本来一直都是写\(7\)次的\(MTT\)的……然后被\(shadowice\)巨巨调教了一通之后只好去学一下\(4\)次的了……
简单来说就是我们现在需要处理一类模数不为\(NTT\)模数的情况
这里是板子
三模\(NTT\)
跑的很慢而且我也不会，这里就不说了
拆系数\(FFT\)
两个多项式\(P(z),Q(z)\)，我们把它们的系数拆成
\[A(z)=\sum_{i=0}^\infty (P_i>>15)z^i,B(z)=\sum_{i=0}^\infty (P_i\&32767)z^i\]
\[C(z)=\sum_{i=0}^\infty (Q_i>>15)z^i,D(z)=\sum_{i=0}^\infty (Q_i\&32767)z^i\]
那么我们只要计算出\((A\times C)<<30\),\((A\times D+B\times C)<<15\),\((B\times D)\)，然后把它们的系数加起来就可以了
//minamoto
#include<bits/stdc++.h>
#define R register
#define ll long long
#define fp(i,a,b) for(R int i=a,I=b+1;i<I;++i)
#define fd(i,a,b) for(R int i=a,I=b-1;i>I;--i)
#define go(u) for(int i=head[u],v=e[i].v;i;i=e[i].nx,v=e[i].v)
using namespace std;
char buf[1<<21],*p1=buf,*p2=buf;
inline char getc(){return p1==p2&&(p2=(p1=buf)+fread(buf,1,1<<21,stdin),p1==p2)?EOF:*p1++;}
int read(){
R int res,f=1;R char ch;
while((ch=getc())>'9'||ch<'0')(ch=='-')&&(f=-1);
for(res=ch-'0';(ch=getc())>='0'&&ch<='9';res=res*10+ch-'0');
return res*f;
}
char sr[1<<21],z[20];int K=-1,Z=0;
inline void Ot(){fwrite(sr,1,K+1,stdout),K=-1;}
void print(R int x){
if(K>1<<20)Ot();if(x<0)sr[++K]='-',x=-x;
while(z[++Z]=x%10+48,x/=10);
while(sr[++K]=z[Z],--Z);sr[++K]=' ';
}
const int N=5e5+5;const double Pi=acos(-1.0);
struct cp{
double x,y;
cp(double xx=0,double yy=0){x=xx,y=yy;}
inline cp operator +(cp b)const{return cp(x+b.x,y+b.y);}
inline cp operator -(cp b)const{return cp(x-b.x,y-b.y);}
inline cp operator *(cp b)const{return cp(x*b.x-y*b.y,x*b.y+y*b.x);}
inline cp operator *(const double &b)const{return cp(x*b,y*b);}
}A[N],B[N],C[N],D[N],H[N],F[N],G[N],w[N];
int r[N],lim=1,l,n,m,P,x;
void FFT(cp *A,int ty){
fp(i,0,lim-1)if(i<r[i])swap(A[i],A[r[i]]);
for(R int mid=1;mid<lim;mid<<=1)
for(R int j=0;j<lim;j+=(mid<<1))
for(R int k=0;k<mid;++k){
cp x=A[j+k],y=w[mid+k]*A[j+k+mid];
A[j+k]=x+y,A[j+k+mid]=x-y;
}
if(ty==-1){
reverse(A+1,A+lim);
double k=1.0/lim;fp(i,0,lim-1)A[i]=A[i]*k;
}
}
int main(){
// freopen("testdata.in","r",stdin);
n=read(),m=read(),P=read();
while(lim<=n+m)lim<<=1,++l;
fp(i,0,lim-1)r[i]=(r[i>>1]>>1)|((i&1)<<(l-1));
for(R int i=1;i<lim;i<<=1)fp(k,0,i-1)w[i+k]=cp(cos(Pi*k/i),sin(Pi*k/i));
fp(i,0,n)x=read(),A[i].x=x>>15,B[i].x=x&32767;
fp(i,0,m)x=read(),C[i].x=x>>15,D[i].x=x&32767;
FFT(A,1),FFT(B,1),FFT(C,1),FFT(D,1);
fp(i,0,lim-1)
F[i]=A[i]*C[i],G[i]=A[i]*D[i]+B[i]*C[i],H[i]=B[i]*D[i];
FFT(F,-1),FFT(G,-1),FFT(H,-1);
fp(i,0,n+m)print((((ll)(F[i].x+0.5)%P<<30)+((ll)(G[i].x+0.5)<<15)+((ll)(H[i].x+0.5)))%P);
return Ot(),0;
}\(FFT\)的优化
\[ \begin{aligned} P(x)=A(x)+iB(x) \\ Q(x)=A(x)-iB(x) \end{aligned} \]
令\(P'[k]\)和\(Q'[k]\)分别表示\(P(x)\)和\(Q(x)\)进行\(DFT\)的序列
\(P'[k]=P(\omega_n^k),Q'[k]=Q(\omega_n^k)\)，即代入\(n\)次单位根之后的点值
显然有
\[A'[k]={P'[k]+Q'[k]\over 2}\]
\[B'[k]={P'[k]-Q'[k]\over 2i}\]
这有啥用啊不还是两次\(FFT\)么……
但实际上我们是可以通过\(P'(x)\)求出\(Q'(x)\)的
推倒什么的太长了就直接拉过来好了
令\(\text{conj}(x)\)表示\(x\)的共轭复数（实部相等，虚部相反），\(A_i\)表示\(A(x)\)的第\(i\)次项系数
\[ \begin{aligned} P'[k] &= A(\omega_{n}^{k}) + i B(\omega_{n}^{k}) \\ & = \sum_{j=0}^{n-1} A_{j} \omega_{n}^{jk} + i B_{j} \omega_{n}^{jk} \\ & = \sum_{j=0}^{n-1} (A_{j} + i B_{j}) \left(\cos \left(\frac{2 \pi jk}{n}\right) + i \sin \left(\frac{2 \pi jk}{n}\right)\right) \\ \\ Q'[k] &= A(\omega_{n}^{k}) - i B(\omega_{n}^{k}) \\ & = \sum_{j=0}^{n-1} A_{j} \omega_{n}^{jk} - i B_{j} \omega_{n}^{jk} \\ & = \sum_{j=0}^{n-1} (A_{j} - i B_{j}) \left(\cos \left(\frac{2 \pi jk}{n}\right) + i \sin \left(\frac{2 \pi jk}{n}\right)\right) \\ & = \sum_{j=0}^{n-1} \left(A_{j} \cos \left(\frac{2 \pi jk}{n}\right) + B_{j} \sin \left(\frac{2 \pi jk}{n}\right)\right) + i \left(A_{j} \sin \left(\frac{2 \pi jk}{n}\right) - B_{j} \cos \left(\frac{2 \pi jk}{n}\right)\right) \\ & = \text{conj} \left( \sum_{j=0}^{n-1} \left(A_{j} \cos \left(\frac{2 \pi jk}{n}\right) + B_{j} \sin \left(\frac{2 \pi jk}{n}\right)\right) - i \left(A_{j} \sin \left(\frac{2 \pi jk}{n}\right) - B_{j} \cos \left(\frac{2 \pi jk}{n}\right)\right) \right) \\ & = \text{conj} \left( \sum_{j=0}^{n-1} \left(A_{j} \cos \left(\frac{-2 \pi jk}{n}\right) - B_{j} \sin \left(\frac{-2 \pi jk}{n}\right)\right) + i \left(A_{j} \sin \left(\frac{-2 \pi jk}{n}\right) + B_{j} \cos \left(\frac{-2 \pi jk}{n}\right)\right) \right) \\ & = \text{conj} \left( \sum_{j=0}^{n-1} (A_{j} + i B_{j}) \left(\cos \left(\frac{-2 \pi jk}{n}\right) + i \sin \left(\frac{-2 \pi jk}{n}\right)\right)\right) \\ & = \text{conj} \left( \sum_{j=0}^{n-1} (A_{j} + i B_{j}) \omega_{n}^{-jk} \right) \\ & = \text{conj} \left( \sum_{j=0}^{n-1} (A_{j} + i B_{j}) \omega_{n}^{(n-k)j} \right) \\ & = \text{conj} (P'[n-k]) \end{aligned} \]
注意这里是在模\(x^n\)意义下的，所以特殊判断\(Q'[0]=\text{conj}(P'[n])=\text{conj}(P'[0])\)
我们就可以通过\(P(x)\)得到\(Q(x)\)，只要一次\(FFT\)就可以了
\(IDFT\)的话，我们只要把\(P'(x)\)给\(IDFT\)一下，实部和虚部就分别对应\(A(x)\)和\(B(x)\)了
实际上这个东西可以在任意两个多项式做乘法的时候用到，可以优化到\(2\)次\(FFT\)（然而\(NTT\)模数下\(FFT\)比\(NTT\)慢的不是一点点……）
具体细节可以看代码理解
//minamoto
#include<bits/stdc++.h>
#define R register
#define ll long long
#define fp(i,a,b) for(R int i=(a),I=(b)+1;i<I;++i)
#define fd(i,a,b) for(R int i=(a),I=(b)-1;i>I;--i)
#define go(u) for(int i=head[u],v=e[i].v;i;i=e[i].nx,v=e[i].v)
using namespace std;
char buf[1<<21],*p1=buf,*p2=buf;
inline char getc(){return p1==p2&&(p2=(p1=buf)+fread(buf,1,1<<21,stdin),p1==p2)?EOF:*p1++;}
int read(){
R int res,f=1;R char ch;
while((ch=getc())>'9'||ch<'0')(ch=='-')&&(f=-1);
for(res=ch-'0';(ch=getc())>='0'&&ch<='9';res=res*10+ch-'0');
return res*f;
}
char sr[1<<21],z[20];int C=-1,Z=0;
inline void Ot(){fwrite(sr,1,C+1,stdout),C=-1;}
void print(R int x){
if(C>1<<20)Ot();if(x<0)sr[++C]='-',x=-x;
while(z[++Z]=x%10+48,x/=10);
while(sr[++C]=z[Z],--Z);sr[++C]=' ';
}
const int N=(1<<18)+5;const double Pi=acos(-1.0);
struct cp{
double x,y;
inline cp(){}
inline cp(R double xx,R double yy):x(xx),y(yy){}
inline cp operator +(const cp &b)const{return cp(x+b.x,y+b.y);}
inline cp operator -(const cp &b)const{return cp(x-b.x,y-b.y);}
inline cp operator *(const cp &b)const{return cp(x*b.x-y*b.y,x*b.y+y*b.x);}
inline cp operator *(const double &b)const{return cp(x*b,y*b);}
inline cp operator ~()const{return cp(x,-y);}
}w[2][N],a[N],b[N],f[N],g[N];
int r[N],n,m,lim,d,P;double iv;
void FFT(cp *A,int ty){
fp(i,0,lim-1)if(i<r[i])swap(A[i],A[r[i]]);
cp t;
for(R int mid=1;mid<lim;mid<<=1)
for(R int j=0;j<lim;j+=(mid<<1))
fp(k,0,mid-1)
A[j+k+mid]=A[j+k]-(t=w[ty][mid+k]*A[j+k+mid]),
A[j+k]=A[j+k]+t;
if(!ty)fp(i,0,lim-1)A[i]=A[i]*iv;
}
int main(){
// freopen("testdata.in","r",stdin);
n=read(),m=read(),P=read();
lim=1,d=0;while(lim<=n+m)lim<<=1,++d;iv=1.0/lim;
fp(i,0,lim-1)r[i]=(r[i>>1]>>1)|((i&1)<<(d-1));
for(R int i=0,x;i<=n;++i)x=read(),f[i]=cp(x>>15,x&32767);
for(R int i=0,x;i<=m;++i)x=read(),g[i]=cp(x>>15,x&32767);
for(R int i=1;i<lim;i<<=1)fp(k,0,i-1)
w[1][i+k]=cp(cos(Pi*k/i),sin(Pi*k/i)),w[0][i+k]=cp(cos(Pi*k/i),-sin(Pi*k/i));
fp(i,n+1,lim-1)f[i]=cp(0,0);fp(i,m+1,lim-1)g[i]=cp(0,0);
FFT(f,1),FFT(g,1);
fp(i,0,lim-1){
static cp q,f0,f1,g0,g1;
q=~f[i?lim-i:0],f0=(f[i]-q)*cp(0,-0.5),f1=(f[i]+q)*0.5,
q=~g[i?lim-i:0],g0=(g[i]-q)*cp(0,-0.5),g1=(g[i]+q)*0.5;
a[i]=f1*g1,b[i]=f0*g1+f1*g0+f0*g0*cp(0,1);
}
FFT(a,0),FFT(b,0);
fp(i,0,n+m)print((((ll)(a[i].x+0.5)%P<<30)+((ll)(b[i].x+0.5)<<15)+((ll)(b[i].y+0.5)))%P);
return Ot(),0;
}