library

This documentation is automatically generated by online-judge-tools/verification-helper

View the Project on GitHub maspypy/library

:x: poly/coef_of_rational_fps_2d.hpp

Depends on

Verified with

Code

#include "poly/convolution.hpp"
#include "poly/fps_div.hpp"
#include "linalg/transpose.hpp"

// P, Q が x について O(1) 次であるとする
// ANS(x) = [y^N]P/Q を M 次まで
// https://atcoder.jp/contests/nadafes2022_day2/tasks/nadafes2022_day2_p
// (2^n,2^m) 次がたくさん出てきていると定数倍がひどくなってしまう
template <typename mint>
vc<mint> coef_of_rational_fps_2d(vvc<mint> P, vvc<mint> Q, int N, int M) {
  int m = max(len(P[0]) - 1, len(Q[0]) - 1);
  chmin(m, N);
  FOR(i, len(P)) P[i].resize(m + 1);
  FOR(i, len(Q)) Q[i].resize(m + 1);
  if (len(P) > M + 1) P.resize(M + 1);
  if (len(Q) > M + 1) Q.resize(M + 1);
  while (len(P) < len(Q)) P.eb(vc<mint>(m + 1));
  while (len(Q) < len(P)) Q.eb(vc<mint>(m + 1));

  // P(y)Q(-y) の特定の parity, Q(y)Q(-y) の偶数次を取り出す
  while (N > 0) {
    if (len(P[0]) > N + 1) FOR(i, len(P)) P[i].resize(N + 1);
    if (len(Q[0]) > N + 1) FOR(i, len(Q)) Q[i].resize(N + 1);
    if (len(P) > M + 1) P.resize(M + 1);
    if (len(Q) > M + 1) Q.resize(M + 1);

    int H = len(P), W = len(P[0]);
    assert(len(Q) == H && len(Q[0]) == W);
    // (H,W) and (H,W)
    // (i,j) -> 2Wi + j
    ll L = 1;
    while (L < 4 * H * W) L *= 2;
    vc<mint> F(L), G(L);
    FOR(i, H) FOR(j, W) F[2 * W * i + j] = P[i][j];
    FOR(i, H) FOR(j, W) G[2 * W * i + j] = Q[i][j];

    int parity = N & 1;

    ntt(F, false);
    ntt(G, false);
    FOR(i, L / 2) F[2 * i + 0] *= G[2 * i + 1], F[2 * i + 1] *= G[2 * i + 0];
    FOR(i, L / 2) G[i] = G[2 * i + 0] * G[2 * i + 1];
    G.resize(L / 2);
    ntt(F, true); // さぼり
    ntt(G, true);

    vv(mint, nxtP, 2 * H - 1, W);
    vv(mint, nxtQ, 2 * H - 1, W);
    FOR(i, 2 * H - 1) FOR(j, W) { nxtP[i][j] = F[2 * W * i + (2 * j + parity)]; }
    FOR(i, 2 * H - 1) FOR(j, W) { nxtQ[i][j] = G[W * i + j]; }
    swap(P, nxtP), swap(Q, nxtQ);
    N /= 2;
  }
  vc<mint> f(M + 1), g(M + 1);
  FOR(i, M + 1) if (i < len(P)) f[i] = P[i][0];
  FOR(i, M + 1) if (i < len(Q)) g[i] = Q[i][0];
  return fps_div<mint>(f, g);
}

/*
[y^N] F(x,y)/G(x,y) を [x^M] まで。
[y^0] G = 1 を仮定している。
deg G = (3, 1) の (N,M)=(3×10^6,10^6) で 400 ms の実績があるが、かなり疎だからかも。
https://atcoder.jp/contests/agc058/tasks/agc058_d
*/
template <typename mint>
vc<mint> coef_of_rational_fps_2d_old(vector<vector<mint>> F, vector<vector<mint>> G, int N, int M) {
  using poly = vc<mint>;
  assert(G[0][0] == mint(1));
  F = transpose(F), G = transpose(G);
  FOR(j, 1, len(G[0])) assert(G[0][j] == mint(0));

  // x^N mod reverse(G) を計算する
  int m = len(G) - 1;

  auto add_at = [&](poly& f, int i, mint x) -> void {
    if (len(f) <= i) f.resize(i + 1);
    f[i] += x;
  };

  auto simplify = [&](poly& f) -> void {
    if (len(f) >= M + 1) f.resize(M + 1);
    while (len(f) && f.back() == mint(0)) f.pop_back();
  };

  auto dfs = [&](auto& dfs, int n) -> vc<poly> {
    if (n == 0) { return {poly({mint(1)})}; }
    vc<poly> f = dfs(dfs, n / 2);
    // 2 乗したい
    int nf = 0;
    FOR(i, len(f)) chmax(nf, len(f[i]));
    int K = 1;
    while (K < 2 * nf) K *= 2;
    FOR(i, len(f)) {
      f[i].resize(K);
      ntt(f[i], 0);
    }
    vc<poly> g(2 * len(f) - 1);
    FOR(i, 2 * len(f) - 1) g[i].resize(K);
    FOR(i, len(f)) FOR(j, len(f)) { FOR(k, K) g[i + j][k] += f[i][k] * f[j][k]; }
    FOR(i, len(g)) ntt(g[i], 1);
    swap(f, g);

    if (n % 2 == 1) { f.insert(f.begin(), poly({})); }
    FOR_R(i, m, len(f)) {
      FOR(j, 1, m + 1) {
        FOR(ny, len(G[j])) {
          mint cf = -G[j][ny];
          if (cf == mint(0)) continue;
          FOR(k, len(f[i])) {
            mint& x = f[i][k];
            if (x == mint(0)) continue;
            add_at(f[i - j], k + ny, cf * x);
          }
        }
      }
    }
    f.resize(m);
    FOR(i, m) simplify(f[i]);
    return f;
  };

  vc<poly> h = dfs(dfs, N);
  // 線形漸化式の最初の方を求める
  vc<poly> A(m);
  FOR(i, m) {
    A[i] = (i < len(F) ? F[i] : poly());
    FOR(j, 1, i + 1) {
      int k = i - j;
      poly tmp = convolution(A[k], G[j]);
      FOR(t, len(tmp)) { add_at(A[i], t, -tmp[t]); }
    }
  }

  vc<mint> res(M + 1);
  FOR(i, m) {
    vc<mint> f = convolution(h[i], A[i]);
    f.resize(M + 1);
    FOR(i, len(f)) add_at(res, i, f[i]);
  }
  return res;
}
#line 2 "mod/modint_common.hpp"

struct has_mod_impl {
  template <class T>
  static auto check(T &&x) -> decltype(x.get_mod(), std::true_type{});
  template <class T>
  static auto check(...) -> std::false_type;
};

template <class T>
class has_mod : public decltype(has_mod_impl::check<T>(std::declval<T>())) {};

template <typename mint>
mint inv(int n) {
  static const int mod = mint::get_mod();
  static vector<mint> dat = {0, 1};
  assert(0 <= n);
  if (n >= mod) n %= mod;
  while (len(dat) <= n) {
    int k = len(dat);
    int q = (mod + k - 1) / k;
    dat.eb(dat[k * q - mod] * mint::raw(q));
  }
  return dat[n];
}

template <typename mint>
mint fact(int n) {
  static const int mod = mint::get_mod();
  assert(0 <= n && n < mod);
  static vector<mint> dat = {1, 1};
  while (len(dat) <= n) dat.eb(dat[len(dat) - 1] * mint::raw(len(dat)));
  return dat[n];
}

template <typename mint>
mint fact_inv(int n) {
  static vector<mint> dat = {1, 1};
  if (n < 0) return mint(0);
  while (len(dat) <= n) dat.eb(dat[len(dat) - 1] * inv<mint>(len(dat)));
  return dat[n];
}

template <class mint, class... Ts>
mint fact_invs(Ts... xs) {
  return (mint(1) * ... * fact_inv<mint>(xs));
}

template <typename mint, class Head, class... Tail>
mint multinomial(Head &&head, Tail &&... tail) {
  return fact<mint>(head) * fact_invs<mint>(std::forward<Tail>(tail)...);
}

template <typename mint>
mint C_dense(int n, int k) {
  assert(n >= 0);
  if (k < 0 || n < k) return 0;
  static vvc<mint> C;
  static int H = 0, W = 0;
  auto calc = [&](int i, int j) -> mint {
    if (i == 0) return (j == 0 ? mint(1) : mint(0));
    return C[i - 1][j] + (j ? C[i - 1][j - 1] : 0);
  };
  if (W <= k) {
    FOR(i, H) {
      C[i].resize(k + 1);
      FOR(j, W, k + 1) { C[i][j] = calc(i, j); }
    }
    W = k + 1;
  }
  if (H <= n) {
    C.resize(n + 1);
    FOR(i, H, n + 1) {
      C[i].resize(W);
      FOR(j, W) { C[i][j] = calc(i, j); }
    }
    H = n + 1;
  }
  return C[n][k];
}

template <typename mint, bool large = false, bool dense = false>
mint C(ll n, ll k) {
  assert(n >= 0);
  if (k < 0 || n < k) return 0;
  if constexpr (dense) return C_dense<mint>(n, k);
  if constexpr (!large) return multinomial<mint>(n, k, n - k);
  k = min(k, n - k);
  mint x(1);
  FOR(i, k) x *= mint(n - i);
  return x * fact_inv<mint>(k);
}

template <typename mint, bool large = false>
mint C_inv(ll n, ll k) {
  assert(n >= 0);
  assert(0 <= k && k <= n);
  if (!large) return fact_inv<mint>(n) * fact<mint>(k) * fact<mint>(n - k);
  return mint(1) / C<mint, 1>(n, k);
}

// [x^d](1-x)^{-n}
template <typename mint, bool large = false, bool dense = false>
mint C_negative(ll n, ll d) {
  assert(n >= 0);
  if (d < 0) return mint(0);
  if (n == 0) { return (d == 0 ? mint(1) : mint(0)); }
  return C<mint, large, dense>(n + d - 1, d);
}
#line 3 "mod/modint.hpp"

template <int mod>
struct modint {
  static constexpr u32 umod = u32(mod);
  static_assert(umod < u32(1) << 31);
  u32 val;

  static modint raw(u32 v) {
    modint x;
    x.val = v;
    return x;
  }
  constexpr modint() : val(0) {}
  constexpr modint(u32 x) : val(x % umod) {}
  constexpr modint(u64 x) : val(x % umod) {}
  constexpr modint(u128 x) : val(x % umod) {}
  constexpr modint(int x) : val((x %= mod) < 0 ? x + mod : x){};
  constexpr modint(ll x) : val((x %= mod) < 0 ? x + mod : x){};
  constexpr modint(i128 x) : val((x %= mod) < 0 ? x + mod : x){};
  bool operator<(const modint &other) const { return val < other.val; }
  modint &operator+=(const modint &p) {
    if ((val += p.val) >= umod) val -= umod;
    return *this;
  }
  modint &operator-=(const modint &p) {
    if ((val += umod - p.val) >= umod) val -= umod;
    return *this;
  }
  modint &operator*=(const modint &p) {
    val = u64(val) * p.val % umod;
    return *this;
  }
  modint &operator/=(const modint &p) {
    *this *= p.inverse();
    return *this;
  }
  modint operator-() const { return modint::raw(val ? mod - val : u32(0)); }
  modint operator+(const modint &p) const { return modint(*this) += p; }
  modint operator-(const modint &p) const { return modint(*this) -= p; }
  modint operator*(const modint &p) const { return modint(*this) *= p; }
  modint operator/(const modint &p) const { return modint(*this) /= p; }
  bool operator==(const modint &p) const { return val == p.val; }
  bool operator!=(const modint &p) const { return val != p.val; }
  modint inverse() const {
    int a = val, b = mod, u = 1, v = 0, t;
    while (b > 0) {
      t = a / b;
      swap(a -= t * b, b), swap(u -= t * v, v);
    }
    return modint(u);
  }
  modint pow(ll n) const {
    assert(n >= 0);
    modint ret(1), mul(val);
    while (n > 0) {
      if (n & 1) ret *= mul;
      mul *= mul;
      n >>= 1;
    }
    return ret;
  }
  static constexpr int get_mod() { return mod; }
  // (n, r), r は 1 の 2^n 乗根
  static constexpr pair<int, int> ntt_info() {
    if (mod == 120586241) return {20, 74066978};
    if (mod == 167772161) return {25, 17};
    if (mod == 469762049) return {26, 30};
    if (mod == 754974721) return {24, 362};
    if (mod == 880803841) return {23, 211};
    if (mod == 943718401) return {22, 663003469};
    if (mod == 998244353) return {23, 31};
    if (mod == 1004535809) return {21, 582313106};
    if (mod == 1012924417) return {21, 368093570};
    return {-1, -1};
  }
  static constexpr bool can_ntt() { return ntt_info().fi != -1; }
};

#ifdef FASTIO
template <int mod>
void rd(modint<mod> &x) {
  fastio::rd(x.val);
  x.val %= mod;
  // assert(0 <= x.val && x.val < mod);
}
template <int mod>
void wt(modint<mod> x) {
  fastio::wt(x.val);
}
#endif

using modint107 = modint<1000000007>;
using modint998 = modint<998244353>;
#line 2 "mod/mod_inv.hpp"

// long でも大丈夫

// (val * x - 1) が mod の倍数になるようにする

// 特に mod=0 なら x=0 が満たす

ll mod_inv(ll val, ll mod) {
  if (mod == 0) return 0;
  mod = abs(mod);
  val %= mod;
  if (val < 0) val += mod;
  ll a = val, b = mod, u = 1, v = 0, t;
  while (b > 0) {
    t = a / b;
    swap(a -= t * b, b), swap(u -= t * v, v);
  }
  if (u < 0) u += mod;
  return u;
}
#line 2 "mod/crt3.hpp"

constexpr u32 mod_pow_constexpr(u64 a, u64 n, u32 mod) {
  a %= mod;
  u64 res = 1;
  FOR(32) {
    if (n & 1) res = res * a % mod;
    a = a * a % mod, n /= 2;
  }
  return res;
}

template <typename T, u32 p0, u32 p1>
T CRT2(u64 a0, u64 a1) {
  static_assert(p0 < p1);
  static constexpr u64 x0_1 = mod_pow_constexpr(p0, p1 - 2, p1);
  u64 c = (a1 - a0 + p1) * x0_1 % p1;
  return a0 + c * p0;
}

template <typename T, u32 p0, u32 p1, u32 p2>
T CRT3(u64 a0, u64 a1, u64 a2) {
  static_assert(p0 < p1 && p1 < p2);
  static constexpr u64 x1 = mod_pow_constexpr(p0, p1 - 2, p1);
  static constexpr u64 x2 = mod_pow_constexpr(u64(p0) * p1 % p2, p2 - 2, p2);
  static constexpr u64 p01 = u64(p0) * p1;
  u64 c = (a1 - a0 + p1) * x1 % p1;
  u64 ans_1 = a0 + c * p0;
  c = (a2 - ans_1 % p2 + p2) * x2 % p2;
  return T(ans_1) + T(c) * T(p01);
}

template <typename T, u32 p0, u32 p1, u32 p2, u32 p3>
T CRT4(u64 a0, u64 a1, u64 a2, u64 a3) {
  static_assert(p0 < p1 && p1 < p2 && p2 < p3);
  static constexpr u64 x1 = mod_pow_constexpr(p0, p1 - 2, p1);
  static constexpr u64 x2 = mod_pow_constexpr(u64(p0) * p1 % p2, p2 - 2, p2);
  static constexpr u64 x3 = mod_pow_constexpr(u64(p0) * p1 % p3 * p2 % p3, p3 - 2, p3);
  static constexpr u64 p01 = u64(p0) * p1;
  u64 c = (a1 - a0 + p1) * x1 % p1;
  u64 ans_1 = a0 + c * p0;
  c = (a2 - ans_1 % p2 + p2) * x2 % p2;
  u128 ans_2 = ans_1 + c * static_cast<u128>(p01);
  c = (a3 - ans_2 % p3 + p3) * x3 % p3;
  return T(ans_2) + T(c) * T(p01) * T(p2);
}

template <typename T, u32 p0, u32 p1, u32 p2, u32 p3, u32 p4>
T CRT5(u64 a0, u64 a1, u64 a2, u64 a3, u64 a4) {
  static_assert(p0 < p1 && p1 < p2 && p2 < p3 && p3 < p4);
  static constexpr u64 x1 = mod_pow_constexpr(p0, p1 - 2, p1);
  static constexpr u64 x2 = mod_pow_constexpr(u64(p0) * p1 % p2, p2 - 2, p2);
  static constexpr u64 x3 = mod_pow_constexpr(u64(p0) * p1 % p3 * p2 % p3, p3 - 2, p3);
  static constexpr u64 x4 = mod_pow_constexpr(u64(p0) * p1 % p4 * p2 % p4 * p3 % p4, p4 - 2, p4);
  static constexpr u64 p01 = u64(p0) * p1;
  static constexpr u64 p23 = u64(p2) * p3;
  u64 c = (a1 - a0 + p1) * x1 % p1;
  u64 ans_1 = a0 + c * p0;
  c = (a2 - ans_1 % p2 + p2) * x2 % p2;
  u128 ans_2 = ans_1 + c * static_cast<u128>(p01);
  c = static_cast<u64>(a3 - ans_2 % p3 + p3) * x3 % p3;
  u128 ans_3 = ans_2 + static_cast<u128>(c * p2) * p01;
  c = static_cast<u64>(a4 - ans_3 % p4 + p4) * x4 % p4;
  return T(ans_3) + T(c) * T(p01) * T(p23);
}
#line 2 "poly/convolution_naive.hpp"

template <class T, typename enable_if<!has_mod<T>::value>::type* = nullptr>
vc<T> convolution_naive(const vc<T>& a, const vc<T>& b) {
  int n = int(a.size()), m = int(b.size());
  if (n > m) return convolution_naive<T>(b, a);
  if (n == 0) return {};
  vector<T> ans(n + m - 1);
  FOR(i, n) FOR(j, m) ans[i + j] += a[i] * b[j];
  return ans;
}

template <class T, typename enable_if<has_mod<T>::value>::type* = nullptr>
vc<T> convolution_naive(const vc<T>& a, const vc<T>& b) {
  int n = int(a.size()), m = int(b.size());
  if (n > m) return convolution_naive<T>(b, a);
  if (n == 0) return {};
  vc<T> ans(n + m - 1);
  if (n <= 16 && (T::get_mod() < (1 << 30))) {
    for (int k = 0; k < n + m - 1; ++k) {
      int s = max(0, k - m + 1);
      int t = min(n, k + 1);
      u64 sm = 0;
      for (int i = s; i < t; ++i) { sm += u64(a[i].val) * (b[k - i].val); }
      ans[k] = sm;
    }
  } else {
    for (int k = 0; k < n + m - 1; ++k) {
      int s = max(0, k - m + 1);
      int t = min(n, k + 1);
      u128 sm = 0;
      for (int i = s; i < t; ++i) { sm += u64(a[i].val) * (b[k - i].val); }
      ans[k] = T::raw(sm % T::get_mod());
    }
  }
  return ans;
}
#line 2 "poly/convolution_karatsuba.hpp"

// 任意の環でできる
template <typename T>
vc<T> convolution_karatsuba(const vc<T>& f, const vc<T>& g) {
  const int thresh = 30;
  if (min(len(f), len(g)) <= thresh) return convolution_naive(f, g);
  int n = max(len(f), len(g));
  int m = ceil(n, 2);
  vc<T> f1, f2, g1, g2;
  if (len(f) < m) f1 = f;
  if (len(f) >= m) f1 = {f.begin(), f.begin() + m};
  if (len(f) >= m) f2 = {f.begin() + m, f.end()};
  if (len(g) < m) g1 = g;
  if (len(g) >= m) g1 = {g.begin(), g.begin() + m};
  if (len(g) >= m) g2 = {g.begin() + m, g.end()};
  vc<T> a = convolution_karatsuba(f1, g1);
  vc<T> b = convolution_karatsuba(f2, g2);
  FOR(i, len(f2)) f1[i] += f2[i];
  FOR(i, len(g2)) g1[i] += g2[i];
  vc<T> c = convolution_karatsuba(f1, g1);
  vc<T> F(len(f) + len(g) - 1);
  assert(2 * m + len(b) <= len(F));
  FOR(i, len(a)) F[i] += a[i], c[i] -= a[i];
  FOR(i, len(b)) F[2 * m + i] += b[i], c[i] -= b[i];
  if (c.back() == T(0)) c.pop_back();
  FOR(i, len(c)) if (c[i] != T(0)) F[m + i] += c[i];
  return F;
}
#line 2 "poly/ntt.hpp"

template <class mint>
void ntt(vector<mint>& a, bool inverse) {
  assert(mint::can_ntt());
  const int rank2 = mint::ntt_info().fi;
  const int mod = mint::get_mod();
  static array<mint, 30> root, iroot;
  static array<mint, 30> rate2, irate2;
  static array<mint, 30> rate3, irate3;

  assert(rank2 != -1 && len(a) <= (1 << max(0, rank2)));

  static bool prepared = 0;
  if (!prepared) {
    prepared = 1;
    root[rank2] = mint::ntt_info().se;
    iroot[rank2] = mint(1) / root[rank2];
    FOR_R(i, rank2) {
      root[i] = root[i + 1] * root[i + 1];
      iroot[i] = iroot[i + 1] * iroot[i + 1];
    }
    mint prod = 1, iprod = 1;
    for (int i = 0; i <= rank2 - 2; i++) {
      rate2[i] = root[i + 2] * prod;
      irate2[i] = iroot[i + 2] * iprod;
      prod *= iroot[i + 2];
      iprod *= root[i + 2];
    }
    prod = 1, iprod = 1;
    for (int i = 0; i <= rank2 - 3; i++) {
      rate3[i] = root[i + 3] * prod;
      irate3[i] = iroot[i + 3] * iprod;
      prod *= iroot[i + 3];
      iprod *= root[i + 3];
    }
  }

  int n = int(a.size());
  int h = topbit(n);
  assert(n == 1 << h);
  if (!inverse) {
    int len = 0;
    while (len < h) {
      if (h - len == 1) {
        int p = 1 << (h - len - 1);
        mint rot = 1;
        FOR(s, 1 << len) {
          int offset = s << (h - len);
          FOR(i, p) {
            auto l = a[i + offset];
            auto r = a[i + offset + p] * rot;
            a[i + offset] = l + r;
            a[i + offset + p] = l - r;
          }
          rot *= rate2[topbit(~s & -~s)];
        }
        len++;
      } else {
        int p = 1 << (h - len - 2);
        mint rot = 1, imag = root[2];
        for (int s = 0; s < (1 << len); s++) {
          mint rot2 = rot * rot;
          mint rot3 = rot2 * rot;
          int offset = s << (h - len);
          for (int i = 0; i < p; i++) {
            u64 mod2 = u64(mod) * mod;
            u64 a0 = a[i + offset].val;
            u64 a1 = u64(a[i + offset + p].val) * rot.val;
            u64 a2 = u64(a[i + offset + 2 * p].val) * rot2.val;
            u64 a3 = u64(a[i + offset + 3 * p].val) * rot3.val;
            u64 a1na3imag = (a1 + mod2 - a3) % mod * imag.val;
            u64 na2 = mod2 - a2;
            a[i + offset] = a0 + a2 + a1 + a3;
            a[i + offset + 1 * p] = a0 + a2 + (2 * mod2 - (a1 + a3));
            a[i + offset + 2 * p] = a0 + na2 + a1na3imag;
            a[i + offset + 3 * p] = a0 + na2 + (mod2 - a1na3imag);
          }
          rot *= rate3[topbit(~s & -~s)];
        }
        len += 2;
      }
    }
  } else {
    mint coef = mint(1) / mint(len(a));
    FOR(i, len(a)) a[i] *= coef;
    int len = h;
    while (len) {
      if (len == 1) {
        int p = 1 << (h - len);
        mint irot = 1;
        FOR(s, 1 << (len - 1)) {
          int offset = s << (h - len + 1);
          FOR(i, p) {
            u64 l = a[i + offset].val;
            u64 r = a[i + offset + p].val;
            a[i + offset] = l + r;
            a[i + offset + p] = (mod + l - r) * irot.val;
          }
          irot *= irate2[topbit(~s & -~s)];
        }
        len--;
      } else {
        int p = 1 << (h - len);
        mint irot = 1, iimag = iroot[2];
        FOR(s, (1 << (len - 2))) {
          mint irot2 = irot * irot;
          mint irot3 = irot2 * irot;
          int offset = s << (h - len + 2);
          for (int i = 0; i < p; i++) {
            u64 a0 = a[i + offset + 0 * p].val;
            u64 a1 = a[i + offset + 1 * p].val;
            u64 a2 = a[i + offset + 2 * p].val;
            u64 a3 = a[i + offset + 3 * p].val;
            u64 x = (mod + a2 - a3) * iimag.val % mod;
            a[i + offset] = a0 + a1 + a2 + a3;
            a[i + offset + 1 * p] = (a0 + mod - a1 + x) * irot.val;
            a[i + offset + 2 * p] = (a0 + a1 + 2 * mod - a2 - a3) * irot2.val;
            a[i + offset + 3 * p] = (a0 + 2 * mod - a1 - x) * irot3.val;
          }
          irot *= irate3[topbit(~s & -~s)];
        }
        len -= 2;
      }
    }
  }
}
#line 8 "poly/convolution.hpp"

template <class mint>
vector<mint> convolution_ntt(vector<mint> a, vector<mint> b) {
  if (a.empty() || b.empty()) return {};
  int n = int(a.size()), m = int(b.size());
  int sz = 1;
  while (sz < n + m - 1) sz *= 2;

  // sz = 2^k のときの高速化。分割統治的なやつで損しまくるので。

  if ((n + m - 3) <= sz / 2) {
    auto a_last = a.back(), b_last = b.back();
    a.pop_back(), b.pop_back();
    auto c = convolution(a, b);
    c.resize(n + m - 1);
    c[n + m - 2] = a_last * b_last;
    FOR(i, len(a)) c[i + len(b)] += a[i] * b_last;
    FOR(i, len(b)) c[i + len(a)] += b[i] * a_last;
    return c;
  }

  a.resize(sz), b.resize(sz);
  bool same = a == b;
  ntt(a, 0);
  if (same) {
    b = a;
  } else {
    ntt(b, 0);
  }
  FOR(i, sz) a[i] *= b[i];
  ntt(a, 1);
  a.resize(n + m - 1);
  return a;
}

template <typename mint>
vector<mint> convolution_garner(const vector<mint>& a, const vector<mint>& b) {
  int n = len(a), m = len(b);
  if (!n || !m) return {};
  static constexpr int p0 = 167772161;
  static constexpr int p1 = 469762049;
  static constexpr int p2 = 754974721;
  using mint0 = modint<p0>;
  using mint1 = modint<p1>;
  using mint2 = modint<p2>;
  vc<mint0> a0(n), b0(m);
  vc<mint1> a1(n), b1(m);
  vc<mint2> a2(n), b2(m);
  FOR(i, n) a0[i] = a[i].val, a1[i] = a[i].val, a2[i] = a[i].val;
  FOR(i, m) b0[i] = b[i].val, b1[i] = b[i].val, b2[i] = b[i].val;
  auto c0 = convolution_ntt<mint0>(a0, b0);
  auto c1 = convolution_ntt<mint1>(a1, b1);
  auto c2 = convolution_ntt<mint2>(a2, b2);
  vc<mint> c(len(c0));
  FOR(i, n + m - 1) { c[i] = CRT3<mint, p0, p1, p2>(c0[i].val, c1[i].val, c2[i].val); }
  return c;
}

vector<ll> convolution(vector<ll> a, vector<ll> b) {
  int n = len(a), m = len(b);
  if (!n || !m) return {};
  if (min(n, m) <= 2500) return convolution_naive(a, b);

  ll mi_a = MIN(a), mi_b = MIN(b);
  for (auto& x: a) x -= mi_a;
  for (auto& x: b) x -= mi_b;
  assert(MAX(a) * MAX(b) <= 1e18);

  auto Ac = cumsum<ll>(a), Bc = cumsum<ll>(b);
  vi res(n + m - 1);
  for (int k = 0; k < n + m - 1; ++k) {
    int s = max(0, k - m + 1);
    int t = min(n, k + 1);
    res[k] += (t - s) * mi_a * mi_b;
    res[k] += mi_a * (Bc[k - s + 1] - Bc[k - t + 1]);
    res[k] += mi_b * (Ac[t] - Ac[s]);
  }

  static constexpr u32 MOD1 = 1004535809;
  static constexpr u32 MOD2 = 1012924417;
  using mint1 = modint<MOD1>;
  using mint2 = modint<MOD2>;

  vc<mint1> a1(n), b1(m);
  vc<mint2> a2(n), b2(m);
  FOR(i, n) a1[i] = a[i], a2[i] = a[i];
  FOR(i, m) b1[i] = b[i], b2[i] = b[i];

  auto c1 = convolution_ntt<mint1>(a1, b1);
  auto c2 = convolution_ntt<mint2>(a2, b2);

  FOR(i, n + m - 1) { res[i] += CRT2<u64, MOD1, MOD2>(c1[i].val, c2[i].val); }
  return res;
}

template <typename mint>
vc<mint> convolution(const vc<mint>& a, const vc<mint>& b) {
  int n = len(a), m = len(b);
  if (!n || !m) return {};
  if (mint::can_ntt()) {
    if (min(n, m) <= 50) return convolution_karatsuba<mint>(a, b);
    return convolution_ntt(a, b);
  }
  if (min(n, m) <= 200) return convolution_karatsuba<mint>(a, b);
  return convolution_garner(a, b);
}
#line 2 "poly/fps_div.hpp"

#line 2 "poly/count_terms.hpp"
template<typename mint>
int count_terms(const vc<mint>& f){
  int t = 0;
  FOR(i, len(f)) if(f[i] != mint(0)) ++t;
  return t;
}
#line 4 "poly/fps_inv.hpp"

template <typename mint>
vc<mint> fps_inv_sparse(const vc<mint>& f) {
  int N = len(f);
  vc<pair<int, mint>> dat;
  FOR(i, 1, N) if (f[i] != mint(0)) dat.eb(i, f[i]);
  vc<mint> g(N);
  mint g0 = mint(1) / f[0];
  g[0] = g0;
  FOR(n, 1, N) {
    mint rhs = 0;
    for (auto&& [k, fk]: dat) {
      if (k > n) break;
      rhs -= fk * g[n - k];
    }
    g[n] = rhs * g0;
  }
  return g;
}

template <typename mint>
vc<mint> fps_inv_dense_ntt(const vc<mint>& F) {
  vc<mint> G = {mint(1) / F[0]};
  ll N = len(F), n = 1;
  G.reserve(N);
  while (n < N) {
    vc<mint> f(2 * n), g(2 * n);
    FOR(i, min(N, 2 * n)) f[i] = F[i];
    FOR(i, n) g[i] = G[i];
    ntt(f, false), ntt(g, false);
    FOR(i, 2 * n) f[i] *= g[i];
    ntt(f, true);
    FOR(i, n) f[i] = 0;
    ntt(f, false);
    FOR(i, 2 * n) f[i] *= g[i];
    ntt(f, true);
    FOR(i, n, min(N, 2 * n)) G.eb(-f[i]);
    n *= 2;
  }
  return G;
}

template <typename mint>
vc<mint> fps_inv_dense(const vc<mint>& F) {
  if (mint::can_ntt()) return fps_inv_dense_ntt(F);
  const int N = len(F);
  vc<mint> R = {mint(1) / F[0]};
  vc<mint> p;
  int m = 1;
  while (m < N) {
    p = convolution(R, R);
    p.resize(m + m);
    vc<mint> f = {F.begin(), F.begin() + min(m + m, N)};
    p = convolution(p, f);
    R.resize(m + m);
    FOR(i, m + m) R[i] = R[i] + R[i] - p[i];
    m += m;
  }
  R.resize(N);
  return R;
}

template <typename mint>
vc<mint> fps_inv(const vc<mint>& f) {
  assert(f[0] != mint(0));
  int n = count_terms(f);
  int t = (mint::can_ntt() ? 160 : 820);
  return (n <= t ? fps_inv_sparse<mint>(f) : fps_inv_dense<mint>(f));
}
#line 5 "poly/fps_div.hpp"

// f/g. f の長さで出力される.
template <typename mint, bool SPARSE = false>
vc<mint> fps_div(vc<mint> f, vc<mint> g) {
  if (SPARSE || count_terms(g) < 200) return fps_div_sparse(f, g);
  int n = len(f);
  g.resize(n);
  g = fps_inv<mint>(g);
  f = convolution(f, g);
  f.resize(n);
  return f;
}

// f/g ただし g は sparse
template <typename mint>
vc<mint> fps_div_sparse(vc<mint> f, vc<mint>& g) {
  if (g[0] != mint(1)) {
    mint cf = g[0].inverse();
    for (auto&& x: f) x *= cf;
    for (auto&& x: g) x *= cf;
  }

  vc<pair<int, mint>> dat;
  FOR(i, 1, len(g)) if (g[i] != mint(0)) dat.eb(i, -g[i]);
  FOR(i, len(f)) {
    for (auto&& [j, x]: dat) {
      if (i >= j) f[i] += x * f[i - j];
    }
  }
  return f;
}
#line 1 "linalg/transpose.hpp"
template <typename VC>
vc<VC> transpose(const vc<VC>& A, int H = -1, int W = -1) {
  if (H == -1) { H = len(A), W = (len(A) == 0 ? 0 : len(A[0])); }
  if (H == 0) return {};
  vc<VC> B(W, VC(H, A[0][0]));
  FOR(x, H) FOR(y, W) B[y][x] = A[x][y];
  return B;
}
#line 4 "poly/coef_of_rational_fps_2d.hpp"

// P, Q が x について O(1) 次であるとする
// ANS(x) = [y^N]P/Q を M 次まで
// https://atcoder.jp/contests/nadafes2022_day2/tasks/nadafes2022_day2_p
// (2^n,2^m) 次がたくさん出てきていると定数倍がひどくなってしまう
template <typename mint>
vc<mint> coef_of_rational_fps_2d(vvc<mint> P, vvc<mint> Q, int N, int M) {
  int m = max(len(P[0]) - 1, len(Q[0]) - 1);
  chmin(m, N);
  FOR(i, len(P)) P[i].resize(m + 1);
  FOR(i, len(Q)) Q[i].resize(m + 1);
  if (len(P) > M + 1) P.resize(M + 1);
  if (len(Q) > M + 1) Q.resize(M + 1);
  while (len(P) < len(Q)) P.eb(vc<mint>(m + 1));
  while (len(Q) < len(P)) Q.eb(vc<mint>(m + 1));

  // P(y)Q(-y) の特定の parity, Q(y)Q(-y) の偶数次を取り出す
  while (N > 0) {
    if (len(P[0]) > N + 1) FOR(i, len(P)) P[i].resize(N + 1);
    if (len(Q[0]) > N + 1) FOR(i, len(Q)) Q[i].resize(N + 1);
    if (len(P) > M + 1) P.resize(M + 1);
    if (len(Q) > M + 1) Q.resize(M + 1);

    int H = len(P), W = len(P[0]);
    assert(len(Q) == H && len(Q[0]) == W);
    // (H,W) and (H,W)
    // (i,j) -> 2Wi + j
    ll L = 1;
    while (L < 4 * H * W) L *= 2;
    vc<mint> F(L), G(L);
    FOR(i, H) FOR(j, W) F[2 * W * i + j] = P[i][j];
    FOR(i, H) FOR(j, W) G[2 * W * i + j] = Q[i][j];

    int parity = N & 1;

    ntt(F, false);
    ntt(G, false);
    FOR(i, L / 2) F[2 * i + 0] *= G[2 * i + 1], F[2 * i + 1] *= G[2 * i + 0];
    FOR(i, L / 2) G[i] = G[2 * i + 0] * G[2 * i + 1];
    G.resize(L / 2);
    ntt(F, true); // さぼり
    ntt(G, true);

    vv(mint, nxtP, 2 * H - 1, W);
    vv(mint, nxtQ, 2 * H - 1, W);
    FOR(i, 2 * H - 1) FOR(j, W) { nxtP[i][j] = F[2 * W * i + (2 * j + parity)]; }
    FOR(i, 2 * H - 1) FOR(j, W) { nxtQ[i][j] = G[W * i + j]; }
    swap(P, nxtP), swap(Q, nxtQ);
    N /= 2;
  }
  vc<mint> f(M + 1), g(M + 1);
  FOR(i, M + 1) if (i < len(P)) f[i] = P[i][0];
  FOR(i, M + 1) if (i < len(Q)) g[i] = Q[i][0];
  return fps_div<mint>(f, g);
}

/*
[y^N] F(x,y)/G(x,y) を [x^M] まで。
[y^0] G = 1 を仮定している。
deg G = (3, 1) の (N,M)=(3×10^6,10^6) で 400 ms の実績があるが、かなり疎だからかも。
https://atcoder.jp/contests/agc058/tasks/agc058_d
*/
template <typename mint>
vc<mint> coef_of_rational_fps_2d_old(vector<vector<mint>> F, vector<vector<mint>> G, int N, int M) {
  using poly = vc<mint>;
  assert(G[0][0] == mint(1));
  F = transpose(F), G = transpose(G);
  FOR(j, 1, len(G[0])) assert(G[0][j] == mint(0));

  // x^N mod reverse(G) を計算する
  int m = len(G) - 1;

  auto add_at = [&](poly& f, int i, mint x) -> void {
    if (len(f) <= i) f.resize(i + 1);
    f[i] += x;
  };

  auto simplify = [&](poly& f) -> void {
    if (len(f) >= M + 1) f.resize(M + 1);
    while (len(f) && f.back() == mint(0)) f.pop_back();
  };

  auto dfs = [&](auto& dfs, int n) -> vc<poly> {
    if (n == 0) { return {poly({mint(1)})}; }
    vc<poly> f = dfs(dfs, n / 2);
    // 2 乗したい
    int nf = 0;
    FOR(i, len(f)) chmax(nf, len(f[i]));
    int K = 1;
    while (K < 2 * nf) K *= 2;
    FOR(i, len(f)) {
      f[i].resize(K);
      ntt(f[i], 0);
    }
    vc<poly> g(2 * len(f) - 1);
    FOR(i, 2 * len(f) - 1) g[i].resize(K);
    FOR(i, len(f)) FOR(j, len(f)) { FOR(k, K) g[i + j][k] += f[i][k] * f[j][k]; }
    FOR(i, len(g)) ntt(g[i], 1);
    swap(f, g);

    if (n % 2 == 1) { f.insert(f.begin(), poly({})); }
    FOR_R(i, m, len(f)) {
      FOR(j, 1, m + 1) {
        FOR(ny, len(G[j])) {
          mint cf = -G[j][ny];
          if (cf == mint(0)) continue;
          FOR(k, len(f[i])) {
            mint& x = f[i][k];
            if (x == mint(0)) continue;
            add_at(f[i - j], k + ny, cf * x);
          }
        }
      }
    }
    f.resize(m);
    FOR(i, m) simplify(f[i]);
    return f;
  };

  vc<poly> h = dfs(dfs, N);
  // 線形漸化式の最初の方を求める
  vc<poly> A(m);
  FOR(i, m) {
    A[i] = (i < len(F) ? F[i] : poly());
    FOR(j, 1, i + 1) {
      int k = i - j;
      poly tmp = convolution(A[k], G[j]);
      FOR(t, len(tmp)) { add_at(A[i], t, -tmp[t]); }
    }
  }

  vc<mint> res(M + 1);
  FOR(i, m) {
    vc<mint> f = convolution(h[i], A[i]);
    f.resize(M + 1);
    FOR(i, len(f)) add_at(res, i, f[i]);
  }
  return res;
}
Back to top page