Type inference

30 07 2008

Intro

The type inference is the ability to a programming language deduct the data types of all functions of a program. It is a feature present in many strongly typed languages, such as Haskell. Where is not mandatory write the signature of the functions. What is great because it increases the production of code, and security because if the inference algorithm fail means that we have an error of types in our code, all this in compilation time.

As I said in previous post, one of the upgrades of Pointfree calculator is the type inference. After reading some scientific articles about Damas-Milner algorithm, also known as W algorithm, I began to imagine a way to implement this in Java, which is the mother language of Pointfree calculator. I started to do some sketches on paper and, after talking with professor José Nuno Oliveira, I realize that the algorithm isn’t that hard.

Remainings

Definition of (either in Haskell):

NOTE: in types means Either in Haskell



Definition of :

Type signature of Left and Right:


Talking same language

I will use a different notation to represent the domain and codomain of functions in order to help the explanation of the algorithm.

For function we have the type:

I will write that as:

Remember the definition of , we receive two functions, f and g. Because the notation is in pointfree, we represent also de domain and codomain of function in front of that, like we do for f and g.
In fact the type of is represented as:

I will also use the symbol , to say that type a unify with type b, that means, informally, that .

Let’s infer!

I will explain the algorithm to infer the type of function f:

The first step of the algorithm is attribute to all functions polymorphic types, so I will call the first type and the last

Because, have type , we conclude ;
Also, because have the type , we can conclude ;
Same thing to , that have the type , we can conclude and , so we have:

Because, the definition of : , we can say that the domain of f is equal to codomain of g, and so we can conclude , as we replace a type that is used in the codomain of first Right, we must also conclude , so:

As I explain before, the function , have the following type: , so:
and ;
Because have the type: , so and :

Because the definition of is , we need the same codomain in both functions, so we conclude , as both type trees have the same structure, we can conclude even more: , so:

And now we have the function, just with the needed types to simplify:

.

Now we just need to unify: and ,

.

We infer the type for function , .
Or if you prefer; in Haskell:

f :: Either (Either a b) c -> Either a (Either b c)




Pointfree Calculator

29 07 2008

Intro

In this semester had to do this project with my friend, João Moura, under supervision of professor Alcino Cunha.

The project was to make a program that make automatic the process of proving an equality of functions written in the pointfree style. Not completely automatic, but so iterative. The program load a file with rules and the user inserts an equality of functions in pointfree and can surf the words to find a sub-expression which want to run a particular rule. From the side are shown all the rules that may apply to selected sub-expression.

What is pointfree?

Pointfree is a style of programming that helps us to compose functions. Even more, it helps us later to prove equality between functions.

How can someone prove equality between functions? We can do so, because we have rules. Let’s see an analogy with algebraic calculus.







We have proved that .

Pointfree calculus as is done now, on paper, is boring and repetitive. A proof is this image of someone trying to make a proof.

 

Abstract representation

If swap is a functions for exchange the elements of a pair:

swap (a,b) = (b,a) --in pointwise
swap = split snd fst --in pointfree

And we want to prove that: . This is clearly obvious, but I will use this example just to show you the representation that we use to see these proofs.

So, here we have the representation of :

If you make a inorder passage it make sense.

As you can see the operator have arity n. Because we want that this operator became flat, we never want to use the law:
.

Function f is variable, and swap is constant, we know their definition.

Imagine now that we load the following file of rules:





Path

In order to know that sub-expression are to select a term, we have the Path which behaves as follows:

If we select 4 in the following expression , the Path becomes: .
If we select 2, the Path becomes: .

the list in front of the Path is to select a range of elements in an operator with arity n. Thus, if we want select in , the Path becomes: .

Proof of

This process is pretty always the same, but I suggest that you follow with some attention.

In this first tree I will show the Path already selected.

Here I will show the rule that will apply in our expression, and the instantiation of that rule to our selected sub-expression.

  • Selected function:
  • Path:
  • Rule:
  • Instantiation of the rule:

So, the resulting tree will be:

Resulting function: .

Proof of

  • Selected function:
  • Path:
  • Rule:
  • Instantiation of the rule:

So, the resulting tree will be:

Resulting function: .

Proof of

  • Selected function:
  • Path:
  • Rule:
  • Instantiation of the rule:

So, the resulting tree will be:

Resulting function: .

Proof of

  • Selected function:
  • Path:
  • Rule:
  • Instantiation of the rule:

So, the resulting tree will be:

Resulting function: .

Proof of

  • Selected function:
  • Path:
  • Rule:
  • Instantiation of the rule:

So, the resulting tree will be:

Resulting function: .

Proof of

  • Selected function:
  • Path:
  • Rule:
  • Instantiation of the rule:

So, the resulting tree will be:

Resulting function: .

Proof of

  • Selected function:
  • Path:
  • Rule:
  • Instantiation of the rule:

So, the resulting tree will be:

Resulting function: .

Proof of

  • Selected function:
  • Path:
  • Rule:
  • Instantiation of the rule:

So, the resulting tree will be:

Resulting function: .

From the rule of equality that we have is true and therefore it is true

Presentation

Time to show the software interface.

The program’s interface is divided into 3 parts:

Input

Here the user can insert something that want to prove, for example:



and so on…

When you hit enter, you start having fun calculating 🙂

Rules that may apply

Here you can navigate in the rules pressing PgUp or PgDn, and enter to aply the selected rule to the selected sub-expression.

Proof state

Here you can browse the expression imagining it as the trees that showed earlier.

  • up/down – to navigate into the levels
  • left/right – no navigate into childs
  • backspace – to make undo in the proof

We can also save and load proofs into XML, and also save proofs to PDF.

Conclusion

Me and João still improving the pointfree calculator, and in the next year it will be used in one course here in Universidade do Minho.
The next stage we will implement type inference in the pointfree calculator, to make it powerfull. A lot of things have to be done, and in the next month we will start doing that, including make the first public release of the software.

If you understand Portuguese and want to see the presentation that we give in Department of Informatic in University of Minho: