# Find largest and second largest elements of the array

So here is a HW problem I have been working. On I was wondering if anybody could give me a hint of what I am doing wrong. I don't want to be given the answer just hints and advice on how to solve it.

I am given this simple algorithm that finds the greatest value and the second greatest value in a array.

  If A < A,
largest = A, Second = A
else
largest = A, Second = A;
For i in range 3 to n {
If A[i] > largest {
second = largest;
largest = A[i];
}
else
if A[i] > second
second = A[i]
}


My first task was to find the number of comparisons it makes not in big-oh terms. So I said 2n-3 because it has one in the beginning and worst case the largest element is in the first two elements of the array A so from index 3 to n it has two comparisons. So 1 + 2(n-2) = 2n - 3 comparisons.

My second task was to develop an algorithm that would use less comparisons. So I broke it down with divide and conquer and said this.

 // start is the first index inclusive
// finish is the last index inclusive
// A is an array of elements
function getGreatestAndSecondGreatestValue(A, start, finish) {
T[] // T is an array of size 2
if(finish - start == 1) {
if(A[start] > A[finish]) {
T = A[start]
T = A[finish]
} else {
T = A[finish]
T = A[start]
}
return T
}

Array1 = getGreatestAndSecondGreatestValue(A,start, start + floor((last-first)/2))
Array2 = getGreatestAndSecondGreatestValue(A, start + ceil((last-first)/2), finish)
return merge(Array1, Array2);
// The greast value will be in the 0th index of the array returned and the second
// greatest value will be in the 1 index of the array.
}

// this merges two arrays keeping the greatest value on the left side of the new array and
// keeping the second greatest value on the right side of the array
function merge(Array1, Array2) {
MergedArray[] // array that will hold the merged results
if(Array1 > Array2) {
MergedArray = Array1
MergedArray = max(Array, Array2)
} else {
MergedArray = Array2
MergedArray = max(Array1, Array2)
}
return MergedArray
}


For example this say we had the array

[1,4,5,2,7,9,10,19]

then is divides it

[1, 4, 5, 2] [7, 9, 10, 19]

then when it reaches a length 2 is places the highest value in index 0 for arrays the size 2

[4, 1] [5, 2] [9, 7] [19, 10]

I would then proceed to merge these

[5, 4] [19, 10]

[19, 10]

Then the greatest value is 19 and the second greatest was 10.

I then proceeded to find the number of comparisons with recurrences.

T(n) = 2T(n) + 2

There are n leaf nodes because $2^{log_{2}(n)} = n$ where $log_{2}(n)$ is the height of the tree. There are then n - 1 internal nodes so n + n - 1 is 2n - 1. 2 is added to every node so then it would be 4n - 2.

For some reason after using a divide and conquer algorithm I am getting a higher number of comparisons and I am not sure why. Could somebody give me some advice on why this didn't work?

• I am a little confused as to what you mean by we can find the greatest value for each pair in one comparison. The recursion is done in such a way that the left side of those pairs will always be the largest values?
– SPD
Oct 26 '17 at 5:49
• @ryan I messed around with some things but I am not sure if it is what you were thinking.The new solution works and requires less comparisons. If you want could you post your solution? I would like to see some solutions for insight if anybody wants to.
– SPD
Oct 26 '17 at 7:40
• Divide and conquer is promising only when the run time increases faster than the problem size ($O(n^2)$, in many cases). Oct 26 '17 at 7:57
• If the array elements are arranged in random order, you reduce the average number of comparisons by about half by comparing with the second largest element so far first. Jul 27 '20 at 22:13

Here I present an optimal solution minimizing the number of comparisons.

The thing to take advantage of is the first step of the original algorithm:

if A < A,
largest = A, second = A
else
largest = A, second = A


We can get the largest and second largest of sets of two in one comparison! So the first thing we should do is divide our array into $$n/2$$ pairs of two. Namely: $$p_1 = (a_1, a_2), p_2 = (a_3, a_4), \ldots, p_i = (a_{2i - 1}, a_{2i}), \ldots, p_{n/2} = (a_{n-1}, a_n)$$ For every pair $$p_i$$, we can determine it's largest and second largest value in one comparison. We'll call these $$p_i.L$$ and $$p_i.S$$ respectively. So it takes $$n/2$$ comparisons to determine all of thee values for every pair.

Let's call the overall largest and second largest elements $$\mathbb{L}$$ and $$\mathbb{S}$$ respectively. We know that $$\mathbb{L}$$ will be among the $$\{p_i.L\}$$'s because it will be the largest of these. We also know $$\mathbb{S}$$ will be among $$\{p_i.L\} \setminus \mathbb{L}$$ or it could be the second largest value in the pair that $$\mathbb{L}$$ came from.

An example: Let our original array be: $$A = [1,4,5,2,7,9,10,19]$$ We then break it up into pairs: $$P = [(1,4)(5,2)(7,9)(10,19)]$$ We then know $$\mathbb{L}$$ will be in: $$W = \forall i.\; \{p_i.L\} = \{4, 5, 9, 19\}$$ So now we just need to find three values: $$W.\mathbb{L}$$ (this will be $$A.\mathbb{L}$$), $$W.\mathbb{S}$$ (this could potentially be $$A.\mathbb{S}$$) and $$p_i.S$$ where $$p_i$$ is the pair that $$W.\mathbb{L}$$ came from (this could also potentially be $$A.\mathbb{S}$$).

From our example we see $$A.\mathbb{L} = W.\mathbb{L} = 19$$. We then see $$A.\mathbb{S}$$ is either $$W.\mathbb{S} = 9$$ or $$10$$ (from the pair containing $$W.\mathbb{L}$$), and it is $$10$$.

So this turns our problem of $$A$$ into a subproblem of size $$n/2$$, $$W$$. Then we can run the original algorithm on this subproblem and do some extra comparisons to get $$A.\mathbb{S}$$ and this totals about $$\approx 1.5n$$ comparisons. But we can do better!

We use recursion on this subproblem instead of just using the original algorithm. So the algorithm looks like this, and will return the indices of the $$A.\mathbb{L}$$ and $$A.\mathbb{S}$$. (assume $$n = 2^k$$ and this is 1-indexed). The algorithm will modify the input array so we will assume $$A$$ to be a copy of the input.

L_and_S (array of length n):
Let A = copy of input
// base case
if n == 2:
if A > A:
return (1, 2)
else:
return (2, 1)

// recursive case
Let W be new array length n/2 (keeps track of largest elements)
Let M be new array length n/2 (keeps track of swaps)
for i in 1 ... n/2:
// swap larger element into first position of pair
// set M[i] to 1 if pair was swapped, else 0
// move larger element into W
if A[2*i - 1] < A[2*i]:
temp = A[2*i - 1]
A[2*i - 1] = A[2*i]
A[2*i] = temp
M[i] = 1
else:
M[i] = 0
W[i] = A[2*i - 1]

// now recurse on W
(j, k) = L_and_S(W)
// map back from W to A
j_swap = M[j]
k_swap = M[k]
j = 2*j - 1
k = 2*k - 1
if A[j + 1] > A[k]:
k = j + 1
k_swap = 0 - j_swap
return (j + j_swap, k + k_swap) // account for the swap


We can run through the example above:

L_and_S([1,4,5,2,7,9,10,19]):
// do n/2 comparisons
A = [4,1,5,2,9,7,19,10]
W = [4,5,9,19]
M = [1,0,1,1]
(j, k) = L_and_S([4,5,9,19]):
// do n/4 comparisons
A = [5,4,19,9]
W = [5,19]
M = [1,1]
(j, k) = L_and_S([5,19]):
// do 1 comparison (5 > 19)
return (2,1)
(j_swap, k_swap) = (1, 1)
(j, k) = (3, 1)
// do 1 comparison (9 > 5)
k = 4
k_swap = -1
return (3 + 1, 4 - 1) = (4, 3)
(j_swap, k_swap) = (1, 1)
(j, k) = (7, 5)
// do 1 comparison (10 > 9)
k = 8
k_swap = -1
return (7 + 1, 8 - 1) = (8, 7)


We see overall this takes $$n/2 + n/4 + 1 + 1 + 1 = 9 = n + 1$$ comparisons! These are comparisons of the array elements, ignoring the base case comparison in the beginning.

The exact recurrence relation for number of comparisons is the following (assume $$n = 2^k$$): \begin{align} T(2) & = 1\\ T(n) & = T(\frac{n}{2}) + \frac{n}{2} + 1\\ & = 1 + \sum_{1}^{\log_2 n - 1} \frac{n}{2^i} + 1\\ & = \log_2 n + \sum_{1}^{\log_2 n - 1} \frac{n}{2^i}\\ & = \log_2 n + n - 2\\ & = n + \log_2 n - 2 \end{align} So we get the total number of comparisons in the worst case is $$n + \log_2 n - 2$$, which is pretty good. In fact, this is a lower bound! (see here) If you include the base case comparison it adds another $$\log_2 n$$ to the comparisons. Here's a table for the first few $$n$$:

$$\begin{array}{|c|c|} \hline n & n + \log_2 n - 2 & 2n - 3 \\ \hline 2 & 1 & 1 \\ \hline 4 & 4 & 5 \\ \hline 8 & 9 & 13 \\ \hline 16 & 18 & 29 \\ \hline 32 & 35 & 61 \\ \hline 64 & 68 & 125 \\ \hline 128 & 133 & 253 \\ \hline 256 & 262 & 509 \\ \hline 512 & 519 & 1021 \\ \hline 1024 & 1032 & 2045 \\ \hline \end{array}$$

Implementation of the code in python can be found here, with test cases at the bottom.

• A brilliantly written answer. The proof and code walkthrough have been priceless. The infinite sum of the geometric series did not come to me naturally as they should have, though. Oct 25 '18 at 19:09
• Just late to the game. I have test several implementation of this n+logn-2 procedure and found that it is slower than (at least a factor of two) the simple implementation (2n comparisons) consisting of one loop and 2 comparison in each loop. I have tested with arrays of 2 million integers. I think the reason is that the simple method only access each element of the array once and incur no extra overhead. Could someone please point me a test case where the n + logn -2 method is actually worth it? Thanks Apr 8 '19 at 14:07
• @Hoang-NganNguyen, in general the $n + \log n - 2$ solution is probably going to be slower because of the overhead. I would recommend just using the $2n - 3$ method. Also, notice that I'm only counting the number of comparisons not the number of operations. In fact, the optimal solution probably does way more operations because of the recursive stack and implementation details.
– ryan
Apr 8 '19 at 15:31
• @ryan: yes, the recursive stack add a lot of overhead. I removed it by using the array M in your code but with double the size to replace the recursive stack. However, the swapping operation only is already more expensive than the 2n-3 method. Apr 8 '19 at 19:43
• You don’t have to find the largest and second largest if 2 items, you can find say the largest and second largest of 8. Much less overhead, but still fewer comparisons. Jul 28 '20 at 8:37

I couldn't think of a way to code the n + log n - 2 solution, but I did see a general algorithm which I describe as follows: It takes n - 1 comparisons to find the largest element because (assume $$n=2^k$$) $$n/2^1 + n/2^2 + ... + n/n = n - 1.$$ Now the second largest must have, at some point, been compared to the largest or it wouldn't have been eliminated. Log n is the number of elements that were compared to the largest, so we have only to find the largest of a set with log n elements. Using the same method as above we find that this takes log n-1 comparisons. Then adding the two gives $$(n-1)+(\log n-1) = n+\log n-2.$$

• Consider putting dollar signs \$ around your formulas to make them show. Also you can use \log for log.
– 6005
Apr 3 '20 at 1:46