#include<bits/stdc++.h>
#define god dimasi5eks
#pragma GCC optimize("O3")
#define fi first
#define se second
#define pb push_back
#define pf push_front
#define mod 1000000007
#define dancila 3.14159265359
#define eps 1e-9
// #define fisier 1
using namespace std;
typedef long long ll;
int add(int a, int b)
{
int x = a+b;
if(x >= mod)
x -= mod;
if(x < 0)
x += mod;
return x;
}
ll mul(ll a, ll b)
{
return (a*b) % mod;
}
ll pw(ll a, ll b)
{
ll ans = 1;
while(b)
{
if(b & 1)
ans = (ans * a) % mod;
a = (a * a) % mod;
b >>= 1;
}
return ans;
}
mt19937 rng(chrono::steady_clock::now().time_since_epoch().count());
long long rand_seed()
{
long long a = rng();
return a;
}
ll dp[101][101][1001][3];
/*
dp[i][j][k][l] :
i - number of numbers placed
j - number of connected components
k - total sum currently (filling empty spaces with a_{i} (0-indexed)
l - number of endpoints that are filled
*/
ll v[101];
int main()
{
ios_base::sync_with_stdio(false);
cin.tie(NULL);
int n, l;
cin >> n >> l;
for(int i = 0; i < n; i++)
cin >> v[i];
sort(v, v + n);
if(n == 1)
{
cout << 1;
return 0;
}
v[n] = (1<<20);
if(v[1] - v[0] <= l)
dp[1][1][v[1] - v[0]][1] = 2; //fill a[0] at one of the endpoints, there are 2 endpoints to fill.
if(2 * (v[1] - v[0]) <= l)
dp[1][1][2 * (v[1] - v[0])][0] = 1; //fill a[0] in the middle, positions doesn't matter.
for(int i = 1; i < n; i++)
for(int j = 1; j <= i; j++)
for(int k = 0; k <= l; k++)
for(int z = 0; z < 3; z++)
{
if(!dp[i][j][k][z])
continue;
int diff = v[i + 1] - v[i];
//First, we try to fill one of the ends
if(z < 2 && k + diff * (2 * j - z - 1) <= l) //there are 2*j - z - 1 positions that we're supposed to "upgrade" (-1 because one of the positions is merged with the endpoints after this move)
{
if(i == n - 1)
dp[i + 1][j][k + diff * (2 * j - z - 1)][z + 1] = add(dp[i + 1][j][k + diff * (2 * j - z - 1)][z + 1], mul(dp[i][j][k][z], (2-z) * j)); //we have j con. comp. to choose to merge with
else
if(z == 0 || j > 1) //otherwise this coincides with i == n - 1
dp[i + 1][j][k + diff * (2 * j - z - 1)][z + 1] = add(dp[i + 1][j][k + diff * (2 * j - z - 1)][z + 1], mul(dp[i][j][k][z], (2-z)*(j-z))); //can only merge with the con comp. that are not connected to ends.
if(k + diff * (2 * j - z + 1) <= l) //now we create a new cc.
dp[i + 1][j + 1][k + diff * (2 * j - z + 1)][z + 1] = add(dp[i + 1][j + 1][k + diff * (2 * j - z + 1)][z + 1], mul(dp[i][j][k][z], (2-z))); //we can choose one of the ends to create
}
//Next, we dont fill the ends.
//Part 1 : Create new cc
if(k + diff*(2*j - z + 2) <= l) //2 new positions to "upgrade"
{
dp[i + 1][j + 1][k + diff*(2*j - z + 2)][z] = add(dp[i + 1][j + 1][k + diff*(2*j - z + 2)][z], dp[i][j][k][z]); //nothing new happens
}
//Part 2 : Stick to one cc
if(k + diff*(2*j - z) <= l) //no new positions to "upgrade"
{
dp[i + 1][j][k + diff*(2*j - z)][z] = add(dp[i + 1][j][k + diff*(2*j - z)][z], mul(dp[i][j][k][z], (2*j - z))); //we can merge in 2*j - z possible positions
}
//Part 3 : Merge two ccs together
if((k + diff*(2*j - z - 2) <= l) && (j >= 2) && (i == n - 1 || j > 2 || z < 2))
{
if(z == 0)
{
dp[i + 1][j - 1][k + diff*(2*j - z - 2)][z] = add(dp[i + 1][j - 1][k + diff*(2*j - z - 2)][z], mul(dp[i][j][k][z], j*(j-1))); //there are jP2 possible merges
}
if(z == 1)
{
dp[i + 1][j - 1][k + diff*(2*j - z - 2)][z] = add(dp[i + 1][j - 1][k + diff*(2*j - z - 2)][z], mul(dp[i][j][k][z], (j-1)*(j-1))); //there are (j-1)P2+(j-1) merges
}
if(z == 2)
{
if(i == n - 1)
{
dp[i + 1][j - 1][k + diff*(2*j - z - 2)][z] = add(dp[i + 1][j - 1][k + diff*(2*j - z - 2)][z], dp[i][j][k][z]); //there's only 1 place it can go.
}
else
{
dp[i + 1][j - 1][k + diff*(2*j - z - 2)][z] = add(dp[i + 1][j - 1][k + diff*(2*j - z - 2)][z], mul(dp[i][j][k][z], (j-2)*(j-1))); //there're (j-2)P2 + 2(j-2) possiblilities
}
}
}
}
ll answer = 0;
for(int i = 0; i <= l; i++)
answer = add(answer, dp[n][1][i][2]);
cout << answer << '\n';
return 0;
}
# |
Verdict |
Execution time |
Memory |
Grader output |
1 |
Correct |
5 ms |
376 KB |
Output is correct |
2 |
Correct |
5 ms |
504 KB |
Output is correct |
3 |
Correct |
5 ms |
380 KB |
Output is correct |
4 |
Correct |
5 ms |
376 KB |
Output is correct |
5 |
Correct |
5 ms |
504 KB |
Output is correct |
6 |
Correct |
5 ms |
504 KB |
Output is correct |
7 |
Correct |
5 ms |
504 KB |
Output is correct |
8 |
Correct |
5 ms |
504 KB |
Output is correct |
9 |
Correct |
5 ms |
636 KB |
Output is correct |
10 |
Correct |
5 ms |
504 KB |
Output is correct |
# |
Verdict |
Execution time |
Memory |
Grader output |
1 |
Correct |
5 ms |
632 KB |
Output is correct |
2 |
Correct |
5 ms |
632 KB |
Output is correct |
3 |
Correct |
5 ms |
632 KB |
Output is correct |
4 |
Correct |
5 ms |
632 KB |
Output is correct |
5 |
Correct |
5 ms |
504 KB |
Output is correct |
6 |
Correct |
5 ms |
632 KB |
Output is correct |
7 |
Correct |
5 ms |
504 KB |
Output is correct |
8 |
Correct |
5 ms |
632 KB |
Output is correct |
9 |
Correct |
5 ms |
760 KB |
Output is correct |
10 |
Correct |
5 ms |
632 KB |
Output is correct |
# |
Verdict |
Execution time |
Memory |
Grader output |
1 |
Correct |
5 ms |
376 KB |
Output is correct |
2 |
Correct |
5 ms |
504 KB |
Output is correct |
3 |
Correct |
5 ms |
380 KB |
Output is correct |
4 |
Correct |
5 ms |
376 KB |
Output is correct |
5 |
Correct |
5 ms |
504 KB |
Output is correct |
6 |
Correct |
5 ms |
504 KB |
Output is correct |
7 |
Correct |
5 ms |
504 KB |
Output is correct |
8 |
Correct |
5 ms |
504 KB |
Output is correct |
9 |
Correct |
5 ms |
636 KB |
Output is correct |
10 |
Correct |
5 ms |
504 KB |
Output is correct |
11 |
Correct |
5 ms |
632 KB |
Output is correct |
12 |
Correct |
5 ms |
632 KB |
Output is correct |
13 |
Correct |
5 ms |
632 KB |
Output is correct |
14 |
Correct |
5 ms |
632 KB |
Output is correct |
15 |
Correct |
5 ms |
504 KB |
Output is correct |
16 |
Correct |
5 ms |
632 KB |
Output is correct |
17 |
Correct |
5 ms |
504 KB |
Output is correct |
18 |
Correct |
5 ms |
632 KB |
Output is correct |
19 |
Correct |
5 ms |
760 KB |
Output is correct |
20 |
Correct |
5 ms |
632 KB |
Output is correct |
21 |
Correct |
6 ms |
1272 KB |
Output is correct |
22 |
Correct |
109 ms |
23928 KB |
Output is correct |
23 |
Correct |
53 ms |
8056 KB |
Output is correct |
24 |
Correct |
60 ms |
12152 KB |
Output is correct |
25 |
Correct |
58 ms |
9336 KB |
Output is correct |
26 |
Correct |
54 ms |
8568 KB |
Output is correct |
27 |
Correct |
35 ms |
9720 KB |
Output is correct |
28 |
Correct |
44 ms |
11896 KB |
Output is correct |
29 |
Correct |
72 ms |
16504 KB |
Output is correct |
30 |
Correct |
56 ms |
9464 KB |
Output is correct |