We will hold AtCoder Grand Contest 067. This contest counts for GP30 scores.
- Contest URL: https://atcoder.jp/contests/agc067
- Start Time: http://www.timeanddate.com/worldclock/fixedtime.html?iso=20240818T2100&p1=248
- Duration: 180 minutes
- Number of Tasks: 5
- Writer: Kubic, JohnVictor
- Tester: HIR180, orzdevinwang
- Rated range: 2000 -
The point values will be 600-1100-1100-1500-1900.
We are looking forward to your participation!
Why doesn't Atcoder official announce this? Just out of curiosity :)
I believe ARCs and AGCs are announced by Maroonrk, while ABCs are announced by atcoder_official
Score distribution be like:
A:
problem I can actually solve in 3 hours
B and C:
problem I can only upsolve after editorial is published
D:
problem that I wouldn't be able to solve even with a help of editorial
E:
Stanford-level PhD research
I have seen LGMs skipping A
The A of AGCs are smarter than me. Maybe I can actually struggle on A for 3 hours >_<
Same here. Still have no idea how to solve it.
UPD: Just have read the editorial and it describes exactly the solution I've submitted in the last few minutes of contest. Looks like I have a bug somewhere in my implementation.
UPD: Forgot return call in DFS. Stupid mistake, but at least I found the right observation in 3 hours. Not that bad.
1100 for B :(, I guess it would be at least 2600+
Is the 2000+ rated range intentional? Most other AGC seem to be 1200+, other than world finals stuff
It's intended, see This post
they changed it
ARCs are now 1200+ as well
Rated Range is $$$ [2000, inf) $$$ ?
Problem B was too hard.
The decision version of problem D was 1348F - Phoenix and Memory. Knowing this was only a little bit helpful though, I ended up with an $$$O(n^3)$$$ solution which is disappointing.
My best AGC performance in over 3 years and of course it had to be the one where I chose unrated registration because half an hour before the contest electricity in my whole building (and presumably even in the whole town) went down and had no idea when it will come back :]
That's disappointing :( But at least you got some GP30 points towards AWTF rating!
Thanks for the round! the conclusion of A is concise and the prove is not hard, I like it.
(though I can't solve anything after that qq)
It seems I have a completely different solution to D. Let me describe it. (As Radewoosh pointed out to me, maybe it's not completely different, but I think the insight is different enough for it to be worth describing anyway)
Key lemma: If the matching is unique then there exists $$$i$$$ such that it is contained in only one interval
It follows from the following lemma that was previously known to me: Let $$$G$$$ be a bipartite graph, where vertices on one side have degree $$$\ge d$$$. If $$$G$$$ has at least one matching then it has at least $$$d!$$$ of them.
The condition from this lemma has to be satisfied and after removing that interval and element, it has to be satisfied recursively in both halves, so we get an equivalent condition determining whether a set of intervals is good.
Now, how do we count them? With inclusion-exclusion principle. Let's say we are given $$$k$$$ points $$$p_1, \ldots, p_k$$$ that are supposed to be covered by only one interval. Obviously, these intervals have to be distinct. Moreover, the left end of interval containing $$$p_c$$$ is an arbitrary number from $$$[p_{c-1}+1, p_c]$$$ (or $$$[1, p_c]$$$ if $$$c=1$$$), similarly for the right end. For the remaining intervals in between, we can solve each of them in an already computed through dp number of ways. We multiply all of these and get a value to sum for the specific set of uniquely covered points. All of these can be efficiently summed through dp in $$$O(n^2)$$$.
(C[a] is the solution for a elements, dp[a] is the corresponding partial value for sets of uniquely covered points, where the last one was a)
How to prove this?
We will use a crucial lemma:
Lemma: In every bipartite graph with at least one perfect matching, there is a vertex on the left side such that all of its edges belong to some perfect matching.
The proof of our final lemma is an obvious consequence of this one through induction.
How do we prove this lemma?
Let G be our bipartite graph with its sides denoted as L and R and M be any of its perfect matchings. Let's create a directed graph D on the same vertex set as G in the following way. For every edge lr, where $$$l \in L$$$ put into D a directed edge from l to r. Moreover, if $$$lr \in M$$$, then additionally put it in the reverse direction as well.
Now, the edge ab belongs to some perfect matching if and only if a and b belong to the same strongly connected component of D (easy exercise on alternating cycles). Let's take a DAG of SCCs of D and any terminal vertex of it, i.e. a SCC such that there is no edge outgoing from it. Any vertex from L from this SCC satisfies our conditions (and there is one since for every edge in M, its both ends belong to the same SCC and M covers all vertices)
My approach is roughly the same. This is reminiscent of the classical inclusion-exclusion recurrence formula for counting DAGs.
Is the tutorial for C unfinished? It seems like it starts from the middle of a full editorial.
I don't see what you mean. Seems finished to me?
Below are some details based on my understanding.
$$$f(n):=a_n$$$
If there exists $$$x<y$$$ such that $$$f(x)=f(y)$$$,then $$$f(y)|f(x) \Rightarrow y|x$$$, contradiction.
Let $$$P$$$ be the set of all prime numbers.
$$$x|y \iff \forall p\in P: v_p(x)\leq v_p(y)$$$
$$$n|m \iff f(n)|f(m) \Rightarrow v_p(f(n))\leq v_p(f(m))$$$
$$$f(2u)/f(u)\geq 2$$$, so it contains some prime factor at least $$$p_1=2$$$.
If $$$f(2^{n+1})=b_n\cdot f(2^n)\neq 2\cdot f(2^n)$$$, then $$$b_n>2$$$.
Since $$$f(x)\leq Cx$$$, there are only a finite number of $$$b_n\neq 2$$$.
So $$$\exists n_1: \forall n>n_1 : f(2^{n+1})=2\cdot f(2^n)$$$.
So $$$\exists N_1,c_1: \forall n>N_1 : f(2^{n})=2^{n-c_1}\cdot f(2^{c_1})$$$.
If for $$$p_1,p_2,...,p_{k-1}$$$ we have
(1) $$$f(p_i\cdot u)/f(u)$$$ contains some prime factor $$$\geq p_i$$$, and
(2) $$$\exists N_i,c_i$$$ such that $$$f(p_i^n)=p_i^{n-c_i}\cdot f(p_i^{c_i})$$$. ($$$\star$$$)
Let $$$X$$$ be a product of powers of some of $$$p_1,...,p_{k-1}$$$,
so $$$p_k\cdot u \not| X\cdot u \Rightarrow f(p_k\cdot u)\not| f(X\cdot u) \Rightarrow \frac{f(p_k\cdot u)}{f(u)}\not| \frac{f(X\cdot u)}{f(u)}$$$
We can pick $$$p_i^{t_i}$$$ for $$$X$$$ s.t. $$$v_{p_i}(f(p_i^{t_i}))>v_{p_i}(f(p_k\cdot u))$$$.
Then $$$\forall i<k: v_{p_i}(f(X\cdot u))\geq v_{p_i}(f(p_i^{t_i}))\geq v_{p_i}(f(p_k\cdot u))$$$
Then $$$\forall i<k: v_{p_i}(\frac{f(X\cdot u)}{f(u)})\geq v_{p_i}(\frac{f(p_k\cdot u)}{f(u)})$$$
So as $$$\frac{f(p_k\cdot u)}{f(u)}\not| \frac{f(X\cdot u)}{f(u)}$$$, there must exists an $$$i\geq k$$$ s.t. $$$v_{p_i}(\frac{f(p_k\cdot u)}{f(u)})>v_{p_i}(\frac{f(X\cdot u)}{f(u)})$$$
So $$$v_{p_i}(\frac{f(p_k\cdot u)}{f(u)})\geq 1$$$, $$$\frac{f(p_k\cdot u)}{f(u)}$$$ contains some prime factor at least $$$p_k$$$.
Then $$$f(p_k^{n+1})/f(p_k^n)\geq p_k$$$, due to the restriction of $$$C$$$, only a finite number of $$$n$$$ can achieve $$$f(p_k^{n+1})/f(p_k^n)>p_k$$$.
So $$$\exists n_k: \forall n>n_k: f(p_k^{n+1})=p_k\cdot f(p_k^n)$$$.
So $$$\exists N_k,c_k$$$ such that $$$f(p_k^n)=p_k^{n-c_k}\cdot f(p_k^{c_k})$$$.
Now we have completed the induction ($$$\star$$$). We have:
(1) For all $$$p\in P: f(pu)/f(u)$$$ contains some prime factor $$$\geq p$$$, and
(2) $$$\forall p_i: \exists N_i,c_i$$$ such that $$$f(p_i^n)=p_i^{n-c_i}\cdot f(p_i^{c_i})$$$.
Then we want to prove: If $$$v_p(x)<v_p(y)$$$ then $$$v_p(f(x))<v_p(f(y))$$$. (*)
Consider what happens if $$$v_p(t)=0$$$ in the $$$f(p\cdot u)=t\cdot f(u)$$$.
Pick all the prime factors of $$$t\cdot f(u)$$$ except $$$p$$$ and let $$$X$$$ be a product of powers of them.
For $$$p_i$$$ we can pick $$$p_i^{t_i}$$$ for $$$X$$$ s.t. $$$v_{p_i}(f(p_i^{t_i}))\geq v_{p_i}(t\cdot f(u))$$$.
Then $$$v_{p_i}(f(X\cdot u))\geq v_{p_i}(f(p_i^{t_i}))\geq v_{p_i}(t\cdot f(u))$$$.
Now $$$f(p\cdot u)=t\cdot f(u)|f(X\cdot u)$$$, but $$$p\cdot u\not| X\cdot u$$$, contradiction.
So $$$p|\frac{f(p\cdot u)}{f(u)}$$$.
If there exists $$$x$$$ and $$$y$$$ s.t. $$$v_p(x)<v_p(y)$$$ but $$$v_p(f(x))\geq v_p(f(y))$$$.
Take the min $$$x$$$ and $$$y$$$.
$$$p^{v_p(x)+1}|y \Rightarrow f(p^{v_p(x)+1})|f(y)\Rightarrow v_p(f(p^{v_p(x)+1}))\leq v_p(f(y))\leq v_p(f(x))$$$.
So $$$y=p^{v_p(x)+1}$$$. Let $$$a=v_p(x)$$$, then $$$y=p^{a+1}, x=p^a\cdot B$$$.
Pick prime factors of $$$f(y)$$$ except $$$p$$$ to make an $$$X$$$ s.t. $$$f(y)|f(x*X)$$$, but $$$y=p^{a+1}\not|x*X$$$, contradiction.
Proved (*).
So we can check $$$v_p(i)$$$ and $$$v_p(A_i)$$$ for each $$$p\leq N$$$, and make sure $$$v_p(f(x))<v_p(f(y))$$$ when $$$v_p(x)<v_p(y)$$$. This guarantees the case where $$$n\not|m$$$. And check $$$f(n)|f(m)$$$ when $$$n|m$$$.
For the infinite sequence, let $$$v_p(f(n))=\max\lbrace v_p(f(x))+v_p(n)-v_p(x): v_p(x)\leq v_p(n)\rbrace$$$. It is good.
Despite not being able to solve C (I lacked to the intuition to see that we can't have $$$f(3^k) \sim 2^k$$$ for example), I love this problem a lot and wish we could have more of this "convoluted math-paper-proof-style journey" problem type. I guess here in particular you could get pretty far by guessing (I'm curious if anyone managed to do that? I wouldn't know), but in spirit this feels like refreshing new ground to me.
This problem had a strong vibe of "IMO P3/6" style to me :p
I proved (or think I proved) everything in a pretty straightforward way as only condition 2 at first.
If we ignore condition 1, then it turns out you need sequences $$$S_p = (r_{p^i}, e_{p^i})$$$ where $$$p$$$ is prime, with the following meaning: $$$A_j$$$ is divisible by $$$(r^e)_{p^i}$$$ iff $$$j$$$ is divisible by $$$p^i$$$. Very importantly, all the sets $$${S_p}$$$ must be disjoint. Of course, for the same value of $$$r$$$, the exponents must be increasing, while there's no extra connection between $$$e$$$ of different $$$r$$$.
Condition 1 just says that all the values (there's at least one) occurring in $$$S_p$$$ an infinite number of times must be $$$\le p$$$. With the disjointness, that means $$$r_{p^i} = p$$$ and checking the condition is super simple.
I found necessary condition $$$j|i \to (\frac{i}{j}|\frac{a_i}{a_j})$$$, and guessed solution from that: just expanded it for all pairs $$$(i,j)$$$ to be like $$$\frac{i}{gcd(i,j)}|\frac{a_i}{gcd(a_i,a_j)}$$$, and it turned out to be sufficient for some reason. submission
The data of the problem A is weak. As you see,I can passed the problem with the wrong code. Here is Submissions.
Is the answer 'no' to all?
Your linked code appears to be correct.
As you see,if n>4000,I will skip the m edges that will be read in and get the wrong n and m.
But it passed.
Ah, I see what you're referring to now. Maybe they only have single/weak testcases for the cases where n>4000.
:D
Why are the editorials so short :( i feel a lot of details have been skipped.
Can anyone give us a detailed introduction of problem B. I've read the editorial but don't understand what exactly should be done?
The final value(color) $$$a_i$$$ of each position is determined by the last operation interval contains $$$i$$$.
For each array $$$a$$$, it can be generated if there exists an operation sequence $$$(p:w) = [L_{p_1},R_{p_1}]:w_1, [L_{p_2},R_{p_2}]:w_2, ..., [L_{p_M},R_{p_M}]:w_M$$$ to generate it.
Try to select the interval for the operation sequence(which generates $$$a$$$) from the last to the first:
For operation $$$M$$$, final colors of elements in the chosen interval $$$[L_{p_M},R_{p_M}]$$$ must be the same, we choose $$$w_M$$$ be this color. If $$$a$$$ can be generated, then there exists an operation sequence $$$(q,w') = [L_{q_1},R_{q_1}]:w_1', [L_{q_2},R_{q_2}]:w_2', ..., [L_{q_M},R_{q_M}]:w_M'$$$. If it is not end with $$$[L_{p_M},R_{p_M}]$$$, we move the $$$[L_{q_t},R_{q_t}]=[L_{p_M},R_{p_M}]:w_t'=w_M$$$ to the end, then the final result outside the interval $$$[L_{q_t},R_{q_t}]$$$ will not change, and inside the interval is just the color we need. So there exists an operation sequence end with $$$[L_{p_M},R_{p_M}]$$$.
For operation $$$j$$$, elements in the chosen interval $$$[L_{p_j},R_{p_j}]$$$ that haven't been determined(covered) by operation $$$>j$$$ must be the same. We can pick any interval satisfies this for operation $$$j$$$. And if we know there exists an operation sequence $$$(q:w')$$$ end with $$$[L_{p_{j+1}},R_{p_{j+1}}]:w_{j+1}, ..., [L_{p_M},R_{p_M}]:w_M$$$, move $$$[L_{q_t},R_{q_t}]=[L_{p_j},R_{p_j}]$$$ to $$$j$$$ will not change $$$a$$$, we can deduce that there exists an operation sequence $$$(q:w')$$$ end with $$$[L_{p_{j}},R_{p_{j}}]:w_{j}, ..., [L_{p_M},R_{p_M}]:w_M$$$.
So the sequence $$$a$$$ can be generated if and only if it can be covered "greedily". That is, any current choice in the reverse operation sequence will not affect whether $$$a$$$ can be generated.
If an interval doesn't cover new elements, we can ignore this operation. So we can just chooce the "valid" interval that some new element is going to be determined by it.
Consider the last valid interval $$$[l,r]$$$ in the reverse operation sequence: Before it works, we have covered $$$[l_1,r_1],[l_2,r_2],...,[l_k,r_k]$$$. Then $$$[l,r]$$$ covers all the gaps so that $$$[1,n]$$$ is completely covered. We can cover each interval independently since covering $$$[l_i,r_i]$$$ only needs intervals within it.
Let $$$f[l][r]$$$ be the number of different subarrays $$$a[l..r]$$$ that can be generated by intervals $$$\subseteq [l,r]$$$. Then $$$f[1][n]$$$ will be the answer.
For any $$$a[l..r]\in C^{r-l+1}$$$, cover it greedily until no interval $$$\subseteq [l,r]$$$ is valid. Assume we end up with $$$[l_1,r_1],[l_2,r_2],...,[l_k,r_k]$$$ covered(The interval set is uniquely determined by $$$a[l..r]$$$ as we can move the curent operation interval in the other operation sequence to the same position of our selected operation sequence each time and finally make two operation sequence consistent). Then $$$f[l][r]$$$ is the number of $$$a[l..r]$$$ end up with $$$[l..r]$$$. All other interval sets have some element uncovered. Let $$$g[l][x][r]$$$ be the number of subarrays $$$a[l..r]$$$ that if we cover it greedily by intervals $$$\subseteq [l,r)$$$, then $$$[x,r]$$$ is the widest range with the right endpoint $$$r$$$ that all uncovered elements in it are the same.
To calculate $$$g[l][x][r]$$$, consider the nearest uncovered element $$$a[i]$$$ on the left of $$$a[r]$$$:
(1) If $$$x\leq i< r$$$, then $$$a[i]=a[r]$$$. Extend $$$[l..x..i]$$$ to $$$[l..x..r]$$$. There are $$$g[l][x][i]*f[i+1][r-1]$$$ extended subarrays. And if there exists an interval $$$[L_t,R_t]\subseteq [l,r)$$$ covers some new element, it must covers $$$i$$$. Then $$$L_t\in [x,i]$$$ as $$$a[x-1]\neq a[i]$$$. Check whether there exists $$$[L_t,R_t]\in [x,i]\times[i,r)$$$ to determine whether these subarrays should be counted in $$$g[l][x][r]$$$;
(2) If $$$l\leq i<x$$$, then $$$i=x-1$$$. Extend $$$[l..j..x-1]$$$ to $$$[l..x..r]$$$. There are $$$g[l][j][x-1=i]*f[x=i+1][r-1]*(C-1)$$$ extended subarrays. Check whether there exists $$$[L_t,R_t]\in [j,i]\times[i,r)$$$;
(3) If no such $$$i$$$, then $$$x=l$$$. $$$g[l][l][r]+=f[l][r-1]*C$$$.
To calculate $$$f[l][r]$$$, just substract all the $$$a[l..r]$$$ that counldn't be covered completely from $$$C^{r-l+1}$$$. Enumerate the rightmost uncovered position $$$i$$$, extend $$$[l..x..i]$$$ to $$$[l..x..r]$$$. There are $$$g[l][x][i]*f[i+1][r]$$$ extended subarrays. Check whether there exists $$$[L_t,R_t]\in [x,i]\times[i,r]$$$.