#include <bits/stdc++.h>
using namespace std;
using ll = long long; using pii = pair<ll,ll>;
using ld = double; using pld = pair<ld,ld>;
const ll INF = 1e18;
int main() {
ios_base::sync_with_stdio(false); cin.tie(0);
cout << setprecision(20);
ll N,K; cin >> N >> K;
vector<pii> vba;
for (ll i=0;i<N;i++) {
ll a,b; cin >> a >> b;
if (b==-1) {
b=INF;
}
vba.push_back({b,a});
}
sort(vba.begin(),vba.end());
ll A[N]; ll B[N];
for (ll i=0;i<N;i++) {
B[i]=vba[i].first;
A[i]=vba[i].second;
}
ll psf[N+1]; //prefix sum for finals: first block of length L (=idx)
//-> take bottom K-L values of remaining N-L
multiset<ll> rnum;
psf[N]=0;
for (ll i=K;i<N;i++) {
rnum.insert(A[i]);
psf[i]=0;
}
for (ll i=(K-1);i>=0;i--) {
rnum.insert(A[i]);
ll v = *(rnum.begin()); rnum.erase(rnum.begin());
psf[i]=psf[i+1]+v;
}
ld ans = psf[0];
for (ll bf=0;bf<=K;bf++) {
vector<ld> dp0(N+1,INF); //bcurr -> minvalue
dp0[0]=0;
for (ll i=0;i<K;i++) {
vector<ld> dp1(N+1,INF);
for (ll bc=0;bc<bf;bc++) { //if bc=bf already you could just use the global feature (on prev turn)
if (dp0[bc]<INF) {
//take A
dp1[bc]=min(dp1[bc],dp0[bc]+((ld)A[i])/((ld)(bf+1)));
//take B
if (B[i]!=INF) {
dp1[bc+1]=min(dp1[bc+1],dp0[bc]+((ld)(B[i]))/((ld)(bc+1)));
}
}
}
dp0=dp1;
//now consider bc=bf term: ready to fly out
ans = min(ans,dp0[bf]+((ld)psf[i+1])/((ld)(bf+1)));
}
}
cout << ans << "\n";
}
