Bostan-Mori
(fps/bostan-mori.hpp)

• View this file on GitHub
• Last update: 2025-10-21 21:13:36+09:00
• Include: #include "fps/bostan-mori.hpp"

Bostan-Mori 法．

高々 $m$ 次のFPS $f,g$ に対し $[x^n]\frac{f(x)}{g(x)}$ を $O(m\log m\log n)$ 時間で求める．

アルゴリズム

以下の変形をする．

\[\frac{f(x)}{g(x)}=\frac{f(x)g(-x)}{g(x)g(-x)}\]

$g(x)g(-x)$ は偶関数であるから $g(x)g(-x)=g’(x^2)$ とおける．

$n$ が偶数のとき，$f(x)g(-x)$ の偶数次の部分を $f’(x^2)$ とすれば以下が成り立つ．

\[[x^n]\frac{f(x)}{g(x)}=[x^n]\frac{f'(x^2)}{g'(x^2)}=[x^{n/2}]\frac{f'(x)}{g'(x)}\]

$n$ が奇数のときも同様．

定数倍高速化

$\max(\deg f,\deg g)+1$ 以上の最小の二冪を $m$ とし，$1$ の原始 $2m$ 乗根 $\omega$ をとる．

$g’(\omega^{2i})=g(\omega^i)g(\omega^{i+m})$ であるので，$g$ に長さ $2m$ の FFT をし，適当に積を取って長さ $m$ の IFTT をすれば $g’$ が求められる．

また FFT のダブリングを用いればさらに定数倍が改善できる．

Depends on

• fps/formal-power-series.hpp

Required by

• 線形漸化式用 (fps/linearly-recurrent-sequence.hpp)

Verified with

• verify/fps/LC_consecutive_terms_of_linear_recurrent_sequence.test.cpp
• verify/fps/LC_kth_term_of_linearly_recurrent_sequence.test.cpp

Code

#pragma once
#include "fps/formal-power-series.hpp"

// [x^n]f(x)/g(x)
template <class mint>
mint BostanMori(FormalPowerSeries<mint> f, FormalPowerSeries<mint> g, long long n) {
  g.shrink();
  mint ret = 0;
  if (f.size() >= g.size()) {
    auto q = f / g;
    if (n < q.size()) ret += q[n];
    f -= q * g;
    f.shrink();
  }
  if (f.empty()) return ret;
  FormalPowerSeries<mint>::set_ntt();
  if (!FormalPowerSeries<mint>::ntt_ptr) {
    f.resize(g.size() - 1);
    for (; n > 0; n >>= 1) {
      auto g1 = g;
      for (int i = 1; i < g1.size(); i += 2) g1[i] = -g1[i];
      auto p = f * g1, q = g * g1;
      if (n & 1) {
        for (int i = 0; i < f.size(); i++) f[i] = p[(i << 1) | 1];
      } else {
        for (int i = 0; i < f.size(); i++) f[i] = p[i << 1];
      }
      for (int i = 0; i < g.size(); i++) g[i] = q[i << 1];
    }
    return ret + f[0] / g[0];
  } else {
    int m = 1, log = 0;
    while (m < g.size()) m <<= 1, log++;
    mint wi = mint(FormalPowerSeries<mint>::ntt_root()).inv().pow((mint::get_mod() - 1) >> (log + 1));
    vector<int> rev(m);
    for (int i = 0; i < rev.size(); i++) rev[i] = (rev[i / 2] / 2) | ((i & 1) << (log - 1));
    vector<mint> pow(m, 1);
    for (int i = 1; i < m; i++) pow[rev[i]] = pow[rev[i - 1]] * wi;
    f.resize(m), g.resize(m);
    f.ntt(), g.ntt();
    mint inv2 = mint(2).inv();
    for (; n >= m; n >>= 1) {
      f.ntt_doubling(), g.ntt_doubling();
      if (n & 1) {
        for (int i = 0; i < m; i++) f[i] = (f[i << 1] * g[(i << 1) | 1] - f[(i << 1) | 1] * g[i << 1]) * inv2 * pow[i];
      } else {
        for (int i = 0; i < m; i++) f[i] = (f[i << 1] * g[(i << 1) | 1] + f[(i << 1) | 1] * g[i << 1]) * inv2;
      }
      for (int i = 0; i < m; i++) g[i] = g[i << 1] * g[(i << 1) | 1];
      f.resize(m), g.resize(m);
    }
    f.intt(), g.intt();
    return ret + (f * g.inv())[n];
  }
}
/**
 * @brief Bostan-Mori
 * @docs docs/fps/bostan-mori.md
 */

#line 2 "fps/formal-power-series.hpp"

template <class mint>
struct FormalPowerSeries : vector<mint> {
  using vector<mint>::vector;
  using FPS = FormalPowerSeries;
  FPS &operator+=(const FPS &r) {
    if (r.size() > this->size()) this->resize(r.size());
    for (int i = 0; i < (int)r.size(); i++) (*this)[i] += r[i];
    return *this;
  }
  FPS &operator+=(const mint &r) {
    if (this->empty()) this->resize(1);
    (*this)[0] += r;
    return *this;
  }
  FPS &operator-=(const FPS &r) {
    if (r.size() > this->size()) this->resize(r.size());
    for (int i = 0; i < (int)r.size(); i++) (*this)[i] -= r[i];
    return *this;
  }
  FPS &operator-=(const mint &r) {
    if (this->empty()) this->resize(1);
    (*this)[0] -= r;
    return *this;
  }
  FPS &operator*=(const mint &v) {
    for (int k = 0; k < (int)this->size(); k++) (*this)[k] *= v;
    return *this;
  }
  FPS &operator/=(const FPS &r) {
    if (this->size() < r.size()) {
      this->clear();
      return *this;
    }
    int n = this->size() - r.size() + 1;
    if ((int)r.size() <= 64) {
      FPS f(*this), g(r);
      g.shrink();
      mint coeff = g.at(g.size() - 1).inv();
      for (auto &x : g) x *= coeff;
      int deg = (int)f.size() - (int)g.size() + 1;
      int gs = g.size();
      FPS quo(deg);
      for (int i = deg - 1; i >= 0; i--) {
        quo[i] = f[i + gs - 1];
        for (int j = 0; j < gs; j++) f[i + j] -= quo[i] * g[j];
      }
      *this = quo * coeff;
      this->resize(n, mint(0));
      return *this;
    }
    return *this = ((*this).rev().pre(n) * r.rev().inv(n)).pre(n).rev();
  }
  FPS &operator%=(const FPS &r) {
    *this -= *this / r * r;
    shrink();
    return *this;
  }
  FPS operator+(const FPS &r) const { return FPS(*this) += r; }
  FPS operator+(const mint &v) const { return FPS(*this) += v; }
  FPS operator-(const FPS &r) const { return FPS(*this) -= r; }
  FPS operator-(const mint &v) const { return FPS(*this) -= v; }
  FPS operator*(const FPS &r) const { return FPS(*this) *= r; }
  FPS operator*(const mint &v) const { return FPS(*this) *= v; }
  FPS operator/(const FPS &r) const { return FPS(*this) /= r; }
  FPS operator%(const FPS &r) const { return FPS(*this) %= r; }
  FPS operator-() const {
    FPS ret(this->size());
    for (int i = 0; i < (int)this->size(); i++) ret[i] = -(*this)[i];
    return ret;
  }
  void shrink() {
    while (this->size() && this->back() == mint(0)) this->pop_back();
  }
  FPS rev() const {
    FPS ret(*this);
    reverse(begin(ret), end(ret));
    return ret;
  }
  FPS dot(FPS r) const {
    FPS ret(min(this->size(), r.size()));
    for (int i = 0; i < (int)ret.size(); i++) ret[i] = (*this)[i] * r[i];
    return ret;
  }
  FPS pre(int sz) const {
    return FPS(begin(*this), begin(*this) + min((int)this->size(), sz));
  }
  FPS operator>>=(int sz) {
    assert(sz >= 0);
    if ((int)this->size() <= sz) return {};
    this->erase(this->begin(), this->begin() + sz);
    return *this;
  }
  FPS operator>>(int sz) const {
    if ((int)this->size() <= sz) return {};
    FPS ret(*this);
    ret.erase(ret.begin(), ret.begin() + sz);
    return ret;
  }
  FPS operator<<=(int sz) {
    assert(sz >= 0);
    this->insert(this->begin(), sz, mint(0));
    return *this;
  }
  FPS operator<<(int sz) const {
    FPS ret(*this);
    ret.insert(ret.begin(), sz, mint(0));
    return ret;
  }
  FPS diff() const {
    const int n = (int)this->size();
    FPS ret(max(0, n - 1));
    mint one(1), coeff(1);
    for (int i = 1; i < n; i++) {
      ret[i - 1] = (*this)[i] * coeff;
      coeff += one;
    }
    return ret;
  }
  FPS integral() const {
    const int n = (int)this->size();
    FPS ret(n + 1);
    ret[0] = mint(0);
    if (n > 0) ret[1] = mint(1);
    auto mod = mint::get_mod();
    for (int i = 2; i <= n; i++) ret[i] = (-ret[mod % i]) * (mod / i);
    for (int i = 0; i < n; i++) ret[i + 1] *= (*this)[i];
    return ret;
  }
  mint eval(mint x) const {
    mint r = 0, w = 1;
    for (auto &v : *this) r += w * v, w *= x;
    return r;
  }
  FPS log(int deg = -1) const {
    assert((*this)[0] == mint(1));
    if (deg == -1) deg = (int)this->size();
    return (this->diff() * this->inv(deg)).pre(deg - 1).integral();
  }
  FPS pow(int64_t k, int deg = -1) const {
    const int n = (int)this->size();
    if (deg == -1) deg = n;
    if (k == 0) {
      FPS ret(deg);
      if (deg) ret[0] = 1;
      return ret;
    }
    for (int i = 0; i < n; i++) {
      if ((*this)[i] != mint(0)) {
        mint rev = mint(1) / (*this)[i];
        FPS ret = (((*this * rev) >> i).log(deg) * k).exp(deg);
        ret *= (*this)[i].pow(k);
        ret = (ret << (i * k)).pre(deg);
        if ((int)ret.size() < deg) ret.resize(deg, mint(0));
        return ret;
      }
      if (__int128_t(i + 1) * k >= deg) return FPS(deg, mint(0));
    }
    return FPS(deg, mint(0));
  }

  static void *ntt_ptr;
  static void set_ntt();
  FPS &operator*=(const FPS &r);
  FPS middle_product(const FPS &r) const;
  void ntt();
  void intt();
  void ntt_doubling();
  static int ntt_root();
  FPS inv(int deg = -1) const;
  FPS exp(int deg = -1) const;
};
template <typename mint>
void *FormalPowerSeries<mint>::ntt_ptr = nullptr;
#line 3 "fps/bostan-mori.hpp"

// [x^n]f(x)/g(x)
template <class mint>
mint BostanMori(FormalPowerSeries<mint> f, FormalPowerSeries<mint> g, long long n) {
  g.shrink();
  mint ret = 0;
  if (f.size() >= g.size()) {
    auto q = f / g;
    if (n < q.size()) ret += q[n];
    f -= q * g;
    f.shrink();
  }
  if (f.empty()) return ret;
  FormalPowerSeries<mint>::set_ntt();
  if (!FormalPowerSeries<mint>::ntt_ptr) {
    f.resize(g.size() - 1);
    for (; n > 0; n >>= 1) {
      auto g1 = g;
      for (int i = 1; i < g1.size(); i += 2) g1[i] = -g1[i];
      auto p = f * g1, q = g * g1;
      if (n & 1) {
        for (int i = 0; i < f.size(); i++) f[i] = p[(i << 1) | 1];
      } else {
        for (int i = 0; i < f.size(); i++) f[i] = p[i << 1];
      }
      for (int i = 0; i < g.size(); i++) g[i] = q[i << 1];
    }
    return ret + f[0] / g[0];
  } else {
    int m = 1, log = 0;
    while (m < g.size()) m <<= 1, log++;
    mint wi = mint(FormalPowerSeries<mint>::ntt_root()).inv().pow((mint::get_mod() - 1) >> (log + 1));
    vector<int> rev(m);
    for (int i = 0; i < rev.size(); i++) rev[i] = (rev[i / 2] / 2) | ((i & 1) << (log - 1));
    vector<mint> pow(m, 1);
    for (int i = 1; i < m; i++) pow[rev[i]] = pow[rev[i - 1]] * wi;
    f.resize(m), g.resize(m);
    f.ntt(), g.ntt();
    mint inv2 = mint(2).inv();
    for (; n >= m; n >>= 1) {
      f.ntt_doubling(), g.ntt_doubling();
      if (n & 1) {
        for (int i = 0; i < m; i++) f[i] = (f[i << 1] * g[(i << 1) | 1] - f[(i << 1) | 1] * g[i << 1]) * inv2 * pow[i];
      } else {
        for (int i = 0; i < m; i++) f[i] = (f[i << 1] * g[(i << 1) | 1] + f[(i << 1) | 1] * g[i << 1]) * inv2;
      }
      for (int i = 0; i < m; i++) g[i] = g[i << 1] * g[(i << 1) | 1];
      f.resize(m), g.resize(m);
    }
    f.intt(), g.intt();
    return ret + (f * g.inv())[n];
  }
}
/**
 * @brief Bostan-Mori
 * @docs docs/fps/bostan-mori.md
 */

Back to top page