Algorithm to generate all unique permutations of fixed-length integer partitions? - algorithm

I'm searching for an algorithm that generates all permutations of fixed-length partitions of an integer. Order does not matter.
For example, for n=4 and length L=3:
[(0, 2, 2), (2, 0, 2), (2, 2, 0),
(2, 1, 1), (1, 2, 1), (1, 1, 2),
(0, 1, 3), (0, 3, 1), (3, 0, 1), (3, 1, 0), (1, 3, 0), (1, 0, 3),
(0, 0, 4), (4, 0, 0), (0, 4, 0)]
I bumbled about with integer partitions + permutations for partitions whose length is lesser than L; but that was too slow because I got the same partition multiple times (because [0, 0, 1] may be a permutation of [0, 0, 1] ;-)
Any help appreciated, and no, this isn't homework -- personal interest :-)

Okay. First, forget about the permutations and just generate the partitions of length L (as suggested by #Svein Bringsli). Note that for each partition, you may impose an ordering on the elements, such as >. Now just "count," maintaining your ordering. For n = 4, k = 3:
(4, 0, 0)
(3, 1, 0)
(2, 2, 0)
(2, 1, 1)
So, how to implement this? It looks like: while subtracting 1 from position i and adding it to the next position maintains our order, subtract 1 from position i, add 1 to position i + 1, and move to the next position. If we're in the last position, step back.
Here's a little python which does just that:
def partition_helper(l, i, result):
if i == len(l) - 1:
return
while l[i] - 1 >= l[i + 1] + 1:
l[i] -= 1
l[i + 1] += 1
result.append(list(l))
partition_helper(l, i + 1, result)
def partition(n, k):
l = [n] + [0] * (k - 1)
result = [list(l)]
partition_helper(l, 0, result)
return result
Now you have a list of lists (really a list of multisets), and generating all permutations of each multiset of the list gives you your solution. I won't go into that, there's a recursive algorithm which basically says, for each position, choose each unique element in the multiset and append the permutations of the multiset resulting from removing that element from the multiset.

Given that you ask this out of interest, you would probably be interested an authorative answer! It can be found in "7.2.1.2 - Generating all permutations" of Knuth's The Art of Computer Programming (subvolume 4A).
Also, 3 concrete algorithms can be found here.

As noted by #pbarranis, the code by #rlibby does not include all lists when n equals k. Below is Python code which does include all lists. This code is non-recursive, which may be more efficient with respect to memory usage.
def successor(n, l):
idx = [j for j in range(len(l)) if l[j] < l[0]-1]
if not idx:
return False
i = idx[0]
l[1:i+1] = [l[i]+1]*(len(l[1:i+1]))
l[0] = n - sum(l[1:])
return True
def partitions(n, k):
l = [0]*k
l[0] = n
results = []
results.append(list(l))
while successor(n, l):
results.append(list(l))
return results
The lists are created in colexicographic order (algorithm and more description here).

I found that using a recursive function was not good for larger lengths and integers because it chews up too much RAM, and using a generator / resumable-function (that 'yields' values) was too slow and required a large library to make it cross-platform.
So here's a non-recursive solution in C++ that produces the partitions in sorted order (which is ideal for permutations too). I've found this to be over 10 times faster than seemingly clever and concise recursive solutions I tried for partition lengths of 4 or greater, but for lengths of 1-3 the performance is not necessarily better (and I don't care about short lengths because they're fast with either approach).
// Inputs
unsigned short myInt = 10;
unsigned short len = 3;
// Partition variables.
vector<unsigned short> partition(len);
unsigned short last = len - 1;
unsigned short penult = last - 1;
short cur = penult; // Can dip into negative value when len is 1 or 2. Can be changed to unsigned if len is always >=3.
unsigned short sum = 0;
// Prefill partition with 0.
fill(partition.begin(), partition.end(), 0);
do {
// Calculate remainder.
partition[last] = max(0, myInt - sum); // Would only need "myInt - sum" if partition vector contains signed ints.
/*
*
* DO SOMETHING WITH "partition" HERE.
*
*/
if (partition[cur + 1] <= partition[cur] + 1) {
do {
cur--;
} while (
cur > 0 &&
accumulate(partition.cbegin(), partition.cbegin() + cur, 0) + (len - cur) * (partition[cur] + 1) > myInt
);
// Escape if seeked behind too far.
// I think this if-statement is only useful when len is 1 or 2, can probably be removed if len is always >=3.
if (cur < 0) {
break;
}
// Increment the new cur position.
sum++;
partition[cur]++;
// The value in each position must be at least as large as the
// value in the previous position.
for (unsigned short i = cur + 1; i < last; ++i) {
sum = sum - partition[i] + partition[i - 1];
partition[i] = partition[i - 1];
}
// Reset cur for next time.
cur = penult;
}
else {
sum++;
partition[penult]++;
}
} while (myInt - sum >= partition[penult]);
Where I've written DO SOMETHING WITH "partition" HERE. is where you would actually consume the value. (On the last iteration the code will continue to execute the remainder of the loop but I found this to be better than constantly checking for exit conditions - it's optimised for larger operations)
0,0,10
0,1,9
0,2,8
0,3,7
0,4,6
0,5,5
1,1,8
1,2,7
1,3,6
1,4,5
2,2,6
2,3,5
2,4,4
3,3,4
Oh I've used "unsigned short" because I know my length and integer won't exceed certain limits, change that if it's not suitable for you :) Check the comments; one variable there (cur) had to be signed to handle lengths of 1 or 2 and there's a corresponding if-statement that goes with that, and I've also noted in a comment that if your partition vector has signed ints there is another line that can be simplified.
To get all the compositions, in C++ I would use this simple permutation strategy which thankfully does not produce any duplicates:
do {
// Your code goes here.
} while (next_permutation(partition.begin(), partition.end()));
Nest that in the DO SOMETHING WITH "partition" HERE spot, and you're good to go.
An alternative to finding the compositions (based on the Java code here https://www.nayuki.io/page/next-lexicographical-permutation-algorithm) is as follows. I've found this to perform better than next_permutation().
// Process lexicographic permutations of partition (compositions).
composition = partition;
do {
// Your code goes here.
// Find longest non-increasing suffix
i = last;
while (i > 0 && composition[i - 1] >= composition[i]) {
--i;
}
// Now i is the head index of the suffix
// Are we at the last permutation already?
if (i <= 0) {
break;
}
// Let array[i - 1] be the pivot
// Find rightmost element that exceeds the pivot
j = last;
while (composition[j] <= composition[i - 1])
--j;
// Now the value array[j] will become the new pivot
// Assertion: j >= i
// Swap the pivot with j
temp = composition[i - 1];
composition[i - 1] = composition[j];
composition[j] = temp;
// Reverse the suffix
j = last;
while (i < j) {
temp = composition[i];
composition[i] = composition[j];
composition[j] = temp;
++i;
--j;
}
} while (true);
You'll notice some undeclared variables there, just declare them earlier in the code before all your do-loops: i, j, pos, and temp (unsigned shorts), and composition (same type and length as partition). You can reuse the declaration of i for it's use in a for-loop in the partitions code snippet. Also note variables like last being used which assume this code is nested within the partitions code given earlier.
Again "Your code goes here" is where you consume the composition for your own purposes.
For reference here are my headers.
#include <vector> // for std::vector
#include <numeric> // for std::accumulate
#include <algorithm> // for std::next_permutation and std::max
using namespace std;
Despite the massive increase in speed using these approaches, for any sizeable integers and partition lengths this will still make you mad at your CPU :)

Like I mentioned above, I couldn't get #rlibby's code to work for my needs, and I needed code where n=l, so just a subset of your need. Here's my code below, in C#. I know it's not perfectly an answer to the question above, but I believe you'd only have to modify the first method to make it work for different values of l; basically add the same code #rlibby did, making the array of length l instead of length n.
public static List<int[]> GetPartitionPermutations(int n)
{
int[] l = new int[n];
var results = new List<int[]>();
GeneratePermutations(l, n, n, 0, results);
return results;
}
private static void GeneratePermutations(int[] l, int n, int nMax, int i, List<int[]> results)
{
if (n == 0)
{
for (; i < l.Length; ++i)
{
l[i] = 0;
}
results.Add(l.ToArray());
return;
}
for (int cnt = Math.Min(nMax, n); cnt > 0; --cnt)
{
l[i] = cnt;
GeneratePermutations(l, (n - cnt), cnt, i + 1, results);
}
}

A lot of searching led to this question. Here is an answer that includes the permutations:
#!/usr/bin/python
from itertools import combinations_with_replacement as cr
def all_partitions(n, k):
"""
Return all possible combinations that add up to n
i.e. divide n objects in k DISTINCT boxes in all possible ways
"""
all_part = []
for div in cr(range(n+1), k-1):
counts = [div[0]]
for i in range(1, k-1):
counts.append(div[i] - div[i-1])
counts.append(n-div[-1])
all_part.append(counts)
return all_part
For instance, all_part(4, 3) as asked by OP gives:
[[0, 0, 4],
[0, 1, 3],
[0, 2, 2],
[0, 3, 1],
[0, 4, 0],
[1, 0, 3],
[1, 1, 2],
[1, 2, 1],
[1, 3, 0],
[2, 0, 2],
[2, 1, 1],
[2, 2, 0],
[3, 0, 1],
[3, 1, 0],
[4, 0, 0]]

Related

Minimum transfer to make array equal

This question is asked in the interview. I am still not able to find what should be right approach to attempt this problem.
Given an array = [7,2,2] find the minimum number of transfer required to make array elements almost equal. If this is not possible the larger elements should come to the left side.
In above example the final state of array would be [4,4,3] and the answer will be 2+ 1 =3.
We are transfering 2 from 7 to first 2 and then we are transfering another 1 from 7 to 2.
If the input is [2,2,7] then the answer will be 4 since we need to keep bigger elements on the left side.
final state = [4,4,3]
2 transfered from 7 to both 2 to make the final count as 4.
The minimum amount of transfers done 1 unit at a time is half the total amount by which the input differs from the desired array. "Almost equal" doesn't seem to mean any complication according to what you've given.
The solution is to imagine what the target array will be. This target array will depend only on the sum of the values in the original array, and the length of the array (which obviously must remain the same).
If the sum of the values is a multiple of the array length, then in the target array all values will be the same. If however there is a remainder, that remainder represents the number of array values that will be one more than some of the value(s) at the end of the array.
We don't actually have to store that target array. It is implicitly defined by the quotient and the remainder of the division of the sum by the array length.
The output of the function is the sum of differences with the actual input array value and the expected value at any array index. We should only count positive differences (i.e. transfers out of a value) as otherwise we would count transfers twice -- once on the outgoing side and again on the incoming side.
Here is an implementation in basic JavaScript:
function solve(arr) {
// Sum all array values
let sum = 0;
for (let i = 0; i < arr.length; i++) {
sum += arr[i];
}
// Get the integer quotient and remainder
let quotient = Math.floor(sum / arr.length);
let remainder = sum % arr.length;
// Determine the target value until the remainder is completely consumed:
let expected = quotient + 1;
// Collect all the positive differences with the expected value
let result = 0;
for (let i = 0; i < arr.length; i++) {
// If we have consumed the remainder, reduce the expected value
if (i == remainder) {
expected = quotient;
}
let transfer = arr[i] - expected;
// Only account for positive transfers to avoid double counting
if (transfer > 0) {
result += transfer;
}
}
return result;
}
let array = [7,2,2];
console.log(solve(array)); // 6
Let's start form target array. What is it?
Having {7, 2, 2} we want to obtain {4, 4, 3}. So every item is at least 3 and some top items are 3 + 1 == 4.
The algorithm is
let sum = sum(original)
let rem = sum(original) % length(original) # here % stands for remainder
target[i] = sum / length(original) + (i < rem ? 1 : 0)
Having original and target
original: 7 2 2
target: 4 4 3
transfer: 3 2 1 (6 in total)
note, that
transfer[i] is just an absolute difference: abs(original[i] - target[i])
we count each transfer twice: once we subtract and then we add.
So the answer is
sum(transfer[i]) / 2 == sum(abs(original[i] - target[i])) / 2
Code (c#):
private static int Solve(int[] initial) {
// Don't forget about degenerated cases
if (initial is null || initial.Length <= 0)
return 0;
int sum = initial.Sum();
int rem = sum % initial.Length;
int result = 0;
for (int i = 0; i < initial.Length; ++i)
result += Math.Abs(sum / initial.Length + ((i < rem) ? 1 : 0) - initial[i]);
return result / 2;
}
Demo: (Fiddle)
int[][] tests = new int[][] {
new int[] {7, 2, 2},
new int[] {2, 2, 7},
new int[] {},
new int[] {2, 2, 2},
new int[] {1, 2, 3},
};
string report = string.Join(Environment.NewLine, tests
.Select(test => $"[{string.Join(", ", test)}] => {Solve(test)}"));
Console.Write(report);
Outcome:
[7, 2, 2] => 3
[2, 2, 7] => 4
[] => 0
[2, 2, 2] => 0
[1, 2, 3] => 1
Seems to me like a simple problem that can be solved with greedy approach.
Steps:
Sum up the input array-elements S, divide by its length n. Lets say, the quotient is Q and remainder (mod) is R. Then, final array target will have 1st R elements with value = Q+1. Rest of the elements will be Q.
Number of transfers will be half of the sum of absolute difference at each (corresponding) position in input and target arrays.
Example:
Input [7, 2, 2]
S=11 n=3 Q=11/3=3 R=11%3=2
Target [3+1, 3+1, 3]
Answer = (abs(7-4) + abs(2-4) + abs(2-3)) / 2 = 3

Interview Question - Which numbers shows up most times in a list of intervals

I only heard of this question, so I don't know the exact limits. You are given a list of positive integers. Each two consecutive values form a closed interval. Find the number that appears in most intervals. If two values appear the same amount of times, select the smallest one.
Example: [4, 1, 6, 5] results in [1, 4], [1, 6], [5, 6] with 1, 2, 3, 4, 5 each showing up twice. The correct answer would be 1 since it's the smallest.
I unfortunately have no idea how this can be done without going for an O(n^2) approach. The only optimisation I could think of was merging consecutive descending or ascending intervals, but this doesn't really work since [4, 3, 2] would count 3 twice.
Edit: Someone commented (but then deleted) a solution with this link http://www.zrzahid.com/maximum-number-of-overlapping-intervals/. I find this one the most elegant, even though it doesn't take into account the fact that some elements in my input would be both the beginning and end of some intervals.
Sort intervals based on their starting value. Then run a swipe line from left (the global smallest value) to the right (the global maximum value) value. At each meeting point (start or end of an interval) count the number of intersection with the swipe line (in O(log(n))). Time complexity of this algorithm would be O(n log(n)) (n is the number of intervals).
The major observation is that the result will be one of the numbers in the input (proof left to the reader as simple exercise, yada yada).
My solution will be inspired by #Prune's solution. The important step is mapping the input numbers to their order within all different numbers in the input.
I will work with C++ std. We can first load all the numbers into a set. We can then create map from that, which maps a number to its order within all numbers.
int solve(input) {
set<int> vals;
for (int n : input) {
vals.insert(n);
}
map<int, int> numberOrder;
int order = 0;
for (int n : vals) { // values in a set are ordered
numberOrder[n] = order++;
}
We then create process array (similar to #Prune's solution).
int process[map.size() + 1]; // adding past-the-end element
int curr = input[0];
for (int i = 0; i < input.size(); ++i) {
last = curr;
curr = input[i];
process[numberOrder[min(last, curr)]]++;
process[numberOrder[max(last, curr)] + 1]--;
}
int appear = 0;
int maxAppear = 0;
for (int i = 0; i < process.size(); ++i) {
appear += process[i];
if (appear > maxAppear) {
maxAppear = appear;
maxOrder = i;
}
}
Last, we need to find our found value in the map.
for (pair<int, int> a : numberOrder) {
if (a.second == maxOrder) {
return a.first;
}
}
}
This solution has O(n * log(n)) time complexity and O(n) space complexity, which is independent on maximum input number size (unlike other solutions).
If the maximum number in the range array is less than the maximum size limit of an array, my solution will work with complexity o(n).
1- I created a new array to process ranges and use it to find the
numbers that appears most in all intervals. For simplicity let's use
your example. the input = [1, 4], [1, 6], [5, 6]. let's call the new
array process and give it length 6 and it is initialized with 0s
process = [0,0,0,0,0,0].
2-Then loop through all the intervals and mark the start with (+1) and
the cell immediately after my range end with (-1)
for range [1,4] process = [1,0,0,0,-1,0]
for range [1,6] process = [2,0,0,0,-1,0]
for range [5,6] process = [2,0,0,0,0,0]
3- The p rocess array will work as accumulative array. initialize a
variable let's call it appear = process[0] which will be equal to 2
in our case. Go through process and keep accumulating what can you
notice? elements 1,2,3,4,5,6 will have appear =2 because each of
them appeared twice in the given ranges .
4- Maximize while you loop through process array you will find the
solution
public class Test {
public static void main(String[] args) {
int[] arr = new int[] { 4, 1, 6, 5 };
System.out.println(solve(arr));
}
public static int solve(int[] range) {
// I assume that the max number is Integer.MAX_VALUE
int size = (int) 1e8;
int[] process = new int[size];
// fill process array
for (int i = 0; i < range.length - 1; ++i) {
int start = Math.min(range[i], range[i + 1]);
int end = Math.max(range[i], range[i + 1]);
process[start]++;
if (end + 1 < size)
process[end + 1]--;
}
// Find the number that appears in most intervals (smallest one)
int appear = process[0];
int max = appear;
int solu = 0;
for (int i = 1; i < size; ++i) {
appear += process[i];
if (appear > max){
solu = i;
max = appear;
}
}
return solu;
}
}
Think of these as parentheses: ( to start and interval, ) to end. Now check the bounds for each pair [a, b], and tally interval start/end markers for each position: the lower number gets an interval start to the left; the larger number gets a close interval to the right. For the given input:
Process [4, 1]
result: [0, 1, 0, 0, 0, -1]
Process [1, 6]
result: [0, 2, 0, 0, 0, -1, 0, -1]
Process [6, 5]
result: [0, 2, 0, 0, 0, -1, 1, -2]
Now, merely make a cumulative sum of this list; the position of the largest value is your desired answer.
result: [0, 2, 0, 0, 0, -1, 1, -2]
cumsum: [0, 2, 2, 2, 2, 1, 2, 0]
Note that the final sum must be 0, and can never be negative. The largest value is 2, which appears first at position 1. Thus, 1 is the lowest integer that appears the maximum (2) quantity.
No that's one pass on the input, and one pass on the range of numbers. Note that with a simple table of values, you can save storage. The processing table would look something like:
[(1, 2)
(4, -1)
(5, 1)
(6, -2)]
If you have input with intervals both starting and stopping at a number, then you need to handle the starts first. For instance, [4, 3, 2] would look like
[(2, 1)
(3, 1)
(3, -1)
(4, -1)]
NOTE: maintaining a sorted insert list is O(n^2) time on the size of the input; sorting the list afterward is O(n log n). Either is O(n) space.
My first suggestion, indexing on the number itself, is O(n) time, but O(r) space on the range of input values.
[

In an array of randomly generated booleans, change 'k' falses to trues to create the largest continuous chain of trues

I was asked in an interview the following question:
In an array of randomly generated booleans, such as : T F T T F F F T F F F F T
Write an algorithm to determine which false values to change to true, to maximize the largest continuous chunk of trues. In the above example, suppose that k = 3. One of the solutions would be:
T F T T T* T* T* T F F F F T
Where T* denotes a value that has been changed.
Besides simple bruteforce, one of the methods I came up with was to find the largest continuous chunk of False values, and compare it with k. If it is less, then we replace the entire chunk with True and continue with the 'k' that is remaining. However, it turns out this method didn't always guarantee the correct answer.
Another more complicated method I thought of is this: for every chunk of falses inbetween chunks of trues, compute how big of a chunk can be built by flipping the falses inbetween the trues. Then it comes down to selecting the best combination of chunks inbetween trues to flip.
What is the optimal algorithm for this problem?
Many thanks.
Find the largest range that contains k false values. You can do this in linear time by keeping a running window.
You really can do this with a sliding window. I actually think that even though conceptually it's not a difficult problem, it is tricky to get the indexing right for the edge cases especially with the pressure of an interview.
Here's one way to do it:
Set two index variable to zero (start and end). Scan ahead incrementing end to right before the k+1 'F' (or the end of the array) putting the indexes of the 'F's in an array. This is your initial best guess and location of the 'F's.
Increment end to the next 'F', and move start to the next index in your array of F locations. Test if it's longer and repeat. You can keep track of the best start which will be the initial 'F' you'll need to change.
It's a little easier to show an example than explain, but it's basically a moving window while keeping track of the best run and best initial 'F' to change. Here's a quick and dirty JS implementation:
function findBestFlips(k, arr) {
let start, end, max, best_start, n;
start = end = max = best_start_index = n = 0;
let fs = [];
for (end = 0; end <= arr.length; end++) {
if (arr[end] == 0) {
fs.push(end)
if (fs.length <= k + 1) {
max = end; // set initial max from start of array to right before (k+1)th false value
continue // fill up fs with k+1 values if possible
}
if (max < (end - (fs[start] + 1))) {
max = end - (fs[start] + 1)
best_start_index = start + 1
}
start++
}
}
/* The above while loop stopped with potentially more ‘T’s at the end of the array.
push on to the end of the array */
if (max < arr.length - (fs[start] + 1)) {
max = arr.length - (fs[start] + 1)
best_start_index = start + 1
}
/* fs should have the index of all the false values
best_start through k + best_start_index are the indexes we need to change
to get the best_run */
if (fs.length <= k) max = arr.length
return {
flip_indexes: fs.slice(best_start_index, k + best_start_index),
best_run: max
}
}
let arr = [1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1]
let k = 3;
console.log(findBestFlips(k, arr))
// edge cases
arr = [1, 0, 1, 1, 1, 1]
k = 3;
console.log(findBestFlips(k, arr))
arr = [0, 0, 0]
k = 3;
console.log(findBestFlips(k, arr))
arr = [0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 1, 1, 1, 0, 0, 0]
k = 3;
console.log(findBestFlips(k, arr))

Given an array of ints and a number n, calculate the number of ways to sum to n using the ints

I saw this problem in my interview preparation.
Given an array of ints and a number n, calculate the number of ways to
sum to n using the ints
Following code is my solution. I tried to solve this by recursion. Subproblem is for each int in the array, we can either pick it or not.
public static int count(List<Integer> list, int n) {
System.out.print(list.size() + ", " + n);
System.out.println();
if (n < 0 || list.size() == 0)
return 0;
if (list.get(0) == n)
return 1;
int e = list.remove(0);
return count(list, n) + count(list, n - e);
}
I tried to use [10, 1, 2, 7, 6, 1, 5] for ints, and set n to 8. The result should be 4. However, I got 0. I tried to print what I have on each layer of stack to debug as showed in the code. Following is what I have:
7, 8
6, 8
5, 8
4, 8
3, 8
2, 8
1, 8
0, 8
0, 3
0, 7
0, 2
0, 1
0, 6
0, 7
0, -2
This result confuses me. I think it looks right from beginning to (0, 3). Starting from (0, 7), it looks wrong to me. I expect (1, 7) there. Because if I understand correctly, this is for count(list, n - e) call on second to the bottom layer on the stack. The list operation on the lower layer shouldn't impact the list on the current layer.
So my questions are:
why is it (0, 7) instead of (1, 7) based on my current code?
what adjustment should I do to my current code to get the correct result?
Thanks!
The reason why your algorithm is not working is because you are using one list that is being modified before the recursive calls.
Since the list is passed by reference, what ends up happening is that you recursively call remove until there is nothing in the list any more and then all of your recursive calls are going to return 0
What you could do is create two copies of the list on every recursive step. However, this would be way too inefficient.
A better way would be to use an index i that marks the element in the list that is being looked at during the call:
public static int count(List<Integer> list, int n, int i) {
//System.out.print(list.size() + ", " + n);
//System.out.println();
if (n < 0 || i <= 0)
return 0;
int e = list.get(i); // e is the i-th element in the list
if (e == n)
return 1 + count(list, n, i-1); // Return 1 + check for more possibilities without picking e
return count(list, n, i-1) + count(list, n - e, i-1); // Result if e is not picked + result if e is picked
}
You would then pass yourList.size() - 1 for i on the initial function call.
One more point is that when you return 1, you still have to add the number of possibilities for when your element e is not picked to be part of a sum. Otherwise, if - for example - your last element in the list was n, the recursion would end on the first step only returning 1 and not checking for more possible number combinations.
Finally, you might want to rewrite the algorithm using a dynamic approach, since that would give you a way better running time.

Algorithm to find two repeated numbers in an array, without sorting

There is an array of size n (numbers are between 0 and n - 3) and only 2 numbers are repeated. Elements are placed randomly in the array.
E.g. in {2, 3, 6, 1, 5, 4, 0, 3, 5} n=9, and repeated numbers are 3 and 5.
What is the best way to find the repeated numbers?
P.S. [You should not use sorting]
There is a O(n) solution if you know what the possible domain of input is. For example if your input array contains numbers between 0 to 100, consider the following code.
bool flags[100];
for(int i = 0; i < 100; i++)
flags[i] = false;
for(int i = 0; i < input_size; i++)
if(flags[input_array[i]])
return input_array[i];
else
flags[input_array[i]] = true;
Of course there is the additional memory but this is the fastest.
OK, seems I just can't give it a rest :)
Simplest solution
int A[N] = {...};
int signed_1(n) { return n%2<1 ? +n : -n; } // 0,-1,+2,-3,+4,-5,+6,-7,...
int signed_2(n) { return n%4<2 ? +n : -n; } // 0,+1,-2,-3,+4,+5,-6,-7,...
long S1 = 0; // or int64, or long long, or some user-defined class
long S2 = 0; // so that it has enough bits to contain sum without overflow
for (int i=0; i<N-2; ++i)
{
S1 += signed_1(A[i]) - signed_1(i);
S2 += signed_2(A[i]) - signed_2(i);
}
for (int i=N-2; i<N; ++i)
{
S1 += signed_1(A[i]);
S2 += signed_2(A[i]);
}
S1 = abs(S1);
S2 = abs(S2);
assert(S1 != S2); // this algorithm fails in this case
p = (S1+S2)/2;
q = abs(S1-S2)/2;
One sum (S1 or S2) contains p and q with the same sign, the other sum - with opposite signs, all other members are eliminated.
S1 and S2 must have enough bits to accommodate sums, the algorithm does not stand for overflow because of abs().
if abs(S1)==abs(S2) then the algorithm fails, though this value will still be the difference between p and q (i.e. abs(p - q) == abs(S1)).
Previous solution
I doubt somebody will ever encounter such a problem in the field ;)
and I guess, I know the teacher's expectation:
Lets take array {0,1,2,...,n-2,n-1},
The given one can be produced by replacing last two elements n-2 and n-1 with unknown p and q (less order)
so, the sum of elements will be (n-1)n/2 + p + q - (n-2) - (n-1)
the sum of squares (n-1)n(2n-1)/6 + p^2 + q^2 - (n-2)^2 - (n-1)^2
Simple math remains:
(1) p+q = S1
(2) p^2+q^2 = S2
Surely you won't solve it as math classes teach to solve square equations.
First, calculate everything modulo 2^32, that is, allow for overflow.
Then check pairs {p,q}: {0, S1}, {1, S1-1} ... against expression (2) to find candidates (there might be more than 2 due to modulo and squaring)
And finally check found candidates if they really are present in array twice.
You know that your Array contains every number from 0 to n-3 and the two repeating ones (p & q). For simplicity, lets ignore the 0-case for now.
You can calculate the sum and the product over the array, resulting in:
1 + 2 + ... + n-3 + p + q = p + q + (n-3)(n-2)/2
So if you substract (n-3)(n-2)/2 from the sum of the whole array, you get
sum(Array) - (n-3)(n-2)/2 = x = p + q
Now do the same for the product:
1 * 2 * ... * n - 3 * p * q = (n - 3)! * p * q
prod(Array) / (n - 3)! = y = p * q
Your now got these terms:
x = p + q
y = p * q
=> y(p + q) = x(p * q)
If you transform this term, you should be able to calculate p and q
Insert each element into a set/hashtable, first checking if its are already in it.
You might be able to take advantage of the fact that sum(array) = (n-2)*(n-3)/2 + two missing numbers.
Edit: As others have noted, combined with the sum-of-squares, you can use this, I was just a little slow in figuring it out.
Check this old but good paper on the topic:
Finding Repeated Elements (PDF)
Some answers to the question: Algorithm to determine if array contains n…n+m? contain as a subproblem solutions which you can adopt for your purpose.
For example, here's a relevant part from my answer:
bool has_duplicates(int* a, int m, int n)
{
/** O(m) in time, O(1) in space (for 'typeof(m) == typeof(*a) == int')
Whether a[] array has duplicates.
precondition: all values are in [n, n+m) range.
feature: It marks visited items using a sign bit.
*/
assert((INT_MIN - (INT_MIN - 1)) == 1); // check n == INT_MIN
for (int *p = a; p != &a[m]; ++p) {
*p -= (n - 1); // [n, n+m) -> [1, m+1)
assert(*p > 0);
}
// determine: are there duplicates
bool has_dups = false;
for (int i = 0; i < m; ++i) {
const int j = abs(a[i]) - 1;
assert(j >= 0);
assert(j < m);
if (a[j] > 0)
a[j] *= -1; // mark
else { // already seen
has_dups = true;
break;
}
}
// restore the array
for (int *p = a; p != &a[m]; ++p) {
if (*p < 0)
*p *= -1; // unmark
// [1, m+1) -> [n, n+m)
*p += (n - 1);
}
return has_dups;
}
The program leaves the array unchanged (the array should be writeable but its values are restored on exit).
It works for array sizes upto INT_MAX (on 64-bit systems it is 9223372036854775807).
suppose array is
a[0], a[1], a[2] ..... a[n-1]
sumA = a[0] + a[1] +....+a[n-1]
sumASquare = a[0]*a[0] + a[1]*a[1] + a[2]*a[2] + .... + a[n]*a[n]
sumFirstN = (N*(N+1))/2 where N=n-3 so
sumFirstN = (n-3)(n-2)/2
similarly
sumFirstNSquare = N*(N+1)*(2*N+1)/6 = (n-3)(n-2)(2n-5)/6
Suppose repeated elements are = X and Y
so X + Y = sumA - sumFirstN;
X*X + Y*Y = sumASquare - sumFirstNSquare;
So on solving this quadratic we can get value of X and Y.
Time Complexity = O(n)
space complexity = O(1)
I know the question is very old but I suddenly hit it and I think I have an interesting answer to it.
We know this is a brainteaser and a trivial solution (i.e. HashMap, Sort, etc) no matter how good they are would be boring.
As the numbers are integers, they have constant bit size (i.e. 32). Let us assume we are working with 4 bit integers right now. We look for A and B which are the duplicate numbers.
We need 4 buckets, each for one bit. Each bucket contains numbers which its specific bit is 1. For example bucket 1 gets 2, 3, 4, 7, ...:
Bucket 0 : Sum ( x where: x & 2 power 0 == 0 )
...
Bucket i : Sum ( x where: x & 2 power i == 0 )
We know what would be the sum of each bucket if there was no duplicate. I consider this as prior knowledge.
Once above buckets are generated, a bunch of them would have values more than expected. By constructing the number from buckets we will have (A OR B for your information).
We can calculate (A XOR B) as follows:
A XOR B = Array[i] XOR Array[i-1] XOR ... 0, XOR n-3 XOR n-2 ... XOR 0
Now going back to buckets, we know exactly which buckets have both our numbers and which ones have only one (from the XOR bit).
For the buckets that have only one number we can extract the number num = (sum - expected sum of bucket). However, we should be good only if we can find one of the duplicate numbers so if we have at least one bit in A XOR B, we've got the answer.
But what if A XOR B is zero?
Well this case is only possible if both duplicate numbers are the same number, which then our number is the answer of A OR B.
Sorting the array would seem to be the best solution. A simple sort would then make the search trivial and would take a whole lot less time/space.
Otherwise, if you know the domain of the numbers, create an array with that many buckets in it and increment each as you go through the array. something like this:
int count [10];
for (int i = 0; i < arraylen; i++) {
count[array[i]]++;
}
Then just search your array for any numbers greater than 1. Those are the items with duplicates. Only requires one pass across the original array and one pass across the count array.
Here's implementation in Python of #eugensk00's answer (one of its revisions) that doesn't use modular arithmetic. It is a single-pass algorithm, O(log(n)) in space. If fixed-width (e.g. 32-bit) integers are used then it is requires only two fixed-width numbers (e.g. for 32-bit: one 64-bit number and one 128-bit number). It can handle arbitrary large integer sequences (it reads one integer at a time therefore a whole sequence doesn't require to be in memory).
def two_repeated(iterable):
s1, s2 = 0, 0
for i, j in enumerate(iterable):
s1 += j - i # number_of_digits(s1) ~ 2 * number_of_digits(i)
s2 += j*j - i*i # number_of_digits(s2) ~ 4 * number_of_digits(i)
s1 += (i - 1) + i
s2 += (i - 1)**2 + i**2
p = (s1 - int((2*s2 - s1**2)**.5)) // 2
# `Decimal().sqrt()` could replace `int()**.5` for really large integers
# or any function to compute integer square root
return p, s1 - p
Example:
>>> two_repeated([2, 3, 6, 1, 5, 4, 0, 3, 5])
(3, 5)
A more verbose version of the above code follows with explanation:
def two_repeated_seq(arr):
"""Return the only two duplicates from `arr`.
>>> two_repeated_seq([2, 3, 6, 1, 5, 4, 0, 3, 5])
(3, 5)
"""
n = len(arr)
assert all(0 <= i < n - 2 for i in arr) # all in range [0, n-2)
assert len(set(arr)) == (n - 2) # number of unique items
s1 = (n-2) + (n-1) # s1 and s2 have ~ 2*(k+1) and 4*(k+1) digits
s2 = (n-2)**2 + (n-1)**2 # where k is a number of digits in `max(arr)`
for i, j in enumerate(arr):
s1 += j - i
s2 += j*j - i*i
"""
s1 = (n-2) + (n-1) + sum(arr) - sum(range(n))
= sum(arr) - sum(range(n-2))
= sum(range(n-2)) + p + q - sum(range(n-2))
= p + q
"""
assert s1 == (sum(arr) - sum(range(n-2)))
"""
s2 = (n-2)**2 + (n-1)**2 + sum(i*i for i in arr) - sum(i*i for i in range(n))
= sum(i*i for i in arr) - sum(i*i for i in range(n-2))
= p*p + q*q
"""
assert s2 == (sum(i*i for i in arr) - sum(i*i for i in range(n-2)))
"""
s1 = p+q
-> s1**2 = (p+q)**2
-> s1**2 = p*p + 2*p*q + q*q
-> s1**2 - (p*p + q*q) = 2*p*q
s2 = p*p + q*q
-> p*q = (s1**2 - s2)/2
Let C = p*q = (s1**2 - s2)/2 and B = p+q = s1 then from Viete theorem follows
that p and q are roots of x**2 - B*x + C = 0
-> p = (B + sqrtD) / 2
-> q = (B - sqrtD) / 2
where sqrtD = sqrt(B**2 - 4*C)
-> p = (s1 + sqrt(2*s2 - s1**2))/2
"""
sqrtD = (2*s2 - s1**2)**.5
assert int(sqrtD)**2 == (2*s2 - s1**2) # perfect square
sqrtD = int(sqrtD)
assert (s1 - sqrtD) % 2 == 0 # even
p = (s1 - sqrtD) // 2
q = s1 - p
assert q == ((s1 + sqrtD) // 2)
assert sqrtD == (q - p)
return p, q
NOTE: calculating integer square root of a number (~ N**4) makes the above algorithm non-linear.
Since a range is specified, you can perform radix sort. This would sort your array in O(n). Searching for duplicates in a sorted array is then O(n)
You can use simple nested for loop
int[] numArray = new int[] { 1, 2, 3, 4, 5, 7, 8, 3, 7 };
for (int i = 0; i < numArray.Length; i++)
{
for (int j = i + 1; j < numArray.Length; j++)
{
if (numArray[i] == numArray[j])
{
//DO SOMETHING
}
}
*OR you can filter the array and use recursive function if you want to get the count of occurrences*
int[] array = { 1, 2, 3, 4, 5, 4, 4, 1, 8, 9, 23, 4, 6, 8, 9, 1,4 };
int[] myNewArray = null;
int a = 1;
void GetDuplicates(int[] array)
for (int i = 0; i < array.Length; i++)
{
for (int j = i + 1; j < array.Length; j++)
{
if (array[i] == array[j])
{
a += 1;
}
}
Console.WriteLine(" {0} occurred {1} time/s", array[i], a);
IEnumerable<int> num = from n in array where n != array[i] select n;
myNewArray = null;
a = 1;
myNewArray = num.ToArray() ;
break;
}
GetDuplicates(myNewArray);
answer to 18..
you are taking an array of 9 and elements are starting from 0..so max ele will be 6 in your array. Take sum of elements from 0 to 6 and take sum of array elements. compute their difference (say d). This is p + q. Now take XOR of elements from 0 to 6 (say x1). Now take XOR of array elements (say x2). x2 is XOR of all elements from 0 to 6 except two repeated elements since they cancel out each other. now for i = 0 to 6, for each ele of array, say p is that ele a[i] so you can compute q by subtracting this ele from the d. do XOR of p and q and XOR them with x2 and check if x1==x2. likewise doing for all elements you will get the elements for which this condition will be true and you are done in O(n). Keep coding!
check this out ...
O(n) time and O(1) space complexity
for(i=0;i< n;i++)
xor=xor^arr[i]
for(i=1;i<=n-3;i++)
xor=xor^i;
So in the given example you will get the xor of 3 and 5
xor=xor & -xor //Isolate the last digit
for(i = 0; i < n; i++)
{
if(arr[i] & xor)
x = x ^ arr[i];
else
y = y ^ arr[i];
}
for(i = 1; i <= n-3; i++)
{
if(i & xor)
x = x ^ i;
else
y = y ^ i;
}
x and y are your answers
For each number: check if it exists in the rest of the array.
Without sorting you're going to have a keep track of numbers you've already visited.
in psuedocode this would basically be (done this way so I'm not just giving you the answer):
for each number in the list
if number not already in unique numbers list
add it to the unique numbers list
else
return that number as it is a duplicate
end if
end for each
How about this:
for (i=0; i<n-1; i++) {
for (j=i+1; j<n; j++) {
if (a[i] == a[j]) {
printf("%d appears more than once\n",a[i]);
break;
}
}
}
Sure it's not the fastest, but it's simple and easy to understand, and requires
no additional memory. If n is a small number like 9, or 100, then it may well be the "best". (i.e. "Best" could mean different things: fastest to execute, smallest memory footprint, most maintainable, least cost to develop etc..)
In c:
int arr[] = {2, 3, 6, 1, 5, 4, 0, 3, 5};
int num = 0, i;
for (i=0; i < 8; i++)
num = num ^ arr[i] ^i;
Since x^x=0, the numbers that are repeated odd number of times are neutralized. Let's call the unique numbers a and b.We are left with a^b. We know a^b != 0, since a != b. Choose any 1 bit of a^b, and use that as a mask ie.choose x as a power of 2 so that x & (a^b) is nonzero.
Now split the list into two sublists -- one sublist contains all numbers y with y&x == 0, and the rest go in the other sublist. By the way we chose x, we know that the pairs of a and b are in different buckets. So we can now apply the same method used above to each bucket independently, and discover what a and b are.
I have written a small programme which finds out the number of elements not repeated, just go through this let me know your opinion, at the moment I assume even number of elements are even but can easily extended for odd numbers also.
So my idea is to first sort the numbers and then apply my algorithm.quick sort can be use to sort this elements.
Lets take an input array as below
int arr[] = {1,1,2,10,3,3,4,5,5,6,6};
the number 2,10 and 4 are not repeated ,but they are in sorted order, if not sorted use quick sort to first sort it out.
Lets apply my programme on this
using namespace std;
main()
{
//int arr[] = {2, 9, 6, 1, 1, 4, 2, 3, 5};
int arr[] = {1,1,2,10,3,3,4,5,5,6,6};
int i = 0;
vector<int> vec;
int var = arr[0];
for(i = 1 ; i < sizeof(arr)/sizeof(arr[0]); i += 2)
{
var = var ^ arr[i];
if(var != 0 )
{
//put in vector
var = arr[i-1];
vec.push_back(var);
i = i-1;
}
var = arr[i+1];
}
for(int i = 0 ; i < vec.size() ; i++)
printf("value not repeated = %d\n",vec[i]);
}
This gives the output:
value not repeated= 2
value not repeated= 10
value not repeated= 4
Its simple and very straight forward, just use XOR man.
for(i=1;i<=n;i++) {
if(!(arr[i] ^ arr[i+1]))
printf("Found Repeated number %5d",arr[i]);
}
Here is an algorithm that uses order statistics and runs in O(n).
You can solve this by repeatedly calling SELECT with the median as parameter.
You also rely on the fact that After a call to SELECT,
the elements that are less than or equal to the median are moved to the left of the median.
Call SELECT on A with the median as the parameter.
If the median value is floor(n/2) then the repeated values are right to the median. So you continue with the right half of the array.
Else if it is not so then a repeated value is left to the median. So you continue with the left half of the array.
You continue this way recursively.
For example:
When A={2, 3, 6, 1, 5, 4, 0, 3, 5} n=9, then the median should be the value 4.
After the first call to SELECT
A={3, 2, 0, 1, <3>, 4, 5, 6, 5} The median value is smaller than 4 so we continue with the left half.
A={3, 2, 0, 1, 3}
After the second call to SELECT
A={1, 0, <2>, 3, 3} then the median should be 2 and it is so we continue with the right half.
A={3, 3}, found.
This algorithm runs in O(n+n/2+n/4+...)=O(n).
What about using the https://en.wikipedia.org/wiki/HyperLogLog?
Redis does http://redis.io/topics/data-types-intro#hyperloglogs
A HyperLogLog is a probabilistic data structure used in order to count unique things (technically this is referred to estimating the cardinality of a set). Usually counting unique items requires using an amount of memory proportional to the number of items you want to count, because you need to remember the elements you have already seen in the past in order to avoid counting them multiple times. However there is a set of algorithms that trade memory for precision: you end with an estimated measure with a standard error, in the case of the Redis implementation, which is less than 1%. The magic of this algorithm is that you no longer need to use an amount of memory proportional to the number of items counted, and instead can use a constant amount of memory! 12k bytes in the worst case, or a lot less if your HyperLogLog (We'll just call them HLL from now) has seen very few elements.
Well using the nested for loop and assuming the question is to find the number occurred only twice in an array.
def repeated(ar,n):
count=0
for i in range(n):
for j in range(i+1,n):
if ar[i] == ar[j]:
count+=1
if count == 1:
count=0
print("repeated:",ar[i])
arr= [2, 3, 6, 1, 5, 4, 0, 3, 5]
n = len(arr)
repeated(arr,n)
Why should we try out doing maths ( specially solving quadratic equations ) these are costly op . Best way to solve this would be t construct a bitmap of size (n-3) bits , i.e, (n -3 ) +7 / 8 bytes . Better to do a calloc for this memory , so every single bit will be initialized to 0 . Then traverse the list & set the particular bit to 1 when encountered , if the bit is set to 1 already for that no then that is the repeated no .
This can be extended to find out if there is any missing no in the array or not.
This solution is O(n) in time complexity

Resources