(Analysis by Dhruv Rohatgi )

Suppose that Farmer John doesn't remove any cows from the line. Then we can simply simulate the actions of the cows in order, keeping track of which boxes of cereal are left. This yields an $O(N^2)$ solution to the original problem: for each $i$, simulate the last $N-i$ cows.

There are several ways to speed up this solution. One way is to start with $i = N$ and add cows one by one. Suppose we have solved the problem for the last $N-i$ cows: that is, if the last $N-i$ cows line up in order, we know which box each cow takes. We want to efficiently update this outcome to the outcome if instead the last $N+1-i$ cows were to line up. Thus, we need to handle the operation "add cow to front of line".

Suppose the new cow has preferences $(f, s)$. Then the new cow will certainly get cereal $f$. If the last $N-i$ cows didn't take cereal $f$, then nothing will change: all of those cows have the same outcomes after adding the new cow. But if some cow $j$ had taken cereal $f$, then after adding the new cow, $j$ no longer gets $f$. If $f$ is $j$'s second choice, then now $j$ gets nothing---and all other cows have the same outcomes. If $f$ is $j$'s first choice, and her second choice was taken by some cow earlier in line, then again $j$ gets nothing, and all other cows have the same outcome. But if $j$'s second choice was not taken by some cow earlier in line, then $j$ will take it. Unfortunately, this case may cause a cascade: we need to recurse on $j$ and figure out whether $j$ stole her second-choice cereal from someone later in line, and so forth.

At each step of the recursion, a cow $j$ can only "steal" from one later cow: the cow who originally took the cereal which $j$ is now taking. So the above algorithm doesn't blow up exponentially or anything. But each addition of a new cow seems like it could cause a recursion of depth $O(N)$, which would imply an overall time complexity of $O(N^2)$!

Fortunately, this is not the case: we can show that the sum of the recursion depths is actually $O(N)$. The reason is that every time we recurse, a cow is either kicked from first-preference to second-preference or from second-preference to nothing. Moreover, as we add cows, the reverse can never happen: a cow who was getting nothing cannot suddenly get something when we add a cow to the front of the line. It follows that the total depth of all the recursion is at most $2N$.

To implement the algorithm efficiently (with a constant-time update at each recursion step) we just need to keep track of which cow currently gets each box of cereal ($occ[pos]$ stores the index of the cow that gets cereal $pos$). The final algorithm runs in time $O(N)$. See my code below:

#include <iostream>
#include <fstream>
#include <algorithm>
using namespace std;
#define MAXN 100001

int N,M;
int f[MAXN], s[MAXN];
int occ[MAXN];
int ans[MAXN];

int main()
{
	ifstream fin("cereal.in");
	fin >> N >> M;
	for(int i=1;i<=N;i++)
		fin >> f[i] >> s[i];
	int cnt = 0;
	for(int i=N;i>=1;i--)
	{
		int j = i;
		int pos = f[i];
		while(1)
		{
			if(occ[pos] == 0)
			{
				occ[pos] = j;
				cnt++;
				break;
			}
			else if(occ[pos] < j)
				break;
			else
			{
				int k = occ[pos];
				occ[pos] = j;
				if(pos == s[k])
					break;
				j = k;
				pos = s[k];
			}
		}
		ans[i] = cnt;
	}
	ofstream fout ("cereal.out");
	for(int i=1;i<=N;i++)
		fout << ans[i] << '\n';
}