Chapter 9 of Test-Driven Development with Idris presents the following data type and removeElem function.
import Data.Vect
data MyElem : a -> Vect k a -> Type where
MyHere : MyElem x (x :: xs)
MyThere : (later : MyElem x xs) -> MyElem x (y :: xs)
-- I slightly modified the definition of this function from the text.
removeElem : (value : a) -> (xs : Vect (S n) a) -> (prf : MyElem value xs) -> Vect n a
removeElem value (value :: ys) MyHere = ys
removeElem value (y :: ys) (MyThere later) = removeElem value (y :: ys) (MyThere later)
The following works:
*lecture> removeElem 1 [1,2,3] MyHere
[2, 3] : Vect 2 Integer
But, the following call is still running after a few minutes:
*lecture> removeElem 2 [1,2,3] (MyThere MyHere)
Why is this, I'm assuming, compilation so slow?
The second case of your removeElem reads
removeElem value (y :: ys) (MyThere later) = removeElem value (y :: ys) (MyThere later)
The right-hand side is exactly the same as the left-hand side; so your recursion diverges. This is why evaluation hangs.
Note that Idris would have caught this error if you declared that removeElem should be total:
total removeElem : (value : a) -> (xs : Vect (S n) a) -> (prf : MyElem value xs) -> Vect n a
removeElem value (value :: ys) MyHere = ys
removeElem value (y :: ys) (MyThere later) = removeElem value (y :: ys) (MyThere later)
which results in the compile-time error
RemoveElem.idr line 9 col 0:
Main.removeElem is possibly not total due to recursive path Main.removeElem
Related
cong and injective allow you to apply and unapply functions to equalities:
cong : (f : a -> b) -> x = y -> f x = f y
injective : Injective f => f x = f y -> x = y
Both of these fail for indexed vectors with different lengths, for obvious reasons.
How can I prove that two equal vectors have the same length? I.e.
sameLen : {xs : Vect n a} -> {ys : Vect m b} -> xs = ys -> n = m
I can't just do
sameLen pf = cong length pf
because length on xs has type Vect n a -> Nat and length on ys has type Vect m b -> Nat. (In fact, I'm not even sure how to prove the same thing for two regular Lists, due to the differing type arguments, never mind with the added indices).
Going the other way, how would I prove something like
data Rose a = V a | T (Vect n (Rose a))
Injective T where
injective Refl = Refl
unwrap : {xs : Vect n (Rose a)} -> {ys : Vect m (Rose b)} -> T xs = T ys -> xs = ys
Again, I can't just do
unwrap pf = injective pf
due to the differing types of T (one with m and one with n). And even if I had a proof m=n, how could I use that to convince Idris that the two applications of T are the same?
Got the answer from the Idris Discord - if you pattern match on Refl then it unifies a and b automatically:
sameLen : {xs : List a} -> {ys : List b} -> xs = ys -> length xs = length ys
sameLen Refl = Refl
sameLen' : {xs : Vect n a} -> {ys : Vect m b} -> xs = ys -> n = m
sameLen' Refl = Refl
Context: I have been trying to implement the unification algorithm (the algorithm to find the most general unifier of two abstract syntax trees). Since a unifier is a substitution, algorithm requires defining composition of substitutions.
To be specific, given a type treeSigma dependent on another type X, a substitution is a function of type:
X -> treeSigma X
and the function substitute takes a substitution as an input and has type
substitute: (X-> (treeSigma X))-> (treeSigma X) -> (treeSigma X)
I need to define a function to compose two substitutions:
compose_kleisli (rho1 rho2: X->(treeSigma X)) : (treeSigma X) := ...
such that,
forall tr: treeSigma X,
substitute (compose_kleisli rho1 rho2) tr = substitute rho1 (substitute rho2 tr).
I am fairly new to coq and have been stuck with defining this composition.
How can I define this composition?
I tried to define it using Record like this:
Record compose {X s} (rho1 rho2: X-> treeSigma X):= mkCompose{
RHO: X-> treeSigma X;
CONDITION: forall t, substitute RHO t = substitute rho2 (substitute rho1 t)
}.
but along with this, I would need to prove the result that the composition can be defined for any two substitutions. Something like:
Theorem composeTotal: forall {X s} (rho1 rho2: X-> treeSigma s X), exists rho3,
forall t, substitute rho3 t = substitute rho2 (substitute rho1 t).
Proving this would require a construction of rho3 which circles back to the same problem of defining compose.
treeSigma is defined as:
(* Signature *)
Record sigma: Type := mkSigma {
symbol : Type;
arity : symbol -> nat
}.
Record sigmaLeaf (s:sigma): Type := mkLeaf {
cLeaf: symbol s;
condLeaf: arity s cLeaf = 0
}.
Record sigmaNode (s:sigma): Type := mkNode {
fNode: symbol s;
condNode: arity s fNode <> 0
}.
(* Sigma Algebra *)
Record sigAlg (s:sigma) (X:Type) := mkAlg {
Carrier: Type;
meaning: forall f:(sigmaNode s), (Vector.t Carrier (arity s (fNode s f))) -> Carrier;
meanLeaf: forall f:(sigmaLeaf s), Vector.t Carrier 0 -> Carrier
}.
(* Abstract tree on arbitrary signature. *)
Inductive treeSigma (s:sigma) (X:Type):=
| VAR (x:X)
| LEAF (c: sigmaLeaf s)
| NODE (f: sigmaNode s) (sub: Vector.t (treeSigma s X) (arity s (fNode s f)) ).
(* Defining abstract syntax as a sigma algebra. *)
Definition meanTreeNode {s X} (f:sigmaNode s) (sub: Vector.t (treeSigma s X) (s.(arity)
(fNode s f))): treeSigma s X:= NODE s X f sub.
Definition meanTreeLeaf {s X} (c:sigmaLeaf s) (sub: Vector.t (treeSigma s X) 0) := LEAF s X c.
Definition treeSigAlg {s X} := mkAlg s X (treeSigma s X) meanTreeNode meanTreeLeaf.
The substitution function is defined as:
Fixpoint homoSigma1 {X:Type} {s} (A: sigAlg s X) (rho: X-> (Carrier s X A))
(wft: (treeSigma s X)) {struct wft}: (Carrier s X A) :=
match wft with
| VAR _ _ x => rho x
| LEAF _ _ c => meanLeaf s X A c []
| NODE _ _ f l2 => meanNode s X A f (
(fix homoSigVec k (l2:Vector.t _ k):= match l2 with
| [] => []
| t::l2s => (homoSigma1 A rho t):: (homoSigVec (vlen _ l2s) l2s)
end)
(arity s (fNode s f)) l2)
end.
Definition substitute {X s} (rho: X-> treeSigma s X) (t: treeSigma s X) := #homoSigma1 X s treeSigAlg rho t.
To be particular, a substitution is the homomorphic extension of rho (which is a variable valuation).
Definitions like this are challenging to work with because the tree type occurs recursively inside of another inductive type. Coq has trouble generating induction principles for these types on its own, so you need to help it a little bit. Here is a possible solution, for a slightly simplified set up:
Require Import Coq.Vectors.Vector.
Import VectorNotations.
Set Implicit Arguments.
Unset Strict Implicit.
Unset Printing Implicit Defensive.
Section Dev.
Variable symbol : Type.
Variable arity : symbol -> nat.
Record alg := Alg {
alg_sort :> Type;
alg_op : forall f : symbol, Vector.t alg_sort (arity f) -> alg_sort;
}.
Arguments alg_op {_} f _.
(* Turn off the automatic generation of induction principles.
This tree type does not distinguish between leaves and nodes,
since they only differ in their arity. *)
Unset Elimination Schemes.
Inductive treeSigma (X:Type) :=
| VAR (x:X)
| NODE (f: symbol) (args : Vector.t (treeSigma X) (arity f)).
Arguments NODE {X} _ _.
Set Elimination Schemes.
(* Manual definition of a custom induction principle for treeSigma.
HNODE is the inductive case for the NODE constructor; the vs argument is
saying that the induction hypothesis holds for each tree in the vector of
arguments. *)
Definition treeSigma_rect (X : Type) (T : treeSigma X -> Type)
(HVAR : forall x, T (VAR x))
(HNODE : forall f (ts : Vector.t (treeSigma X) (arity f))
(vs : Vector.fold_right (fun t V => T t * V)%type ts unit),
T (NODE f ts)) :
forall t, T t :=
fix loopTree (t : treeSigma X) : T t :=
match t with
| VAR x => HVAR x
| NODE f ts =>
let fix loopVector n (ts : Vector.t (treeSigma X) n) :
Vector.fold_right (fun t V => T t * V)%type ts unit :=
match ts with
| [] => tt
| t :: ts => (loopTree t, loopVector _ ts)
end in
HNODE f ts (loopVector (arity f) ts)
end.
Definition treeSigma_ind (X : Type) (T : treeSigma X -> Prop) :=
#treeSigma_rect X T.
Definition treeSigma_alg (X:Type) : alg := {|
alg_sort := treeSigma X;
alg_op := #NODE X;
|}.
Fixpoint homoSigma {X : Type} {Y : alg} (ρ : X -> Y) (t : treeSigma X) : Y :=
match t with
| VAR x => ρ x
| NODE f xs => alg_op f (Vector.map (homoSigma ρ) xs)
end.
Definition substitute X (ρ : X -> treeSigma X) (t : treeSigma X) : treeSigma X :=
#homoSigma X (treeSigma_alg X) ρ t.
(* You can define composition simply by applying using substitution. *)
Definition compose X (ρ1 ρ2 : X -> treeSigma X) : X -> treeSigma X :=
fun x => substitute ρ1 (ρ2 x).
(* The property you are looking for follows by induction on the tree. Note
that this requires a nested induction on the vector of arguments. *)
Theorem composeP X (ρ1 ρ2 : X -> treeSigma X) t :
substitute (compose ρ1 ρ2) t = substitute ρ1 (substitute ρ2 t).
Proof.
unfold compose, substitute.
induction t as [x|f ts IH]; trivial.
simpl; f_equal.
induction ts as [|n t ts IH']; trivial.
simpl.
destruct IH as [e IH].
rewrite e.
f_equal.
now apply IH'.
Qed.
End Dev.
In order to do this you need to use the operations of the monad, typically:
Set Implicit Arguments.
Unset Strict Implicit.
Unset Printing Implicit Defensive.
Section MonadKleisli.
(* Set Universe Polymorphism. // Needed for real use cases *)
Variable (M : Type -> Type).
Variable (Ma : forall A B, (A -> B) -> M A -> M B).
Variable (η : forall A, A -> M A).
Variable (μ : forall A, M (M A) -> M A).
(* Compose: o^* *)
Definition oStar A B C (f : A -> M B) (g: B -> M C) : A -> M C :=
fun x => μ (Ma g (f x)).
(* Bind *)
Definition bind A B (x : M A) (f : A -> M B) : M B := oStar (fun _ => x) f tt.
End MonadKleisli.
Depending on how you organize your definitions, proving your desired properties will likely require functional extensionality, not a big deal usually but something to keep in ind.
Assume I have the following idris source code:
module Source
import Data.Vect
--in order to avoid compiler confusion between Prelude.List.(++), Prelude.String.(++) and Data.Vect.(++)
infixl 0 +++
(+++) : Vect n a -> Vect m a -> Vect (n+m) a
v +++ w = v ++ w
--NB: further down in the question I'll assume this definition isn't needed because the compiler
-- will have enough context to disambiguate between these and figure out that Data.Vect.(++)
-- is the "correct" one to use.
lemma : reverse (n :: ns) +++ (n :: ns) = reverse ns +++ (n :: n :: ns)
lemma {ns = []} = Refl
lemma {ns = n' :: ns} = ?lemma_rhs
As shown, the base case for lemma is trivially Refl. But I can't seem to find a way to prove the inductive case: the repl "just" spits out the following
*source> :t lemma_rhs
phTy : Type
n1 : phTy
len : Nat
ns : Vect len phTy
n : phTy
-----------------------------------------
lemma_rhs : Data.Vect.reverse, go phTy
(S (S len))
(n :: n1 :: ns)
[n1, n]
ns ++
n :: n1 :: ns =
Data.Vect.reverse, go phTy (S len) (n1 :: ns) [n1] ns ++
n :: n :: n1 :: ns
I understand that phTy stands for "phantom type", the implicit type of the vectors I'm considering. I also understand that go is the name of the function defined in the where clause for the definition of the library function reverse.
Question
How can I continue the proof? Is my inductive strategy sound? Is there a better one?
Context
This has came up in one of my toy projects, where I try to define arbitrary tensors; specifically, this seems to be needed in order to define "full index contraction". I'll elaborate a little bit on that:
I define tensors in a way that's roughly equivalent to
data Tensor : (rank : Nat) -> (shape : Vector rank Nat) -> Type where
Scalar : a -> Tensor Z [] a
Vector : Vect n (Tensor rank shape a) -> Tensor (S rank) (n :: shape) a
glossing over the rest of the source code (since it isn't relevant, and it's quite long and uninteresting as of now), I was able to define the following functions
contractIndex : Num a =>
Tensor (r1 + (2 + r2)) (s1 ++ (n :: n :: s2)) a ->
Tensor (r1 + r2) (s1 ++ s2) a
tensorProduct : Num a =>
Tensor r1 s1 a ->
Tensor r2 s2 a ->
Tensor (r1 + r2) (s1 ++ s2) a
contractProduct : Num a =>
Tensor (S r1) s1 a ->
Tensor (S r2) ((last s1) :: s2) a ->
Tensor (r1 + r2) ((take r1 s1) ++ s2) a
and I'm working on this other one
fullIndexContraction : Num a =>
Tensor r (reverse ns) a ->
Tensor r ns a ->
Tensor 0 [] a
fullIndexContraction {r = Z} {ns = []} t s = t * s
fullIndexContraction {r = S r} {ns = n :: ns} t s = ?rhs
that should "iterate contractProduct as much as possible (that is, r times)"; equivalently, it could be possible to define it as tensorProduct composed with as many contractIndex as possible (again, that amount should be r).
I'm including all this becuse maybe it's easier to just solve this problem without proving the lemma above: if that were the case, I'd be fully satisfied as well. I just thought the "shorter" version above might be easier to deal with, since I'm pretty sure I'll be able to figure out the missing pieces myself.
The version of idris i'm using is 1.3.2-git:PRE (that's what the repl says when invoked from the command line).
Edit: xash's answer covers almost everything, and I was able to write the following functions
nreverse_id : (k : Nat) -> nreverse k = k
contractAllIndices : Num a =>
Tensor (nreverse k + k) (reverse ns ++ ns) a ->
Tensor Z [] a
contractAllProduct : Num a =>
Tensor (nreverse k) (reverse ns) a ->
Tensor k ns a ->
Tensor Z []
I also wrote a "fancy" version of reverse, let's call it fancy_reverse, that automatically rewrites nreverse k = k in its result. So I tried to write a function that doesn't have nreverse in its signature, something like
fancy_reverse : Vect n a -> Vect n a
fancy_reverse {n} xs =
rewrite sym $ nreverse_id n in
reverse xs
contract : Num a =>
{auto eql : fancy_reverse ns1 = ns2} ->
Tensor k ns1 a ->
Tensor k ns2 a ->
Tensor Z [] a
contract {eql} {k} {ns1} {ns2} t s =
flip contractAllProduct s $
rewrite sym $ nreverse_id k in
?rhs
now, the inferred type for rhs is Tensor (nreverse k) (reverse ns2) and I have in scope a rewrite rule for k = nreverse k, but I can't seem to wrap my head around how to rewrite the implicit eql proof to make this type check: am I doing something wrong?
The prelude Data.Vect.reverse is hard to reason about, because AFAIK the go helper function won't be resolved in the typechecker. The usual approach is to define oneself an easier reverse that doesn't need rewrite in the type level. Like here for example:
%hide Data.Vect.reverse
nreverse : Nat -> Nat
nreverse Z = Z
nreverse (S n) = nreverse n + 1
reverse : Vect n a -> Vect (nreverse n) a
reverse [] = []
reverse (x :: xs) = reverse xs ++ [x]
lemma : {xs : Vect n a} -> reverse (x :: xs) = reverse xs ++ [x]
lemma = Refl
As you can see, this definition is straight-forward enough, that this equivalent lemma can be solved without further work. Thus you can probably just match on the reverse ns in fullIndexContraction like in this example:
data Foo : Vect n Nat -> Type where
MkFoo : (x : Vect n Nat) -> Foo x
foo : Foo a -> Foo (reverse a) -> Nat
foo (MkFoo []) (MkFoo []) = Z
foo (MkFoo $ x::xs) (MkFoo $ reverse xs ++ [x]) =
x + foo (MkFoo xs) (MkFoo $ reverse xs)
To your comment: first, len = nreverse len must sometimes be used, but if you had rewrite on the type level (through the usual n + 1 = 1 + n shenanigans) you had the same problem (if not even with more complicated proofs, but this is just a guess.)
vectAppendAssociative is actually enough:
lemma2 : Main.reverse (n :: ns1) ++ ns2 = Main.reverse ns1 ++ (n :: ns2)
lemma2 {n} {ns1} {ns2} = sym $ vectAppendAssociative (reverse ns1) [n] ns2
I'm a beginner in Idris and trying to make a code valid.
Could you let me know the better place for noob questions on Idris?
filter : (elem -> Bool) -> Vect len elem -> (p: (Fin len) ** Vect (finToNat p) elem)
filter {len=S l} p xs = ((FZ {k=l}) ** [])
filter {len=S l} p (x::xs) =
let (a ** tail) = filter {len=l} p xs
in if p x then
((FS a) ** x::tail)
else
((weaken a) ** tail)
I wrote another filter which can't pass the type check yet.
This new filter's type implies that the filtered vector cannot be longer than the original one.
However, Idris saids
...
Specifically:
Type mismatch between
finToNat a
and
finToNat (weaken a)
We know those two terms always have the same value.
How can I describe the fact and let Idris say ok?
You have to show that finToNat a = finToNat (weaken a). tail has type Vect (finToNat a) elem, but you need Vect (finToNat (weaken a)) elem for the second component in the last line, because you wrote weaken a in the first pair component.
lemma : {n : _} -> (a : Fin n) -> finToNat (weaken a) = finToNat a
lemma FZ = Refl
lemma (FS x) = rewrite lemma x in Refl
filter : (elem -> Bool) -> Vect len elem -> (p: (Fin len) ** Vect (finToNat p) elem)
filter {len=S l} p xs = ((FZ {k=l}) ** [])
filter {len=S l} p (x::xs) =
let (a ** tail) = Main.filter {len=l} p xs
in if p x then
((FS a) ** x::tail)
else
(weaken a ** (rewrite lemma a in tail))
Suppose I have some natural numbers d ≥ 2 and n > 0; in this case, I can split off the d's from n and get n = m * dk, where m is not divisible by d.
I'd like to use this repeated removal of the d-divisible parts as a recursion scheme; so I thought I'd make a datatype for the Steps leading to m:
import Data.Nat.DivMod
data Steps: (d : Nat) -> {auto dValid: d `GTE` 2} -> (n : Nat) -> Type where
Base: (rem: Nat) -> (rem `GT` 0) -> (rem `LT` d) -> (quot : Nat) -> Steps d {dValid} (rem + quot * d)
Step: Steps d {dValid} n -> Steps d {dValid} (n * d)
and write a recursive function that computes the Steps for a given pair of d and n:
total lemma: x * y `GT` 0 -> x `GT` 0
lemma {x = Z} LTEZero impossible
lemma {x = Z} (LTESucc _) impossible
lemma {x = (S k)} prf = LTESucc LTEZero
steps : (d : Nat) -> {auto dValid: d `GTE` 2} -> (n : Nat) -> {auto nValid: n `GT` 0} -> Steps d {dValid} n
steps Z {dValid = LTEZero} _ impossible
steps Z {dValid = (LTESucc _)} _ impossible
steps (S d) {dValid} n {nValid} with (divMod n d)
steps (S d) (q * S d) {nValid} | MkDivMod q Z _ = Step (steps (S d) {dValid} q {nValid = lemma nValid})
steps (S d) (S rem + q * S d) | MkDivMod q (S rem) remSmall = Base (S rem) (LTESucc LTEZero) remSmall q
However, steps is not accepted as total since there's no apparent reason why the recursive call is well-founded (there's no structural relationship between q and n).
But I also have a function
total wf : (S x) `LT` (S x) * S (S y)
with a boring proof.
Can I use wf in the definition of steps to explain to Idris that steps is total?
Here is one way of using well-founded recursion to do what you're asking. I'm sure though, that there is a better way. In what follows I'm going to use the standard LT function, which allows us to achieve our goal, but there some obstacles we will need to work around.
Unfortunately, LT is a function, not a type constructor or a data constructor, which means we cannot define an implementation of the
WellFounded
typeclass for LT. The following code is a workaround for this situation:
total
accIndLt : {P : Nat -> Type} ->
(step : (x : Nat) -> ((y : Nat) -> LT y x -> P y) -> P x) ->
(z : Nat) -> Accessible LT z -> P z
accIndLt {P} step z (Access f) =
step z $ \y, lt => accIndLt {P} step y (f y lt)
total
wfIndLt : {P : Nat -> Type} ->
(step : (x : Nat) -> ((y : Nat) -> LT y x -> P y) -> P x) ->
(x : Nat) -> P x
wfIndLt step x = accIndLt step x (ltAccessible x)
We are going to need some helper lemmas dealing with the less than relation, the lemmas can be found in this gist (Order module). It's a subset of my personal library which I recently started. I'm sure the proofs of the helper lemmas can be minimized, but it wasn't my goal here.
After importing the Order module, we can solve the problem (I slightly modified the original code, it's not difficult to change it or write a wrapper to have the original type):
total
steps : (n : Nat) -> {auto nValid : 0 `LT` n} -> (d : Nat) -> Steps (S (S d)) n
steps n {nValid} d = wfIndLt {P = P} step n d nValid
where
P : (n : Nat) -> Type
P n = (d : Nat) -> (nV : 0 `LT` n) -> Steps (S (S d)) n
step : (n : Nat) -> (rec : (q : Nat) -> q `LT` n -> P q) -> P n
step n rec d nV with (divMod n (S d))
step (S r + q * S (S d)) rec d nV | (MkDivMod q (S r) prf) =
Base (S r) (LTESucc LTEZero) prf q
step (Z + q * S (S d)) rec d nV | (MkDivMod q Z _) =
let qGt0 = multLtNonZeroArgumentsLeft nV in
let lt = multLtSelfRight (S (S d)) qGt0 (LTESucc (LTESucc LTEZero)) in
Step (rec q lt d qGt0)
I modeled steps after the divMod function from the contrib/Data/Nat/DivMod/IteratedSubtraction.idr module.
Full code is available here.
Warning: the totality checker (as of Idris 0.99, release version) does not accept steps as total. It has been recently fixed and works for our problem (I tested it with Idris 0.99-git:17f0899c).