Improving a solution - c++

Question

The description of a task goes like this:
We have n numbers, and we have to find quantity of unique sums of all the pairs in the array.
For example:
3 2 5 6 3
The sums of all the pairs(non-repeated) are 5 9 8 6 8 7 5 11 9 8
Unique are 5 9 8 6 7 11
Therefore output is 6
I have come up with this really primitive, and time-consuming (meaning complexity) solution:
int n = 0;
cin >> n;
vector<int> vec(n);
for (int i = 0; i < n; i++)
{
cin >> vec[i];
}
vector<int> sum;
for (int i = 0; i < n; i++)
{
for (int j = i+1; j < n; j++)
{
sum.push_back(vec[i] + vec[j]);
}
}
sort(sum.begin(), sum.end());
for (int i = 0; i < sum.size()-1;)
{
if (sum[i] == sum[i + 1]) sum.erase(sum.begin() + i);
else i++;
}
cout << endl << sum.size();
I feel like there could be a solution using Combinatorics or something easier. I have thought a lot and couldn't think of anything. So my request is if anyone can improve the solution.

Answer 1

As mentioned above what you need it is difficult to do this without computing the sum of all pairs, so I am not going to handle that, I am just going to advise about efficient data structures.
Analysis of your solution
Your code adds everything in advance O(n^2) then sorts O(n^2 log(n)), then remove duplicates. But since you are erasing from a vector, that ultimately has complexity linear with the number of elements to the end of the list. It means that the second loop will make the complexity of your algorithm O(n^4).
You can count the unique elements in a sorted array without removing
int count = 0;
for (int i = 0; i < sum.size()-1; ++i)
{
if (sum[i] != sum[i + 1]) ++count
}
This change alone makes your algorithm complexity O(n^2 log n).
Alternatives without sorting.
Here are alternatives that O(n^2) and storage depending on the range of the input values instead of the length of the vector (except for the last).
I am testing with 1000 elements smaller between 0 and 10000
vector<int> vec;
for(int i = 0; i < 1000; ++i){
vec.push_back(rand() % 10000);
}
Your implementation sum_pairs1(vec) (18 seconds)
int sum_pairs1(const vector<int> &vec){
vector<int> sum;
int n = vec.size();
for (int i = 0; i < n; i++)
{
for (int j = i+1; j < n; j++)
{
sum.push_back(vec[i] + vec[j]);
}
}
sort(sum.begin(), sum.end());
for (int i = 0; i < sum.size()-1;)
{
if (sum[i] == sum[i + 1]) sum.erase(sum.begin() + i);
else i++;
}
return sum.size();
}
If you know the range for the sum of the values you can use a bitset, efficient use of memory sum_pairs2<20000>(vec) (0.016 second).
template<size_t N>
int sum_pairs2(const vector<int> &vec){
bitset<N> seen;
int n = vec.size();
for (int i = 0; i < n; i++)
{
for (int j = i+1; j < n; j++)
{
seen[vec[i] + vec[j]] = true;
}
}
return seen.count();
}
If you know that the maximum sum is not so high (the vector is not very sparse), but you don't know at compilation time you can use a vector, you can keep track of minimum and maximum to allocate the minimum possible and also supporting negative values.
int sum_pairs2b(const vector<int> &vec){
int VMAX = vec[0];
int VMIN = vec[0]
for(auto v : vec){
if(VMAX < v) VMAX = v;
else if(VMIN > v) VMIN = v;
}
vector<bool> seen(2*(VMAX - VMIN) + 1);
int n = vec.size();
for (int i = 0; i < n; i++)
{
for (int j = i+1; j < n; j++)
{
seen[vec[i] + vec[j] - 2*VMIN] = true;
}
}
int count = 0;
for(auto c : seen){
if(c) ++count;
}
return count;
}
And If you want a more general solution that works well with sparse data sum_pairs3<int>(vec) (0.097 second)
template<typename T>
int sum_pairs3(const vector<T> &vec){
unordered_set<T> seen;
int n = vec.size();
for (int i = 0; i < n; i++)
{
for (int j = i+1; j < n; j++)
{
seen.insert(vec[i] + vec[j]);
}
}
return seen.size();
}