int f(const std::vector<int>& v) {
int result = 0;
for (int i = 0; i < v.size(); ++i) { O(N)
for (int j = v.size(); j >= 0; j -= 2) { O(N/2)
result += v.at(i) * j;
}
}
return result;
}
The inner for loop is O(N/2), however I am wondering why this is because
For example, if v.size() is 10, then
10 >= 0 ✓
8 >= 0 ✓
6 >= 0 ✓
4 >= 0 ✓
2>= 0 ✓
0 >= 0 ✓
-2 Fails
The inner for loop could be executed 6 times with an input size of 10
What am I missing?
EDIT* I understand that only highest magnitude is taken into consideration. This question was more about coming up with the original O(N/2 + 1)
Complexity gives you a way to assess the magnitude of time it would take an input of certain size to complete, not the accurate time it would perform with.
Therefore, when dealing with complexity, you should only consider the highest magnitude, without constant multipliers:
O(N/2 + 1) = O(N/2) = O(N)
In a comment, you said:
I understand this, but I am just curious as to how O(N/2) is obtained
Take a look at the following table:
Size of vector Number of time the inner loop is executed:
0 1
1 1
2 2
3 2
...
100 51
101 51
...
2x x + 1
2x + 1 x + 1
If you take the constant 1 out of that equation, the inner loop is O(N/2).
Related
Here is the link to the question. Essentially, it asks to find the kth number having digit sum as 10. I have tried multiple solutions and also looked upon solutions online. Specifically this one (also shared below). The one with constant time talks about outliers in Arithmetic Progression and uses it to find the nth number having sum as 10. Obviously, the code is incorrect as it fails for test cases when k=1000 etc.
#include <bits/stdc++.h>
using namespace std;
int findNth(int n)
{
int nthElement = 19 + (n - 1) * 9;
int outliersCount = (int)log10(nthElement) - 1;
// find the nth perfect number
nthElement += 9 * outliersCount;
return nthElement;
}
int main()
{
cout << findNth(5) << endl;
return 0;
}
Eventually, I ended up writing combination of Arithmetic Progression + brute force as below
#include <bits/stdc++.h>
using namespace std;
#define ll unsigned long long
int main() {
int n;
cin >> n;
int count = 0;
ll i = 19;
for (; ; i += 9) {
int curr = i;
int localSum = 0;
while (curr) {
localSum += curr%10;
curr /= 10;
}
if (localSum == 10) {
count += 1;
}
if (count == n) {
break;
}
}
cout << i << endl;
return 0;
}
I am wondering, if there is no constant time or better algorithm that does not require me to calculate the sum, but my algorithm always hops in a way that I have number whose digit sum is 10?
Here is a Python solution that you can translate into C++.
cached_count_ds_l = {}
def count_digit_sum_length (s, l):
k = (s, l)
if k not in cached_count_ds_l:
if l < 2:
if s == 0:
return 1
elif l == 1 and s < 10:
return 1
else:
return 0
else:
ans = 0
for i in range(min(10, s+1)):
ans += count_digit_sum_length(s-i, l-1)
cached_count_ds_l[k] = ans
return cached_count_ds_l[k]
def nth_of_sum (s, n):
l = 0
while count_digit_sum_length(s, l) < n:
l += 1
digits = []
while 0 < l:
for i in range(10):
if count_digit_sum_length(s-i, l-1) < n:
n -= count_digit_sum_length(s-i, l-1)
else:
digits.append(str(i))
s -= i
l -= 1
break
return int("".join(digits))
print(nth_of_sum(10, 1000))
The idea is to use dynamic programming to find how many numbers there are of a given maximum length with a given digit sum. And then to use that to cross off whole blocks of numbers on the way to finding the right one.
The main logic goes like this:
0 numbers of length 0 sum to 10
- need longer
0 numbers of length 1 sum to 10
- need longer
9 numbers of length 2 sum to 10
- need longer
63 numbers of length 3 sum to 10
- need longer
282 numbers of length 4 sum to 10
- need longer
996 numbers of length 5 sum to 10
- need longer
2997 numbers of length 6 sum to 10
- answer has length 6
Looking for 1000th number of length 6 that sums to 10
- 996 with a leading 0 sum to 10
- Need the 4th past 99999
- 715 with a leading 1 sum to 10
- Have a leading 1
Looking for 4th number of length 5 that sums to 9
- 495 with a leading 0 sum to 9
- Have a leading 10
Looking for 4th number of length 4 that sums to 9
- 220 with a leading 0 sum to 9
- Have a leading 100
Looking for 4th number of length 3 that sums to 9
- 55 with a leading 0 sum to 9
- Have a leading 1000
Looking for 4th number of length 2 that sums to 9
- 1 with a leading 0 sum to 9
- Need the 3rd past 9
- 1 with a leading 1 sum to 9
- Need the 2nd past 19
- 1 with a leading 2 sum to 9
- Need the 1st past 29
- 1 with a leading 3 sum to 9
- Have a leading 10003
Looking for 1st number of length 1 that sums to 6
- 0 with a leading 0 sum to 6
- Need the 1st past 0
- 0 with a leading 1 sum to 6
- Need the 1st past 1
- 0 with a leading 2 sum to 6
- Need the 1st past 2
- 0 with a leading 3 sum to 6
- Need the 1st past 3
- 0 with a leading 4 sum to 6
- Need the 1st past 4
- 0 with a leading 5 sum to 6
- Need the 1st past 5
- 1 with a leading 6 sum to 6
- Have a leading 100036
And it finishes in a fraction of a second.
Incidentally the million'th is 20111220000010, the billionth is 10111000000002000000010000002100, and the trillionth is 10000000100000100000100000000000001000000000000100000000010110001000.
I have a word corpus of say 3000 words such as [hello, who, this ..].
I want to find the nth 3 word combination from this corpus.I am fine with any order as long as the algorithm gives consistent output.
What would be the time complexity of the algorithm.
I have seen this answer but was looking for something simple.
(Note that I will be using 1-based indexes and ranks throughout this answer.)
To generate all combinations of 3 elements from a list of n elements, we'd take all elements from 1 to n-2 as the first element, then for each of these we'd take all elements after the first element up to n-1 as the second element, then for each of these we'd take all elements after the second element up to n as the third element. This gives us a fixed order, and a direct relation between the rank and a specific combination.
If we take element i as the first element, there are (n-i choose 2) possibilities for the second and third element, and thus (n-i choose 2) combinations with i as the first element. If we then take element j as the second element, there are (n-j choose 1) = n-j possibilities for the third element, and thus n-j combinations with i and j as the first two elements.
Linear search in tables of binomial coefficients
With tables of these binomial coefficients, we can quickly find a specific combination, given its rank. Let's look at a simplified example with a list of 10 elements; these are the number of combinations with element i as the first element:
i
1 C(9,2) = 36
2 C(8,2) = 28
3 C(7,2) = 21
4 C(6,2) = 15
5 C(5,2) = 10
6 C(4,2) = 6
7 C(3,2) = 3
8 C(2,2) = 1
---
120 = C(10,3)
And these are the number of combinations with element j as the second element:
j
2 C(8,1) = 8
3 C(7,1) = 7
4 C(6,1) = 6
5 C(5,1) = 5
6 C(4,1) = 4
7 C(3,1) = 3
8 C(2,1) = 2
9 C(1,1) = 1
So if we're looking for the combination with e.g. rank 96, we look at the number of combinations for each choice of first element i, until we find which group of combinations the combination ranked 96 is in:
i
1 36 96 > 36 96 - 36 = 60
2 28 60 > 28 60 - 28 = 32
3 21 32 > 21 32 - 21 = 11
4 15 11 <= 15
So we know that the first element i is 4, and that within the 15 combinations with i=4, we're looking for the eleventh combination. Now we look at the number of combinations for each choice of second element j, starting after 4:
j
5 5 11 > 5 11 - 5 = 6
6 4 6 > 4 6 - 4 = 2
7 3 2 <= 3
So we know that the second element j is 7, and that the third element is the second combination with j=7, which is k=9. So the combination with rank 96 contains the elements 4, 7 and 9.
Binary search in tables of running total of binomial coefficients
Instead of creating a table of the binomial coefficients and then performing a linear search, it is of course more efficient to create a table of the running total of the binomial coefficient, and then perform a binary search on it. This will improve the time complexity from O(N) to O(logN); in the case of N=3000, the two look-ups can be done in log2(3000) = 12 steps.
So we'd store:
i
1 36
2 64
3 85
4 100
5 110
6 116
7 119
8 120
and:
j
2 8
3 15
4 21
5 26
6 30
7 33
8 35
9 36
Note that when finding j in the second table, you have to subtract the sum corresponding with i from the sums. Let's walk through the example of rank 96 and combination [4,7,9] again; we find the first value that is greater than or equal to the rank:
3 85 96 > 85
4 100 96 <= 100
So we know that i=4; we then subtract the previous sum next to i-1, to get:
96 - 85 = 11
Now we look at the table for j, but we start after j=4, and subtract the sum corresponding to 4, which is 21, from the sums. then again, we find the first value that is greater than or equal to the rank we're looking for (which is now 11):
6 30 - 21 = 9 11 > 9
7 33 - 21 = 12 11 <= 12
So we know that j=7; we subtract the previous sum corresponding to j-1, to get:
11 - 9 = 2
So we know that the second element j is 7, and that the third element is the second combination with j=7, which is k=9. So the combination with rank 96 contains the elements 4, 7 and 9.
Hard-coding the look-up tables
It is of course unnecessary to generate these look-up tables again every time we want to perform a look-up. We only need to generate them once, and then hard-code them into the rank-to-combination algorithm; this should take only 2998 * 64-bit + 2998 * 32-bit = 35kB of space, and make the algorithm incredibly fast.
Inverse algorithm
The inverse algorithm, to find the rank given a combination of elements [i,j,k] then means:
Finding the index of the elements in the list; if the list is sorted (e.g. words sorted alphabetically) this can be done with a binary search in O(logN).
Find the sum in the table for i that corresponds with i-1.
Add to that the sum in the table for j that corresponds with j-1, minus the sum that corresponds with i.
Add to that k-j.
Let's look again at the same example with the combination of elements [4,7,9]:
i=4 -> table_i[3] = 85
j=7 -> table_j[6] - table_j[4] = 30 - 21 = 9
k=9 -> k-j = 2
rank = 85 + 9 + 2 = 96
Look-up tables for N=3000
This snippet generates the look-up table with the running total of the binomial coefficients for i = 1 to 2998:
function C(n, k) { // binomial coefficient (Pascal's triangle)
if (k < 0 || k > n) return 0;
if (k > n - k) k = n - k;
if (! C.t) C.t = [[1]];
while (C.t.length <= n) {
C.t.push([1]);
var l = C.t.length - 1;
for (var i = 1; i < l / 2; i++)
C.t[l].push(C.t[l - 1][i - 1] + C.t[l - 1][i]);
if (l % 2 == 0)
C.t[l].push(2 * C.t[l - 1][(l - 2) / 2]);
}
return C.t[n][k];
}
for (var total = 0, x = 2999; x > 1; x--) {
total += C(x, 2);
document.write(total + ", ");
}
This snippet generates the look-up table with the running total of the binomial coefficients for j = 2 to 2999:
for (var total = 0, x = 2998; x > 0; x--) {
total += x;
document.write(total + ", ");
}
Code example
Here's a quick code example, unfortunately without the full hardcoded look-up tables, because of the size restriction on answers on SO. Run the snippets above and paste the results into the arrays iTable and jTable (after the leading zeros) to get the faster version with hard-coded look-up tables.
function combinationToRank(i, j, k) {
return iTable[i - 1] + jTable[j - 1] - jTable[i] + k - j;
}
function rankToCombination(rank) {
var i = binarySearch(iTable, rank, 1);
rank -= iTable[i - 1];
rank += jTable[i];
var j = binarySearch(jTable, rank, i + 1);
rank -= jTable[j - 1];
var k = j + rank;
return [i, j, k];
function binarySearch(array, value, first) {
var last = array.length - 1;
while (first < last - 1) {
var middle = Math.floor((last + first) / 2);
if (value > array[middle]) first = middle;
else last = middle;
}
return (value <= array[first]) ? first : last;
}
}
var iTable = [0]; // append look-up table values here
var jTable = [0, 0]; // and here
// remove this part when using hard-coded look-up tables
function C(n,k){if(k<0||k>n)return 0;if(k>n-k)k=n-k;if(!C.t)C.t=[[1]];while(C.t.length<=n){C.t.push([1]);var l=C.t.length-1;for(var i=1;i<l/2;i++)C.t[l].push(C.t[l-1][i-1]+C.t[l-1][i]);if(l%2==0)C.t[l].push(2*C.t[l-1][(l-2)/2])}return C.t[n][k]}
for (var iTotal = 0, jTotal = 0, x = 2999; x > 1; x--) {
iTable.push(iTotal += C(x, 2));
jTable.push(jTotal += x - 1);
}
document.write(combinationToRank(500, 1500, 2500) + "<br>");
document.write(rankToCombination(1893333750) + "<br>");
I have come across a problem where we want to tell the maximum size of the longest increasing sub-sequence.
an array A consisting of N integers.
M queries (Li, Ri)
for each query we wants to find the length of the longest increasing subsequence in
array A[Li], A[Li + 1], ..., A[Ri].
I implemented finding the sub-sequence using dp approach
// mind the REPN, LLD, these are macros I use for programming
// LLD = long long int
// REPN(i, a, b) = for (int i = a; i < b; ++i)
LLD a[n], dp[n];
REPN(i, 0, n)
{
scanf("%lld", &a[i]);
dp[i] = 1;
}
REPN(i, 1, n)
{
REPN(j, 0, i)
{
if(a[i] > a[j])
dp[i] = std::max(dp[j] + 1, dp[i]);
}
}
For example:
Array: 1 3 8 9 7 2 4 5 10 6
dplis: 1 2 3 4 3 1 3 4 5 5
max: 5
But if it was for range Li=2 & Ri=9
Then:
Array: 3 8 9 7 2 4 5 10
dplis: 1 2 3 2 1 2 3 4
max: 4
How can i determine the maximum longest increasing sub-sequence in a sub array?
PS: I don't want to recompute the whole dplis array, I want to use the original one because too much computation will kill the purpose of the question.
One of the approaches was to construct a complete 2D DP array that consists of sub-sequence from position i where range of i is from 0 to n, but it fails on many cases due to TLE(Time limit exceeded)
REPN(k,0,n) {
REPN(i,k+1,n) {
REPN(j,k,i) {
if(a[i]>a[j]) dp[k][i]=std::max(dp[k][j]+1, dp[k][i]);
}
}
}
REPN(i,0,q) {
read(l); read(r);
LLD max=-1;
REPN(i,0,r) {
if(max<dp[l-1][i]) max=dp[l-1][i];
}
printf("%lld\n", max);
}
If you have any new logic/implementation, I will gladly study it in-depth. Cheers.
Full disclosure - I was inspired by Is x += a quicker than x = x + a?
That aside, I decided to test += vs -=. Simple tests reveal they're about the same. Then I tried something similar to:
std::vector<int> x;
for (int i = 0 ; i < 10000 ; i++)
x.push_back(rand()%10);
and call += and -= proportionally to a given number:
long long sum = 0;
for ( each number in the array )
if ( x[j] < k )
sum += x[j];
else
sum -= x[j];
so, if k is, say, small, -= would get called more often (duuuh). I tried with k = 2 which would give a higher proportion of -= called, and with k = 5, which should yield about the same number of -= and +=.
The punchline: calling -= is about twice as faster than calling +=. Why would it be more efficient in this case?
I'm gonna jump in before Mysticial gets a hold of this and guess: branch prediction.
So, it's not the -= vs +=.
The condition x[j] < k can be better predicted when it's almost always true or false than it can be when there's about the same number of numbers for which it can evaluate to either.
For k = 2, one in 10 will evaluate to false.
For k = 5, they'll be about the same and distributed randomly, so harder to predict.
EDIT: See http://ideone.com/1PYMl - all extra stuff is there to prevent unused code optimization (the couts).
tl;dr: Results for varying k:
k: 1 Time: 280
k: 2 Time: 360
k: 3 Time: 440
k: 4 Time: 520
k: 5 Time: 550
k: 6 Time: 510
k: 7 Time: 450
k: 8 Time: 360
k: 9 Time: 260
as you can see, the closer k gets to a chaotically varying condition, the program takes more. Towards the ends, it takes about half the time.
How is this code working for multiple of 5
bool isMultipleof5(int n)
{
/* If n is a multiple of 5 then we make sure that last
digit of n is 0 */
if ( (n&1) == 1 )
n <<= 1;
float x = n;
x = ( (int)(x*0.1) )*10;
/* If last digit of n is 0 then n will be equal to (int)x */
if ( (int)x == n )
return true;
return false;
}
It first makes n divisable by 2.
Next, it checks if it is divisable by 10 by multiplying with 0.1 and again with 10. The idea that if it is divisable by 10, you will get back to the original, and only then.
So, if the modifies n is divisable by 10 - it is certainly divisable by 5 as well, and since modified n is always divisable by 2, if it is divisable by 5 it will be divisable by 10, and the algorithm works.
NOTE: This is very unsuggested and especially might break with large values due to floating point precision issues. using the % operator should be prefered: return (n % 5) == 0
This is how the code works with some examples.
if ( (n&1) == 1 ) //Checks if the number is odd
n <<= 1; //Multiplies the number by 2 if odd
x = ( (int)(x * 0.1) //Divides the number 10 then truncates any decimal places
* 10 ) //Multiplies it back by 10
if ( (int)x == n ) //If the floating point value equals the (semi) original value its divisible by 5
return true;
return false; //Other wise false
Example:
15 & 1 == 1 //15 is odd
15 <<= 1; //n is now 30
30 / 10 = 3;
3 * 10 = 30; //x is now 30
30 == 30 //15 is a multiple of 5
17 & 1 == 1 //17 is odd
17 <<= 1; //n is now 34
34 / 10 = 3.4;
((int)3.4 = 3) * 10 = 30; //x is now 30
30 != 34 //17 is not a multiple of 5.
As others said though just simply use the mod operator %.
This is how it works:
Double the number. Now anything ending in 5 will be divisible 10 (and also divisible by 5). n <<= 1; (the check for oddness is unnecessary (n&1) == 1)
Divide it by 10, and cast away the fractional part. (int)(x*0.1)
Multiply it by 10, so now we have the same number as in step 1 only if the number in step 1 was already divisible by 10.
The use of floating point to divide by 10 makes this algorithm dangerous and probably incorrect for large values.
Try this
bool isMultipleof5(int n)
{
return (n%5) == 0;
}
A simpler way would be
bool isMultipleof5(int n)
{
return 0 == ( n % 5 ) ;
}
#define IS_MULTIPLE_OF_5(n) (((n)%5) ? 0 : 1)
I'd agree that (n % 5) == 0 would be an ideal solution, but that wasn't really the question.
This code works because it first checks if the input is odd. If it is, it multiplies by two. Since all odd multiples of 5 end with a 5, multiplying by 2 gives a number that ends with 0.
Then it checks if the last digit is 0. This can only happen if it started as a 0 (i.e. was even, we didn't change it) or if it was odd and ended in a 5 (we multiplied by 2). So, if it ends in 0 then the input must have been divisible by 5.
I'd add that this is also an awkward way to check the value of the last digit. I'd suggest n % 10 == 0 instead, but like others mentioned... you could have just used n % 5 == 0 in the first place ;).