Exercise 1

Haskell functions can be polymorphic if we use type variables in their definitions. Write a function

permutations :: [a] -> [[a]]

taking a list of elements of type a and returning a list of all its permutations.

We will use the same approach as in Racket. First, we define a function

interleave :: a -> [a] -> [[a]]

taking an element x of type a and a list ys of elements of type a and returning a list of lists where x is plugged into ys by all possible ways. E.g.

> interleave 0 [1,2,3]
[[0,1,2,3],[1,0,2,3],[1,2,0,3],[1,2,3,0]]

The base case for the interleave function is simple as there is only a single way to plug x into the empty list [], namely [x]. If the list is of the form y:ys, then one way to plug x into it is to prepend x (i.e., x:y:ys). The remaining possibilities can be computed recursively by calling interleave x ys and prepending y.

Solution: interleave

interleave :: a -> [a] -> [[a]]
interleave x [] = [[x]]
interleave x (y:ys) = (x:y:ys) : [y:xs | xs <- interleave x ys]

Now we can easily define the permutations function. The base case for the empty list is trivial. For a nonempty list x:xs we can recursively compute permutations of xs and interleave x into all such permutations. Finally, we must concatenate the results into a single list of permutations. This can be done by the function concat implemented in Prelude (you also saw how to implement such a function in the lecture. I called that function flatten).

Solution: permutations

permutations :: [a] -> [[a]]
permutations [] = [[]]
permutations (x:xs) = concat [interleave x p | p <- permutations xs]

Exercise 2

Use the function permutations from the previous exercise to write a function findHamiltonian, which finds all Hamiltonian paths in a given graph.

We first have to represent graphs in a data structure. To be general, we define a graph data structure over any data type a. First, we define a type for edges as pairs of values of type a. Second, we define a parametric algebraic data type Graph a as a record consisting of a list of vertices and a list of edges. We also make this type an instance of the type class Show by automatic derivation.

type Edge a = (a,a)
data Graph a = Graph {vertices :: [a], edges :: [Edge a]} deriving Show

Now it is possible to define graphs, for instance as follows:

gr :: Graph Int
gr = Graph {vertices=[1..6], edges=[(1, 2), (1, 5), (2, 3), (2, 5), (3, 4), (4, 5), (4, 6)]}

> gr
Graph {vertices = [1,2,3,4,5,6], edges = [(1,2),(1,5),(2,3),(2,5),(3,4),(4,5),(4,6)]}

Moreover, we have automatically functions vertices :: Graph a -> [a] and edges :: Graph a -> [Edge a]:

> vertices gr
[1,2,3,4,5,6]

> edges gr
[(1,2),(1,5),(2,3),(2,5),(3,4),(4,5),(4,6)]

Recall that a Hamiltonian path in a graph is a path going through all the vertices exactly once. To solve the task, we will use brute force, generating all possible permutations of vertices and checking whether they form a path. First, we define a helper function isEdge taking a pair of vertices of type a and a graph over a and returning True if those vertices are connected and False otherwise. To test the membership of an element in a list, we can use the function elem. Note the type declaration of isEdge. As the function is polymorphic, we have to assume that a is an instance of the class Eq so that we test the membership by the elem function.

Solution: isEdge

isEdge :: Eq a => Edge a -> Graph a -> Bool
isEdge (a,b) g = (a,b) `elem` edgs || (b,a) `elem` edgs where
    edgs = edges g

Next, we define a function testing whether a given list of vertices is a path in a given graph. That can be easily done by list comprehension generating all indexes but the last one. Note the use of the function and. It can be applied to a list of type [Bool] performing logical conjunction of all its members.

Solution: isPath

isPath :: Eq a => [a] -> Graph a -> Bool
isPath vs g = and [ isEdge pair g | pair <- zip vs (tail vs) ]

Finally, we take all permutations of vertices and test which form a path in a given graph. Collecting all such paths is the list of all Hamiltonian paths.

Solution: findHamiltonian

findHamiltonian :: Eq a => Graph a -> [[a]]
findHamiltonian g = [p | p <- perms, isPath p g]
    where perms = permutations (vertices g)

Exercise 3

The following exercise focuses on operator overloading. With the boom of neural nets, finding algorithms computing efficiently and precisely derivatives of functions used to construct network layers became essential. There are three approaches. The first one is the syntactic derivation manipulating symbolic expressions. The second one is the approximation via the limit defining the derivative of a function at a point. Third, computing derivatives via dual numbers. We will discuss the last approach.

Dual numbers are numbers of the form $a+bϵ$ for $a,b\in \mathbb{R}$ and $ϵ$ is something like the complex unit $i$ but instead of $i^2=-1$ we have ${ϵ}^2=0$.^[1] Based on the above definition, we can define over the set of dual numbers algebraic operations like addition and multiplication:

Dual numbers are numbers of the form $a+bϵ$ for $a,b\in \mathbb{R}$ and $ϵ$ is something like the complex unit $i$ but instead of $i^2=-1$ we have ${ϵ}^2=0$.^[1] Based on the above definition, we can define over the set of dual numbers algebraic operations like addition and multiplication:

$$\begin{array}{rl}(a+bϵ)+(c+dϵ)& =(a+c)+(b+d)ϵ,\\ (a+bϵ)\times (c+dϵ)& =ac+(ad+bc)ϵ+bd{ϵ}^2=ac+(ad+bc)ϵ.\end{array}$$

Dual numbers can be used to compute the derivative of a function at a point. Consider first a polynomial $p(x)=b_0+b_{1}x+b_2{x}^2$. Let us compute its value at $a+ϵ$.

$$\begin{array}{rl}p(a+ϵ)& =b_0+b_1(a+ϵ)+b_2(a+ϵ{)}^2\\ & =b_0+b_{1}a+b_2{a}^2+b_1ϵ+b_{2}2aϵ\\ & =p(a)+p^{\prime }(a)ϵ\end{array}$$

So you see that the resulting dual number contains the value of $p(a)$ and also its derivative $p^{\prime }(a)$. This can be generalized to any analytic function by taking its Taylor expansion at a point $a$:

$$f(x)=\sum _{i=0}^{\mathrm{\infty }}\frac{f^{(n)}(a)}{n!}(x-a{)}^n$$

If we evaluate this series at $a+ϵ$, we get

$$f(a+ϵ)=\sum _{i=0}^{\mathrm{\infty }}\frac{f^{(n)}(a)}{n!}(a+ϵ-a{)}^n=f(a)+f^{\prime }(a)ϵ$$

as ${ϵ}^n=0$ for $n\ge 2$. Thus we again computed the value $f(a)$ and also $f^{\prime }(a)$ by evaluating a function $f(x)$ at $a+ϵ$.

data DualNum a = DN a a deriving (Eq, Ord)

We also define our own instance of Show so that, e.g., DN 3 10 is displayed as 3 + 10eps.

instance Show a => Show (DualNum a) where
    show (DN x x') = show x ++ " + " ++ show x' ++ "eps"

In order to be able to evaluate a function at a dual number, we define DualNum a as an instance of Num . Then we can compute $f(a)$ and $f^{\prime }(a)$ for $f$ defined as any composition of functions from the Num definition.

In order to be able to evaluate a function at a dual number, we define DualNum a as an instance of Num . Then we can compute $f(a)$ and $f^{\prime }(a)$ for $f$ defined as any composition of functions from the Num definition.

I should likely comment on the above definition a bit. Addition, subtraction, and multiplication are straightforward. The function fromInteger :: Num a => Integer -> a embeds integers into the set of dual numbers. So it maps $a$ to $a+0ϵ$. The last two definitions are not mathematically correct. We pretend that signum function has the derivative $0$ everywhere, which is not true at $0$. Similarly abs has no derivative at $0$.

I should likely comment on the above definition a bit. Addition, subtraction, and multiplication are straightforward. The function fromInteger :: Num a => Integer -> a embeds integers into the set of dual numbers. So it maps $a$ to $a+0ϵ$. The last two definitions are not mathematically correct. We pretend that signum function has the derivative $0$ everywhere, which is not true at $0$. Similarly abs has no derivative at $0$.

Dual numbers can be also divided if the first component of the divisor is non-zero as follows:

$$\frac{x+x^{\prime }ϵ}{y+y^{\prime }ϵ}=\frac{(x+x^{\prime }ϵ)(y-y^{\prime }ϵ)}{(y+y^{\prime }ϵ)(y-y^{\prime }ϵ)}=\frac{xy+(x^{\prime }y-x{y}^{\prime })ϵ}{y^2}=\frac{x}{y}+\frac{x^{\prime }y-x{y}^{\prime }}{y^2}ϵ$$

instance Fractional a => Fractional (DualNum a) where
    (DN x x') / (DN y y') = DN (x/y) ((x'*y - x*y') / (y*y))
    fromRational r = DN (fromRational r) 0

Now we can define a function $f$ and evaluate it at a dual number $a+ϵ$ to compute $f(a)$ and $f^{\prime }(a)$. The function has to be polymorphically working for any instance a of the class Num or Fractional.

Now we can define a function $f$ and evaluate it at a dual number $a+ϵ$ to compute $f(a)$ and $f^{\prime }(a)$. The function has to be polymorphically working for any instance a of the class Num or Fractional.

Indeed, for $f(x)=x^2+1$ we have $f(5)=5^2+1=26$ and $f^{\prime }(x)=2x$, $f^{\prime }(5)=10$. Another example

Indeed, for $f(x)=x^2+1$ we have $f(5)=5^2+1=26$ and $f^{\prime }(x)=2x$, $f^{\prime }(5)=10$. Another example

because $g^{\prime }(x)=\frac{x^2-2x+2}{(x-1{)}^2}$, i.e., $g(0)=2$ and $g^{\prime }(0)=2$.

because $g^{\prime }(x)=\frac{x^2-2x+2}{(x-1{)}^2}$, i.e., $g(0)=2$ and $g^{\prime }(0)=2$.

sqr :: (Fractional a, Ord a) => a -> a
sqr x = convAbs $ iterate improve 1
  where improve r = (r + x/r) / 2
        convAbs (x1:x2:xs) | abs (x1-x2) < 1e-10 = x2
                           | otherwise = convAbs xs

I think that it is quite impressive that we can compute a derivative of the function sqr at a point even though it is defined by an iterative computation.

> sqr (DN 9 1)
3.0 + 0.16666666666666666eps

Indeed, for $sqr(x)=\sqrt{x}$ we have $sq{r}^{\prime }(x)=\frac{1}{2\sqrt{x}}$ so that $sq{r}^{\prime }(9)=\frac{1}{6}=0.1\bar{6}$.

Indeed, for $sqr(x)=\sqrt{x}$ we have $sq{r}^{\prime }(x)=\frac{1}{2\sqrt{x}}$ so that $sq{r}^{\prime }(9)=\frac{1}{6}=0.1\bar{6}$.

Bonus exercise

For those interested, we can go even further and make DualNum a an instance of Floating. If you recall the chain rule, i.e., for a compose function $f(x)=h(g(x))$, we have $f^{\prime }(x)=h^{\prime }(g(x))g^{\prime }(x)$), then it is easy to decode the following definitions.

Write a function merge :: Ord b => (a -> b) -> [a] -> [a] -> [a] taking a function f :: a -> b where b is supposed to be an orderable type and two lists of elements of type a . Suppose that these two lists are sorted via f , i.e., for [a1,a2,a3,...] we have f a1 <= f a2 <= f a3 <= .... As a result, it returns a merged sorted list.

Once you have the function merge , implement a function subseqs :: [a] -> [[a]] which takes a list and returns all its sublists (i.e., subsequences of its elements) sorted by their length.

Hint

The subsequences can be generated recursively because subsequences of x:xs are just subsequences of xs together with subsequences of xs extended by x. To produce the sorted result, use the merge function.

Solution

Solutions will be revealed after labs on 4/24/2025.

Solution

Solutions will be revealed after labs on 4/24/2025.

---

1. For those who like algebra, it can be constructed by taking the ring of univariate polynomials $\mathbb{R}[ϵ]$ and taking its quotient by the ideal generated by ${ϵ}^2$, i.e., $\mathbb{R}[ϵ]/⟨{ϵ}^2⟩$. You can safely ignore the last sentence if you do not understand it. As I like algebra, I could not resist the temptation to explain it in algebraic language. ↩︎