Related
This question already has answers here:
minimum number of steps to reduce number to 1
(10 answers)
Closed 6 years ago.
I am working on a code challenge now. My solution got "time exceed" even I have optimized it. I am seeking for help about more efficient solution or optimizing my solution a step more.
The description of the problem is :
Write a function which takes a positive integer as a string and returns the minimum number of operations needed to transform the number to 1. The number is up to 309 digits long, so there won't too many character than you can express in that many digits.
The transform process is limited to three operations:
1. Add 1
2. Subtract 1
3. Divide the number by 2 (only even number allow here)
My idea is to use DFS to traverse all possible solution with memorization to speed it up. But it does exceed the time limitation. The problem can not use dp because dp require a very large array to memorize. Below is my code:
private static int dfs(String num, int step,Map<String,Integer> memory){
if(num.equals("1")){
return step;
}
Integer size = memory.get(num);
if(size != null && size < step){
return Integer.MAX_VALUE;
}
memory.put(num, step);
int min = Integer.MAX_VALUE;
int lastDigit = num.charAt(num.length() - 1) - '0';
if(lastDigit % 2 == 0){
min = Math.min(min, dfs(divideBy2(num), step + 1, memory));
}else{
min = Math.min(min, dfs(divideBy2(num), step + 2, memory));
min = Math.min(min, dfs(divideBy2(plusOne(num)), step + 2, memory));
}
return min;
}
private static String plusOne(String num){
StringBuilder sb = new StringBuilder();
int carry = 1;
for(int i = num.length() - 1; i >=0; i--){
int d = (carry + num.charAt(i) - '0') % 10;
carry = (carry + num.charAt(i) - '0') / 10;
sb.insert(0, d);
}
if(carry == 1){
sb.insert(0, carry);
}
return sb.toString();
}
private static String divideBy2(String num){
StringBuilder sb = new StringBuilder();
int x = 0;
for(int i = 0; i < num.length(); i++){
int d = (x * 10 + num.charAt(i) - '0') / 2 ;
x = (num.charAt(i) - '0') % 2 ;
if( i > 0 || (i == 0 && d != 0))
sb.append(d);
}
return sb.toString();
}
Note: After test several cases: I got some sense but can not generalize the rule.
If the current number is odd. we got two choices here: plus 1 or subtract 1. The number after the operation can be divided by 2 more times, the steps will be shorter.
Update: Hi, guys, I work all the night and find a solution to pass the test. The idea is divide the problem into 2 sub-problem: 1. if the number is even, just divide it by two. 2. if the number is odd, choose the way let the number has more tailing zeros in its bit representation. I will explain more about the odd situation: if the number is odd, the last two bit can be "01" or "11". When it is "01", decrease it by 1 , which let the last two bit become to "00". If it is "11", increase it by 1, which generate "00". By doing this, the next even number generated by the odd number can be divided more times, which is really fast in practice. Below is my code, if you have some questions about the implementation, feel free to send me a message:
public static int answer(String n) {
// Your code goes here.
int count = 0;
while(!n.equals("1")){
if((n.charAt(n.length() - 1) - '0') % 2 == 0){
n = divideBy2(n);
}else if(n.equals("3") || lastTwoBit(n)){
n = subtractOne(n);
}else{
n = plusOne(n);
}
count++;
}
return count;
}
private static boolean lastTwoBit(String num){
int n = -1;
if(num.length() == 1){
n = Integer.valueOf(num);
}else{
n = Integer.valueOf(num.substring(num.length() - 2, num.length()));
}
if(((n >>> 1) & 1) == 0){
return true;
}
return false;
}
private static String subtractOne(String num){
if(num.equals("1")){
return "0";
}
StringBuilder sb = new StringBuilder();
int carry = -1;
for(int i = num.length() - 1; i >= 0; i--){
int d = carry + num.charAt(i) - '0';
if(d < 0){
carry = -1;
sb.insert(0, '9');
}else if((d == 0 && i != 0) || d > 0){
carry = 0;
sb.insert(0, d );
}
}
return sb.toString();
}
private static String plusOne(String num){
StringBuilder sb = new StringBuilder();
int carry = 1;
int i = 0;
for(i = num.length() - 1; i >=0; i--){
if(carry == 0){
break;
}
int d = (carry + num.charAt(i) - '0') % 10;
carry = (carry + num.charAt(i) - '0') / 10;
sb.insert(0, d);
}
if(carry ==0){
sb.insert(0, num.substring(0, i + 1));
}
if(carry == 1){
sb.insert(0, carry);
}
return sb.toString();
}
private static String divideBy2(String num){
StringBuilder sb = new StringBuilder();
int x = 0;
for(int i = 0; i < num.length(); i++){
int d = (x * 10 + num.charAt(i) - '0') / 2 ;
x = (num.charAt(i) - '0') % 2 ;
if( i > 0 || (i == 0 && d != 0))
sb.append(d);
}
return sb.toString();
}
While not at 1...
if Odd... Subtract 1 => even
if Even.. Divide by 2.
just sum the ops and return.
e.g. 5593
5593 -1 = 5592 /2 = 2796 /2 = 1398 /2 = 699 -1 = 698 /2 = 349 -1 = 348 /2 = 174 /2 = 87 -1 = 86 /2 = 43 -1 = 42 /2 = 21 -1 = 20 /2 = 10 /2 = 5 -1 = 4 /2 = 2 /2 = 1
19 Operations -///-/-//-/-/-//-//
Edit: Time complexity is O(logN) for we divide the number by two / subtract and then divide.
and Space is O(1)
public int make1(string s)
{
int n = 0;
while(s != "1")
{
switch(s[s.Length-1])
{
case '0':
case '2':
case '4':
case '6':
case '8':
s = div2(s);
++n;
break;
case '1':
case '3':
case '5':
case '7':
case '9':
s = minus1(s);
s = div2(s);
n += 2;
}
}
return n;
}
Closed. This question needs to be more focused. It is not currently accepting answers.
Want to improve this question? Update the question so it focuses on one problem only by editing this post.
Closed 4 years ago.
Improve this question
The problem comes from Codility programming training and it sounds as follows:
we have an array (A[]) with n (ranging from 1 to 100,000) elements and these are our parameters. The elements of the array are integers from −2,147,483,648 to 2,147,483,647, and we need to find smallest positive integer that is NOT in the array. Of course this could be done easily in O(n*log n) by sorting them all and going through the sorted array, looking for the missing posiitve number (this last operation has O(n) worst time complexity in my solution). But according to Codility, this ENTIRE problem can be done in O(n), and I cannot see any way to do that. Could someone give some tips to let me get un-stuck?
PS Here is a link to detailed description of the problem which I'm not allowed to copy - https://codility.com/c/intro/demo35UEXH-EAT
By pigeonhole principle, at least one of the numbers 1, 2, ..., n+1 is not in the array.
Let us create a boolean array b of size n+1 to store whether each of these numbers is present.
Now, we process the input array. If we find a number from 1 to n+1, we mark the corresponding entry in b. If the number we see does not fit into these bounds, just discard it and proceed to the next one. Both cases are O(1) per input entry, total O(n).
After we are done processing the input, we can find the first non-marked entry in our boolean array b trivially in O(n).
Simple solution 100% in Java.
Please note it is O(nlogn) solution but gives 100% result in codility.
public static int solution(final int[] A)
{
Arrays.sort(A);
int min = 1;
// Starting from 1 (min), compare all elements, if it does not match
// that would the missing number.
for (int i : A) {
if (i == min) {
min++;
}
}
return min;
}
wrote this today and got 100/100. not the most elegant solution, but easy to understand -
public int solution(int[] A) {
int max = A.length;
int threshold = 1;
boolean[] bitmap = new boolean[max + 1];
//populate bitmap and also find highest positive int in input list.
for (int i = 0; i < A.length; i++) {
if (A[i] > 0 && A[i] <= max) {
bitmap[A[i]] = true;
}
if (A[i] > threshold) {
threshold = A[i];
}
}
//find the first positive number in bitmap that is false.
for (int i = 1; i < bitmap.length; i++) {
if (!bitmap[i]) {
return i;
}
}
//this is to handle the case when input array is not missing any element.
return (threshold+1);
}
public int solutionMissingInteger(int[] A) {
int solution = 1;
HashSet<Integer> hashSet = new HashSet<>();
for(int i=0; i<A.length; ++i){
if(A[i]<1) continue;
if(hashSet.add(A[i])){
//this int was not handled before
while(hashSet.contains(solution)){
solution++;
}
}
}
return solution;
}
Simple Java soution. Scored 100/100 in correctness and performance.
public int solution(int[] A) {
int smallestMissingInteger = 1;
if (A.length == 0) {
return smallestMissingInteger;
}
Set<Integer> set = new HashSet<Integer>();
for (int i = 0; i < A.length; i++) {
if (A[i] > 0) {
set.add(A[i]);
}
}
while (set.contains(smallestMissingInteger)) {
smallestMissingInteger++;
}
return smallestMissingInteger;
}
Build a hash table of all the values. For the numbers 1 to n + 1, check if they are in the hash table. At least one of them is not. Print out the lowest such number.
This is O(n) expected time (you can get with high probability). See #Gassa's answer for how to avoid the hash table in favor of a lookup table of size O(n).
JavaScript 100%
function solution(A) {
let sortedOb = {};
let biggest = 0;
A.forEach(el => {
if (el > 0) {
sortedOb[el] = 0;
biggest = el > biggest ? el : biggest;
}
});
let arr = Object.keys(sortedOb).map(el => +el);
if (arr.length == 0) return 1;
for(let i = 1; i <= biggest; i++) {
if (sortedOb[i] === undefined) return i;
}
return biggest + 1;
}
100% Javascript
function solution(A) {
// write your code in JavaScript (Node.js 4.0.0)
var max = 0;
var array = [];
for (var i = 0; i < A.length; i++) {
if (A[i] > 0) {
if (A[i] > max) {
max = A[i];
}
array[A[i]] = 0;
}
}
var min = max;
if (max < 1) {
return 1;
}
for (var j = 1; j < max; j++) {
if (typeof array[j] === 'undefined') {
return j
}
}
if (min === max) {
return max + 1;
}
}
C# scored 100%,
Explanation: use of lookup table where we store already seen values from input array, we only care about values that are greater than 0 and lower or equal than length on input array
public static int solution(int[] A)
{
var lookUpArray = new bool[A.Length];
for (int i = 0; i < A.Length; i++)
if (A[i] > 0 && A[i] <= A.Length)
lookUpArray[A[i] - 1] = true;
for (int i = 0; i < lookUpArray.Length; i++)
if (!lookUpArray[i])
return i + 1;
return A.Length + 1;
}
This is my solution is Swift 4
public func solution(_ A: inout [Int]) -> Int {
var minNum = 1
var hashSet = Set<Int>()
for int in A {
if int > 0 {
hashSet.insert(int)
}
}
while hashSet.contains(minNum) {
minNum += 1
}
return minNum
}
var array = [1,3,6]
solution(&array)
// Answer: 2
100%: the Python sort routine is not regarded as cheating...
def solution(A):
"""
Sort the array then loop till the value is higher than expected
"""
missing = 1
for elem in sorted(A):
if elem == missing:
missing += 1
if elem > missing:
break
return missing
It worked for me. It is not O(n), but little simpler:
import java.util.stream.*;
class Solution {
public int solution(int[] A) {
A = IntStream.of(A)
.filter(x->x>0)
.distinct()
.sorted()
.toArray();
int min = 1;
for(int val : A)
{
if(val==min)
min++;
else
return min;
}
return min;
}
}
My solution. 100%. In Java.
import java.util.ArrayList;
import java.util.Arrays;
import java.util.List;
public class Solution {
public int solution(int[] A) {
Arrays.sort(A);
ArrayList<Integer> positive = new ArrayList<>();
for (int i = 0; i < A.length; i++) {
if(A[i] > 0)
positive.add(A[i]);
}
if(positive.isEmpty()) return 1;
if(positive.get(0) > 1) return 1;
for(int i = 0; i < positive.size() - 1; i++) {
if(positive.get(i + 1) - positive.get(i) > 1)
return positive.get(i) + 1;
}
return positive.get(positive.size() - 1) + 1;
}
public static void main(String[] args) {
Solution solution = new Solution();
int[] A = {-5,1,2,3,4,6,7,8,9,5};
System.out.println(solution.solution(A));
}
}
javascript 100% 100%
first sort the array, you just need to scan positive elements so find index of 1 (if there is no 1 in array then answer is 1). then search elements after 1 till find missing number.
function solution(A) {
// write your code in JavaScript (Node.js 6.4.0)
var missing = 1;
// sort the array.
A.sort(function(a, b) { return a-b });
// try to find the 1 in sorted array if there is no 1 so answer is 1
if ( A.indexOf(1) == -1) { return 1; }
// just search positive numbers to find missing number
for ( var i = A.indexOf(1); i < A.length; i++) {
if ( A[i] != missing) {
missing++;
if ( A[i] != missing ) { return missing; }
}
}
// if cant find any missing number return next integer number
return missing + 1;
}
I believe the solution is more involved than 'marking' corresponding values using a boolean array of n (100,000) elements. The boolean array of size n will not 'directly' map to the possible range of values (−2,147,483,648 to 2,147,483,647).
This Java example I wrote attempts to map the 100K rows by mapping the value based on their offset from the max value. It also performs a modulus to reduce the resulting array to the same size as the sample element length.
/**
*
* This algorithm calculates the values from the min value and mods this offset with the size of the 100K sample size.
* This routine performs 3 scans.
* 1. Find the min/max
* 2. Record the offsets for the positive integers
* 3. Scan the offsets to find missing value.
*
* #author Paul Goddard
*
*/
public class SmallestPositiveIntMissing {
static int ARRAY_SIZE = 100000;
public static int solve(int[] array) {
int answer = -1;
Maxmin maxmin = getMaxmin(array);
int range = maxmin.max - maxmin.min;
System.out.println("min: " + maxmin.min);
System.out.println("max: " + maxmin.max);
System.out.println("range: " + range);
Integer[] values = new Integer[ARRAY_SIZE];
if (range == ARRAY_SIZE) {
System.out.println("No gaps");
return maxmin.max + 1;
}
for (int val: array) {
if (val > 0) {
int offset = val - maxmin.min;
int index = offset % ARRAY_SIZE;
values[index] = val;
}
}
for (int i = 0; i < ARRAY_SIZE; i++) {
if (values[i] == null) {
int missing = maxmin.min + i;
System.out.println("Missing: " + missing);
answer = missing;
break;
}
}
return answer;
}
public static Maxmin getMaxmin(int[] array) {
int max = Integer.MIN_VALUE;
int min = Integer.MAX_VALUE;
for (int val:array) {
if (val >=0) {
if (val > max) max = val;
if (val < min) min = val;
}
}
return new Maxmin(max,min);
}
public static void main(String[] args) {
int[] A = arrayBuilder();
System.out.println("Min not in array: " + solve(A));
}
public static int[] arrayBuilder() {
int[] array = new int[ARRAY_SIZE];
Random random = new Random();
System.out.println("array: ");
for (int i=0;i < ARRAY_SIZE; i++) {
array[i] = random.nextInt();
System.out.print(array[i] + ", ");
}
System.out.println(" array done.");
return array;
}
}
class Maxmin {
int max;
int min;
Maxmin(int max, int min) {
this.max = max;
this.min = min;
}
}
Sweet Swift version. 100% correct
public func solution(inout A : [Int]) -> Int {
//Create a Hash table
var H = [Int:Bool]()
// Create the minimum possible return value
var high = 1
//Iterate
for i in 0..<A.count {
// Get the highest element
high = A[i] > high ? A[i] : high
// Fill hash table
if (A[i] > 0){
H[A[i]] = true
}
}
// iterate through possible values on the hash table
for j in 1...high {
// If you could not find it on the hash, return it
if H[j] != true {
return j
} else {
// If you went through all values on the hash
// and can't find it, return the next higher value
// e.g.: [1,2,3,4] returns 5
if (j == high) {
return high + 1
}
}
}
return high
}
int[] copy = new int[A.length];
for (int i : A)
{
if (i > 0 && i <= A.length)
{
copy[i - 1] = 1;
}
}
for (int i = 0; i < copy.length; i++)
{
if (copy[i] == 0)
{
return i + 1;
}
}
return A.length + 1;
Swift 3 - 100%
public func solution(_ A : inout [Int]) -> Int {
// write your code in Swift 3.0 (Linux)
var solution = 1
var hashSet = Set<Int>()
for int in A
{
if int > 0
{
hashSet.insert(int)
while hashSet.contains(solution)
{
solution += 1
}
}
}
return solution
}
Thanks to Marian's answer above.
This is my solution using python:
def solution(A):
m = max(A)
if m <= 0:
return 1
if m == 1:
return 2
# Build a sorted list with all elements in A
s = sorted(list(set(A)))
b = 0
# Iterate over the unique list trying to find integers not existing in A
for i in xrange(len(s)):
x = s[i]
# If the current element is lte 0, just skip it
if x <= 0:
continue;
b = b + 1
# If the current element is not equal to the current position,
# it means that the current position is missing from A
if x != b:
return b
return m + 1
Scored 100%/100% https://codility.com/demo/results/demoDCU7CA-SBR/
Create a binary array bin of N+1 length (C uses 0 based indexing)
Traverse the binary array O(n)
If A[i] is within the bounds of bin then mark bin entry at index A[i] as present or true.
Traverse the binary array again
Index of any bin entry that is not present or false is your missing integer
~
#include<stdio.h>
#include<stdlib.h>
#include<stdbool.h>
int solution(int A[], int N) {
// write your code in C99 (gcc 6.2.0)
int i;
bool *bin = (bool *)calloc((N+1),sizeof(bool));
for (i = 0; i < N; i++)
{
if (A[i] > 0 && A[i] < N+1)
{
bin[A[i]] = true;
}
}
for (i = 1; i < N+1; i++)
{
if (bin[i] == false)
{
break;
}
}
return i;
}
May be helpful, I am using arithmetic progression to calculate the sum, and using binary searach the element is fetched. checked with array of couple of hundred values works good. As there is one for loop and itression in step of 2, O(n/2) or less
def Missingelement (A):
B = [x for x in range(1,max(A)+1,1)]
n1 = len(B) - 1
begin = 0
end = (n1)//2
result = 0
print(A)
print(B)
if (len(A) < len(B)):
for i in range(2,n1,2):
if BinSum(A,begin,end) > BinSum(B,begin,end) :
end = (end + begin)//2
if (end - begin) <= 1 :
result=B[begin + 1 ]
elif BinSum(A,begin,end) == BinSum(B,begin,end):
r = end - begin
begin = end
end = (end + r)
if begin == end :
result=B[begin + 1 ]
return result
def BinSum(C,begin,end):
n = (end - begin)
if end >= len(C):
end = len(C) - 1
sum = n*((C[begin]+C[end])/2)
return sum
def main():
A=[1,2,3,5,6,7,9,10,11,12,14,15]
print ("smallest number missing is ",Missingelement(A))
if __name__ == '__main__': main()
Code for C, in fact, this can be used for any programming language without any change in the logic.
Logic is sum of N number is N*(N+1)/2.
int solution(int A[], int N) {
// write your code in C99
long long sum=0;
long long i;
long long Nsum=0;
for(i=0;i<N;i++){
sum=sum + (long long)A[i];
}
if (N%2==0){
Nsum= (N+1)*((N+2)/2);
return (int)(Nsum-sum);
}
else{
Nsum= ((N+1)/2)*(N+2);
return (int)(Nsum-sum);
}
}
This gave the 100/100 score.
This solution gets 100/100 on the test:
class Solution {
public int solution(int[] A) {
int x = 0;
while (x < A.length) {
// Keep swapping the values into the matching array positions.
if (A[x] > 0 && A[x] <= A.length && A[A[x]-1] != A[x]) {
swap(A, x, A[x] - 1);
} else {
x++; // Just need to increment when current element and position match.
}
}
for (int y=0; y < A.length; y++) {
// Find first element that doesn't match position.
// Array is 0 based while numbers are 1 based.
if (A[y] != y + 1) {
return y + 1;
}
}
return A.length + 1;
}
private void swap (int[] a, int i, int j) {
int t = a[i];
a[i] = a[j];
a[j] = t;
}
}
100% in PHP https://codility.com/demo/results/trainingKFXWKW-56V/
function solution($A){
$A = array_unique($A);
sort($A);
if (empty($A)) return 1;
if (max($A) <= 0) return 1;
if (max($A) == 1) return 2;
if (in_array(1, $A)) {
$A = array_slice($A, array_search(1, $A)); // from 0 to the end
array_unshift($A, 0); // Explanation 6a
if ( max($A) == array_search(max($A), $A)) return max($A) + 1; // Explanation 6b
for ($i = 1; $i <= count($A); $i++){
if ($A[$i] != $i) return $i; // Explanation 6c
}
} else {
return 1;
}
}
// Explanation
remove all duplicates
sort from min to max
if the array is empty return 1
if max of array is zero or less, return 1
if max of array is 1, return 2 // next positive integer
all other cases:
6a) split the array from value 1 to the end and add 0 before first number
6b) if the value of last element of array is the max of array, then the array is ascending so we return max + 1 // next positive integer
6c) if the array is not ascending, we find a missing number by a function for: if key of element is not as value the element but it should be (A = [0=>0, 1=>1,2=>3,...]), we return the key, because we expect the key and value to be equal.
Here is my solution, it Yields 88% in evaluation- Time is O(n), Correctness 100%, Performance 75%. REMEMBER - it is possible to have an array of all negative numbers, or numbers that exceed 100,000. Most of the above solutions (with actual code) yield much lower scores, or just do not work. Others seem to be irrelevant to the Missing Integer problem presented on Codility.
int compare( const void * arg1, const void * arg2 )
{
return *((int*)arg1) - *((int*)arg2);
}
solution( int A[], int N )
{
// Make a copy of the original array
// So as not to disrupt it's contents.
int * A2 = (int*)malloc( sizeof(int) * N );
memcpy( A2, A1, sizeof(int) * N );
// Quick sort it.
qsort( &A2[0], N, sizeof(int), compare );
// Start out with a minimum of 1 (lowest positive number)
int min = 1;
int i = 0;
// Skip past any negative or 0 numbers.
while( (A2[i] < 0) && (i < N )
{
i++;
}
// A variable to tell if we found the current minimum
int found;
while( i < N )
{
// We have not yet found the current minimum
found = 0;
while( (A2[i] == min) && (i < N) )
{
// We have found the current minimum
found = 1;
// move past all in the array that are that minimum
i++;
}
// If we are at the end of the array
if( i == N )
{
// Increment min once more and get out.
min++;
break;
}
// If we found the current minimum in the array
if( found == 1 )
{
// progress to the next minimum
min++;
}
else
{
// We did not find the current minimum - it is missing
// Get out - the current minimum is the missing one
break;
}
}
// Always free memory.
free( A2 );
return min;
}
My 100/100 solution
public int solution(int[] A) {
Arrays.sort(A);
for (int i = 1; i < 1_000_000; i++) {
if (Arrays.binarySearch(A, i) < 0){
return i;
}
}
return -1;
}
static int spn(int[] array)
{
int returnValue = 1;
int currentCandidate = 2147483647;
foreach (int item in array)
{
if (item > 0)
{
if (item < currentCandidate)
{
currentCandidate = item;
}
if (item <= returnValue)
{
returnValue++;
}
}
}
return returnValue;
}
For a given n and m I iterate over all n by m partial circulant matrices with entries that are either 0 or 1. I want to find if there is a matrix such that there are no two subsets of the columns that give the same sum. Here when we add columns we just do it elementwise. My current code uses constraint programming via ortools. However it is not as fast I would like. For n = 7 and m = 12 it takes over 3 minutes and for n = 10, m = 18 it doesn't terminate even though there are only 2^18 = 262144 different matrices to consider. Here is my code.
from scipy.linalg import circulant
import numpy as np
import itertools
from ortools.constraint_solver import pywrapcp as cs
n = 7
m = 12
def isdetecting(matrix):
X = np.array([solver.IntVar(values) for i in range(matrix.shape[1])])
X1 = X.tolist()
for row in matrix:
x = X[row].tolist()
solver.Add(solver.Sum(x) == 0)
db = solver.Phase(X1, solver.INT_VAR_DEFAULT, solver.INT_VALUE_DEFAULT)
solver.NewSearch(db)
count = 0
while (solver.NextSolution() and count < 2):
solution = [x.Value() for x in X1]
count += 1
solver.EndSearch()
if (count < 2):
return True
values = [-1,0,1]
solver = cs.Solver("scip")
for row in itertools.product([0,1],repeat = m):
M = np.array(circulant(row)[0:n], dtype=bool)
if isdetecting(M):
print M.astype(int)
break
Can this problem be solved fast enough so that n = 10, m = 18 can be solved?
One problem is that you are declaring the "solver" variable globally and it seems to confuse or-tools to reuse it many times. When moving it inside "isdetecting", then the (7,12) problem is solved much faster, in about 7 seconds (compared to 2:51 minutes for the original model). I haven't checked it for the larger problem, though.
Also, it might be good idea to test different labelings (instead of solver.INT_VAR_DEFAULT and solver.INT_VALUE_DEFAULT), though binary value tend to be not very sensitive to different labelings. See the code for another labeling.
def isdetecting(matrix):
solver = cs.Solver("scip") # <----
X = np.array([solver.IntVar(values) for i in range(matrix.shape[1])])
X1 = X.tolist()
for row in matrix:
x = X[row].tolist()
solver.Add(solver.Sum(x) == 0)
# db = solver.Phase(X1, solver.INT_VAR_DEFAULT, solver.INT_VALUE_DEFAULT)
db = solver.Phase(X1, solver.CHOOSE_FIRST_UNBOUND, solver.ASSIGN_CENTER_VALUE)
solver.NewSearch(db)
count = 0
while (solver.NextSolution() and count < 2):
solution = [x.Value() for x in X1]
count += 1
solver.EndSearch()
if (count < 2):
print "FOUND"
return True
Edit: Here are constraints to remove the all-0 solutions as mentioned in the comments. What I know of, it require a separate list. It now takes a little longer (10.4s vs 7s).
X1Abs = [solver.IntVar(values, 'X1Abs[%i]' % i) for i in range(X1_len)]
for i in range(X1_len):
solver.Add(X1Abs[i] == abs(X1[i]))
solver.Add(solver.Sum(X1Abs) > 0)
Something like this is what I had in mind. I'd estimate the running time for command line parameters 10 18 at less than 8 hours on my machine.
public class Search {
public static void main(String[] args) {
int n = Integer.parseInt(args[0]);
int m = Integer.parseInt(args[1]);
int row = search(n, m);
if (row >= 0) {
printRow(m, row);
}
}
private static int search(int n, int m) {
if (n < 0 || m < n || m >= 31 || powOverflows(m + 1, n)) {
throw new IllegalArgumentException();
}
long[] column = new long[m];
long[] sums = new long[1 << m];
int row = 1 << m;
while (row-- > 0) {
System.err.println(row);
for (int j = 0; j < m; j++) {
column[j] = 0;
for (int i = 0; i < n; i++) {
column[j] = (column[j] * (m + 1)) + ((row >> ((i + j) % m)) & 1);
}
}
for (int subset = 0; subset < (1 << m); subset++) {
long sum = 0;
for (int j = 0; j < m; j++) {
if (((subset >> j) & 1) == 1) {
sum += column[j];
}
}
sums[subset] = sum;
}
java.util.Arrays.sort(sums);
boolean duplicate = false;
for (int k = 1; k < (1 << m); k++) {
if (sums[k - 1] == sums[k]) {
duplicate = true;
break;
}
}
if (!duplicate) {
break;
}
}
return row;
}
private static boolean powOverflows(long b, int e) {
if (b <= 0 || e < 0) {
throw new IllegalArgumentException();
}
if (e == 0) {
return false;
}
long max = Long.MAX_VALUE;
while (e > 1) {
if (b > Integer.MAX_VALUE) {
return true;
}
if ((e & 1) == 1) {
max /= b;
}
b *= b;
e >>= 1;
}
return b > max;
}
private static void printRow(int m, int row) {
for (int j = 0; j < m; j++) {
System.out.print((row >> j) & 1);
}
System.out.println();
}
}
I am solving a problem to find out all the 4 digit Vampire numbers.
A Vampire Number v=x*y is defined as a number with 'n' even number of digits formed by multiplying a pair of 'n/2'-digit numbers (where the digits are taken from the original number in any order)x and y together. If v is a vampire number, then x&y and are called its "fangs."
Examples of vampire numbers are:
1. 1260=21*60
2. 1395=15*93
3. 1530=30*51
I have tried the brute force algorithm to combine different digits of a given number and multiply them together . But this method is highly inefficient and takes up a lot of time.
Is there a more efficient algorithmic solution to this problem?
Or you can use a property of vampire numbers described on this page (linked from Wikipedia) :
An important theoretical result found by Pete Hartley:
If x·y is a vampire number then x·y == x+y (mod 9)
Proof: Let mod be the binary modulo operator and d(x) the sum of the decimal
digits of x. It is well-known that d(x) mod 9 = x mod 9, for all x.
Assume x·y is a vampire. Then it contains the same digits as x and y,
and in particular d(x·y) = d(x)+d(y). This leads to:
(x·y) mod 9 = d(x·y) mod 9 = (d(x)+d(y)) mod 9 = (d(x) mod 9 + d(y) mod 9) mod 9
= (x mod 9 + y mod 9) mod 9 = (x+y) mod 9
The solutions to the congruence are (x mod 9, y mod 9) in {(0,0),
(2,2), (3,6), (5,8), (6,3), (8,5)}
So your code could look like this :
for(int i=18; i<100; i=i+9){ // 18 is the first multiple of 9 greater than 10
for(int j=i; j<100; j=j+9){ // Start at i because as #sh1 said it's useless to check both x*y and y*x
checkVampire(i,j);
}
}
for(int i=11; i<100; i=i+9){ // 11 is the first number greater than 10 which is = 2 mod 9
for(int j=i; j<100; j=j+9){
checkVampire(i,j);
}
}
for(int i=12; i<100; i=i+9){
for(int j=i+3; j<100; j=j+9){
checkVampire(i,j);
}
}
for(int i=14; i<100; i=i+9){
for(int j=i+3; j<100; j=j+9){
checkVampire(i,j);
}
}
// We don't do the last 2 loops, again for symmetry reasons
Since they are 40 elements in each of the sets like {(x mod 9, y mod 9) = (0,0); 10 <= x <= y <= 100}, you only do 4*40 = 160 iterations, when a brute-force gives you 10ˆ4 iterations. You can do even less operations if you take into account the >= 1000 constraint, for instance you can avoid checking if j < 1000/i.
Now you can easily scale up to find vampires with more than 4 digits =)
Iterate over all possible fangs (100 x 100 = 10000 possibilities), and find if their product has the same digits as the fangs.
Yet another brute force (C) version, with a free bubble sort to boot...
#include <stdio.h>
static inline void bubsort(int *p)
{ while (1)
{ int s = 0;
for (int i = 0; i < 3; ++i)
if (p[i] > p[i + 1])
{ s = 1;
int t = p[i]; p[i] = p[i + 1]; p[i + 1] = t;
}
if (!s) break;
}
}
int main()
{ for (int i = 10; i < 100; ++i)
for (int j = i; j < 100; ++j)
{ int p = i * j;
if (p < 1000) continue;
int xd[4];
xd[0] = i % 10;
xd[1] = i / 10;
xd[2] = j % 10;
xd[3] = j / 10;
bubsort(xd);
int x = xd[0] + xd[1] * 10 + xd[2] * 100 + xd[3] * 1000;
int yd[4];
yd[0] = p % 10;
yd[1] = (p / 10) % 10;
yd[2] = (p / 100) % 10;
yd[3] = (p / 1000);
bubsort(yd);
int y = yd[0] + yd[1] * 10 + yd[2] * 100 + yd[3] * 1000;
if (x == y)
printf("%2d * %2d = %4d\n", i, j, p);
}
return 0;
}
Runs pretty much instantaneously. Variable names aren't too descriptive, but should be pretty obvious...
The basic idea is to start with two potential fangs, break them down into digits, and sort the digits for easy comparison. Then we do the same with the product - break it down to digits and sort. Then we re-constitute two integers from the sorted digits, and if they're equal, we have a match.
Possible improvements: 1) start j at 1000 / i instead of i to avoid having to do if (p < 1000) ..., 2) maybe use insertion sort instead of bubble sort (but who's gonna notice those 2 extra swaps?), 3) use a real swap() implementation, 4) compare the arrays directly rather than building a synthetic integer out of them. Not sure any of those would make any measurable difference, though, unless you run it on a Commodore 64 or something...
Edit: Just out of curiosity, I took this version and generalized it a bit more to work for the 4, 6 and 8 digit cases - without any major optimization, it can find all the 8-digit vampire numbers in < 10 seconds...
This is an ugly hack (brute force, manual checking for permutations, unsafe buffer operations, produces dupes, etc.) but it does the job. Your new exercise is to improve it :P
Wikipedia claims that there are 7 vampire numbers which are 4 digits long. The full code has found them all, even some duplicates.
Edit: Here's a slightly better comparator function.
Edit 2: Here's a C++ version that uniques results (therefore it avoids duplicates) using an std::map (and stores the last occurrence of the particular vampire number along with its factors in it). It also meets the criterion that at least one of the factors should not end with 0, i. e. a number is not a vampire number if both of the multiplicands are divisible by then. This test looks for 6-digit vampire numbers and it does indeed find exactly 148 of them, in accordance with what Wikipedia sates.
The original code:
#include <stdio.h>
void getdigits(char buf[], int n)
{
while (n) {
*buf++ = n % 10;
n /= 10;
}
}
int is_vampire(const char n[4], const char i[2], const char j[2])
{
/* maybe a bit faster if unrolled manually */
if (i[0] == n[0]
&& i[1] == n[1]
&& j[0] == n[2]
&& j[1] == n[3])
return 1;
if (i[0] == n[1]
&& i[1] == n[0]
&& j[0] == n[2]
&& j[1] == n[3])
return 1;
if (i[0] == n[0]
&& i[1] == n[1]
&& j[0] == n[3]
&& j[1] == n[2])
return 1;
if (i[0] == n[1]
&& i[1] == n[0]
&& j[0] == n[3]
&& j[1] == n[2])
return 1;
// et cetera, the following 20 repetitions are redacted for clarity
// (this really should be a loop, shouldn't it?)
return 0;
}
int main()
{
for (int i = 10; i < 100; i++) {
for (int j = 10; j < 100; j++) {
int n = i * j;
if (n < 1000)
continue;
char ndigits[4];
getdigits(ndigits, n);
char idigits[2];
char jdigits[2];
getdigits(idigits, i);
getdigits(jdigits, j);
if (is_vampire(ndigits, idigits, jdigits))
printf("%d * %d = %d\n", i, j, n);
}
}
return 0;
}
I wouldn't have given up so easily on brute force. You have distinct set of numbers, 1000 to 9999 that you must run through. I would divide up the set into some number of subsets, and then spin up threads to handle each subset.
You could further divide the work be coming up with the various combinations of each number; IIRC my discrete math, you have 4*3*2 or 24 combinations for each number to try.
A producer / consumer approach might be worthwhile.
Iteration seems fine to me, since you only need to do this once to find all the values and you can just cache them afterwards. Python (3) version that takes about 1.5 seconds:
# just some setup
from itertools import product, permutations
dtoi = lambda *digits: int(''.join(str(digit) for digit in digits))
gen = ((dtoi(*digits), digits) for digits in product(range(10), repeat=4) if digits[0] != 0)
l = []
for val, digits in gen:
for check1, check2 in ((dtoi(*order[:2]), dtoi(*order[2:])) for order in permutations(digits) if order[0] > 0 and order[2] > 0):
if check1 * check2 == val:
l.append(val)
break
print(l)
Which will give you [1260, 1395, 1435, 1530, 1827, 2187, 6880]
EDIT: full brute force that weeds out identical X and Y values...
import java.util.ArrayList;
import java.util.Collections;
import java.util.List;
public class Vampire {
public static void main(String[] args) {
for (int x = 10; x < 100; x++) {
String sx = String.valueOf(x);
for (int y = x; y < 100; y++) {
int v = x * y;
String sy = String.valueOf(y);
String sv = String.valueOf(v);
if (sortVampire(sx + sy).equals(sortVampire(sv))) {
System.out.printf("%d * %d = %d%n", x, y, v);
}
}
}
}
private static List<Character> sortVampire(String v) {
List<Character> vc = new ArrayList<Character>();
for (int j = 0; j < v.length(); j++) {
vc.add(v.charAt(j));
}
Collections.sort(vc);
return vc;
}
}
Brute force version in C# with LINQ:
class VampireNumbers
{
static IEnumerable<int> numberToDigits(int number)
{
while(number > 0)
{
yield return number % 10;
number /= 10;
}
}
static bool isVampire(int first, int second, int result)
{
var resultDigits = numberToDigits(result).OrderBy(x => x);
var vampireDigits = numberToDigits(first)
.Concat(numberToDigits(second))
.OrderBy(x => x);
return resultDigits.SequenceEqual(vampireDigits);
}
static void Main(string[] args)
{
var vampires = from fang1 in Enumerable.Range(10, 89)
from fang2 in Enumerable.Range(10, 89)
where fang1 < fang2
&& isVampire(fang1, fang2, fang1 * fang2)
select new { fang1, fang2 };
foreach(var vampire in vampires)
{
Console.WriteLine(vampire.fang1 * vampire.fang2
+ " = "
+ vampire.fang1
+ " * "
+ vampire.fang2);
}
}
}
Similar to someone mentioned above, my method is to first find all permutations of a number, then split them in half to form two 2-digit numbers, and test if their product equal to the original number.
Another interesting discussion above is how many permutations a number can have. Here is my opinion:
(1) a number whose four digitals are the same has 1 permutation;
(2) a number who has only two different digits has 6 permutations (it doesn't matter if it contains zeros, because we don't care after permutation if it is still a 4-digit number);
(3) a number who has three different digits has 12 permutations;
(4) a number with all four different digits has 24 permutations.
public class VampireNumber {
// method to find all permutations of a 4-digit number
public static void permuta(String x, String s, int v)
{for(int i = 0; i < s.length(); i++)
{permuta( x + s.charAt(i), s.substring(0,i) + s.substring(i+1), v);
if (s.length() == 1)
{x = x + s;
int leftpart = Integer.parseInt(x.substring(0,2));
int rightpart = Integer.parseInt(x.substring(2));
if (leftpart*rightpart == v)
{System.out.println("Vampir = " + v);
}
}
}
}
public static void main(String[] args){
for (int i = 1000; i < 10000; i++) {
permuta("", Integer.toString(i), i); //convert the integer to a string
}
}
}
The approach I would try would be to loop through each number in [1000, 9999], and test if any permutation of its digits (split in the middle) multiplied to make it.
This will require (9999 - 1000) * 24 = 215,976 tests, which should execute acceptably fast on a modern machine.
I would definitely store the digits separately, so you can avoid having to do something like a bunch of division to extract the digits from a single integer.
If you write your code such that you're only ever doing integer addition and multiplication (and maybe the occasional division to carry), it should be pretty fast. You could further increase the speed by skipping two-digit pairs which "obviously" won't work - e.g., ones with leading zeros (note that the largest product than can be produced by a one digit number and a two digit number is 9 * 99, or 891).
Also note that this approach is embarassingly parallel (http://en.wikipedia.org/wiki/Embarrassingly_parallel), so if you really need to speed it up even more then you should look into testing the numbers in separate threads.
<?php
for ($i = 10; $i <= 99; $j++) {
// Extract digits
$digits = str_split($i);
// Loop through 2nd number
for ($j = 10; $j <= 99; $j++) {
// Extract digits
$j_digits = str_split($j);
$digits[2] = $j_digits[0];
$digits[3] = $j_digits[1];
$product = $i * $j;
$product_digits = str_split($product);
// check if fangs
$inc = 0;
while (in_array($digits[$inc], $product_digits)) {
// Remove digit from product table
/// So AAAA -> doesnt match ABCD
unset($product_digits[$array_serach($digits[$inc], $product_digits)]);
$inc++;
// If reached 4 -> vampire number
if ($inc == 4) {
$vampire[] = $product;
break;
}
}
}
}
// Print results
print_r($vampire);
?>
Took less than a second on PHP. couldn't even tell it had to run 8100 computations... computers are fast!
Results:
Gives you all the 4 digits plus some are repeated. You can further process the data and remove duplicates.
It seems to me that to perform the fewest possible tests without relying on any particularly abstract insights, you probably want to iterate over the fangs and cull any obviously pointless candidates.
For example, since x*y == y*x about half your search space can be eliminated by only evaluating cases where y > x. If the largest two-digit fang is 99 then the smallest which can make a four-digit number is 11, so don't start lower than 11.
EDIT:
OK, throwing everything I thought of into the mix (even though it looks silly against the leading solution).
for (x = 11; x < 100; x++)
{
/* start y either at x, or if x is too small then 1000 / x */
for (y = (x * x < 1000 ? 1000 / x : x); y < 100; y++)
{
int p = x * y;
/* if sum of digits in product is != sum of digits in x+y, then skip */
if ((p - (x + y)) % 9 != 0)
continue;
if (is_vampire(p, x, y))
printf("%d\n", p);
}
}
and the test, since I haven't seen anyone use a histogram, yet:
int is_vampire(int p, int x, int y)
{
int h[10] = { 0 };
int i;
for (i = 0; i < 4; i++)
{
h[p % 10]++;
p /= 10;
}
for (i = 0; i < 2; i++)
{
h[x % 10]--;
h[y % 10]--;
x /= 10;
y /= 10;
}
for (i = 0; i < 10; i++)
if (h[i] != 0)
return 0;
return 1;
}
1260 1395 1435 1530 1827 2187 6880 is vampire
I am new to programming... But there are only 12 combinations in finding all 4-digit vampire numbers. My poor answer is:
public class VampNo {
public static void main(String[] args) {
for(int i = 1000; i < 10000; i++) {
int a = i/1000;
int b = i/100%10;
int c = i/10%10;
int d = i%10;
if((a * 10 + b) * (c * 10 + d) == i || (b * 10 + a) * (d * 10 + c) == i ||
(a * 10 + d) * (b * 10 + c) == i || (d * 10 + a) * (c * 10 + b) == i ||
(a * 10 + c) * (b * 10 + d) == i || (c * 10 + a) * (d * 10 + b) == i ||
(a * 10 + b) * (d * 10 + c) == i || (b * 10 + a) * (c * 10 + d) == i ||
(b * 10 + c) * (d * 10 + a) == i || (c * 10 + b) * (a * 10 + d) == i ||
(a * 10 + c) * (d * 10 + b) == i || (c * 10 + a) * (b * 10 + d) == i)
System.out.println(i + " is vampire");
}
}
}
The main task now is to simplify boolean expression in If() block
I've edited Owlstead's algorithm a bit to make it more understandable to Java beginners/learners.
import java.util.Arrays;
public class Vampire {
public static void main(String[] args) {
for (int x = 10; x < 100; x++) {
String sx = Integer.toString(x);
for (int y = x; y < 100; y++) {
int v = x * y;
String sy = Integer.toString(y);
String sv = Integer.toString(v);
if( Arrays.equals(sortVampire(sx + sy), sortVampire(sv)))
System.out.printf("%d * %d = %d%n", x, y, v);
}
}
}
private static char[] sortVampire (String v){
char[] sortedArray = v.toCharArray();
Arrays.sort(sortedArray);
return sortedArray;
}
}
This python code run very fast (O(n2))
result = []
for i in range(10,100):
for j in range(10, 100):
list1 = []
list2 = []
k = i * j
if k < 1000 or k > 10000:
continue
else:
for item in str(i):
list1.append(item)
for item in str(j):
list1.append(item)
for item in str(k):
list2.append(item)
flag = 1
for each in list1:
if each not in list2:
flag = 0
else:
list2.remove(each)
for each in list2:
if each not in list1:
flag = 0
if flag == 1:
if k not in result:
result.append(k)
for each in result:
print(each)
And here is my code. To generate zombie numbers we need to use Random class :)
import java.io.PrintStream;
import java.util.Set;
import java.util.HashSet;
import java.util.Iterator;
class VampireNumbers {
static PrintStream p = System.out;
private static Set<Integer> findVampireNumber() {
Set<Integer> vampireSet = new HashSet<Integer>();
for (int y = 1000; y <= 9999; y++) {
char[] numbersSeparately = ("" + y).toCharArray();
int numberOfDigits = numbersSeparately.length;
for (int i = 0; i < numberOfDigits; i++) {
for (int j = 0; j < numberOfDigits; j++) {
if (i != j) {
int value1 = Integer.valueOf("" + numbersSeparately[i] + numbersSeparately[j]);
int ki = -1;
for (int k = 0; k < numberOfDigits; k++) {
if (k != i && k != j) {
ki = k;
}
}
int kj = -1;
for (int t = 0; t < numberOfDigits; t++) {
if (t != i && t != j && t != ki) {
kj = t;
}
}
int value21 = Integer.valueOf("" + numbersSeparately[ki] + numbersSeparately[kj]);
int value22 = Integer.valueOf("" + numbersSeparately[kj] + numbersSeparately[ki]);
if (value1 * value21 == y && !(numbersSeparately[j] == 0 && numbersSeparately[kj] == 0)
|| value1 * value22 == y
&& !(numbersSeparately[j] == 0 && numbersSeparately[ki] == 0)) {
vampireSet.add(y);
}
}
}
}
}
return vampireSet;
}
public static void main(String[] args) {
Set<Integer> vampireSet = findVampireNumber();
Iterator<Integer> i = vampireSet.iterator();
int number = 1;
while (i.hasNext()) {
p.println(number + ": " + i.next());
number++;
}
}
}
Given an array A of N integers we draw N discs in a 2D plane, such that i-th disc has center in (0,i) and a radius A[i]. We say that k-th disc and j-th disc intersect, if k-th and j-th discs have at least one common point.
Write a function
int number_of_disc_intersections(int[] A);
which given an array A describing N discs as explained above, returns the number of pairs of intersecting discs. For example, given N=6 and
A[0] = 1
A[1] = 5
A[2] = 2
A[3] = 1
A[4] = 4
A[5] = 0
there are 11 pairs of intersecting discs:
0th and 1st
0th and 2nd
0th and 4th
1st and 2nd
1st and 3rd
1st and 4th
1st and 5th
2nd and 3rd
2nd and 4th
3rd and 4th
4th and 5th
so the function should return 11.
The function should return -1 if the number of intersecting pairs exceeds 10,000,000. The function may assume that N does not exceed 10,000,000.
O(N) complexity and O(N) memory solution.
private static int Intersections(int[] a)
{
int result = 0;
int[] dps = new int[a.length];
int[] dpe = new int[a.length];
for (int i = 0, t = a.length - 1; i < a.length; i++)
{
int s = i > a[i]? i - a[i]: 0;
int e = t - i > a[i]? i + a[i]: t;
dps[s]++;
dpe[e]++;
}
int t = 0;
for (int i = 0; i < a.length; i++)
{
if (dps[i] > 0)
{
result += t * dps[i];
result += dps[i] * (dps[i] - 1) / 2;
if (10000000 < result) return -1;
t += dps[i];
}
t -= dpe[i];
}
return result;
}
So you want to find the number of intersections of the intervals [i-A[i], i+A[i]].
Maintain a sorted array (call it X) containing the i-A[i] (also have some extra space which has the value i+A[i] in there).
Now walk the array X, starting at the leftmost interval (i.e smallest i-A[i]).
For the current interval, do a binary search to see where the right end point of the interval (i.e. i+A[i]) will go (called the rank). Now you know that it intersects all the elements to the left.
Increment a counter with the rank and subtract current position (assuming one indexed) as we don't want to double count intervals and self intersections.
O(nlogn) time, O(n) space.
Python 100 / 100 (tested) on codility, with O(nlogn) time and O(n) space.
Here is #noisyboiler's python implementation of #Aryabhatta's method with comments and an example.
Full credit to original authors, any errors / poor wording are entirely my fault.
from bisect import bisect_right
def number_of_disc_intersections(A):
pairs = 0
# create an array of tuples, each containing the start and end indices of a disk
# some indices may be less than 0 or greater than len(A), this is fine!
# sort the array by the first entry of each tuple: the disk start indices
intervals = sorted( [(i-A[i], i+A[i]) for i in range(len(A))] )
# create an array of starting indices using tuples in intervals
starts = [i[0] for i in intervals]
# for each disk in order of the *starting* position of the disk, not the centre
for i in range(len(starts)):
# find the end position of that disk from the array of tuples
disk_end = intervals[i][1]
# find the index of the rightmost value less than or equal to the interval-end
# this finds the number of disks that have started before disk i ends
count = bisect_right(starts, disk_end )
# subtract current position to exclude previous matches
# this bit seemed 'magic' to me, so I think of it like this...
# for disk i, i disks that start to the left have already been dealt with
# subtract i from count to prevent double counting
# subtract one more to prevent counting the disk itsself
count -= (i+1)
pairs += count
if pairs > 10000000:
return -1
return pairs
Worked example: given [3, 0, 1, 6] the disk radii would look like this:
disk0 ------- start= -3, end= 3
disk1 . start= 1, end= 1
disk2 --- start= 1, end= 3
disk3 ------------- start= -3, end= 9
index 3210123456789 (digits left of zero are -ve)
intervals = [(-3, 3), (-3, 9), (1, 1), (1,3)]
starts = [-3, -3, 1, 1]
the loop order will be: disk0, disk3, disk1, disk2
0th loop:
by the end of disk0, 4 disks have started
one of which is disk0 itself
none of which could have already been counted
so add 3
1st loop:
by the end of disk3, 4 disks have started
one of which is disk3 itself
one of which has already started to the left so is either counted OR would not overlap
so add 2
2nd loop:
by the end of disk1, 4 disks have started
one of which is disk1 itself
two of which have already started to the left so are either counted OR would not overlap
so add 1
3rd loop:
by the end of disk2, 4 disks have started
one of which is disk2 itself
two of which have already started to the left so are either counted OR would not overlap
so add 0
pairs = 6
to check: these are (0,1), (0,2), (0,2), (1,2), (1,3), (2,3),
Well, I adapted Falk Hüffner's idea to c++, and made a change in the range.
Opposite to what is written above, there is no need to go beyond the scope of the array (no matter how large are the values in it).
On Codility this code received 100%.
Thank you Falk for your great idea!
int number_of_disc_intersections ( const vector<int> &A ) {
int sum=0;
vector<int> start(A.size(),0);
vector<int> end(A.size(),0);
for (unsigned int i=0;i<A.size();i++){
if ((int)i<A[i]) start[0]++;
else start[i-A[i]]++;
if (i+A[i]>=A.size()) end[A.size()-1]++;
else end[i+A[i]]++;
}
int active=0;
for (unsigned int i=0;i<A.size();i++){
sum+=active*start[i]+(start[i]*(start[i]-1))/2;
if (sum>10000000) return -1;
active+=start[i]-end[i];
}
return sum;
}
This can even be done in linear time [EDIT: this is not linear time, see comments]. In fact, it becomes easier if you ignore the fact that there is exactly one interval centered at each point, and just treat it as a set of start- and endpoints of intervals. You can then just scan it from the left (Python code for simplicity):
from collections import defaultdict
a = [1, 5, 2, 1, 4, 0]
start = defaultdict(int)
stop = defaultdict(int)
for i in range(len(a)):
start[i - a[i]] += 1
stop[i + a[i]] += 1
active = 0
intersections = 0
for i in range(-len(a), len(a)):
intersections += active * start[i] + (start[i] * (start[i] - 1)) / 2
active += start[i]
active -= stop[i]
print intersections
Here's a O(N) time, O(N) space algorithm requiring 3 runs across the array and no sorting, verified scoring 100%:
You're interested in pairs of discs. Each pair involves one side of one disc and the other side of the other disc. Therefore we won't have duplicate pairs if we handle one side of each disc. Let's call the sides right and left (I rotated the space while thinking about it).
An overlap is either due to a right side overlapping another disc directly at the center (so pairs equal to the radius with some care about the array length) or due to the number of left sides existing at the rightmost edge.
So we create an array that contains the number of left sides at each point and then it's a simple sum.
C code:
int solution(int A[], int N) {
int C[N];
int a, S=0, t=0;
// Mark left and middle of disks
for (int i=0; i<N; i++) {
C[i] = -1;
a = A[i];
if (a>=i) {
C[0]++;
} else {
C[i-a]++;
}
}
// Sum of left side of disks at location
for (int i=0; i<N; i++) {
t += C[i];
C[i] = t;
}
// Count pairs, right side only:
// 1. overlaps based on disk size
// 2. overlaps based on disks but not centers
for (int i=0; i<N; i++) {
a = A[i];
S += ((a<N-i) ? a: N-i-1);
if (i != N-1) {
S += C[((a<N-i) ? i+a: N-1)];
}
if (S>10000000) return -1;
}
return S;
}
I got 100 out of 100 with this C++ implementation:
#include <map>
#include <algorithm>
inline bool mySortFunction(pair<int,int> p1, pair<int,int> p2)
{
return ( p1.first < p2.first );
}
int number_of_disc_intersections ( const vector<int> &A ) {
int i, size = A.size();
if ( size <= 1 ) return 0;
// Compute lower boundary of all discs and sort them in ascending order
vector< pair<int,int> > lowBounds(size);
for(i=0; i<size; i++) lowBounds[i] = pair<int,int>(i-A[i],i+A[i]);
sort(lowBounds.begin(), lowBounds.end(), mySortFunction);
// Browse discs
int nbIntersect = 0;
for(i=0; i<size; i++)
{
int curBound = lowBounds[i].second;
for(int j=i+1; j<size && lowBounds[j].first<=curBound; j++)
{
nbIntersect++;
// Maximal number of intersections
if ( nbIntersect > 10000000 ) return -1;
}
}
return nbIntersect;
}
A Python answer
from bisect import bisect_right
def number_of_disc_intersections(li):
pairs = 0
# treat as a series of intervals on the y axis at x=0
intervals = sorted( [(i-li[i], i+li[i]) for i in range(len(li))] )
# do this by creating a list of start points of each interval
starts = [i[0] for i in intervals]
for i in range(len(starts)):
# find the index of the rightmost value less than or equal to the interval-end
count = bisect_right(starts, intervals[i][1])
# subtract current position to exclude previous matches, and subtract self
count -= (i+1)
pairs += count
if pairs > 10000000:
return -1
return pairs
100/100 c#
class Solution
{
class Interval
{
public long Left;
public long Right;
}
public int solution(int[] A)
{
if (A == null || A.Length < 1)
{
return 0;
}
var itervals = new Interval[A.Length];
for (int i = 0; i < A.Length; i++)
{
// use long to avoid data overflow (eg. int.MaxValue + 1)
long radius = A[i];
itervals[i] = new Interval()
{
Left = i - radius,
Right = i + radius
};
}
itervals = itervals.OrderBy(i => i.Left).ToArray();
int result = 0;
for (int i = 0; i < itervals.Length; i++)
{
var right = itervals[i].Right;
for (int j = i + 1; j < itervals.Length && itervals[j].Left <= right; j++)
{
result++;
if (result > 10000000)
{
return -1;
}
}
}
return result;
}
}
I'm offering one more solution because I did not find the counting principle of the previous solutions easy to follow. Though the results are the same, an explanation and more intuitive counting procedure seems worth presenting.
To begin, start by considering the O(N^2) solution that iterates over the discs in order of their center points, and counts the number of discs centered to the right of the current disc's that intersect the current disc, using the condition current_center + radius >= other_center - radius. Notice that we could get the same result counting discs centered to the left of the current disc using the condition current_center - radius <= other_center + radius.
def simple(A):
"""O(N^2) solution for validating more efficient solution."""
N = len(A)
unique_intersections = 0
# Iterate over discs in order of their center positions
for j in range(N):
# Iterate over discs whose center is to the right, to avoid double-counting.
for k in range(j+1, N):
# Increment cases where edge of current disk is at or right of the left edge of another disk.
if j + A[j] >= k - A[k]:
unique_intersections += 1
# Stop early if we have enough intersections.
# BUT: if the discs are small we still N^2 compare them all and time out.
if unique_intersections > 10000000:
return -1
return unique_intersections
We can go from O(N^2) to O(N) if we could only "look up" the number of discs to the right (or to the left!) that intersect the current disc. The key insight is to reinterpret the intersection condition as "the right edge of one disc overlaps the left edge of another disc", meaning (a ha!) the centers don't matter, only the edges.
The next insight is to try sorting the edges, taking O(N log N) time. Given a sorted array of the left edges and a sorted array of the right edges, as we scan our way from left to right along the number line, the number of left or right edges to the left of the current location point is simply the current index into left_edges and right_edges respectively: a constant-time deduction.
Finally, we use the "right edge > left edge" condition to deduce that the number of intersections between the current disc and discs that start only to the left of the current disc (to avoid duplicates) is the number of left edges to the left of the current edge, minus the number of right edges to the left of the current edge. That is, the number of discs starting to left of this one, minus the ones that closed already.
Now for this code, tested 100% on Codility:
def solution(A):
"""O(N log N) due to sorting, with O(N) pass over sorted arrays"""
N = len(A)
# Left edges of the discs, in increasing order of position.
left_edges = sorted([(p-r) for (p,r) in enumerate(A)])
# Right edges of the discs, in increasing order of position.
right_edges = sorted([(p+r) for (p,r) in enumerate(A)])
#print("left edges:", left_edges[:10])
#print("right edges:", right_edges[:10])
intersections = 0
right_i = 0
# Iterate over the discs in order of their leftmost edge position.
for left_i in range(N):
# Find the first right_edge that's right of or equal to the current left_edge, naively:
# right_i = bisect.bisect_left(right_edges, left_edges[left_i])
# Just scan from previous index until right edge is at or beyond current left:
while right_edges[right_i] < left_edges[left_i]:
right_i += 1
# Count number of discs starting left of current, minus the ones that already closed.
intersections += left_i - right_i
# Return early if we find more than 10 million intersections.
if intersections > 10000000:
return -1
#print("correct:", simple(A))
return intersections
Java 2*100%.
result is declared as long for a case codility doesn't test, namely 50k*50k intersections at one point.
class Solution {
public int solution(int[] A) {
int[] westEnding = new int[A.length];
int[] eastEnding = new int[A.length];
for (int i=0; i<A.length; i++) {
if (i-A[i]>=0) eastEnding[i-A[i]]++; else eastEnding[0]++;
if ((long)i+A[i]<A.length) westEnding[i+A[i]]++; else westEnding[A.length-1]++;
}
long result = 0; //long to contain the case of 50k*50k. codility doesn't test for this.
int wests = 0;
int easts = 0;
for (int i=0; i<A.length; i++) {
int balance = easts*wests; //these are calculated elsewhere
wests++;
easts+=eastEnding[i];
result += (long) easts*wests - balance - 1; // 1 stands for the self-intersection
if (result>10000000) return -1;
easts--;
wests-= westEnding[i];
}
return (int) result;
}
}
Swift 4 Solution 100% (Codility do not check the worst case for this solution)
public func solution(_ A : inout [Int]) -> Int {
// write your code in Swift 4.2.1 (Linux)
var count = 0
let sortedA = A.sorted(by: >)
if sortedA.isEmpty{ return 0 }
let maxVal = sortedA[0]
for i in 0..<A.count{
let maxIndex = min(i + A[i] + maxVal + 1,A.count)
for j in i + 1..<maxIndex{
if j - A[j] <= i + A[i]{
count += 1
}
}
if count > 10_000_000{
return -1
}
}
return count
}
Here my JavaScript solution, based in other solutions in this thread but implemented in other languages.
function solution(A) {
let circleEndpoints = [];
for(const [index, num] of Object.entries(A)) {
circleEndpoints.push([parseInt(index)-num, true]);
circleEndpoints.push([parseInt(index)+num, false]);
}
circleEndpoints = circleEndpoints.sort(([a, openA], [b, openB]) => {
if(a == b) return openA ? -1 : 1;
return a - b;
});
let openCircles = 0;
let intersections = 0;
for(const [endpoint, opening] of circleEndpoints) {
if(opening) {
intersections += openCircles;
openCircles ++;
} else {
openCircles --;
}
if(intersections > 10000000) return -1;
}
return intersections;
}
count = 0
for (int i = 0; i < N; i++) {
for (int j = i+1; j < N; j++) {
if (i + A[i] >= j - A[j]) count++;
}
}
It is O(N^2) so pretty slow, but it works.
This is a ruby solution that scored 100/100 on codility. I'm posting it now because I'm finding it difficult to follow the already posted ruby answer.
def solution(a)
end_points = []
a.each_with_index do |ai, i|
end_points << [i - ai, i + ai]
end
end_points = end_points.sort_by { |points| points[0]}
intersecting_pairs = 0
end_points.each_with_index do |point, index|
lep, hep = point
pairs = bsearch(end_points, index, end_points.size - 1, hep)
return -1 if 10000000 - pairs + index < intersecting_pairs
intersecting_pairs += (pairs - index)
end
return intersecting_pairs
end
# This method returns the maximally appropriate position
# where the higher end-point may have been inserted.
def bsearch(a, l, u, x)
if l == u
if x >= a[u][0]
return u
else
return l - 1
end
end
mid = (l + u)/2
# Notice that we are searching in higher range
# even if we have found equality.
if a[mid][0] <= x
return bsearch(a, mid+1, u, x)
else
return bsearch(a, l, mid, x)
end
end
Probably extremely fast. O(N). But you need to check it out. 100% on Codility.
Main idea:
1. At any point of the table, there are number of circles "opened" till the right edge of the circle, lets say "o".
2. So there are (o-1-used) possible pairs for the circle in that point. "used" means circle that have been processed and pairs for them counted.
public int solution(int[] A) {
final int N = A.length;
final int M = N + 2;
int[] left = new int[M]; // values of nb of "left" edges of the circles in that point
int[] sleft = new int[M]; // prefix sum of left[]
int il, ir; // index of the "left" and of the "right" edge of the circle
for (int i = 0; i < N; i++) { // counting left edges
il = tl(i, A);
left[il]++;
}
sleft[0] = left[0];
for (int i = 1; i < M; i++) {// counting prefix sums for future use
sleft[i]=sleft[i-1]+left[i];
}
int o, pairs, total_p = 0, total_used=0;
for (int i = 0; i < N; i++) { // counting pairs
ir = tr(i, A, M);
o = sleft[ir]; // nb of open till right edge
pairs = o -1 - total_used;
total_used++;
total_p += pairs;
}
if(total_p > 10000000){
total_p = -1;
}
return total_p;
}
private int tl(int i, int[] A){
int tl = i - A[i]; // index of "begin" of the circle
if (tl < 0) {
tl = 0;
} else {
tl = i - A[i] + 1;
}
return tl;
}
int tr(int i, int[] A, int M){
int tr; // index of "end" of the circle
if (Integer.MAX_VALUE - i < A[i] || i + A[i] >= M - 1) {
tr = M - 1;
} else {
tr = i + A[i] + 1;
}
return tr;
}
There are a lot of great answers here already, including the great explanation from the accepted answer. However, I wanted to point out a small observation about implementation details in the Python language.
Originally, I've came up with the solution shown below. I was expecting to get O(N*log(N)) time complexity as soon as we have a single for-loop with N iterations, and each iteration performs a binary search that takes at most log(N).
def solution(a):
import bisect
if len(a) <= 1:
return 0
cuts = [(c - r, c + r) for c, r in enumerate(a)]
cuts.sort(key=lambda pair: pair[0])
lefts, rights = zip(*cuts)
n = len(cuts)
total = 0
for i in range(n):
r = rights[i]
pos = bisect.bisect_right(lefts[i+1:], r)
total += pos
if total > 10e6:
return -1
return total
However, I've get O(N**2) and a timeout failure. Do you see what is wrong here? Right, this line:
pos = bisect.bisect_right(lefts[i+1:], r)
In this line, you are actually taking a copy of the original list to pass it into binary search function, and it totally ruins the efficiency of the proposed solution! It makes your code just a bit more consice (i.e., you don't need to write pos - i - 1) but heavily undermies the performance. So, as it was shown above, the solution should be:
def solution(a):
import bisect
if len(a) <= 1:
return 0
cuts = [(c - r, c + r) for c, r in enumerate(a)]
cuts.sort(key=lambda pair: pair[0])
lefts, rights = zip(*cuts)
n = len(cuts)
total = 0
for i in range(n):
r = rights[i]
pos = bisect.bisect_right(lefts, r)
total += (pos - i - 1)
if total > 10e6:
return -1
return total
It seems that sometimes one could be too eager about making slices and copies because Python allows you to do it so easily :) Probably not a great insight, but for me it was a good lesson to pay more attention to these "technical" moments when converting ideas and algorithms into real-word solutions.
I know that this is an old questions but it is still active on codility.
private int solution(int[] A)
{
int openedCircles = 0;
int intersectCount = 0;
We need circles with their start and end values. For that purpose I have used Tuple.
True/False indicates if we are adding Circle Starting or Circle Ending value.
List<Tuple<decimal, bool>> circles = new List<Tuple<decimal, bool>>();
for(int i = 0; i < A.Length; i ++)
{
// Circle start value
circles.Add(new Tuple<decimal, bool>((decimal)i - (decimal)A[i], true));
// Circle end value
circles.Add(new Tuple<decimal, bool>((decimal)i + (decimal)A[i], false));
}
Order "circles" by their values.
If one circle is ending at same value where other circle is starting, it should be counted as intersect (because of that "opening" should be in front of "closing" if in same point)
circles = circles.OrderBy(x => x.Item1).ThenByDescending(x => x.Item2).ToList();
Counting and returning counter
foreach (var circle in circles)
{
// We are opening new circle (within existing circles)
if(circle.Item2 == true)
{
intersectCount += openedCircles;
if (intersectCount > 10000000)
{
return -1;
}
openedCircles++;
}
else
{
// We are closing circle
openedCircles--;
}
}
return intersectCount;
}
Javascript solution 100/100 based on this video https://www.youtube.com/watch?v=HV8tzIiidSw
function sortArray(A) {
return A.sort((a, b) => a - b)
}
function getDiskPoints(A) {
const diskStarPoint = []
const diskEndPoint = []
for(i = 0; i < A.length; i++) {
diskStarPoint.push(i - A[i])
diskEndPoint.push(i + A[i])
}
return {
diskStarPoint: sortArray(diskStarPoint),
diskEndPoint: sortArray(diskEndPoint)
};
}
function solution(A) {
const { diskStarPoint, diskEndPoint } = getDiskPoints(A)
let index = 0;
let openDisks = 0;
let intersections = 0;
for(i = 0; i < diskStarPoint.length; i++) {
while(diskStarPoint[i] > diskEndPoint[index]) {
openDisks--
index++
}
intersections += openDisks
openDisks++
}
return intersections > 10000000 ? -1 : intersections
}
so, I was doing this test in Scala and I would like to share here my example. My idea to solve is:
Extract the limits to the left and right of each position on the array.
A[0] = 1 --> (0-1, 0+1) = A0(-1, 1)
A[1] = 5 --> (1-5, 1+5) = A1(-4, 6)
A[2] = 2 --> (2-2, 2+2) = A2(0, 4)
A[3] = 1 --> (3-1, 3+1) = A3(2, 4)
A[4] = 4 --> (4-4, 4+4) = A4(0, 8)
A[5] = 0 --> (5-0, 5+0) = A5(5, 5)
Check if there is intersections between any two positions
(A0_0 >= A1_0 AND A0_0 <= A1_1) OR // intersection
(A0_1 >= A1_0 AND A0_1 <= A1_1) OR // intersection
(A0_0 <= A1_0 AND A0_1 >= A1_1) // one circle contain inside the other
if any of these two checks is true count one intersection.
object NumberOfDiscIntersections {
def solution(a: Array[Int]): Int = {
var count: Long = 0
for (posI: Long <- 0L until a.size) {
for (posJ <- (posI + 1) until a.size) {
val tupleI = (posI - a(posI.toInt), posI + a(posI.toInt))
val tupleJ = (posJ - a(posJ.toInt), posJ + a(posJ.toInt))
if ((tupleI._1 >= tupleJ._1 && tupleI._1 <= tupleJ._2) ||
(tupleI._2 >= tupleJ._1 && tupleI._2 <= tupleJ._2) ||
(tupleI._1 <= tupleJ._1 && tupleI._2 >= tupleJ._2)) {
count += 1
}
}
}
count.toInt
}
}
This got 100/100 in c#
class CodilityDemo3
{
public static int GetIntersections(int[] A)
{
if (A == null)
{
return 0;
}
int size = A.Length;
if (size <= 1)
{
return 0;
}
List<Line> lines = new List<Line>();
for (int i = 0; i < size; i++)
{
if (A[i] >= 0)
{
lines.Add(new Line(i - A[i], i + A[i]));
}
}
lines.Sort(Line.CompareLines);
size = lines.Count;
int intersects = 0;
for (int i = 0; i < size; i++)
{
Line ln1 = lines[i];
for (int j = i + 1; j < size; j++)
{
Line ln2 = lines[j];
if (ln2.YStart <= ln1.YEnd)
{
intersects += 1;
if (intersects > 10000000)
{
return -1;
}
}
else
{
break;
}
}
}
return intersects;
}
}
public class Line
{
public Line(double ystart, double yend)
{
YStart = ystart;
YEnd = yend;
}
public double YStart { get; set; }
public double YEnd { get; set; }
public static int CompareLines(Line line1, Line line2)
{
return (line1.YStart.CompareTo(line2.YStart));
}
}
}
Thanks to Falk for the great idea! Here is a ruby implementation that takes advantage of sparseness.
def int(a)
event = Hash.new{|h,k| h[k] = {:start => 0, :stop => 0}}
a.each_index {|i|
event[i - a[i]][:start] += 1
event[i + a[i]][:stop ] += 1
}
sorted_events = (event.sort_by {|index, value| index}).map! {|n| n[1]}
past_start = 0
intersect = 0
sorted_events.each {|e|
intersect += e[:start] * (e[:start]-1) / 2 +
e[:start] * past_start
past_start += e[:start]
past_start -= e[:stop]
}
return intersect
end
puts int [1,1]
puts int [1,5,2,1,4,0]
#include <stdio.h>
#include <stdlib.h>
void sortPairs(int bounds[], int len){
int i,j, temp;
for(i=0;i<(len-1);i++){
for(j=i+1;j<len;j++){
if(bounds[i] > bounds[j]){
temp = bounds[i];
bounds[i] = bounds[j];
bounds[j] = temp;
temp = bounds[i+len];
bounds[i+len] = bounds[j+len];
bounds[j+len] = temp;
}
}
}
}
int adjacentPointPairsCount(int a[], int len){
int count=0,i,j;
int *bounds;
if(len<2) {
goto toend;
}
bounds = malloc(sizeof(int)*len *2);
for(i=0; i< len; i++){
bounds[i] = i-a[i];
bounds[i+len] = i+a[i];
}
sortPairs(bounds, len);
for(i=0;i<len;i++){
int currentBound = bounds[i+len];
for(j=i+1;a[j]<=currentBound;j++){
if(count>100000){
count=-1;
goto toend;
}
count++;
}
}
toend:
free(bounds);
return count;
}
An Implementation of Idea stated above in Java:
public class DiscIntersectionCount {
public int number_of_disc_intersections(int[] A) {
int[] leftPoints = new int[A.length];
for (int i = 0; i < A.length; i++) {
leftPoints[i] = i - A[i];
}
Arrays.sort(leftPoints);
// System.out.println(Arrays.toString(leftPoints));
int count = 0;
for (int i = 0; i < A.length - 1; i++) {
int rpoint = A[i] + i;
int rrank = getRank(leftPoints, rpoint);
//if disk has sifnificant radius, exclude own self
if (rpoint > i) rrank -= 1;
int rank = rrank;
// System.out.println(rpoint+" : "+rank);
rank -= i;
count += rank;
}
return count;
}
public int getRank(int A[], int num) {
if (A==null || A.length == 0) return -1;
int mid = A.length/2;
while ((mid >= 0) && (mid < A.length)) {
if (A[mid] == num) return mid;
if ((mid == 0) && (A[mid] > num)) return -1;
if ((mid == (A.length - 1)) && (A[mid] < num)) return A.length;
if (A[mid] < num && A[mid + 1] >= num) return mid + 1;
if (A[mid] > num && A[mid - 1] <= num) return mid - 1;
if (A[mid] < num) mid = (mid + A.length)/2;
else mid = (mid)/2;
}
return -1;
}
public static void main(String[] args) {
DiscIntersectionCount d = new DiscIntersectionCount();
int[] A =
//{1,5,2,1,4,0}
//{0,0,0,0,0,0}
// {1,1,2}
{3}
;
int count = d.number_of_disc_intersections(A);
System.out.println(count);
}
}
Here is the PHP code that scored 100 on codility:
$sum=0;
//One way of cloning the A:
$start = array();
$end = array();
foreach ($A as $key=>$value)
{
$start[]=0;
$end[]=0;
}
for ($i=0; $i<count($A); $i++)
{
if ($i<$A[$i])
$start[0]++;
else
$start[$i-$A[$i]]++;
if ($i+$A[$i] >= count($A))
$end[count($A)-1]++;
else
$end[$i+$A[$i]]++;
}
$active=0;
for ($i=0; $i<count($A);$i++)
{
$sum += $active*$start[$i]+($start[$i]*($start[$i]-1))/2;
if ($sum>10000000) return -1;
$active += $start[$i]-$end[$i];
}
return $sum;
However I dont understand the logic. This is just transformed C++ code from above. Folks, can you elaborate on what you were doing here, please?
A 100/100 C# implementation as described by Aryabhatta (the binary search solution).
using System;
class Solution {
public int solution(int[] A)
{
return IntersectingDiscs.Execute(A);
}
}
class IntersectingDiscs
{
public static int Execute(int[] data)
{
int counter = 0;
var intervals = Interval.GetIntervals(data);
Array.Sort(intervals); // sort by Left value
for (int i = 0; i < intervals.Length; i++)
{
counter += GetCoverage(intervals, i);
if(counter > 10000000)
{
return -1;
}
}
return counter;
}
private static int GetCoverage(Interval[] intervals, int i)
{
var currentInterval = intervals[i];
// search for an interval starting at currentInterval.Right
int j = Array.BinarySearch(intervals, new Interval { Left = currentInterval.Right });
if(j < 0)
{
// item not found
j = ~j; // bitwise complement (see Array.BinarySearch documentation)
// now j == index of the next item larger than the searched one
j = j - 1; // set index to the previous element
}
while(j + 1 < intervals.Length && intervals[j].Left == intervals[j + 1].Left)
{
j++; // get the rightmost interval starting from currentInterval.Righ
}
return j - i; // reduce already processed intervals (the left side from currentInterval)
}
}
class Interval : IComparable
{
public long Left { get; set; }
public long Right { get; set; }
// Implementation of IComparable interface
// which is used by Array.Sort().
public int CompareTo(object obj)
{
// elements will be sorted by Left value
var another = obj as Interval;
if (this.Left < another.Left)
{
return -1;
}
if (this.Left > another.Left)
{
return 1;
}
return 0;
}
/// <summary>
/// Transform array items into Intervals (eg. {1, 2, 4} -> {[-1,1], [-1,3], [-2,6]}).
/// </summary>
public static Interval[] GetIntervals(int[] data)
{
var intervals = new Interval[data.Length];
for (int i = 0; i < data.Length; i++)
{
// use long to avoid data overflow (eg. int.MaxValue + 1)
long radius = data[i];
intervals[i] = new Interval
{
Left = i - radius,
Right = i + radius
};
}
return intervals;
}
}
100% score in Codility.
Here is an adaptation to C# of Толя solution:
public int solution(int[] A)
{
long result = 0;
Dictionary<long, int> dps = new Dictionary<long, int>();
Dictionary<long, int> dpe = new Dictionary<long, int>();
for (int i = 0; i < A.Length; i++)
{
Inc(dps, Math.Max(0, i - A[i]));
Inc(dpe, Math.Min(A.Length - 1, i + A[i]));
}
long t = 0;
for (int i = 0; i < A.Length; i++)
{
int value;
if (dps.TryGetValue(i, out value))
{
result += t * value;
result += value * (value - 1) / 2;
t += value;
if (result > 10000000)
return -1;
}
dpe.TryGetValue(i, out value);
t -= value;
}
return (int)result;
}
private static void Inc(Dictionary<long, int> values, long index)
{
int value;
values.TryGetValue(index, out value);
values[index] = ++value;
}
Here's a two-pass C++ solution that doesn't require any libraries, binary searching, sorting, etc.
int solution(vector<int> &A) {
#define countmax 10000000
int count = 0;
// init lower edge array
vector<int> E(A.size());
for (int i = 0; i < (int) E.size(); i++)
E[i] = 0;
// first pass
// count all lower numbered discs inside this one
// mark lower edge of each disc
for (int i = 0; i < (int) A.size(); i++)
{
// if disc overlaps zero
if (i - A[i] <= 0)
count += i;
// doesn't overlap zero
else {
count += A[i];
E[i - A[i]]++;
}
if (count > countmax)
return -1;
}
// second pass
// count higher numbered discs with edge inside this one
for (int i = 0; i < (int) A.size(); i++)
{
// loop up inside this disc until top of vector
int jend = ((int) E.size() < (long long) i + A[i] + 1 ?
(int) E.size() : i + A[i] + 1);
// count all discs with edge inside this disc
// note: if higher disc is so big that edge is at or below
// this disc center, would count intersection in first pass
for (int j = i + 1; j < jend; j++)
count += E[j];
if (count > countmax)
return -1;
}
return count;
}
My answer in Swift; gets a 100% score.
import Glibc
struct Interval {
let start: Int
let end: Int
}
func bisectRight(intervals: [Interval], end: Int) -> Int {
var pos = -1
var startpos = 0
var endpos = intervals.count - 1
if intervals.count == 1 {
if intervals[0].start < end {
return 1
} else {
return 0
}
}
while true {
let currentLength = endpos - startpos
if currentLength == 1 {
pos = startpos
pos += 1
if intervals[pos].start <= end {
pos += 1
}
break
} else {
let middle = Int(ceil( Double((endpos - startpos)) / 2.0 ))
let middlepos = startpos + middle
if intervals[middlepos].start <= end {
startpos = middlepos
} else {
endpos = middlepos
}
}
}
return pos
}
public func solution(inout A: [Int]) -> Int {
let N = A.count
var nIntersections = 0
// Create array of intervals
var unsortedIntervals: [Interval] = []
for i in 0 ..< N {
let interval = Interval(start: i-A[i], end: i+A[i])
unsortedIntervals.append(interval)
}
// Sort array
let intervals = unsortedIntervals.sort {
$0.start < $1.start
}
for i in 0 ..< intervals.count {
let end = intervals[i].end
var count = bisectRight(intervals, end: end)
count -= (i + 1)
nIntersections += count
if nIntersections > Int(10E6) {
return -1
}
}
return nIntersections
}
C# solution 100/100
using System.Linq;
class Solution
{
private struct Interval
{
public Interval(long #from, long to)
{
From = #from;
To = to;
}
public long From { get; }
public long To { get; }
}
public int solution(int[] A)
{
int result = 0;
Interval[] intervals = A.Select((value, i) =>
{
long iL = i;
return new Interval(iL - value, iL + value);
})
.OrderBy(x => x.From)
.ToArray();
for (int i = 0; i < intervals.Length; i++)
{
for (int j = i + 1; j < intervals.Length && intervals[j].From <= intervals[i].To; j++)
result++;
if (result > 10000000)
return -1;
}
return result;
}
}