关于循环矩阵的一些理解和应用
问题引入
给定一个长度为\(n\)的环,每次对于一个位置上的数,加上它左边的数乘上\(L\),再加上右边的数乘上\(R\)。\(L,R\)是给定的常数,问执行\(s\)次之后,这个环上的每个位置上的数是多少
(计算时左边和右边的数都是上一次的数)
如果\(n \leq 100, s \leq 2^{30}\)
可以想到矩阵快速幂
构造矩阵
于是开心地快速幂了
#include<cstdio>
#include<cstring>
#include<cctype>
#include<algorithm>
#define RG register
#define file(x) freopen(#x".in", "r", stdin), freopen(#x".out", "w", stdout)
#define clear(x, y) memset(x, y, sizeof(x))
inline int read()
{
int data = 0, w = 1; char ch = getchar();
while(ch != '-' && (!isdigit(ch))) ch = getchar();
if(ch == '-') w = -1, ch = getchar();
while(isdigit(ch)) data = data * 10 + (ch ^ 48), ch = getchar();
return data * w;
}
const int maxn(1010);
int n, s, L, R, _x, Pow[10], a[maxn], Mod;
namespace $
{
const int N(110);
struct Matrix
{
int n, m, a[N][N];
Matrix(int x, int y) : n(x), m(y) { memset(a, 0, sizeof a); }
int *operator [] (const int &id) { return a[id]; }
const int *operator [] (const int &id) const { return a[id]; }
};
Matrix operator * (const Matrix &a, const Matrix &b)
{
Matrix c(a.n, b.m);
for(RG int i = 0; i < a.n; i++)
for(RG int j = 0; j < a.m; j++)
for(RG int k = 0; k < b.m; k++)
c[i][k] = (1ll * c[i][k] + 1ll * a[i][j] * b[j][k]) % Mod;
return c;
}
int main()
{
Matrix S(1, n), T(n, n);
for(RG int i = 0; i < n; i++) S[0][i] = a[i];
for(RG int i = 0; i < n; i++)
{
int pre = (i - 1 + n) % n, suc = (i + 1) % n;
T[i][i] = 1, T[pre][i] = L, T[suc][i] = R;
}
while(s)
{
if(s & 1) S = S * T;
T = T * T; s >>= 1;
}
for(RG int i = 0; i < n; i++) printf("%d ", S[0][i]);
return 0;
}
}
int main()
{
Pow[0] = 1;
for(RG int i = 1; i < 10; i++) Pow[i] = Pow[i - 1] * 10;
n = read(), s = read(), L = read(), R = read(), _x = read();
for(RG int i = 0; i < n; i++) a[i] = read();
Mod = Pow[_x];
return $::main();
}
解决
可是:
\(n \leq 1000, s \leq 2^{30}\)
\(\text{TLE}\)
但仔细看这个矩阵可以发现一个特点:它的每一行都是上一行往右移动一位得到的
不但如此,这个矩阵无论自乘多少次,都满足这个性质,所以理论上只需要维护第一行就好了
令\(k[i]\)表示矩阵第一行的第\(i\)个,那么有(假设是一个\(4 * 4\)的矩阵
可以看出一些规律:
这样复杂度就是\(O(n^2log\;s)\)了
解决啦!!!
#include<cstdio>
#include<cstring>
#include<cctype>
#include<algorithm>
#define RG register
#define file(x) freopen(#x".in", "r", stdin), freopen(#x".out", "w", stdout)
#define clear(x, y) memset(x, y, sizeof(x))
inline int read()
{
int data = 0, w = 1; char ch = getchar();
while(ch != '-' && (!isdigit(ch))) ch = getchar();
if(ch == '-') w = -1, ch = getchar();
while(isdigit(ch)) data = data * 10 + (ch ^ 48), ch = getchar();
return data * w;
}
const int maxn(1010);
int n, s, L, R, _x, Mod = 1;
int S[maxn], T[maxn];
void Mul(int S[], int T[])
{
static int c[maxn]; clear(c, 0);
for(RG int i = 1; i <= n; i++)
for(RG int j = 1; j <= n; j++)
(c[(i + j - 2) % n + 1] += 1ll * S[i] * T[j] % Mod) %= Mod;
std::copy(c + 1, c + n + 1, S + 1);
}
int main()
{
n = read(), s = read(), L = read(), R = read(), _x = read();
for(RG int i = 1; i <= _x; i++) Mod *= 10;
for(RG int i = 1; i <= n; i++) S[i] = read();
T[1] = 1, T[2] = L, T[n] = R;
for(; s; s >>= 1, Mul(T, T)) if(s & 1) Mul(S, T);
for(RG int i = 1; i <= n; i++) printf("%d ", S[i]);
return 0;
}
Extend
\(n\leq 5\times 10^4, s \leq 2^{30}\)
\(\mathrm{TLE}\)
这时我们想到,可不可以\(\mathrm{FFT}\)优化循环矩阵乘法呢??
可以!!!
将前面的下标从\(0\)开始
这™不就是个卷积的形式吗???
证明看参考博客
这里面的知识点一定要好好理(ji)解(jie lun)
这里不再赘述。
所以:
和上面的结论是一样的。
于是用个\(NTT\)或者\(MTT\)即可
我非常友善的\(\mathrm{mod}\;998244353\)
#include<cstdio>
#include<cstring>
#include<cctype>
#include<algorithm>
#define RG register
#define file(x) freopen(#x".in", "r", stdin), freopen(#x".out", "w", stdout)
#define clear(x, y) memset(x, y, sizeof(x))
inline int read()
{
int data = 0, w = 1; char ch = getchar();
while(ch != '-' && (!isdigit(ch))) ch = getchar();
if(ch == '-') w = -1, ch = getchar();
while(isdigit(ch)) data = data * 10 + (ch ^ 48), ch = getchar();
return data * w;
}
const int maxn(200010), Mod(998244353);
int n, s, L, R, N, r[maxn];
int S[maxn], T[maxn];
inline int fastpow(int x, int y)
{
int ans = 1;
while(y)
{
if(y & 1) ans = 1ll * ans * x % Mod;
x = 1ll * x * x % Mod, y >>= 1;
}
return ans;
}
template<int opt> void FFT(int *p)
{
for(RG int i = 0; i < N; i++) if(i < r[i]) std::swap(p[i], p[r[i]]);
for(RG int i = 1; i < N; i <<= 1)
{
int rot = fastpow(3, (Mod - 1) / (i << 1));
for(RG int j = 0; j < N; j += (i << 1))
for(RG int k = 0, w = 1; k < i; ++k, w = 1ll * w * rot % Mod)
{
int x = p[j + k], y = 1ll * w * p[i + j + k] % Mod;
p[j + k] = (x + y) % Mod, p[i + j + k] = (x - y + Mod) % Mod;
}
}
if(opt == -1) std::reverse(p + 1, p + N);
}
void Mul(int *S, int *T)
{
static int a[maxn], b[maxn];
std::fill(a, a + N, 0); std::fill(b, b + N, 0);
std::copy(S, S + n, a); std::copy(T, T + n, b);
FFT<1>(a), FFT<1>(b);
for(RG int i = 0; i < N; i++) a[i] = 1ll * a[i] * b[i] % Mod;
FFT<-1>(a); std::fill(S, S + n, 0);
for(RG int i = 0, inv = fastpow(N, Mod - 2); i < N; i++)
a[i] = 1ll * a[i] * inv % Mod;
for(RG int i = 0; i < N; i++) S[i % n] = (S[i % n] + a[i]) % Mod;
}
int main()
{
int P = -1;
n = read(), s = read(), L = read(), R = read();
for(RG int i = 0; i < n; i++) S[i] = read();
for(N = 1; N <= n + n; N <<= 1, ++P);
for(RG int i = 0; i < N; i++) r[i] = (r[i >> 1] >> 1) | ((i & 1) << P);
T[0] = 1, T[1] = L, T[n - 1] = R;
for(; s; s >>= 1, Mul(T, T)) if(s & 1) Mul(S, T);
for(RG int i = 0; i < n; i++) printf("%d ", S[i]);
return 0;
}
总结
形如\(\begin{bmatrix}c_1&c_2&\cdots&c_{n-1}&c_n\\c_2&c_3&\cdots&c_n&c_1\\\cdots&\cdots&\cdots&\cdots&\cdots\\c_n&c_1&\cdots&c_{n-2}&c_{n-1}\end{bmatrix}\)的矩阵是循环矩阵
循环矩阵的乘积仍是循环矩阵
所以我们只要维护循环矩阵的第一行
就可以\(O(n^2)\)甚至\(O(nlog_2n)\)维护循环矩阵的乘积
通过这样一些神奇的性质,我们可以降低矩阵乘法的复杂度
来出毒瘤题
没了