Hello Codeforces! So my friend Haidora thought that in problem C yesterday it asked for the number of beautiful sets of all lengths (lol) and we were talking now about a solution for this problem and came up with an $$$O(rlogr)$$$ DP solution.
Formal Statement
A set of positive integers $$$S$$$ is called beautiful if, for every two integers $$$x$$$ and $$$y$$$ from this set, either $$$x$$$ divides $$$y$$$ or $$$y$$$ divides $$$x$$$ (or both).
You are given two integers $$$l$$$ and $$$r$$$ . Consider all beautiful sets consisting of integers not less than $$$l$$$ and not greater than $$$r$$$: You have to print their number modulo $$$998244353$$$
Solution
Let $$$dp_x$$$ be the number of beautiful sets we can build such that the minimum number in these sets is $$$x$$$. Now, how do we find $$$dp_x$$$? Let's iterate through all numbers $$$i$$$ such that $$$xi \le r$$$, which means that we'll go through all multiples of $$$x$$$ as long as they are less than $$$r$$$. Now let's add their answers to our $$$dp_x$$$. Formally:
$$$dp_x = 1 + \sum_{i=2}^{xi \le r} dp_{xi} $$$ Now our answer would be summing them all up, because we want all sets, so for all possible minimums we want to know the number of sets possible using it. Formally: $$$ans = \sum_{i=l}^{i \le r} dp_{i} $$$
Note that the $$$1+$$$ is there because choosing $$$x$$$ alone in a set is also valid.
Time complexity
Why is it $$$O(rlogr)$$$? For each $$$x$$$ you are iterating through all number $$$xi \le r$$$ which is $$$r \sum_{i=l}^{i \le r} dp_{1/i} $$$ which is close to $$$O(rlogr)$$$ Thanks to NaimSS for correcting my time complexity.
Implementation
I used recursive DP:
#include <bits/stdc++.h>
using namespace std;
const int N=1e7+5,M=998244353;
int dp[N],l,r;
int rec(int x){
if (x>r) return 0;
if (dp[x]) return dp[x];
dp[x]=1;
for (int i=2;i*x<=r;i++) dp[x]=(dp[x]+rec(x*i))%M;
return dp[x];
}
int main(){
ios_base::sync_with_stdio(false);cin.tie(NULL);cout.tie(NULL);
cin>>l>>r;
int ans=0;
for (int i=l;i<=r;i++) ans=(ans+rec(i))%M;
cout<<ans<<endl;
return 0;
}
This code runs less than 100ms on C++20(64) when $$$l = 1$$$ and $$$r = 10^6$$$ , and less than 2000ms when $$$l = 1$$$ and $$$r = 10^7$$$. Tried on Codeforces custom test.
Iterative DP by Farsh_LE:
#include <bits/stdc++.h>
using namespace std;
const int N=1e7+5,M=998244353;
int dp[N],l,r;
int main(){
ios_base::sync_with_stdio(false);cin.tie(NULL);cout.tie(NULL);
cin>>l>>r;
int ans=0;
for (int x=r;x>=l;x--){
dp[x]=1;
for (int i=2;x*i<=r;i++) dp[x]=(dp[x]+dp[x*i])%M;
ans=(ans+dp[x])%M;
}cout<<ans<<endl;
return 0;
}
Nearly same speed as recursive, a bit faster though.
Final Notes
This is my first blog of this kind, so feel free to share your opinions as it will help me and others for the future. Also, please point out any mistakes you find. And I have a few questions: Could there be a faster solution? Also if there is an iterative DP solution please share it, because I couldn't find any way to do it iteratively. Thanks!
$$$ T=20000 $$$ :- Am i joke to you?
Edit:- i thought it was the same problem but it's a bit different.
I forgot to mention, this version can't have this much testcases :(. Unless there's a faster solution...
You just have to use prefix sums.
How?
Let sum[i] = (rec(1) + rec(2) + ... + rec(i)) % MOD. Now the answer for each query (l, r) is (sum[r] — sum[l-1] + MOD) % MOD.
So this way the problem can be solved using precomputation. Orz
I just tried. It doesn't work. It also logically is false because when you compute p[i]=rec(1)+rec(2)...rec(i) you're computing them for the maximum possible r, so summing up all number answers for the maximum r won't get a correct result when r differs. I hope you understand.
I guess this is an iterative code :
for (int i=r;i>=l;i--) { dp[i]=1;
Any way, good job 3bito , proud of you dude
very nice!
Thanks man! I'll include the implementation in the blog later.
This seems O(NlogN), or rather O(RlogR). You are basically summing $$$\sum_{x=l}^{r} (r/x) = r \sum_{x=l}^{r} \frac{1}{x}$$$. This is known as the [harmonic Sum](https://en.wikipedia.org/wiki/Harmonic_series_(mathematics)) and it grows in O(logN). For it to be O(loglogN) you would need to only iterate over the primes, as $$$\sum \frac{1}{p}$$$ growns in O(loglogN) (reference).
Ok I get it. Thanks sir.
in your rec, you can break early when ur rec(x*i) returns 0, furthermore you only need to recurse on x*2 and x*3; however there are just too many test cases that it will not run in time ):
also your first recursive calls can be [l, 2*l)
this was the sol that I came up with but it was too slow D:
I think you haven't read carefully. This is another version of the problem where you have to find all beautiful aets of all lengths. Also there will never be a return 0 because for each x I'm iterating through all i such that xi<=r
Auto comment: topic has been updated by Abito (previous revision, new revision, compare).
My math solution ran in 140ms and 4KB. The time complexity is O(log r) and space complexity is O(1).
...
Submission link: https://codeforces.me/contest/1796/submission/195535532
Please read the blog carefully. I am talking about another version of the problem.