lean2/doc/lean/tutorial.md
Leonardo de Moura 69bfc682b4 chore(*): replace leodemoura with leanprover
Signed-off-by: Leonardo de Moura <leonardo@microsoft.com>
2014-04-29 15:31:29 -07:00

27 KiB
Raw Blame History

Lean Tutorial

WARNING: This tutoral is for Lean 0.1, before major modifications performed in version 0.2. Version 0.2 is still under development. Here you can find a snapshot of Lean 0.1.

Introduction

Lean is an automatic and interactive theorem prover. It can be used to create specifications, build mathematical libraries, and solve constraints. In this tutorial, we introduce basic concepts, the logic used in Lean, and the main commands.

Getting started

We can use Lean in interactive or batch mode. The following example just displays the message hello world.

        print "hello world"

All we have to do to run your first example is to call the lean executable with the name of the text file that contains the command above. If you saved the above command in the file hello.lean, then you just have to execute

   lean hello.lean

As a more complex example, the next example defines a function that doubles the input value.

      -- defines the double function
      definition double (x : Nat) := x + x

Basics

We can also view Lean as a suite of tools for evaluating and processing expressions representing terms, definitions, and theorems.

Every expression has a unique type in Lean. The command check returns the type of a given expression.

      check double 3
      check double

The last command returns Nat → Nat. That is, the type of double is a function from Nat to Nat, where Nat is the type of the natural numbers.

The command import loads existing libraries and extensions. The following command imports the command find that searches the Lean environment using regular expressions

      import find

      find "Nat"      -- find all object that start with the prefix Nat
      check Nat::ge   -- display the signature of the Nat::ge definition

We say Nat::ge is a hierarchical name comprised of two parts: Nat and ge

The command using creates aliases based on a given prefix. For example, the following command creates aliases for all objects starting with Nat

      using Nat
      check ge       -- display the signature of the Nat::ge definition

The command variable assigns a type to an identifier. The following command postulates/assumes that n, m and o have type Nat.

      variable n : Nat
      variable m : Nat
      variable o : Nat

The command variables n m o : Nat can be used a shorthand for the three commands above.

In Lean, proofs are also expressions, and all functionality provided for manipulating expressions is also available for manipulating proofs. For example, refl n is a proof for n = n. In Lean, refl is the reflexivity theorem.

     check refl n

The command axiom postulates that a given proposition (aka Boolean formula) is true. The following commands postulate two axioms Ax1 and Ax2 that state that n = m and m = o.

      axiom Ax1 : n = m
      axiom Ax2 : m = o

Ax1 and Ax2 are not just names. For example, trans Ax1 Ax2 is a proof that n = o, where trans is the transitivity theorem.

      check trans Ax1 Ax2

The expression trans Ax1 Ax2 is just a function application like any other. Moreover, in Lean, propositions are types. Any Boolean expression P can be used as a type. The elements of type P can be viewed as the proofs of P. Moreover, in Lean, proof checking is type checking. For example, the Lean type checker will reject the type incorrect term trans Ax2 Ax1.

Because we use proposition as types, we must support empty types. For example, the type false must be empty, since we don't have a proof for false.

Most systems based on the propositions as types paradigm are based on constructive logic. Lean on the other hand is based on classical logic. The excluded middle is a theorem in Lean, and em p is a proof for p ¬ p.

    variable p : Bool
    check em p

The commands axiom and variable are essentially the same command. We provide both just to make Lean files more readable. We encourage users to use axiom only for propostions, and variable for everything else.

Similarly, a theorem is just a definition. The following command defines a new theorem called nat_trans3

      theorem nat_trans3 (a b c d : Nat) (H1 : a = b) (H2 : c = b) (H3 : c = d) : a = d
      := trans (trans H1 (symm H2)) H3

The theorem nat_trans3 has 7 parameters, it takes for natural numbers a, b, c and d, and three proofs showing that a = b, c = b and c = d, and returns a proof that a = d. In the example above, symm is the symmetry theorem. Now, we use nat_trans3 in a simple example.

      variables x y z w : Nat
      axiom Hxy : x = y
      axiom Hzy : z = y
      axiom Hzw : z = w
      check nat_trans3 x y z w Hxy Hzy Hzw

The theorem nat_trans3 is somewhat inconvenient to use because it has 7 parameters. However, the first four parameters can be inferred from the last 3. We can use _ as a placeholder that instructs Lean to synthesize this expression. The synthesis process is based on type inference, and it is the most basic form of automation provided by Lean.

      check nat_trans3 _ _ _ _ Hxy Hzy Hzw

Lean also supports implicit arguments. We mark implicit arguments using curly braces instead of parenthesis. In the following example, we define the theorem nat_trans3i using implicit arguments.

      theorem nat_trans3i {a b c d : Nat} (H1 : a = b) (H2 : c = b) (H3 : c = d) : a = d
      := trans (trans H1 (symm H2)) H3

It is identical to nat_trans3, the only difference is the use of curly braces. Lean will (try to) infer the implicit arguments. The idea behind implicit arguments is quite simple, we are just instructing Lean to automatically insert the placeholders _ for us.

      check nat_trans3i Hxy Hzy Hzw

Sometimes, Lean will not be able to infer the parameters automatically. So, whenever we define a theorem/definition/axiom/variable containing implicit arguments, Lean will automatically create an explicit version where all parameters are explicit. The explicit version uses the same name with a @ prefix.

      check @nat_trans3i

The theorems refl, trans and symm all have implicit arguments.

      check @refl
      check @trans
      check @symm

We can also instruct Lean to display all implicit arguments when it prints expressions. This is useful when debugging non-trivial problems.

      set_option pp::implicit true   -- show implicit arguments
      check nat_trans3i Hxy Hzy Hzw
      set_option pp::implicit false  -- hide implicit arguments

In the previous example, the check command stated that nat_trans3i Hxy Hzy Hzw has type @eq x w. The expression x = w is just notational convenience.

We have seen many occurrences of (Type U), where U is a universe variable. In Lean, the type of Nat and Bool is Type.

      check Nat
      check Bool

We say Type is the type of all small types, but what is the type of Type?

      check Type

Lean returns (Type 1). Similarly, the type of (Type 1) is (Type 2). In Lean, we also have universe cumulativity. That is, we can provide an element of type (Type i) where an element of type (Type j) is expected when i ≤ j. This makes the system more convenient to use. Otherwise, we would need a reflexivity theorem for Type (i.e., (Type 0)), Type 1, Type 2, etc. Universe cumulativity improves usability, but it is not enough because we would still have the question: how big should i be? Moreover, if we choose an i that is not big enough we have to go back and correct all libraries. This is not satisfactory and not modular. So, in Lean, we allow users to declare universe variables and simple constraints between them. The Lean kernel defines one universe variable U, and states that U ≥ 1 using the command universe U ≥ 1. The Lean type casting library defines another universe variable called M and states that universe M ≥ 1. In Lean, whenever we declare a new universe V, the system automatically adds the constraint U ≥ V + 1. That is, U the maximal universe in Lean. Lean reports an universe inconsistency if the universe constraints are inconsistent. For example, it will return an error if execute the command universe M ≥ U. We can view universe variables as placeholders, and we can always solve the universe constraints and find and assignment for the universe variables used in our developments. This assignment allows us to automatically generate a Lean specification that is not based on this particular feature.

Propositional logic

To manipulate formulas with a richer logical structure, it is important to master the notation Lean uses for building composite logical expressions out of basic formulas using logical connectives. The logical connectives (and, or, not, etc) are defined in the Lean kernel. The kernel also defines notational convention for rewriting formulas in a natural way. Here is a table showing the notation for the so called propositional (or Boolean) connectives.

Ascii Ascii alt. Unicode Definition
true true
false false
not ¬ not
/\ && and
/ || or
-> implies
<-> iff

true and false are logical constants to denote the true and false propositions. Logical negation is a unary operator just like arithmetical negation on numbers. The other connectives are all binary operators. The meaning of the operators is the usual one. The table above makes clear that Lean supports unicode characters. We can use Ascii or/and unicode versions. Here is a simple example using the connectives above.

    variable q : Bool
    check p → q → p ∧ q
    check ¬ p → p ↔ false
    check p  q → q  p
    -- Ascii version
    check p -> q -> p && q
    check not p -> p <-> false
    check p || q -> q \/ p

Depending on the platform, Lean uses unicode characters by default when printing expressions. The following commands can be used to change this behavior.

    set_option pp::unicode false
    check p → q → p ∧ q
    set_option pp::unicode true
    check p → q → p ∧ q

Note that, it may seem that the symbols -> and are overloaded, and Lean uses them to represent Boolean implication and the type of functions. Actually, they are not overloaded, they are the same symbols. In Lean, the Boolean p → q expression is also the type of the functions that given a proof for p, returns a proof for q. This is very convenient for writing proofs.

   -- Hpq is a function that takes a proof for p and returns a proof for q
   axiom Hpq : p → q
   -- Hq is a proof/certificate for p
   axiom Hp  : p
   -- The expression Hpq Hp is a proof/certificate for q
   check Hpq Hp

In composite expressions, the precedences of the various binary connectives are in order of the above table, with and being the strongest and iff the weakest. For example, a ∧ b → c d ∧ e means (a ∧ b) → (c (d ∧ e)). All of them are right-associative. So, p ∧ q ∧ r means p ∧ (q ∧ r). The actual precedence and fixity of all logical connectives is defined in the Lean kernel definition file.

Finally, not, and, or and iff are the actual names used when defining the Boolean connectives. They can be used as any other function.

   check and
   check or
   check not

Lean supports currying and true is a function from Bool to Bool.

   check and true
   definition id := and true

Functions

There are many variable-binding constructs in mathematics. Lean expresses all of them using just one abstraction, which is a converse operation to function application. Given a variable x, a type A, and a term t that may or may not contain x, one can construct the so-called lambda abstraction fun x : A, t, or using unicode notation λ x : A, t. Here is some simple examples.

   check fun x : Nat, x + 1
   check fun x y : Nat, x + 2 * y
   check fun x y : Bool, not (x ∧ y)
   check λ x : Nat, x + 1
   check λ (x : Nat) (p : Bool), x = 0  p

In many cases, Lean can automatically infer the type of the variable. Actually, In all examples above, the type can be inferred automatically.

   check fun x, x + 1
   check fun x y, x + 2 * y
   check fun x y, not (x ∧ y)
   check λ x, x + 1
   check λ x p, x = 0  p

However, Lean will complain that it cannot infer the type of the variable x in fun x, x because any type would work in this example.

The following example shows how to use lambda abstractions in function applications

   check (fun x y, x + 2 * y) 1
   check (fun x y, x + 2 * y) 1 2
   check (fun x y, not (x ∧ y)) true false

Lambda abstractions are also used to create proofs for propositions of the form A → B. This should be natural since we can "view" A → B as the type of functions that given a proof for A returns a proof for B. For example, a proof for p → p is just fun H : p, H (the identity function).

    check fun H : p, H

Definitional equality

The command eval t computes a normal form for the term t. In Lean, we say two terms are definitionally equal if the have the same normal form. For example, the terms (λ x : Nat, x + 1) a and a + 1 are definitionally equal. The Lean type/proof checker uses the normalizer when checking types/proofs. So, we can prove that two definitionally equal terms are equal using just refl. Here is a simple example.

   theorem def_eq_th (a : Nat) : ((λ x : Nat, x + 1) a) = a + 1
   := refl (a+1)

Provable equality

In the previous examples, we have used nat_trans3 x y z w Hxy Hzy Hzw to show that x = w. In this case, x and w are not definitionally equal, but they are provably equal in the environment that contains nat_trans3 and axioms Hxy, Hzy and Hzw.

Proving

The Lean kernel contains basic theorems for creating proof terms. The basic theorems are useful for creating manual proofs. The are also the basic building blocks used by all automated proof engines available in Lean. The theorems can be broken into three different categories: introduction, elimination, and rewriting. First, we cover the introduction and elimination theorems for the basic Boolean connectives.

And (conjuction)

The expression and_intro H1 H2 creates a proof for a ∧ b using proofs H1 : a and H2 : b. We say and_intro is the and-introduction operation. In the following example we use and_intro for creating a proof for p → q → p ∧ q.

   check fun (Hp : p) (Hq : q), and_intro Hp Hq

The expression and_eliml H creates a proof a from a proof H : a ∧ b. Similarly and_elimr H is a proof for b. We say they are the left/right and-elimination.

   -- Proof for p ∧ q → p
   check fun H : p ∧ q, and_eliml H
   -- Proof for p ∧ q → q
   check fun H : p ∧ q, and_elimr H

Now, we prove p ∧ q → q ∧ p with the following simple proof term.

    check fun H : p ∧ q, and_intro (and_elimr H) (and_eliml H)

Note that the proof term is very similar to a function that just swaps the elements of a pair.

Or (disjuction)

The expression or_introl H1 b creates a proof for a b using a proof H1 : a. Similarly, or_intror a H2 creates a proof for a b using a proof H2 : b. We say they are the left/right or-introduction.

   -- Proof for p → p  q
   check fun H : p, or_introl H q
   -- Proof for q → p  q
   check fun H : q, or_intror p H

The or-elimination rule is slightly more complicated. The basic idea is the following, we can prove c from a b, by showing we can prove c by assuming a or by assuming b. It is essentially a proof by cases. or_elim Hab Hac Hbc takes three arguments Hab : a b, Hac : a → c and Hbc : b → c and produces a proof for c. In the following example, we use or_elim to prove that p v q → q p.

   check fun H : p  q,
           or_elim H
              (fun Hp : p, or_intror q Hp)
              (fun Hq : q, or_introl Hq p)

Not (negation)

not_intro H produces a proof for ¬ a from H : a → false. That is, we obtain ¬ a if we can derive false from a. The expression absurd_elim b Ha Hna produces a proof for b from Ha : a and Hna : ¬ a. That is, we can deduce anything if we have a and ¬ a. We now use not_intro and absurd_elim to produce a proof term for (a → b) → ¬ b → ¬ a

   variables a b : Bool
   check fun (Hab : a → b) (Hnb : ¬ b),
             not_intro (fun Ha : a, absurd_elim false (Hab Ha) Hnb)

Here is the proof term for ¬ a → b → (b → a) → c

   variable c : Bool
   check fun (Hna : ¬ a) (Hb : b) (Hba : b → a),
             absurd_elim c (Hba Hb) Hna

Iff (if-and-only-if)

The expression iff_intro H1 H2 produces a proof for a ↔ b from H1 : a → b and H2 : b → a. iff_eliml H produces a proof for a → b from H : a ↔ b. Similarly, iff_elimr H produces a proof for b → a from H : a ↔ b. Note that, in Lean, a ↔ b is definitionally equal to a = b when a and b have type Bool. Here is the proof term for a ∧ b ↔ b ∧ a

   check iff_intro (fun H : a ∧ b, and_intro (and_elimr H) (and_eliml H))
                   (fun H : b ∧ a, and_intro (and_elimr H) (and_eliml H))

True and False

The expression trivial is a proof term for true, and false_elim a H produces a proof for a from H : false.

Other basic operators used in proof construction are eqt_intro, eqt_elim, eqf_intro and eqf_elim. eqt_intro H produces a proof for a ↔ true from H : a. eqt_elim H produces a proof for a from H : a ↔ true. eqf_intro H produces a proof for a ↔ false from H : ¬ a. eqf_elim H produces a proof for ¬ a from H : a ↔ false.

    check @eqt_intro
    check @eqt_elim
    check @eqf_intro
    check @eqf_elim

Rewrite rules

The Lean kernel also contains many theorems that are meant to be used as rewriting/simplification rules. The conclusion of these theorems is of the form t = s or t ↔ s. For example, and_id a is proof term for a ∧ a ↔ a. The Lean simplifier can use these theorems to automatically create proof terms for us. The expression (by simp [rule-set]) is similar to _, but it tells Lean to synthesize the proof term using the simplifier using the rewrite rule set named [rule-set]. In the following example, we create a simple rewrite rule set and use it to prove a theorem that would be quite tedious to prove by hand.

    -- import module that defines several tactics/strategies including "simp"
    import tactic
    -- create a rewrite rule set with name 'simple'
    rewrite_set simple
    -- add some theorems to the rewrite rule set 'simple'
    add_rewrite and_id and_truer and_truel and_comm and_assoc and_left_comm iff_id : simple
    theorem th1 (a b : Bool) : a ∧ b ∧ true ∧ b ∧ true ∧ b ↔ a ∧ b
    := (by simp simple)

In Lean, we can combine manual and automated proofs in a natural way. We can manually write the proof skeleton and use the by construct to invoke automated proof engines like the simplifier for filling the tedious steps. Here is a very simple example.

   theorem th2 (a b : Bool) : a ∧ b ↔ b ∧ a
   := iff_intro
        (fun H : a ∧ b, (by simp simple))
        (fun H : b ∧ a, (by simp simple))

Dependent functions and quantifiers

Lean supports dependent functions. In type theory, they are also called dependent product types or Pi-types. The idea is quite simple, suppose we have a type A in some universe (Type i), and a family of types B : A → (Type j) which assigns to each a : A a type B a. So a dependent function is a function whose range varies depending on its arguments. In lean, the dependent functions is written as forall a : A, B a, or ∀ x : A, B a using unicode. The proposition as types paradigm is based on dependent functions. In the previous examples, we have seen many examples of dependent functions. The theorems refl, trans and symm, and the equality are all dependent functions,

    check @refl
    check @trans
    check @symm
    check @eq

The universal quantifier is also a dependent function. In Lean, if we have a family of types B : A → Bool, then ∀ x : A, B a has type Bool. This features complicates the Lean set-theoretic model, but it improves usability. Several theorem provers have a forall elimination (aka instantiation) proof rule. In Lean (and other systems based on proposition as types), this rule is just function application. In the following example we add an axiom stating that f x is 0 forall x. Then we instantiate the axiom using function application.

    variable f : Nat → Nat
    axiom fzero : ∀ x, f x = 0
    check fzero 1
    check fzero x

Since we instantiate quantifiers using function application, it is natural to create proof terms for universal quantifiers using lambda abstraction. In the following example, we create a proof term showing that for all x and y, f x = f y.

   check λ x y, trans (fzero x) (symm (fzero y))

We can view the proof term above as a simple function or "recipe" for showing that f x = f y for any x and y. The function "invokes" fzero for creating proof terms for f x = 0 and f y = 0. Then, it uses symmetry symm to create a proof term for 0 = f y. Finally, transitivity is used to combine the proofs for f x = 0 and 0 = f y.

In Lean, the existential quantifier exists x : A, B x is defined as ¬ forall x : A, ¬ B x. We can also write existential quantifiers as ∃ x : A, B x. Actually both versions are just notational convenience for Exists A (fun x : A, B x). That is, the existential quantifier is actually a constant defined in the file kernel.lean. This file also defines the exists_intro and exists_elim theorems. To build a proof for ∃ x : A, B x, we should provide a term w : A and a proof term Hw : B w to exists_intro. We say w is the witness for the existential introduction. In previous examples, nat_trans3i Hxy Hzy Hzw was a proof term for x = w. Then, we can create a proof term for ∃ a : Nat, a = w using

    theorem ex_a_eq_w : exists a, a = w  := exists_intro x (nat_trans3i Hxy Hzy Hzw)
    check ex_a_eq_w

Note that exists_intro also has implicit arguments. For example, Lean has to infer the implicit argument P : A → Bool, a predicate (aka function to Bool). This creates complications. For example, suppose we have Hg : g 0 0 = 0 and we invoke exists_intro 0 Hg. There are different possible values for P. Each possible value corresponds to a different theorem: ∃ x, g x x = x, ∃ x, g x x = 0, ∃ x, g x 0 = x, etc. Lean uses the context where exists_intro occurs to infer the users intent. In the example above, we were trying to prove the theorem ∃ a, a = w. So, we are implicitly telling Lean how to choose P. In the following example, we demonstrate this issue. We ask Lean to display the implicit arguments using the option pp::implicit. We see that each instance of exists_intro 0 Hg has different values for the implicit argument P.

   check @exists_intro
   variable g : Nat → Nat → Nat
   axiom Hg : g 0 0 = 0
   theorem gex1 : ∃ x, g x x = x := exists_intro 0 Hg
   theorem gex2 : ∃ x, g x 0 = x := exists_intro 0 Hg
   theorem gex3 : ∃ x, g 0 0 = x := exists_intro 0 Hg
   theorem gex4 : ∃ x, g x x = 0 := exists_intro 0 Hg
   set_option pp::implicit true  -- display implicit arguments
   print environment 4           -- print the last four theorems
   set_option pp::implicit false -- hide implicit arguments

We can view exists_intro (aka existential introduction) as an information hiding procedure. We are "hiding" what is the witness for some fact. The existential elimination performs the opposite operation. The exists_elim theorem allows us to prove some proposition B from ∃ x : A, B x if we can derive B using an "abstract" witness w and a proof term Hw : B w.

   check @exists_elim

In the following example, we define even a as ∃ b, a = 2*b, and then we show that the sum of two even numbers is an even number.

   definition even (a : Nat) := ∃ b, a = 2*b
   theorem EvenPlusEven {a b : Nat} (H1 : even a) (H2 : even b) : even (a + b)
   := exists_elim H1 (fun (w1 : Nat) (Hw1 : a = 2*w1),
      exists_elim H2 (fun (w2 : Nat) (Hw2 : b = 2*w2),
           exists_intro (w1 + w2)
              (calc a + b  =  2*w1 + b      : { Hw1 }
                      ...  =  2*w1 + 2*w2   : { Hw2 }
                      ...  =  2*(w1 + w2)   : symm (distributer 2 w1 w2))))

The example above also uses calculational proofs to show that a + b = 2*(w1 + w2). The calc construct is just syntax sugar for creating proofs using transitivity and substitution.

The module macros provides notation for making proof terms more readable. For example, it defines the obtain _, from _, _ macro as syntax sugar for exists_elim. With this macro we can write the example above as:

   import macros
   theorem EvenPlusEven2 {a b : Nat} (H1 : even a) (H2 : even b) : even (a + b)
   := obtain (w1 : Nat) (Hw1 : a = 2*w1), from H1,
      obtain (w2 : Nat) (Hw2 : b = 2*w2), from H2,
           exists_intro (w1 + w2)
              (calc a + b  =  2*w1 + b      : { Hw1 }
                      ...  =  2*w1 + 2*w2   : { Hw2 }
                      ...  =  2*(w1 + w2)   : symm (distributer 2 w1 w2))

The module macros also defines take x : A, H and assume x : A, H as syntax sugar for fun x : A, H. This may been silly, but it allows us to simulate Mizar-style declarative proofs in Lean. Using these macros, we can write

   definition Set (A : Type) : Type := A → Bool

   definition element {A : Type} (x : A) (s : Set A) := s x
   infix 60 ∈ : element

   definition subset {A : Type} (s1 : Set A) (s2 : Set A) := ∀ x, x ∈ s1 → x ∈ s2
   infix 50 ⊆ : subset

   theorem subset_trans {A : Type} {s1 s2 s3 : Set A} (H1 : s1 ⊆ s2) (H2 : s2 ⊆ s3) : s1 ⊆ s3
   := take x : A,
        assume Hin : x ∈ s1,
            show x ∈ s3, from
               let L1 : x ∈ s2 := H1 x Hin
               in H2 x L1

Finally, the construct show A, from H means "have" a proof for A using H. It is just syntax sugar for let H_show : A := H in H_show. It is useful to document intermediate steps in manually constructed proofs.