions
//! @f$ f : A \to B @f$ and @f$ g : B \to C @f$, and we must produce a
//! new function @f$ compose(g, f) : A \to C @f$. This composition of
//! functions is associative, which means that
//! @code
//! compose(h, compose(g, f)) == compose(compose(h, g), f)
//! @endcode
//!
//! Also, this composition has an identity element, which is the identity
//! function. This simply means that
//! @code
//! compose(f, id) == compose(id, f) == f
//! @endcode
//!
//! This is probably nothing new if you are reading the `Monad` laws.
//! Now, we can observe that the above is equivalent to saying that
//! functions with the composition operator form a `Monoid`, where the
//! neutral element is the identity function.
//!
//! Given an `Applicative` `F`, what if we wanted to compose two functions
//! @f$ f : A \to F(B) @f$ and @f$ g : B \to F(C) @f$? When the
//! `Applicative` `F` is also a `Monad`, such functions taking normal
//! values but returning monadic values are called _monadic functions_.
//! To compose them, we obviously can't use normal function composition,
//! since the domains and codomains of `f` and `g` do not match properly.
//! Instead, we'll need a new operator -- let's call it `monadic_compose`:
//! @f[
//! \mathtt{monadic\_compose} :
//! (B \to F(C)) \times (A \to F(B)) \to (A \to F(C))
//! @f]
//!
//! How could we go about implementing this function? Well, since we know
//! `F` is an `Applicative`, the only functions we have are `transform`
//! (from `Functor`), and `lift` and `ap` (from `Applicative`). Hence,
//! the only thing we can do at this point while respecting the signatures
//! of `f` and `g` is to set (for `x` of type `A`)
//! @code
//! monadic_compose(g, f)(x) = transform(f(x), g)
//! @endcode
//!
//! Indeed, `f(x)` is of type `F(B)`, so we can map `g` (which takes `B`'s)
//! on it. Doing so will leave us with a result of type `F(F(C))`, but what
//! we wanted was a result of type `F(C)` to respect the signature of
//! `monadic_compose`. If we had a joker of type @f$ F(F(C)) \to F(C) @f$,
//! we could simply set
//! @code
//! monadic_compose(g, f)(x) = joker(transform(f(x), g))
//! @endcode
//!
//! and we would be happy. It turns out that `flatten` is precisely this
//! joker. Now, we'll want our joker to satisfy some properties to make
//! sure this composition is associative, just like our normal composition
//! was. These properties are slightly cumbersome to specify, so we won't
//! do it here. Also, we'll need some kind of neutral element for the
//! composition. This neutral element can't be the usual identity function,
//! because it does not have the right type: our neutral element needs to
//! be a function of type @f$ X \to F(X) @f$ but the identity function has
//! type @f$ X \to X @f$. It is now the right time to observe that `lift`
//! from `Applicative` has exactly the right signature, and so we'll take
//! this for our neutral element.
//!
//! We are now ready to formulate the `Monad` laws using this composition
//! operator. For a `Monad` `M` and functions @f$ f : A \to M(B) @f$,
//! @f$ g : B \to M(C) @f$ and @f$ h : C \to M(D) @f$, the following
//! must be satisfied:
//! @code
//! // associativity
//! monadic_compose(h, monadic_compose(g, f)) == monadic_compose(monadic_compose(h, g), f)
//!
//! // right identity
//! monadic_compose(f, lift<M(A)>) == f
//!
//! // left identity
//! monadic_compose(lift<M(B)>, f) == f
//! @endcode
//!
//! which is to say that `M` along with monadic composition is a Monoid
//! where the neutral element is `lift`.
//!
//!
//! Refined concepts
//! ----------------
//! 1. `Functor`
//! 2. `Applicative` (free implementation of `ap`)\n
//! When the minimal complete definition for `Monad` and `Functor` are
//! both satisfied, it is possible to implement `ap` by setting
//! @code
//! ap(fs, xs) = chain(fs, [](auto f) {
//! return transform(xs, f);
//! })
//! @endcode
//!
//!
//! Concrete models
//! ---------------
//! `hana::lazy`, `hana::optional`, `hana::tuple`
//!
//!
//! [1]: https://byorgey.wordpress.com/2009/01/12/abstraction-intuition-and-the-monad-tutorial-fallacy/
//! [2]: https://wiki.haskell.org/Typeclassopedia#Monad
template <typename M>
struct Monad;
BOOST_HANA_NAMESPACE_END
#endif // !BOOST_HANA_FWD_CONCEPT_MONAD_HPP