liquidhaskell-0.8.10.7: docs/blog/2013-03-04-bounding-vectors.lhs
---
layout: post
title: "Bounding Vectors"
date: 2013-03-04 16:12
author: Ranjit Jhala
published: true
comments: true
external-url:
categories: basic
demo: vectorbounds.hs
---
Today, lets look at a classic use-case for refinement types, namely,
the static verification of **vector access bounds**. Along the way,
we'll see some examples that illustrate how LiquidHaskell reasons
about *recursion*, *higher-order functions*, *data types*, and
*polymorphism*.
<!-- more -->
\begin{code}
module VectorBounds (
safeLookup
, unsafeLookup, unsafeLookup'
, absoluteSum, absoluteSum'
, dotProduct
, sparseProduct, sparseProduct'
) where
import Prelude hiding (length)
import Data.List (foldl')
import Data.Vector hiding (foldl')
\end{code}
Specifying Bounds for Vectors
-----------------------------
One [classical][dmlarray] use-case for refinement types is to verify
the safety of accesses of arrays and vectors and such, by proving that
the indices used in such accesses are *within* the vector bounds.
Lets see how to do this with LiquidHaskell by writing a few short
functions that manipulate vectors, in particular, those from the
popular [vector][vec] library.
First things first. Lets **specify** bounds safety by *refining*
the types for the [key functions][vecspec] exported by the module
`Data.Vector`.
\begin{code}Specifications for `Data.Vector`
module spec Data.Vector where
import GHC.Base
measure vlen :: (Vector a) -> Int
assume length :: x:(Vector a) -> {v:Int | v = (vlen x)}
assume ! :: x:(Vector a) -> {v:Int | 0 <= v && v < (vlen x)} -> a
\end{code}
In particular, we
- **define** a *property* called `vlen` which denotes the size of the vector,
- **assume** that the `length` function *returns* an integer equal to the vector's size, and
- **assume** that the lookup function `!` *requires* an index between `0` and the vector's size.
There are several things worth paying close attention to in the above snippet.
**Measures**
[Recall][listtail] that measures define auxiliary (or so-called **ghost**)
properties of data values that are useful for specification and verification,
but which *don't actually exist at run-time*. Thus, they will
*only appear in specifications*, i.e. inside type refinements, but *never*
inside code. Often we will use helper functions like `length` in this case,
which *pull* or *materialize* the ghost values from the refinement world
into the actual code world.
**Assumes**
We write `assume` because in this scenario we are not *verifying* the
implementation of `Data.Vector`, we are simply *using* the properties of
the library to verify client code. If we wanted to verify the library
itself, we would ascribe the above types to the relevant functions in the
Haskell source for `Data.Vector`.
**Dependent Refinements**
Notice that in the function type (e.g. for `length`) we have *named* the *input*
parameter `x` so that we can refer to it in the *output* refinement.
\begin{code} In this case, the type
assume length :: x:(Vector a) -> {v : Int | v = (vlen x)}
\end{code}
states that the `Int` output is exactly equal to the size of the input `Vector` named `x`.
In other words, the output refinement **depends on** the input value, which
crucially allows us to write properties that *relate* different program values.
**Verifying a Simple Wrapper**
Lets try write some simple functions to sanity check the above specifications.
First, consider an *unsafe* vector lookup function:
\begin{code}
unsafeLookup vec index = vec ! index
\end{code}
If we run this through LiquidHaskell, it will spit back a type error for
the expression `x ! i` because (happily!) it cannot prove that `index` is
between `0` and the `vlen vec`. Of course, we can specify the bounds
requirement in the input type
\begin{code}
{-@ unsafeLookup :: vec:Vector a
-> {v: Int | (0 <= v && v < (vlen vec))}
-> a
@-}
\end{code}
then LiquidHaskell is happy to verify the lookup. Of course, now the burden
of ensuring the index is valid is pushed to clients of `unsafeLookup`.
Instead, we might write a *safe* lookup function that performs the *bounds check*
before looking up the vector:
\begin{code}
{-@ safeLookup :: Vector a -> Int -> Maybe a @-}
safeLookup x i
| 0 <= i && i < length x = Just (x ! i)
| otherwise = Nothing
\end{code}
**Predicate Aliases**
The type for `unsafeLookup` above is rather verbose as we have to spell out
the upper and lower bounds and conjoin them. Just as we enjoy abstractions
when programming, we will find it handy to have abstractions in the
specification mechanism. To this end, LiquidHaskell supports
*predicate aliases*, which are best illustrated by example
\begin{code}
{-@ predicate Btwn Lo I Hi = (Lo <= I && I < Hi) @-}
{-@ predicate InBounds I A = (Btwn 0 I (vlen A)) @-}
\end{code}
Now, we can simplify the type for the unsafe lookup function to
\begin{code}
{-@ unsafeLookup' :: x:Vector a -> {v:Int | (InBounds v x)} -> a @-}
unsafeLookup' :: Vector a -> Int -> a
unsafeLookup' x i = x ! i
\end{code}
Our First Recursive Function
----------------------------
OK, with the tedious preliminaries out of the way, lets write some code!
To start: a vanilla recursive function that adds up the absolute values of
the elements of an integer vector.
\begin{code}
absoluteSum :: Vector Int -> Int
absoluteSum vec = if 0 < n then go 0 0 else 0
where
go acc i
| i /= n = go (acc + abz (vec ! i)) (i + 1)
| otherwise = acc
n = length vec
\end{code}
where the function `abz` is the absolute value function from [before][ref101].
\begin{code}
abz n = if 0 <= n then n else (0 - n)
\end{code}
Digression: Introducing Errors
------------------------------
If you are following along in the demo page -- I heartily
recommend that you try the following modifications,
one at a time, and see what happens.
**What happens if:**
1. You *remove* the check `0 < n` (see `absoluteSumNT` in the demo code)
2. You *replace* the guard with `i <= n`
In the *former* case, LiquidHaskell will *verify* safety, but
in the *latter* case, it will grumble that your program is *unsafe*.
Do you understand why?
(Thanks to [smog_alado][smog_alado] for pointing this out :))
Refinement Type Inference
-------------------------
LiquidHaskell happily verifies `absoluteSum` -- or, to be precise,
the safety of the vector accesses `vec ! i`. The verification works
out because LiquidHaskell is able to automatically infer a suitable
type for `go`. Shuffle your mouse over the identifier above to see
the inferred type. Observe that the type states that the first
parameter `acc` (and the output) is `0 <= V`. That is, the returned
value is non-negative.
More importantly, the type states that the second parameter `i` is
`0 <= V` and `V <= n` and `V <= (vlen vec)`. That is, the parameter `i`
is between `0` and the vector length (inclusive). LiquidHaskell uses these
and the test that `i /= n` to establish that `i` is in fact between `0`
and `(vlen vec)` thereby verifing safety.
In fact, if we want to use the function externally (i.e. in another module)
we can go ahead and strengthen its type to specify that the output is
non-negative.
\begin{code}
{-@ absoluteSum :: Vector Int -> {v: Int | 0 <= v} @-}
\end{code}
**What happens if:** You *replace* the output type for `absoluteSum` with `{v: Int | 0 < v }` ?
Bottling Recursion With a Higher-Order `loop`
---------------------------------------------
Next, lets refactor the above low-level recursive function
into a generic higher-order `loop`.
\begin{code}
loop :: Int -> Int -> a -> (Int -> a -> a) -> a
loop lo hi base f = go base lo
where
go acc i
| i /= hi = go (f i acc) (i + 1)
| otherwise = acc
\end{code}
**Using `loop` to compute `absoluteSum`**
We can now use `loop` to implement `absoluteSum` like so:
\begin{code}
absoluteSum' vec = if 0 < n then loop 0 n 0 body else 0
where body = \i acc -> acc + (vec ! i)
n = length vec
\end{code}
LiquidHaskell verifies `absoluteSum'` without any trouble.
It is very instructive to see the type that LiquidHaskell *infers*
for `loop` -- it looks something like
\begin{code}
{-@ loop :: lo: {v: Int | (0 <= v) }
-> hi: {v: Int | (lo <= v) }
-> a
-> (i: {v: Int | (Btwn lo v hi)} -> a -> a)
-> a
@-}
\end{code}
In english, the above type states that
- `lo` the loop *lower* bound is a non-negative integer
- `hi` the loop *upper* bound is a greater than `lo`,
- `f` the loop *body* is only called with integers between `lo` and `hi`.
Inference is a rather convenient option -- it can be tedious to have to keep
typing things like the above! Of course, if we wanted to make `loop` a
public or exported function, we could use the inferred type to generate
an explicit signature too.
\begin{code}At the call
loop 0 n 0 body
\end{code}
the parameters `lo` and `hi` are instantiated with `0` and `n` respectively
(which, by the way is where the inference engine deduces non-negativity
from) and thus LiquidHaskell concludes that `body` is only called with
values of `i` that are *between* `0` and `(vlen vec)`, which shows the
safety of the call `vec ! i`.
**Using `loop` to compute `dotProduct`**
Here's another use of `loop` -- this time to compute the `dotProduct`
of two vectors.
\begin{code}
dotProduct :: Vector Int -> Vector Int -> Int
dotProduct x y = loop 0 (length x) 0 (\i -> (+ (x ! i) * (y ! i)))
\end{code}
The gimlet-eyed reader will realize that the above is quite unsafe -- what
if `x` is a 10-dimensional vector while `y` has only 3-dimensions?
\begin{code}A nasty
*** Exception: ./Data/Vector/Generic.hs:244 ((!)): index out of bounds ...
\end{code}
*Yech*.
This is precisely the sort of unwelcome surprise we want to do away with at
compile-time. Refinements to the rescue! We can specify that the vectors
have the same dimensions quite easily
\begin{code}
{-@ dotProduct :: x:(Vector Int)
-> y:{v: (Vector Int) | (vlen v) = (vlen x)}
-> Int
@-}
\end{code}
after which LiquidHaskell will gladly verify that the implementation of
`dotProduct` is indeed safe!
Refining Data Types
-------------------
Next, suppose we want to write a *sparse dot product*, that is,
the dot product of a vector and a **sparse vector** represented
by a list of index-value tuples.
**Representing Sparse Vectors**
We can represent the sparse vector with a **refinement type alias**
\begin{code}
{-@ type SparseVector a N = [({v: Int | (Btwn 0 v N)}, a)] @-}
\end{code}
As with usual types, the alias `SparseVector` on the left is just a
shorthand for the (longer) type on the right, it does not actually
define a new type. Thus, the above alias is simply a refinement of
Haskell's `[(Int, a)]` type, with a size parameter `N` that facilitates
easy specification reuse. In this way, refinements let us express
invariants of containers like lists in a straightforward manner.
**Aside:** If you are familiar with the *index-style* length
encoding e.g. as found in [DML][dml] or [Agda][agdavec], then note
that despite appearances, our `SparseVector` definition is *not*
indexed. Instead, we deliberately choose to encode properties
with logical refinement predicates, to facilitate SMT based
checking and inference.
**Verifying Uses of Sparse Vectors**
Next, we can write a recursive procedure that computes the sparse product
\begin{code}
{-@ sparseProduct :: (Num a) => x:(Vector a)
-> SparseVector a (vlen x)
-> a
@-}
sparseProduct x y = go 0 y
where
go sum ((i, v) : y') = go (sum + (x ! i) * v) y'
go sum [] = sum
\end{code}
LiquidHaskell verifies the above by using the specification for `y` to
conclude that for each tuple `(i, v)` in the list, the value of `i` is
within the bounds of the vector `x`, thereby proving the safety of the
access `x ! i`.
Refinements and Polymorphism
----------------------------
The sharp reader will have undoubtedly noticed that the sparse product
can be more cleanly expressed as a [fold][foldl].
\begin{code} Indeed! Let us recall the type of the `foldl` operation
foldl' :: (a -> b -> a) -> a -> [b] -> a
\end{code}
Thus, we can simply fold over the sparse vector, accumulating the `sum`
as we go along
\begin{code}
{-@ sparseProduct' :: (Num a) => x:(Vector a)
-> SparseVector a (vlen x)
-> a
@-}
sparseProduct' x y = foldl' body 0 y
where
body sum (i, v) = sum + (x ! i) * v
\end{code}
LiquidHaskell digests this too, without much difficulty.
The main trick is in how the polymorphism of `foldl'` is instantiated.
1. The GHC type inference engine deduces that at this site, the type variable
`b` from the signature of `foldl'` is instantiated to the Haskell type `(Int, a)`.
2. Correspondingly, LiquidHaskell infers that in fact `b` can be instantiated
to the *refined* type `({v: Int | (Btwn 0 v (vlen x))}, a)`.
Walk the mouse over to `i` to see this inferred type. (You can also hover over
`foldl'`above to see the rather more verbose fully instantiated type.)
Thus, the inference mechanism saves us a fair bit of typing and allows us to
reuse existing polymorphic functions over containers and such without ceremony.
Conclusion
----------
That's all for now folks! Hopefully the above gives you a reasonable idea of
how one can use refinements to verify size related properties, and more
generally, to specify and verify properties of recursive, and polymorphic
functions operating over datatypes. Next time, we'll look at how we can
teach LiquidHaskell to think about properties of recursive structures
like lists and trees.
[smog_alado]: http://www.reddit.com/user/smog_alado
[vecspec]: https://github.com/ucsd-progsys/liquidhaskell/blob/master/include/Data/Vector.spec
[vec]: http://hackage.haskell.org/package/vector
[dml]: http://www.cs.bu.edu/~hwxi/DML/DML.html
[agdavec]: http://code.haskell.org/Agda/examples/Vec.agda
[ref101]: /blog/2013/01/01/refinement-types-101.lhs/
[ref102]: /blog/2013/01/27/refinements-101-reax.lhs/
[foldl]: http://hackage.haskell.org/packages/archive/base/latest/doc/html/src/Data-List.html
[listtail]: /blog/2013/01/31/safely-catching-a-list-by-its-tail.lhs/
[dmlarray]: http://www.cs.bu.edu/~hwxi/academic/papers/pldi98.pdf