Label
利用欧拉函数性质普通莫比乌斯反演+对数论分块求和的筛法优化
Description
T T T组 n ( T , n ≤ 5 × 1 0 5 ) n(T,n\leq5\times10^5) n(T,n≤5×105),求
∑ i = 1 n ∑ j = i n ∑ d ∣ g c d ( i , j ) i j g c d ( i d , j d ) \sum_{i=1}^{n}\sum_{j=i}^{n}\sum_{d|gcd(i,j)}\frac{ij}{gcd(\frac{i}{d},\frac{j}{d})} i=1∑nj=i∑nd∣gcd(i,j)∑gcd(di,dj)ij
Solution
∑ i = 1 n ∑ j = i n ∑ d ∣ ( i , j ) i j ( i d , j d ) \sum_{i=1}^{n}\sum_{j=i}^{n}\sum_{d|(i,j)}\frac{ij}{(\frac{i}{d},\frac{j}{d})} i=1∑nj=i∑nd∣(i,j)∑(di,dj)ij
= ∑ i = 1 n ∑ j = 1 i ∑ d ∣ ( i , j ) i j ( i d , j d ) =\sum_{i=1}^{n}\sum_{j=1}^{i}\sum_{d|(i,j)}\frac{ij}{(\frac{i}{d},\frac{j}{d})} =i=1∑nj=1∑id∣(i,j)∑(di,dj)ij
(由于求和下界是变量的话,会涉及到奇奇怪怪的上取整,故此处作此变换)
= ∑ d = 1 n ∑ i = 1 n ∑ j = 1 i [ d ∣ ( i , j ) ] i j ( i d , j d ) =\sum_{d=1}^{n}\sum_{i=1}^{n}\sum_{j=1}^{i}[d|(i,j)]\frac{ij}{(\frac{i}{d},\frac{j}{d})} =d=1∑ni=1∑nj=1∑i[d∣(i,j)](di,dj)ij
= ∑ d = 1 n d 2 ∑ i = 1 ⌊ n d ⌋ ∑ j = 1 i i j ( i , j ) =\sum_{d=1}^{n}d^2\sum_{i=1}^{\lfloor\frac{n}{d}\rfloor}\sum_{j=1}^{i}\frac{ij}{(i,j)} =d=1∑nd2i=1∑⌊dn⌋j=1∑i(i,j)ij
= ∑ d = 1 n d 2 ∑ i = 1 ⌊ n d ⌋ ∑ j = 1 i l c m ( j , i ) ( 1 ) =\sum_{d=1}^{n}d^2\sum_{i=1}^{\lfloor\frac{n}{d}\rfloor}\sum_{j=1}^{i}lcm(j,i)(1) =d=1∑nd2i=1∑⌊dn⌋j=1∑ilcm(j,i)(1)
为了化简求和号,我们逐层化简:
∑ i = 1 n l c m ( i , n ) \sum_{i=1}^{n}lcm(i,n) i=1∑nlcm(i,n)
(这个式子加上这个题的数据范围就是spoj5971 LCMSUM,但spoj目前交不了题,故将此题推导过程顺便放在此处)
= ∑ i = 1 n n i ( i , n ) = n ∑ d = 1 n 1 d ∑ i = 1 n i [ ( i , n ) = d ] =\sum_{i=1}^{n}\frac{ni}{(i,n)}=n\sum_{d=1}^{n}\frac{1}{d}\sum_{i=1}^{n}i[(i,n)=d] =i=1∑n(i,n)ni=nd=1∑nd1i=1∑ni[(i,n)=d]
= n ∑ d ∣ n ∑ i = 1 n d i [ ( i , n d ) = 1 ] = n ∑ d ∣ n ∑ i = 1 d i [ ( i , d ) = 1 ] =n\sum_{d|n}\sum_{i=1}^{\frac{n}{d}}i[(i,\frac{n}{d})=1]=n\sum_{d|n}\sum_{i=1}^{d}i[(i,d)=1] =nd∣n∑i=1∑dni[(i,dn)=1]=nd∣n∑i=1∑di[(i,d)=1]
由于欧拉函数的一个经典性质: ∑ i = 1 n i [ ( i , n ) = 1 ] = n φ ( n ) + [ n = 1 ] 2 \sum_{i=1}^{n}i[(i,n)=1]=\frac{n\varphi(n)+[n=1]}{2} ∑i=1ni[(i,n)=1]=2nφ(n)+[n=1],故:
n ∑ d ∣ n ∑ i = 1 d i [ ( i , d ) = 1 ] = n ∑ d ∣ n d φ ( d ) + [ d = 1 ] 2 = n ( 1 2 + ∑ d ∣ n d φ ( d ) 2 ) n\sum_{d|n}\sum_{i=1}^{d}i[(i,d)=1]=n\sum_{d|n}\frac{d\varphi(d)+[d=1]}{2}=n(\frac{1}{2}+\sum_{d|n}\frac{d\varphi(d)}{2}) nd∣n∑i=1∑di[(i,d)=1]=nd∣n∑2dφ(d)+[d=1]=n(21+d∣n∑2dφ(d))
= n ( 1 + ∑ d ∣ n d φ ( d ) ) 2 =\frac{n(1+\sum_{d|n}d\varphi(d))}{2} =2n(1+∑d∣ndφ(d))
设 g ( n ) = ∑ d ∣ n d φ ( d ) , f ( n ) = ∑ i = 1 n l c m ( i , n ) , S ( n ) = ∑ i = 1 n f ( i ) g(n)=\sum_{d|n}d\varphi(d),f(n)=\sum_{i=1}^{n}lcm(i,n),S(n)=\sum_{i=1}^{n}f(i) g(n)=∑d∣ndφ(d),f(n)=∑i=1nlcm(i,n),S(n)=∑i=1nf(i),则:
f ( n ) = n ( 1 + g ( n ) ) 2 f(n)=\frac{n(1+g(n))}{2} f(n)=2n(1+g(n))
( 1 ) = ∑ d = 1 n d 2 ∑ i = 1 ⌊ n d ⌋ f ( i ) = ∑ d = 1 n d 2 S ( ⌊ n d ⌋ ) (1)=\sum_{d=1}^{n}d^2\sum_{i=1}^{\lfloor\frac{n}{d}\rfloor}f(i)=\sum_{d=1}^{n}d^2S(\lfloor\frac{n}{d}\rfloor) (1)=d=1∑nd2i=1∑⌊dn⌋f(i)=d=1∑nd2S(⌊dn⌋)
由于 g ( n ) g(n) g(n)也是积性函数,故 g ( n ) g(n) g(n)既可以通过枚举 d d d从而 O ( n l o g n ) O(nlogn) O(nlogn)求出,也可以利用线性筛 O ( n ) O(n) O(n)求出(如何 O ( n ) O(n) O(n)求出:推导过程较麻烦,见另文)。
由于求出 g ( 1 ) ∼ g ( n ) g(1)\sim g(n) g(1)∼g(n)后,我们可以 O ( n ) O(n) O(n)预处理出 f ( 1 ) ∼ f ( n ) f(1)\sim f(n) f(1)∼f(n)与 s ( 1 ) ∼ s ( n ) s(1)\sim s(n) s(1)∼s(n),故 ( 1 ) (1) (1)式的取值集合可以通过数论分块 O ( n n ) O(n\sqrt n) O(nn)求出,总时间复杂度为 O ( n n + n l o g n ) O(n\sqrt n+nlogn) O(nn+nlogn),对于 5 e 5 5e5 5e5的数据规模来讲会TLE,考虑优化。
对于 ( 1 ) (1) (1)式 a n s ( n ) = ∑ d = 1 n d 2 S ( ⌊ n d ⌋ ) ans(n)=\sum_{d=1}^{n}d^2S(\lfloor\frac{n}{d}\rfloor) ans(n)=∑d=1nd2S(⌊dn⌋),我们仍然枚举 d d d,考虑每一个 d d d可能对 a n s ans ans表的哪些值有贡献:我们在枚举 d d d的同时枚举所有可能的 ⌊ n d ⌋ \lfloor\frac{n}{d}\rfloor ⌊dn⌋的取值 i i i,显然应有 i d ≤ m a x n id\leq max{n} id≤maxn,此时满足 ⌊ n d ⌋ = i \lfloor\frac{n}{d}\rfloor=i ⌊dn⌋=i的 n n n的取值显然为 [ i d , ( i + 1 ) d ) [id,(i+1)d) [id,(i+1)d)。所以,对于任意的 d d d和对应的任意 i i i, d d d对于 a n s ( i d ) ∼ a n s ( ( i + 1 ) d − 1 ) ans(id)\sim ans((i+1)d-1) ans(id)∼ans((i+1)d−1)的值会有 d 2 S ( i ) d^2S(i) d2S(i)的贡献。
枚举 d d d与 i ( i d ≤ m a x n ) i(id\leq max{n}) i(id≤maxn)的时间复杂度为 O ( n l o g n ) O(nlogn) O(nlogn),而由于每对 d , i d,i d,i对ans表的贡献是对连续区间的值的贡献,相当于区间加操作。由于我们执行 O ( n l o g n ) O(nlogn) O(nlogn)次区间加操作但中间过程中并不需要查询 a n s ( i ) ans(i) ans(i)的值,故我们利用差分数组维护该操作即可。
算法时间总复杂度: O ( n l o g n ) O(nlogn) O(nlogn)
Code
#include<cstdio>
#include<cmath>
#include<iostream>
#define ri register int
#define ll long long
using namespace std;
const int MAXN=5e5+20;
int T,prime[MAXN],cnt,q[MAXN],N;
ll phi[MAXN],f[MAXN],g[MAXN],ans[MAXN];
bool vis[MAXN];
void EulaSieve(int n)
{
phi[1]=1,vis[1]=true;
for(ri i=2;i<=n;++i)
{
if(!vis[i]) prime[++cnt]=i,phi[i]=(ll)i-1;
for(ri j=1;j<=cnt&&i*prime[j]<=n;++j)
{
vis[i*prime[j]]=true;
if(i%prime[j]==0) phi[i*prime[j]]=phi[i]*(ll)prime[j];
else phi[i*prime[j]]=phi[i]*phi[prime[j]];
if(i%prime[j]==0) break;
}
}
}
void EratSieve(int n)
{
for(ri i=1;i<=n;++i)
for(ri j=i;j<=n;j+=i)
g[j]+=(ll)i*phi[i];
for(ri i=1;i<=n;++i)
f[i]=(ll)i*(g[i]+1)/2LL,f[i]=f[i-1]+f[i];
for(ri d=1;d<=n;++d)
for(ri i=1;i*d<=n;++i)
{
ri sj=(i+1)*d;
if(sj>n) sj=n+1;
ans[i*d]+=(ll)d*(ll)d*f[i],ans[sj]-=(ll)d*(ll)d*f[i];
}
for(ri i=1;i<=n;++i)
ans[i]=ans[i]+ans[i-1];
}
int main()
{
std::ios::sync_with_stdio(false);
cin>>T;
for(ri i=1;i<=T;++i)
{
cin>>q[i];
N=max(q[i],N);
}
EulaSieve(N),EratSieve(N);
for(ri i=1;i<=T;++i)
cout<<"Case #"<<i<<": "<<(unsigned int)ans[q[i]]<<'\n';
return 0;
}