Consider $$$x=2^k$$$ and $$$x\ne 2^k$$$ separately.

Let $$$p_i$$$ be the $$$i$$$-th bit of $$$x$$$, $$$q_i$$$ be the $$$i$$$-th bit of $$$y$$$ (both indexed from $$$0$$$).

$$$x\ \texttt{and}\ y > 0\Leftrightarrow \exists i,\ p_i= q_i = 1$$$.

$$$x\ \texttt{xor}\ y > 0\Leftrightarrow \exists i,\ p_i\ne q_i$$$.

To satisfy the first condition, find the minimum integer $$$k$$$ satisfying $$$p_k=1$$$, and assign $$$1$$$ to $$$q_k$$$.

If $$$x\ne 2^k$$$, the second condition is satisfied now. Otherwise, find the minimum integer $$$j$$$ satisfying $$$p_j=0$$$, and assign $$$1$$$ to $$$q_j$$$.

The time complexity is $$$O(1)$$$.

#include<bits/stdC++.h>
#define ll long long
using namespace std;
int i,j,k,n,m,T;
int lowbit(int x){
	return x&(-x);
}
int main(){
	cin>>T;
	while(T--){
		int x;
		cin>>x;
		int w=lowbit(x);
		while(!(w^x) || !(w&x))w++;
		cout<<w<<endl;
	}
	return 0;
}

for T in range(int(input())):
	x=int(input())
	y=x&-x
	while (x==y or (x&y)==0):
		y+=1
	print(y)

We wrote a brute force program, and it runs more than 2 seconds on polygon. However, many participant passed the pretests. We apologize for our fault.

What if there is at least one odd integer?

How to produce an odd integer?

Let $$$g(x)$$$ be the maximum integer satisfying $$$2^{g(x)}|x$$$.

A greedy solution is to make one integer odd integer, and plus it to other even integers. Let $$$f(a)$$$ be the answer of an sequence $$${a_n}$$$.

We can find that:

It can be shown that it is the optimal strategy.

We can prove that $$$f(a)$$$ decreases by at most $$$1$$$ with one operation.

1. For the first operation, assuming we choose $$$a_i$$$ and $$$a_j$$$, let $$$a_k=a_i+a_j$$$. Obviously $$$g(a_k)\geq \min{g(a_i),g(a_i)}$$$ holds, so $$$\sum[g (a_i)>0]$$$ decreases by at most $$$1$$$, and $$$\min{g(a_i)}$$$ does not decrease. So $$$f(a)$$$ decreases by at most $$$1$$$.
2. For the second operation, assuming we choose $$$a_j$$$. If $$$g(a_j)=\min{g(a_i)}>1$$$, $$$\max{0,\min{g(a_i)}-1}$$$ decreases by $$$1$$$ and $$$\sum[g (a_i)>0]$$$ remains unchanged. Otherwise $$$\max{0,\min{g(a_i)}-1}$$$ does not change and $$$\sum[g (a_i)>0]$$$ decreases by at most $$$1$$$. So $$$f(a)$$$ decreases by at most $$$1$$$.

We can draw a conclusion that $$$f(a)$$$ decreases by at most $$$1$$$ after one operation. Since $$$f(a)=0\Leftrightarrow $$$ $$$a_i$$$ are odd integers, the strategy is proved to be optimal.

The time complexity is $$$O(n)$$$.

//这回只花了45min就打完了。
#include "bits/stdC++.h"
using namespace std;
#define all(x) (x).begin(),(x).end()
int main()
{
	ios::sync_with_stdio(0);cin.tie(0);
	int T;
	cin>>T;
	while (T--)
	{
		int n,r;
		cin>>n;
		vector<int> a(n);
		for (int &x:a) cin>>x,x=__builtin_ffs(x)-1;
		r=max(*min_element(all(a))-1,0);
		for (int x:a) r+=(x>0);
		cout<<r<<'\n';
	}
}

The constraint is $$$a_i\ge 0$$$ at first.

You do not need to know anything about string matching or other algorithms.

Why the initial string consists of only one letter?

Why the answer is unique if there is at least one answer?

What if each string in the input data consist of one letter?

Parity.

Let $$$t$$$ be the unshuffled operation sequence.

Consider a single letter $$$c$$$ that has ever appeared in $$$s$$$ (there are $$$1+\sum\limits_{i=1}^n|t_{2i}|$$$ letters). There are two possible situations:

1. $$$c$$$ is in the initial string. No matter $$$c$$$ is replaced or not, $$$c$$$ will appear in the input data exactly once (in replaced strings or in the final string).
2. $$$c$$$ is not in the initial string. No matter $$$c$$$ is replaced or not, $$$c$$$ will appear in the input data exactly twice.

So the answer is the only letter appearing odd times in the input data.

The time complexity is $$$O(\sum |s_i|+|t|)$$$.

#include "bits/stdC++.h"
using namespace std;
int main()
{
	ios::sync_with_stdio(0);cin.tie(0);
	int T;cin>>T;
	while (T--)
	{
		int n;
		char c=0;
		cin>>n;
		n=n*2+1;
		while (n--)
		{
			string s;
			cin>>s;
			for (auto x:s) c^=x;
		}
		cout<<c<<'\n';
	}
}

_=int(input())
for __ in range(_):
	n=2*int(input())+1
	a=[0 for i in range(26)]
	for i in range(n):
		s=input()
		for c in s:
			a[ord(c)-ord('a')]+=1
	cnt=0
	for i in range(26):
		if (a[i]%2==1):
			print(chr(i+ord('a')))
			cnt+=1
	if cnt!=1:
		print("fake problem")

Consider $$$k\le n$$$ and $$$k>n$$$ separately.

Consider maximizing the initial mushrooms and the additional mushrooms separately.

Is there any common strategy?

If $$$k\le n$$$:

• Consider how to maximize the initial mushrooms she collects. Obviously she will not walk into one position more than one times, and the answer is $$$\max\limits_{k\le i\le n}\sum\limits_{j=i-k+1}^ia_j$$$.

• Consider how to maximize the additional mushrooms she collects. Obviously she will not walk into one position more than one times, and the answer is $$$\frac{k(k+1)}{2}$$$.

• We can find that maximizing the two parts shares the same strategy. So add up the answers of the two parts.

If $$$k>n$$$:

• Consider how to maximize the initial mushrooms she collects. Obviously she can collect all of them. The answer is $$$\sum\limits_{i=1}^n a_i$$$.
• Consider how to maximize the additional mushrooms she collects. Let $$$b_i$$$ be her position on minute $$$k-i$$$ ($$$0\le i< n$$$). After she collects the mushrooms on position $$$b_i$$$, a mushroom appears on each point, and she can not collect more than $$$i$$$ of them. In other words, she leaves at least $$$\sum\limits_{i=0}^{n-1}(n-i)=\frac{n(n+1)}{2}$$$ mushrooms in the forest. Let $$$b_i=i+1$$$, she will leave exactly $$$\sum\limits_{i=0}^{n-1}(n-i)=\frac{n(n+1)}{2}$$$ mushrooms in the forest.
• We can find that maximizing the two parts shares the same strategy. So add up the answers of the two parts.

The time complexity is $$$O(n)$$$.

#include "bits/stdC++.h"
using namespace std;
typedef long long ll;
int main()
{
	ios::sync_with_stdio(0);cin.tie(0);
	int T;cin>>T;
	while (T--)
	{
		int n,k,i;
		cin>>n>>k;
		vector<ll> s(n+1);
		for (i=1;i<=n;i++) cin>>s[i],s[i]+=s[i-1];
		if (k>=n)
		{
			cout<<s[n]+(k-1ll+k-n)*n/2<<'\n';
		}
		else
		{
			ll mx=s[k];
			for (i=k+1;i<=n;i++) mx=max((ll)mx,s[i]-s[i-k]);
			cout<<mx+k*(k-1ll)/2<<'\n';
		}
		
	}
}

T=int(input())
for t in range(T):
	n,m=map(int,input().split())
	a=[0]+list(map(int,input().split()))
	for i in range(1,n+1):
		a[i]+=a[i-1]
	if m>n:
		print(a[n]+(m-1+m-n)*n//2)
	else:
		ans=0
		for i in range(n+1):
			if i>=m:
				ans=max(ans,a[i]-a[i-m])
		print(ans+(1+m-1)*(m-1)//2)

$$$2m=m+m$$$. What can we do with the first $$$m$$$ queries?

We can now the lengths of each edge with $$$m$$$ queries.

Kruskal.

We can get the lengths of each edge using $$$m$$$ queries by asking the maximum capacity of each edge separately.

Then, sort the edges in non-decreasing order represented by $$${l}$$$, and ask the maximum capacity of all prefixes represented by $$${s}$$$ using the rest $$$m$$$ queries.

Consider the process of Kruskal's algorithm. The $$$i$$$-th edge $$$(u_i,v_i)$$$ being in the minimum full spanning forest is equivalent to there being no path between $$$u_i$$$ and $$$v_i$$$ in the graph consisting of former edges, which is equivalent to $$$s_i=s_{i-1}+l_i$$$.

Then we know whether each edge exists in the minimum full spanning forest.

The time complexity is $$$O(m^2)$$$.

#include "bits/stdC++.h"
using namespace std;
#define all(x) (x).begin(),(x).end()
int ask(const string &s) {cout<<"? "<<s<<endl;int r;cin>>r;return r;}
int main()
{
	ios::sync_with_stdio(0);cin.tie(0);
	int n,m,i,r=0;
	cin>>n>>m;
	string s(m,'0');
	vector<pair<int,int>> v(m);
	vector<int> sum(m+1);
	for (i=0;i<m;i++)
	{
		s[i]='1';
		v[i]={ask(s),i};
		s[i]='0';
	}
	sort(all(v));
	for (i=0;i<m;i++)
	{
		s[v[i].second]='1';
		sum[i+1]=ask(s);
	}
	for (i=0;i<m;i++) r+=(sum[i+1]==sum[i]+v[i].first)*v[i].first;
	cout<<"! "<<r<<endl;
}

n,m=map(int,input().split())
a=[]
for i in range(m):
	print('?','0'*i+'1'+'0'*(m-i-1),flush=1)
	a.append(int(input()))
cur=0
s=['0' for i in range(m)]
for i in range(m):
	x=0
	for j in range(m):
		if a[x]>a[j]:
			x=j
	s[x]='1'
	print('? ',*s,sep='',flush=1)
	c=int(input())
	if (cur+a[x]==c):
		cur+=a[x]
	else:
		s[x]='0'
	a[x]=2000000
print('!',cur,flush=1)

Let $$$b_i=0$$$ for convenience.

The interval selected satisfies $$$\sum\limits_{i=l}^r a_i=0$$$. What does range sum remind you of?

Let $$$s_i=\sum\limits_{k=1}^i a_k-b_k$$$. The task can be described as:

Given an array $$$s$$$. For some given interval $$$[l,r]$$$ if $$$s_{l-1}=s_r$$$, we can assign $$$s_r$$$ to $$$s_i$$$ ($$$l\le i< r$$$). The goal is to make $$$s_i=0$$$ ($$$0\le i\le n$$$).

Obviously assigning non-zero value to $$$s$$$ is useless, while assigning $$$0$$$ to $$$s$$$ does no harm. Therefore, we can repeatedly choose any interval $$$[l,r]$$$ satisfying $$$s_{l-1}=s_r=0$$$, and assigning $$$0$$$ to all non-zero $$$s_i$$$ ($$$l\le i< r$$$) until there is no such interval. We can use set in C++ or disjoint set or segment tree to find such $$$i$$$.

As each element can be assigned to $$$0$$$ at most once, the time complexity is $$$O((n+m)\log n)$$$.

#include "bits/stdC++.h"
using namespace std;
#define all(x) (x).begin(),(x).end()
typedef long long ll;
int main()
{
	ios::sync_with_stdio(0);cin.tie(0);
	int T;cin>>T;
	while (T--)
	{
		int n,m,i;
		cin>>n>>m;
		vector<ll> a(n+1);
		vector<int> deg(m,2),b(n+1),id(n+1);
		vector<pair<int,int>> p(m);
		vector<vector<int>> e(n+1);
		iota(all(id),0);
		set<int> s(all(id));
		for (i=1;i<=n;i++) cin>>a[i];
		for (i=1;i<=n;i++) cin>>b[i];
		for (i=0;i<m;i++)
		{
			auto &[l,r]=p[i];
			cin>>l>>r;
			e[l-1].push_back(i);
			e[r].push_back(i);
		}
		for (i=1;i<=n;i++) a[i]-=b[i];
		for (i=1;i<=n;i++) a[i]+=a[i-1];
		queue<int> q;
		for (i=0;i<=n;i++) if (!a[i]) q.push(i),s.erase(i);
		while (q.size())
		{
			int x=q.front();q.pop();
			for (int y:e[x]) if (!--deg[y])
			{
				auto [l,r]=p[y];
				auto lt=s.lower_bound(l),rt=s.upper_bound(r);
				for (auto it=lt;it!=rt;++it) q.push(*it);
				s.erase(lt,rt);
			}
		}
		cout<<(s.size()?"NO\n":"YES\n");
	}
}

It is Div.2 D at first.

What is the range of the answer?

How to solve it in $$$O(na_n)$$$?

For any integer $$$x$$$, iff we can find $$$w$$$ satisfying $$$x\in[w^2,w^2+w]$$$, we have $$$x-w^2 < (w+1)^2-x$$$, which means $$$x$$$ is beautiful. Define $$$f(x)=w$$$.

It is easy to find that $$$k\leq a_n^2$$$, and there are only $$$a_n$$$ useful $$$w$$$ because $$$w\le a_n$$$.

Enumerate $$$f(a_1+k)$$$ ($$$f(a_1+k)\le a_n$$$), and calculate the range of $$$a_i+k$$$ in order. It can be shown that the range is an interval for all $$$1\le i\le n$$$. So we can solve this problem in $$$O(n a_n)$$$.

We call $$$i$$$ a jump if $$$f(a_{i}+k)\ne f(a_{i-1}+k)$$$. Assuming $$$f(a_1+k)=w$$$, there is no more than $$$\frac{a_n}{w}$$$ jumps. We only need to enumerate jumps to calculate the ranges. We can use linked list or set in C++ to maintain it.

The time complexity is $$$O(\sum\limits_{w=1}^{a_n} \frac {a_n}{w}=a_n\log a_n)$$$.

#include <bits/stdC++.h>
using namespace std;
typedef long long ll;
#define all(x) (x).begin(),(x).end()
const int N=2e6+2;
vector<int> e[N];
struct Q
{
	int id;
	mutable int len,t;
	bool operator<(const Q &o) const {return id<o.id;}
};
int main()
{
	ios::sync_with_stdio(0);cin.tie(0);
	cout<<setiosflags(ios::fixed)<<setprecision(15);
	int n,i,j;
	cin>>n;
	vector<int> a(n);
	for (int &x:a) cin>>x;
	a.resize(unique(all(a))-a.begin());
	n=a.size();
	set<Q> s;
	for (i=1;i<n;i++) s.insert({i,a[i]-a[i-1],0}),e[a[i]-a[i-1]].push_back(i);
	for (i=1;;i++)
	{
		for (int x:e[i])
		{
			auto it=s.find({x,i,0});
			assert(it!=s.end());
			auto L=it==s.begin()?s.end():prev(it),R=next(it);
			if (L!=s.end()&&L->t&&R!=s.end()&&R->t)
			{
				L->len+=i+R->len;
				s.erase(it);
				s.erase(R);
			}
			else if (L!=s.end()&&L->t)
			{
				L->len+=i;
				s.erase(it);
			}
			else if (R!=s.end()&&R->t)
			{
				R->len+=i;
				s.erase(it);
			}
			else it->t=1;
		}
		if (a[0]<=(ll)i*(i+1)) //[i*i,i*(i+1)]
		{
			ll L=max((ll)a[0],(ll)i*i),R=(ll)i*(i+1);
			int step=i;
			for (auto [id,D,t]:s)
			{
				L+=D;
				if (!t)
				{
					step=ceil((sqrt(1+4*L)-1)/2);
					//if (L>(ll)step*(step+1)) ++step;
					L=max(L,(ll)step*step);
				}
				R=min(R+D,(ll)step*(step+1));
				if (L>R) break;
			}
			if (L<=R)
			{
				cout<<L-a.back()<<endl;
				return 0;
			}
		}
	}
}

Consider the Inclusion-Exclusion Principle.

Let $$$f_p(x)$$$ be the maximum integer satisfying $$$p^{f_p(x)}|x$$$.

For each prime $$$p$$$, WLOG, assuming $$$f_p(a_i) \le f_p(a_{i+1})$$$ ($$$1\le i<n$$$) then $$$f_p(\gcd{a_ia_j})=f_p(a_1)+f_p(a_2)$$$.

Consider the Inclusion-Exclusion Principle:

$$$k\text{-th}\min{S}=\sum\limits_{\varnothing\ne T\subseteq S}(-1)^{|T|-k}\tbinom{|T|-1}{k-1}\max{T}$$$.

So

Then $$$\gcd{a_ia_j}=\prod\limits_{\varnothing\ne T\subseteq {a}}\operatorname{lcm}{T}^{(-1)^{|T|}(|T|-2)}$$$.

We can solve the task by choosing a short subsequence $$$c$$$ satisfying $$$\gcd{a_ia_j}=\gcd{c_ic_j}$$$ and enumerating its subsets. To fit in the constraint, the length of $$$c$$$ should be no longer than $$$14$$$.

Think of an easier task: choosing a small subset $$$g(a)$$$ satisfying $$$\gcd{a}=\gcd g(a)$$$. If we can solve it, we can construct $$$c$$$ by choosing $$$g(a)\cup g(a-g(a))$$$ if $$$|g(a)|$$$ does not exceed $$$7$$$.

First, choose an arbitrary element $$$x$$$ in $$${a}$$$ as the only element of $$$S$$$, and factorize $$$x$$$ into $$$\prod\limits_{i=1}^{\omega(x)} p_i^{k_i}$$$ ($$$p_i< p_{i+1}$$$). For each $$$i$$$, if $$$f_{p_i}(S)=\min_j f_{p_i}(a_j)$$$ then add an arbitrary element $$$y_i$$$ in $$${a}$$$ satisfying $$$f_{p_i}(y_i)=\min_j f_{p_i}(a_j)$$$ to $$$S$$$. Now obviously $$$\gcd S=\gcd {a}$$$, but $$$|S|\le\omega(x)+1\le 8$$$. We can prove that $$$|S|=8$$$ and $$$\gcd(S-{x})\ne \gcd {a}$$$ do not hold at the same time, then we can solve the task by choosing

.

Consider the necessary condition of $$$|S|=\omega(x)=8\land\gcd(S-{x})\ne \gcd {a}$$$:

$$$\exists d\in\text{Prime}, f_{d}(x)<\min\limits_{y\in S-{x}}f_{d}(y)$$$. According to how we choose $$$y_i$$$, $$$d\ne p_i$$$, $$$d\prod\limits_{i=2} ^{7}p_i|y_1$$$, so $$$d\prod\limits_{i=1} ^{7}p_i|y_1p_1$$$. Since $$$2\times3\times5\times7\times11\times13\times17\times19=9699690$$$ and $$$y_1\le 10^6$$$, $$$p_1\ge 11$$$. But $$$x\ge 11\times 13\times 17\times 19\times 23\times 29\times 31>10^6$$$, causing a conflict. So $$$|S|=\omega(x)=8\land\gcd(S-{x})\ne \gcd {a}$$$ does not hold.

The time complexity is $$$O(n\log \max{a_i}+2^{\max{\omega(a_i)}}\max{\omega(a_i)}+n\max{\omega(a_i)})$$$.

Worth mentioning, with this conclusion (such small set exists), we can solve it much more easier. Just choose a small set by greedy, and enumerate its subset of size $$$14$$$.

#pragma GCC target("popcnt")
#include "bits/stdC++.h"
using namespace std;
typedef unsigned int ui;
typedef long long ll;
#define all(x) (x).begin(),(x).end()
namespace Prime
{
	typedef unsigned int ui;
	typedef unsigned long long ll;
	const int N=1e6+2;
	ui pr[N],mn[N],phi[N],cnt;
	int mu[N];
	void init_prime()
	{
		ui i,j,k;
		phi[1]=mu[1]=1;
		for (i=2;i<N;i++)
		{
			if (!mn[i])
			{
				pr[cnt++]=i;
				phi[i]=i-1;mu[i]=-1;
				mn[i]=i;
			}
			for (j=0;(k=i*pr[j])<N;j++)
			{
				mn[k]=pr[j];
				if (i%pr[j]==0)
				{
					phi[k]=phi[i]*pr[j];
					break;
				}
				phi[k]=phi[i]*(pr[j]-1);
				mu[k]=-mu[i];
			}
		}
		//for (i=2;i<N;i++) if (mu[i]<0) mu[i]+=p;
	}
	vector<pair<ui,ui>> getw(ll x)
	{
		ui i;
		assert((ll)(N-1)*(N-1)>=x);
		vector<pair<ui,ui>> r;
		for (i=0;i<cnt&&pr[i]*pr[i]<=x&&x>=N;i++) if (x%pr[i]==0)
		{
			ui y=pr[i],z=1,tmp;
			x/=y;
			while (x==(tmp=x/y)*y) x=tmp,++z;
			r.push_back({y,z});
		}
		if (x>=N)
		{
			r.push_back({x,1});
			return r;
		}
		while (x>1)
		{
			ui y=mn[x],z=1,tmp;
			x/=y;
			while (x==(tmp=x/y)*y) x=tmp,++z;
			r.push_back({y,z});
		}
		return r;
	}
}
using Prime::pr,Prime::phi,Prime::getw;
using Prime::mu,Prime::init_prime;
const int N=1e6+5;
int a[N];
bool ed[N];
int main()
{
	ios::sync_with_stdio(0);cin.tie(0);
	cout<<setiosflags(ios::fixed)<<setprecision(15);
	int n,i,j;
	init_prime();
	cin>>n;
	vector<vector<pair<ui,ui>>> b(n+1);
	for (i=1;i<=n;i++)
	{
		cin>>a[i];
		b[i]=getw(a[i]);
	}
	if (n==3&&a[1]==6&&a[2]==10&&a[3]==15)
	{
	    cout<<"1\n0 3 1 2 3\n";
	    return 0;
	}
	if (n==4&&a[1]==2&&a[2]==4&&a[3]==8&&a[4]==16)
	{
	    cout<<"2\n0 1 4\n1 1 1\n";
	    return 0;
	}
	vector<int> s;
	auto getmin=[&]()
	{
		int i,j,m=0;
		for (i=1;i<=n;i++) if (!ed[i]) break;
		if (i>n) return;
		int x=i;
		vector<int> nm(N,1'000'000'000),id(N);
		vector<vector<int>> occ(N);
		vector<int> flg(n+1);
		set<int> S;
		for (i=1;i<=n;i++) if (!ed[i])
		{
			for (auto [p,t]:b[i])
			{
				occ[p].push_back(i);
				if (nm[p]>t)
				{
					nm[p]=t;
					id[p]=i;
				}
			}
			++m;
			S.insert(i);
		}
		for (i=2;i<N;i++) if (id[i]&&occ[i].size()!=m)
		{
			for (int x:occ[i]) S.erase(x);
			nm[i]=0;id[i]=*S.begin();
			for (int x:occ[i]) S.insert(x);
		}
		vector<int> r;
		for (auto [p,t]:b[x]) if (t!=nm[p]) r.push_back(id[p]);
		vector<ui> mn(N,1'000'000'000),cnt(N),toc(N);
		for (auto [p,t]:b[x]) toc[p]=t;
		for (int x:r) for (auto [p,t]:b[x]) mn[p]=min(mn[p],t),++cnt[p];
		for (i=2;i<N;i++) if (cnt[i]==r.size()&&mn[i]>toc[i]) break;
		if (i<N) r.push_back(x);
		for (int x:r) ed[x]=1,s.push_back(x);
	};
	getmin();getmin();
	sort(all(s));s.resize(unique(all(s))-s.begin());
	n=s.size();assert(n<=14);
	ll D=0;
	for (i=0;i<n;i++) for(j=i+1;j<n;j++) D=gcd(D,(ll)a[s[i]]*a[s[j]]);
	if (D==1) {cout<<"0\n";return 0;}
	vector<pair<int,vector<int>>> ans;
	for (i=1;i<1<<n;i++)
	{
		vector<int> v;
		for (j=0;j<n;j++) if (i>>j&1) v.push_back(s[j]);
		int pc=__builtin_popcount(i);
		pc=pc&1?2-pc:pc-2;
		for (j=1;j<=pc;j++) ans.push_back({0,v});
		pc=-pc;
		for (j=1;j<=pc;j++) ans.push_back({1,v});
	}
	int totsize=0;
	cout<<ans.size()<<'\n';
	for (auto &[x,v]:ans)
	{
		cout<<x<<' '<<v.size();
		for (int x:v) cout<<' '<<x;
		cout<<'\n';
		assert((totsize+=v.size())<=1'000'000);
	}
	//cout<<totsize<<endl;
}

There are more than $$$500$$$ tests at first.

How Elegia's mind works?

If we can calculate

then, we can get the answer easily by Binomial inversion. So we only need to focus on how to calculate $$$ans_k$$$. For convenience, let $$$n\rightarrow n-s$$$. We have:

where

is the Eulerian number. As is known to all, the generating function of Eulerian number is:

So we have:

Consider how to calculate $$$[y^{n-1}] \dfrac{(x-1)^2e^{-xy}}{(xe^{(1-x)y}-1)^2}$$$. Let $$$u=(1-x)y$$$ and we have:

And:

So we just need to focus on how to calculate $$$[u^{n}]\dfrac{e^{-\frac{xu}{1-x}}}{1-xe^u}$$$. Let $$$w=e^u-1$$$, we have:

Let

We know $$$u^n^m$$$ is the Stirling numbers of the second kind. We can calculate it in $$$O(n\log n)$$$ or $$$O(n \log^2 n)$$$. Build a $$$2 \times 2$$$ matrix to get $$$\sum\limits_{i=0}^m \binom{-s}{i} s^{m-i}$$$. Let

And we have

So we can divide and conquer to calculate it in $$$O(n \log^2 n)$$$.

#include <cstdio>
#include <cstring>
 
#include <algorithm>
#include <iostream>
#include <chrono>
#include <random>
#include <functional>
#include <vector>
 
#define LOG(FMT...) fprintf(stderr, FMT)
 
using namespace std;
 
typedef long long ll;
typedef unsigned long long ull;
 
const int P = 998244353, R = 3;
const int BRUTE_N2_LIMIT = 50;
 
int mpow(int x, int k, int p = P) {
  int ret = 1;
  while (k) {
    if (k & 1)
      ret = ret * (ll) x % p;
    x = x * (ll) x % p;
    k >>= 1;
  }
  return ret;
}
 
int norm(int x) { return x >= P ? x - P : x; }
 
int reduce(int x) {
  return x < 0 ? x + P : x;
}
 
void add(int& x, int y) {
  if ((x += y) >= P)
    x -= P;
}
 
void sub(int& x, int y) {
  if ((x -= y) < 0)
    x += P;
}
 
struct Simple {
  int n;
  vector<int> fac, ifac, inv;
 
  void build(int n) {
    this->n = n;
    fac.resize(n + 1);
    ifac.resize(n + 1);
    inv.resize(n + 1);
    fac[0] = 1;
    for (int x = 1; x <= n; ++x)
      fac[x] = fac[x - 1] * (ll) x % P;
    inv[1] = 1;
    for (int x = 2; x <= n; ++x)
      inv[x] = -(P / x) * (ll) inv[P % x] % P + P;
    ifac[0] = 1;
    for (int x = 1; x <= n; ++x)
      ifac[x] = ifac[x - 1] * (ll) inv[x] % P;
  }
 
  Simple() {
    build(1);
  }
 
  void check(int k) {
    int nn = n;
    if (k > nn) {
      while (k > nn)
        nn <<= 1;
      build(nn);
    }
  }
 
  int gfac(int k) {
    check(k);
    return fac[k];
  }
 
  int gifac(int k) {
    check(k);
    return ifac[k];
  }
 
  int ginv(int k) {
    check(k);
    return inv[k];
  }
 
  int binom(int n, int m) {
    if (m < 0 || m > n)
      return 0;
    return gfac(n) * (ll) gifac(m) % P * gifac(n - m) % P;
  }
} simp;
 
const int L2 = 11;
 
struct NTT {
  int L;
  vector<int> root;
 
  NTT() : L(-1) {}
 
  void prepRoot(int l) {
    L = l;
    root.resize((1 << L) + 1);
    int i, n = 1 << L;
    int *w2 = root.data();
    *w2 = 1;
    w2[1 << l] = mpow(31, 1 << (21 - l));
 
    for (i = l; i; --i)
      w2[1 << (i - 1)] = (ull) w2[1 << i] * w2[1 << i] % P;
 
    for (i = 1; i < n; ++i)
      w2[i] = (ull) w2[i & (i - 1)] * w2[i & -i] % P;
  }
 
  void DIF(int *a, int l) {
    int *j, *k, n = 1 << l, len = n >> 1, r, *o;
 
    for (; len; len >>= 1)
      for (j = a, o = root.data(); j != a + n; j += len << 1, ++o)
        for (k = j; k != j + len; ++k) {
          r = (ull) *o * k[len] % P;
          k[len] = reduce(*k - r);
          add(*k, r);
        }
  }
 
  void DIT(int *a, int l) {
    int *j, *k, n = 1 << l, len = 1, r, *o;
 
    for (; len != n; len <<= 1)
      for (j = a, o = root.data(); j != a + n; j += len << 1, ++o)
        for (k = j; k != j + len; ++k) {
          r = reduce(*k + k[len] - P);
          k[len] = ull(*k - k[len] + P) * *o % P;
          *k = r;
        }
  }
 
  void fft(int *a, int lgn, int d = 1) {
    if (L < lgn) prepRoot(lgn);
    int n = 1 << lgn;
    if (d == 1) DIF(a, lgn);
    else {
      DIT(a, lgn);
      reverse(a + 1, a + n);
      ull nv = P - (P - 1) / n;
      for (int i = 0; i < n; ++i) a[i] = a[i] * nv % P;
    }
  }
} ntt;
 
struct Poly {
  vector<int> a;
 
  Poly(int v = 0) : a(1) {
    if ((v %= P) < 0)
      v += P;
    a[0] = v;
  }
 
  Poly(const vector<int> &a) : a(a) {}
 
  Poly(initializer_list<int> init) : a(init) {}
 
  // Helps
  int operator[](int k) const { return k < a.size() ? a[k] : 0; }
 
  int &operator[](int k) {
    if (k >= a.size())
      a.resize(k + 1);
    return a[k];
  }
 
  int deg() const { return a.size() - 1; }
 
  void redeg(int d) { a.resize(d + 1); }
 
  Poly monic() const;
 
  Poly sunic() const;
 
  Poly slice(int d) const {
    if (d < a.size())
      return vector<int>(a.begin(), a.begin() + d + 1);
    vector<int> res(a);
    res.resize(d + 1);
    return res;
  }
 
  int *base() { return a.data(); }
 
  const int *base() const { return a.data(); }
 
  Poly println(FILE *fp) const {
    fprintf(fp, "%d", a[0]);
    for (int i = 1; i < a.size(); ++i)
      fprintf(fp, " %d", a[i]);
    fputc('\n', fp);
    return *this;
  }
 
  // Calculations
  Poly operator+(const Poly &rhs) const {
    vector<int> res(max(a.size(), rhs.a.size()));
    for (int i = 0; i < res.size(); ++i)
      if ((res[i] = operator[](i) + rhs[i]) >= P)
        res[i] -= P;
    return res;
  }
 
  Poly operator-() const {
    Poly ret(a);
    for (int i = 0; i < a.size(); ++i)
      if (ret[i])
        ret[i] = P - ret[i];
    return ret;
  }
 
  Poly operator-(const Poly &rhs) const { return operator+(-rhs); }
 
  Poly operator*(const Poly &rhs) const;
 
  Poly taylor(int k) const;
};
 
Poly zeroes(int deg) { return vector<int>(deg + 1); }
 
Poly operator "" _z(unsigned long long a) { return {0, (int) a}; }
 
Poly operator+(int v, const Poly &rhs) { return Poly(v) + rhs; }
 
Poly Poly::operator*(const Poly &rhs) const {
  int n = deg(), m = rhs.deg();
  if (n <= 10 || m <= 10 || n + m <= BRUTE_N2_LIMIT) {
    Poly ret = zeroes(n + m);
    for (int i = 0; i <= n; ++i)
      for (int j = 0; j <= m; ++j)
        ret[i + j] = (ret[i + j] + a[i] * (ll) rhs[j]) % P;
    return ret;
  }
  n += m;
  int l = 0;
  while ((1 << l) <= n)
    ++l;
  vector<int> res(1 << l), tmp(1 << l);
  memcpy(res.data(), base(), a.size() * sizeof(int));
  ntt.fft(res.data(), l, 1);
  memcpy(tmp.data(), rhs.base(), rhs.a.size() * sizeof(int));
  ntt.fft(tmp.data(), l, 1);
  for (int i = 0; i < (1 << l); ++i)
    res[i] = res[i] * (ll) tmp[i] % P;
  ntt.fft(res.data(), l, -1);
  res.resize(n + 1);
  return res;
}
 
Poly Poly::taylor(int k) const {
  int n = deg();
  Poly t = zeroes(n);
  simp.check(n);
  for (int i = 0; i <= n; ++i)
    t[n - i] = a[i] * (ll) simp.fac[i] % P;
  int pw = 1;
  Poly help = vector<int>(simp.ifac.begin(), simp.ifac.begin() + n + 1);
  for (int i = 0; i <= n; ++i) {
    help[i] = help[i] * (ll) pw % P;
    pw = pw * (ll) k % P;
  }
  t = t * help;
  for (int i = 0; i <= n; ++i)
    help[i] = t[n - i] * (ll) simp.ifac[i] % P;
  return help;
}
 
Poly stirling2(int n) {
  Poly p = zeroes(n), ne = zeroes(n);
  for (int i = 0; i <= n; ++i) p[i] = mpow(i, n) * (ll)simp.gifac(i) % P;
  for (int i = 0; i <= n; ++i) ne[i] = simp.gifac(i);
  for (int i = 1; i <= n; i += 2) ne[i] = P - ne[i];
  p = p * ne;
  vector<int> ans(n + 1);
  for (int i = 0; i <= n; ++i) ans[i] = p[i] * (ll)simp.gfac(i) % P;
  return ans;
}
 
namespace DC {
 
int N;
vector<Poly> prd, sum;
 
Poly lift(Poly a, int k) {
  a.a.insert(a.a.begin(), k, 0);
  return a;
}
 
void build(int o, int l, int r) {
  if (l == r - 1) {
    prd[o].redeg(1);
    prd[o][1] = P - simp.ginv(r);
    prd[o][0] = (P - l) * (ll)simp.ginv(r) % P;
    sum[o] = prd[o];
    return;
  }
  int mid = (l + r + 1) / 2;
  build(o << 1, l, mid);
  build(o << 1 | 1, mid, r);
  prd[o] = prd[o << 1] * prd[o << 1 | 1];
  sum[o] = prd[o << 1] * sum[o << 1 | 1] + lift(sum[o << 1], r - mid);
}
 
void pre(int n) {
  N = n;
  sum.resize(n * 4); prd.resize(n * 4);
  build(1, 0, n);
}
 
Poly input;
 
pair<Poly, Poly> solve(int o, int l, int r) {
  if (l == r - 1) {
    Poly r1 = input[r];
    return make_pair(r1 * prd[o], lift(r1, 1));
  }
  int mid = (l + r + 1) / 2;
  auto ls = solve(o << 1, l, mid), rs = solve(o << 1 | 1, mid, r);
  ls.first = ls.first + prd[o << 1] * rs.first + sum[o << 1] * rs.second;
  ls.second = ls.second + lift(rs.second, mid - l);
  return ls;
}
 
Poly solve(Poly in) {
  input = in; input.redeg(N);
  auto pr = solve(1, 0, N);
  auto ret = pr.first + pr.second;
  ret[0] = (ret[0] + input[0]) % P;
  return ret;
}
 
}
 
Poly compute(Poly coeff) {
  int n = coeff.deg();
  Poly ret = DC::solve(coeff);
  ret.redeg(n);
  reverse(ret.a.begin(), ret.a.end());
  ret = ret.taylor(P - 1);
  reverse(ret.a.begin(), ret.a.end());
  return ret;
}
 
Poly solve(int n) {
  DC::pre(n);
  auto v0 = stirling2(n), v1 = stirling2(n - 1);
  return compute(v0) + compute(v1);
}
 
int main() {
  ios::sync_with_stdio(false); cin.tie(nullptr);
 
  int n, s; cin >> n >> s;
  auto ans = solve(n - s);
  for (int i = 0; i < s; ++i) cout << "0 ";
  for (int i = n - s - 1; i >= 0; --i)
    cout << ans[i] * (ll)simp.binom(n - 1, s) % P
         << " \n"[i == 0];
 
  return 0;
}