\(\mathcal{P}\text{ortal.}\)
首先想到要把 \(\varphi(ij)\) 拆开,这里有个公式
\[\varphi(ij)=\dfrac{\varphi(i)\varphi(j)\gcd(i,j)}{\varphi(\gcd(i,j))} \]考虑证明,有
\[\begin{aligned} \varphi(i)\varphi(j) &= i\prod\limits_{p|i,p\in \mathtt{prime}}\frac{p-1}{p}\cdot j \prod\limits_{p|j,p\in\mathtt{prime}}\frac{p-1}{p}\\ &= ij\prod\limits_{p|ij,p\in\mathtt{prime}}\frac{p-1}{p}\cdot \prod\limits_{p|\gcd(i,j),p\in\mathtt{prime}}\frac{p-1}{p} \end{aligned} \]也就是 \(i,j\) 的并加上 \(i,j\) 的交。 接下来进入推柿子环节
\[\begin{aligned} \sum\limits_{i=1}^{n}\sum\limits_{j=1}^{m} \varphi(ij) &= \sum\limits_{i=1}^{n}\sum\limits_{j=1}^{m} \frac{\varphi(i)\varphi(j)\gcd(i,j)}{\varphi(\gcd(i,j))} \\&= \sum\limits_{d=1}^{n}\sum\limits_{i=1}^{n}\sum\limits_{j=1}^{m} \frac{\varphi(i)\varphi(j)d[\gcd(i,j)=d]}{\varphi(d)} \\&= \sum\limits_{d=1}^{n}\frac{d}{\varphi(d)}\sum\limits_{i=1}^{n}\sum\limits_{j=1}^{m }\varphi(i)\varphi(j)[\gcd(i,j)=d] \\&= \sum\limits_{d=1}^{n}\frac{d}{\varphi(d)}\sum_{t=1}^{n/d}\mu(t)\cdot \sum\limits_{i=1}^{n/(dt)}\sum\limits_{j=1}^{m/(dt)}\varphi(idt)\varphi(jdt) \\&= \sum\limits_{k=1}^{n}\sum\limits_{d|k} \frac{d\cdot\mu(k/d)}{\varphi(d)} \sum\limits_{i=1}^{n/k }\varphi(ik)\sum\limits_{j=1}^{m/k }\varphi(jk) \end{aligned} \]记 \(\displaystyle f(k)=\sum\limits_{d|k} \frac{d\cdot\mu(k/d)}{\varphi(d)},g(k,n)=\sum\limits_{i=1}^{n}\varphi(ik)\),容易发现这两个函数都容易在调和级数的复杂度内预处理出来。然而 \(\mathcal O(nT)\) 仍然是不现实的。如果是 CF 我就直接上了。
考虑怎么优化,还是将优化重点放在整除分块上
\[\sum_{i=1}^n\sum_{j=1}^m\varphi(ij)=\sum_{k=1}^nf(k)\cdot g(k,n/k)\cdot g(k,m/k) \]记 \(\displaystyle h(a,b,n)=\sum_{k=1}^nf(k)\cdot g(k,a)\cdot g(k,b)\),于是可以整除分块
\[\sum_{i=1}^n\sum_{j=1}^m\varphi(ij)=\sum_{n/l=n/r,\ m/l=m/r}h(n/r,m/r,r)-h(n/r,m/r,l-1) \]问题是,预处理 \(h\) 的复杂度是很高的,这里可以考虑 根号分治 —— 设定阈值 \(B\),将 \(a,b\leqslant B\) 的 \(h\) 预处理出来,不难发现这是 \(\mathcal O(nB^2)\) 的。当查询的时候,若 \(n/r\leqslant B\) 就直接 \(\mathcal O(1)\) 查询,否则有 \(r\leqslant n/B\),干脆不差分直接暴力算就是 \(\mathcal O(n/B)\) 的。
复杂度 \(\mathcal O\left(n\ln n+nB^2+\left(n^{1/2}+n/B\right)T\right)\).
块长可以用均值不等式来算,总复杂度算出来大概是 \(\mathcal O(6\cdot 10^7)\).
# include <cstdio> # include <cctype> # define print(x,y) write(x), putchar(y) template <class T> inline T read(const T sample) { T x=0; char s; bool f=0; while(!isdigit(s=getchar())) f|=(s=='-'); for(; isdigit(s); s=getchar()) x=(x<<1)+(x<<3)+(s^48); return f? -x: x; } template <class T> inline void write(T x) { static int writ[50], w_tp=0; if(x<0) putchar('-'), x=-x; do writ[++w_tp]=x-x/10*10, x/=10; while(x); while(putchar(writ[w_tp--]^48), w_tp); } # include <vector> # include <iostream> using namespace std; const int B = 20; const int maxn = 1e5+5; const int mod = 998244353; int inv(int x,int y=mod-2,int r=1) { for(; y; y>>=1, x=1ll*x*x%mod) if(y&1) r=1ll*r*x%mod; return r; } bool is[maxn]; vector <int> g[maxn]; int h[B+1][B+1][maxn]; int phi[maxn], mu[maxn], pc, p[maxn], f[maxn]; int beelzebul(int n,int m) { int ans=0; if(n>m) swap(n,m); for(int i=1;i<=m/B+1;++i) ans = (ans+1ll*f[i]*g[i][n/i]%mod*g[i][m/i]%mod)%mod; for(int l=m/B+2, r; l<=n; l=r+1) { r = min(n, min(n/(n/l),m/(m/l))); ans = (0ll+ans+h[n/r][m/r][r]-h[n/r][m/r][l-1])%mod; } return (ans+mod)%mod; } int func(int i,int a,int b) { if(a>=g[i].size() || b>=g[i].size()) return 0; return 1ll*f[i]*g[i][a]%mod*g[i][b]%mod; } void sieve() { phi[1]=mu[1]=1; for(int i=2;i<=maxn-5;++i) { if(!is[i]) p[++pc]=i, mu[i]=-1, phi[i]=i-1; for(int j=1; j<=pc && i*p[j]<=maxn-5; ++j) { is[i*p[j]] = true, mu[i*p[j]]=-mu[i]; if(i%p[j]==0) { phi[i*p[j]] = phi[i]*p[j]; mu[i*p[j]]=0; break; } phi[i*p[j]] = phi[i]*(p[j]-1); } } for(int i=1;i<=maxn-5;++i) { const int coe = 1ll*i*inv(phi[i])%mod; g[i].emplace_back(0); for(int j=i;j<=maxn-5;j+=i) f[j] = (1ll*coe*mu[j/i]+f[j])%mod, g[i].emplace_back((g[i][j/i-1]+phi[j])%mod); } for(int a=1;a<=B;++a) for(int b=1;b<=B;++b) for(int i=1;i<=maxn-5;++i) h[a][b][i] = (h[a][b][i-1]+func(i,a,b))%mod; } int main() { sieve(); for(int T=read(9); T; --T) { int n=read(9), m=read(9); print(beelzebul(n,m),'\n'); } return 0; }