Submission #39393831

---

Source Code Expand

#ifndef LOCAL
#pragma GCC optimize ("Ofast")
#pragma GCC optimize ("unroll-loops")
#endif

#include <bits/stdc++.h>
using namespace std;

using ll=long long;
#define int ll

#define rng(i,a,b) for(int i=int(a);i<int(b);i++)
#define rep(i,b) rng(i,0,b)
#define gnr(i,a,b) for(int i=int(b)-1;i>=int(a);i--)
#define per(i,b) gnr(i,0,b)
#define pb push_back
#define eb emplace_back
#define a first
#define b second
#define bg begin()
#define ed end()
#define all(x) x.bg,x.ed
#define si(x) int(x.size())
#ifdef LOCAL
#define dmp(x) cerr<<__LINE__<<" "<<#x<<" "<<x<<endl
#else
#define dmp(x) void(0)
#endif

template<class t,class u> bool chmax(t&a,u b){if(a<b){a=b;return true;}else return false;}
template<class t,class u> bool chmin(t&a,u b){if(b<a){a=b;return true;}else return false;}

template<class t> using vc=vector<t>;
template<class t> using vvc=vc<vc<t>>;

using pi=pair<int,int>;
using vi=vc<int>;

template<class t,class u>
ostream& operator<<(ostream& os,const pair<t,u>& p){
	return os<<"{"<<p.a<<","<<p.b<<"}";
}

template<class t> ostream& operator<<(ostream& os,const vc<t>& v){
	os<<"{";
	for(auto e:v)os<<e<<",";
	return os<<"}";
}

#define mp make_pair
#define mt make_tuple
#define one(x) memset(x,-1,sizeof(x))
#define zero(x) memset(x,0,sizeof(x))
#ifdef LOCAL
void dmpr(ostream&os){os<<endl;}
template<class T,class... Args>
void dmpr(ostream&os,const T&t,const Args&... args){
	os<<t<<" ";
	dmpr(os,args...);
}
#define dmp2(...) dmpr(cerr,__LINE__,##__VA_ARGS__)
#else
#define dmp2(...) void(0)
#endif

using uint=unsigned;
using ull=unsigned long long;

template<class t,size_t n>
ostream& operator<<(ostream&os,const array<t,n>&a){
	return os<<vc<t>(all(a));
}

template<int i,class T>
void print_tuple(ostream&,const T&){
}

template<int i,class T,class H,class ...Args>
void print_tuple(ostream&os,const T&t){
	if(i)os<<",";
	os<<get<i>(t);
	print_tuple<i+1,T,Args...>(os,t);
}

template<class ...Args>
ostream& operator<<(ostream&os,const tuple<Args...>&t){
	os<<"{";
	print_tuple<0,tuple<Args...>,Args...>(os,t);
	return os<<"}";
}

template<class t>
void print(t x,int suc=1){
	cout<<x;
	if(suc==1)
		cout<<"\n";
	if(suc==2)
		cout<<" ";
}

ll read(){
	ll i;
	cin>>i;
	return i;
}

vi readvi(int n,int off=0){
	vi v(n);
	rep(i,n)v[i]=read()+off;
	return v;
}

pi readpi(int off=0){
	int a,b;cin>>a>>b;
	return pi(a+off,b+off);
}

template<class t,class u>
void print(const pair<t,u>&p,int suc=1){
	print(p.a,2);
	print(p.b,suc);
}

template<class t,class u>
void print_offset(const pair<t,u>&p,ll off,int suc=1){
	print(p.a+off,2);
	print(p.b+off,suc);
}

template<class T>
void print(const vector<T>&v,int suc=1){
	rep(i,v.size())
		print(v[i],i==int(v.size())-1?suc:2);
}

template<class T>
void print_offset(const vector<T>&v,ll off,int suc=1){
	rep(i,v.size())
		print(v[i]+off,i==int(v.size())-1?suc:2);
}

template<class T,size_t N>
void print(const array<T,N>&v,int suc=1){
	rep(i,N)
		print(v[i],i==int(N)-1?suc:2);
}

string readString(){
	string s;
	cin>>s;
	return s;
}

template<class T>
T sq(const T& t){
	return t*t;
}

void YES(bool ex=true){
	cout<<"YES\n";
	if(ex)exit(0);
	#ifdef LOCAL
	cout.flush();
	#endif
}
void NO(bool ex=true){
	cout<<"NO\n";
	if(ex)exit(0);
	#ifdef LOCAL
	cout.flush();
	#endif
}
void Yes(bool ex=true){
	cout<<"Yes\n";
	if(ex)exit(0);
	#ifdef LOCAL
	cout.flush();
	#endif
}
void No(bool ex=true){
	cout<<"No\n";
	if(ex)exit(0);
	#ifdef LOCAL
	cout.flush();
	#endif
}
//#define CAPITAL
/*
void yes(bool ex=true){
	#ifdef CAPITAL
	cout<<"YES"<<"\n";
	#else
	cout<<"Yes"<<"\n";
	#endif
	if(ex)exit(0);
	#ifdef LOCAL
	cout.flush();
	#endif
}
void no(bool ex=true){
	#ifdef CAPITAL
	cout<<"NO"<<"\n";
	#else
	cout<<"No"<<"\n";
	#endif
	if(ex)exit(0);
	#ifdef LOCAL
	cout.flush();
	#endif
}*/
void possible(bool ex=true){
	#ifdef CAPITAL
	cout<<"POSSIBLE"<<"\n";
	#else
	cout<<"Possible"<<"\n";
	#endif
	if(ex)exit(0);
	#ifdef LOCAL
	cout.flush();
	#endif
}
void impossible(bool ex=true){
	#ifdef CAPITAL
	cout<<"IMPOSSIBLE"<<"\n";
	#else
	cout<<"Impossible"<<"\n";
	#endif
	if(ex)exit(0);
	#ifdef LOCAL
	cout.flush();
	#endif
}

constexpr ll ten(int n){
	return n==0?1:ten(n-1)*10;
}

const ll infLL=LLONG_MAX/3;

#ifdef int
const int inf=infLL;
#else
const int inf=INT_MAX/2-100;
#endif

int topbit(signed t){
	return t==0?-1:31-__builtin_clz(t);
}
int topbit(ll t){
	return t==0?-1:63-__builtin_clzll(t);
}
int topbit(ull t){
	return t==0?-1:63-__builtin_clzll(t);
}
int botbit(signed a){
	return a==0?32:__builtin_ctz(a);
}
int botbit(ll a){
	return a==0?64:__builtin_ctzll(a);
}
int botbit(ull a){
	return a==0?64:__builtin_ctzll(a);
}
int popcount(signed t){
	return __builtin_popcount(t);
}
int popcount(ll t){
	return __builtin_popcountll(t);
}
int popcount(ull t){
	return __builtin_popcountll(t);
}
bool ispow2(int i){
	return i&&(i&-i)==i;
}
ll mask(int i){
	return (ll(1)<<i)-1;
}
ull umask(int i){
	return (ull(1)<<i)-1;
}
ll minp2(ll n){
	if(n<=1)return 1;
	else return ll(1)<<(topbit(n-1)+1);
}

bool inc(int a,int b,int c){
	return a<=b&&b<=c;
}

template<class t> void mkuni(vc<t>&v){
	sort(all(v));
	v.erase(unique(all(v)),v.ed);
}

ll rand_int(ll l, ll r) { //[l, r]
	//#ifdef LOCAL
	static mt19937_64 gen;
	/*#else
	static mt19937_64 gen(chrono::steady_clock::now().time_since_epoch().count());
	#endif*/
	return uniform_int_distribution<ll>(l, r)(gen);
}

ll rand_int(ll k){ //[0,k)
	return rand_int(0,k-1);
}

template<class t>
void myshuffle(vc<t>&a){
	rep(i,si(a))swap(a[i],a[rand_int(0,i)]);
}

template<class t>
int lwb(const vc<t>&v,const t&a){
	return lower_bound(all(v),a)-v.bg;
}
template<class t>
bool bis(const vc<t>&v,const t&a){
	return binary_search(all(v),a);
}

vvc<int> readGraph(int n,int m){
	vvc<int> g(n);
	rep(i,m){
		int a,b;
		cin>>a>>b;
		//sc.read(a,b);
		a--;b--;
		g[a].pb(b);
		g[b].pb(a);
	}
	return g;
}

vvc<int> readTree(int n){
	return readGraph(n,n-1);
}

vc<ll> presum(const vi&a){
	vc<ll> s(si(a)+1);
	rep(i,si(a))s[i+1]=s[i]+a[i];
	return s;
}
//BIT で数列を管理するときに使う (CF850C)
template<class t>
vc<t> predif(vc<t> a){
	gnr(i,1,si(a))a[i]-=a[i-1];
	return a;
}
template<class t>
vvc<ll> imos(const vvc<t>&a){
	int n=si(a),m=si(a[0]);
	vvc<ll> b(n+1,vc<ll>(m+1));
	rep(i,n)rep(j,m)
		b[i+1][j+1]=b[i+1][j]+b[i][j+1]-b[i][j]+a[i][j];
	return b;
}

//verify してないや
void transvvc(int&n,int&m){
	swap(n,m);
}
template<class t,class... Args>
void transvvc(int&n,int&m,vvc<t>&a,Args&...args){
	assert(si(a)==n);
	vvc<t> b(m,vi(n));
	rep(i,n){
		assert(si(a[i])==m);
		rep(j,m)b[j][i]=a[i][j];
	}
	a.swap(b);
	transvvc(n,m,args...);
}
//CF854E
void rotvvc(int&n,int&m){
	swap(n,m);
}
template<class t,class... Args>
void rotvvc(int&n,int&m,vvc<t>&a,Args&...args){
	assert(si(a)==n);
	vvc<t> b(m,vi(n));
	rep(i,n){
		assert(si(a[i])==m);
		rep(j,m)b[m-1-j][i]=a[i][j];
	}
	a.swap(b);
	rotvvc(n,m,args...);
}

//ソートして i 番目が idx[i]
//CF850C
template<class t>
vi sortidx(const vc<t>&a){
	int n=si(a);
	vi idx(n);iota(all(idx),0);
	sort(all(idx),[&](int i,int j){return a[i]<a[j];});
	return idx;
}
//vs[i]=a[idx[i]]
//例えば sortidx で得た idx を使えば単にソート列になって返ってくる
//CF850C
template<class t>
vc<t> a_idx(const vc<t>&a,const vi&idx){
	int n=si(a);
	assert(si(idx)==n);
	vc<t> vs(n);
	rep(i,n)vs[i]=a[idx[i]];
	return vs;
}
//CF850C
vi invperm(const vi&p){
	int n=si(p);
	vi q(n);
	rep(i,n)q[p[i]]=i;
	return q;
}

template<class t,class s=t>
s SUM(const vc<t>&a){
	return accumulate(all(a),s(0));
}

template<class t>
t MAX(const vc<t>&a){
	return *max_element(all(a));
}

template<class t>
t MIN(const vc<t>&a){
	return *min_element(all(a));
}

template<class t,class u>
pair<t,u> operator+(const pair<t,u>&a,const pair<t,u>&b){
	return mp(a.a+b.a,a.b+b.b);
}

vi vid(int n){
	vi res(n);iota(all(res),0);
	return res;
}

//cost(i,j) が monge
//O(N) で DP する
//stress-tested
//IOI 2016 Alien
//平均 7N 回程度 cost(i,j) が走っていそう
template<class F>
auto&larsch(int n,F cost){
	assert(2<=n);
	using t=typename invoke_result<F,int,int>::type;
	const int L=topbit(n);
	static vc<t> res;res.resize(n);
	static vc<pair<t,int>> val[30];
	static vc<pi> pos[30];
	static int s[30];
	rep(i,L){
		val[i].resize(n>>i);
		pos[i].resize(n>>i);
		s[i]=0;
	}
	s[0]=n;
	rep(i,n)pos[0][i].a=i;
	auto eval=[&](int lv,int i,int j){
		return res[pos[lv][i].a]+cost(pos[lv][i].a,((j+1)<<lv)-1);
	};
	auto setval=[&](int lv,int i,int j){
		val[lv][j]=mp(eval(lv,i,j),i);
	};
	res[0]=t{};
	rng(i,1,n){
		const int l=i,r=i+1;
		int lv=botbit(l);
		if((l>>lv)<(n>>lv)){
			setval(lv,(l>>lv)-1,l>>lv);
			if(l>>(lv+1))
				chmin(val[lv][l>>lv],
				mp(val[lv+1][l>>(lv+1)].a,pos[lv+1][val[lv+1][l>>(lv+1)].b].b));
		}
		while(lv>0){
			lv--;
			if((l>>lv)<(n>>lv)){
				int lw=val[lv][(l>>lv)-1].b;
				int up=(l>>(lv+1))<(n>>(lv+1))?
					pos[lv+1][val[lv+1][l>>(lv+1)].b].b:(l>>lv)-1;
				assert(lw<=up);
				setval(lv,lw,l>>lv);
				rng(j,lw+1,up+1)
					chmin(val[lv][l>>lv],mp(eval(lv,j,l>>lv),j));
			}
		}
		res[i]=val[0][i].a;
		while((r&1<<lv)==0&&lv<L-1){
			lv++;
			rng(b,(r>>(lv-1))-2,r>>(lv-1)){
				int a=pos[lv-1][b].a;
				while(s[lv]>=(r>>lv)){
					if(eval(lv,s[lv]-1,s[lv])>res[a]+cost(a,((s[lv]+1)<<lv)-1))s[lv]--;
					else break;
				}
				if(s[lv]+1<(n>>lv))pos[lv][s[lv]++]=pi(a,b);
			}
		}
	}
	return res;
}

//0->n-1 のジャンプ
//monge コストでちょうど k step でやれ，みたいな最小化問題 
//ans(step) が単調減少だとする(つまり小刻みであればあるほど嬉しい)
//max(ans(step))-max(ans(step)) の上界を dif で与えている
//stress-tested (define ll なし，復元もちゃんと確かめた)
//IOI 2016 Alien
//N=10^5,dif=10^18 で 0.2 sec 強かかる（costはO(1))
template<class F>
ll kstepmin(int n,int k,ll dif,F cost,vi*sol=nullptr){
	assert(inc(1,k,n-1));
	ll lw=-1,up=dif+1,ans=infLL;
	vc<pair<ll,int>> buf(n);
	while(up-lw>1){
		ll mid=(lw+up)/2;
		auto&dp=larsch(n,[&](int i,int j){
			return pair<ll,int>(cost(i,j)+mid,1);
		});
		if(dp[n-1].b<=k){
			up=mid;
			ans=dp[n-1].a-k*mid;
			copy(all(dp),buf.bg);
		}else lw=mid;
	}
	assert(ans<infLL);
	if(sol){
		sol->clear();
		int cur=n-1;
		sol->pb(cur);
		while(cur){
			assert(k);
			k--;
			int nx=cur;
			while(nx--){
				if(buf[nx].b<=k&&buf[nx].a+cost(nx,cur)+up==buf[cur].a){
					sol->pb(cur=nx);
					break;
				}
			}
			assert(nx>=0);
		}
		assert(k==0);
		reverse(sol->bg,sol->ed);
	}
	return ans;
}


void slv(){
	int n,k;cin>>n>>k;
	vi a=readvi(n);
	int m=MAX(a)+1;
	chmin(k,m-1);
	vi cnt(m);
	for(auto v:a)cnt[v]++;
	rep(i,m-1)cnt[i+1]+=cnt[i];
	print(kstepmin(m,k,ten(15),[&](int i,int j){return j*(cnt[j]-cnt[i]);}));
}

signed main(){
	cin.tie(0);
	ios::sync_with_stdio(0);
	cout<<fixed<<setprecision(20);
	
	//int t;cin>>t;rep(_,t)
	slv();
}

Submission Info

Judge Result