codeforces题解:
Let us first find the answer for a fixed sequence of songs. Consider any two songs which are at positions i and j (i < j) in the playlist. If Manao liked song i and disliked song j, then song i will be listened to again. Therefore, with probability
p[i]*(1-p[j])
the process length will increase byL[i]
. The sum ofL[i]*p[i]*(1-p[j])
over all pairs (plus the length of all songs since Manao listens to them at least once) is the expected length for the fixed sequence.So we have that if there are two songs (l1, p1) and (l2, p2), the first one should be placed earlier in the playlist if
l1*p1*(1-p2)>l2*p2*(1-p1)
and later otherwise. This is obviously true if there are only two songs. But suppose that we have more and you ask, why can't there be another song (l3, p3) such that the above inequality rules out that the first song should be after this one and the second song should be before it? Then it's unclear which of the orders results in the maximum expected value. Consider this case in details:l1*p1*(1-p2)>l2*p2*(1-p1) l1*p1*(1-p3)<l3*p3*(1-p1) l2*p2*(1-p3)>l3*p3*(1-p2) <=> l1*p1/(1-p1)>l2*p2/(1-p2) l1*p1/(1-p1)<l3*p3/(1-p3) l2*p2/(1-p2)>l3*p3/(1-p3)
(this is not quite true if any pi is equal to 0 or 1, but that's not important here)
We have a contradicting system of inequations, so such a case is impossible.
Next, let's consider some order of songs
(l[1], p[1]), ..., (l[n], p[n])
, in which there is a pair of neighbouring songs i and i+1, for which the conditionl[i] * p[i] * (1 - p[i + 1]) >= l[i + 1] * p[i + 1] * (1 - p[i])
does not hold, and assume that this order is optimal. Evidently, if we interchange songs i and i+1, the answer will only change by the value contributed by this pair (i.e.l[i] * p[i] * (1 - p[i + 1])
). The rest of the songs keep their order towards song i and song i+1. Butl[i] * p[i] * (1 - p[i + 1]) < l[i + 1] * p[i + 1] * (1 - p[i])
, therefore if we put song i+1 before song i, we obtain a larger value. So we have a contradiction — the order chosen is not optimal.So it turns out that the permutation with maximum possible expected value is obtained when sorting the songs in decreasing order of
l[i]*p[i]/(1-p[i])
. But we still have a problem of computing the answer for a fixed permutation: we only learned how to do this inO(n^2)
, which is too slow with n=50000. We can use an idea which is probably a yet another dynamic programming example. Suppose we have fixed j and are counting the contribution of song j towards the answer if Manao dislikes it. This value is(l1*p1 + l2*p2 + ... + l[j-1]*p[j-1])
. For j+1, the corresponding value will be(l1*p1+...+l[j-1]*p[j-1]+l[j]*p[j])
. It turns out that these values differ in only a single summand, so we can compute each of them inO(1)
if we consider j-th one by one in increasing order. This idea can be expressed as follows:lovedLenExp = 0. answerExp = 0. for j = 1 to N answerExp += l[j] answerExp += lovedLenExp * (1 - p[j]) lovedLenExp += l[j] * p[j]
That's all :)
#include <cstdio>
#include <cstring>
#include <cstdlib>
#include <cmath>
#include <algorithm>
#include <iostream>
#include <map>
#include <set>
#include <vector>
#include <string>
#include <queue>
using namespace std;
typedef long long LL;
typedef double DB;
#define INF 2147483647
#define mms(_x, _y) memset(_x, _y, sizeof(_x));
#define rep(_x, _y, _z) for(int _x = _y; _x <= _z; _x++)
#define per(_x, _y, _z) for(int _x = _y; _x >= _z; _x--)
#define BIGNUM int
const int MOD=1e9+7;
const double pi = acos(-1.0);
const double E = exp(1.0);
const int M = 1e5+9;
struct x{
DB l,p;
}nu[M];
int n;
DB s,an;
int cmp(const void* a,const void* b){
return (*(x*)b).l*(*(x*)b).p*(100-(*(x*)a).p)-(*(x*)a).l*(*(x*)a).p*(100-(*(x*)b).p);
}
void Init(){
scanf("%d",&n);
rep(i,0,n-1)scanf("%lf%lf",&nu[i].l,&nu[i].p);
s=0;
an=0;
}
void Solve(){
qsort(nu,n,sizeof(nu[0]),cmp);
rep(i,0,n-1){
s+=an*(1-nu[i].p/100.0)+nu[i].l;
an+=nu[i].l*nu[i].p/100.0;
}
printf("%.15lf\n",s);
}
int main(){
int _T = 1;
//scanf("%d",&_T);
while(_T--){
Init();
Solve();
}
return 0;
}