library
This documentation is automatically generated by online-judge-tools/verification-helper
test/3_yukicoder/2587_2.test.cpp
- View this file on GitHub
- Last update: 2024-12-26 06:32:57+09:00
- Problem: https://yukicoder.me/problems/no/2587
Depends on
ds/hashmap.hpp
- ds/unionfind/unionfind.hpp
- graph/base.hpp
- graph/characteristic_polynomial_of_tree_adjacency_matrix.hpp
- graph/ds/static_toptree.hpp
- graph/shortest_path/bfs01.hpp
- graph/tree.hpp
- graph/tree_walk_generating_function.hpp
- mod/crt3.hpp
- mod/mod_inv.hpp
- mod/modint.hpp
- mod/modint_common.hpp
- mod/powertable.hpp
- my_template.hpp
- nt/primetable.hpp
- other/io.hpp
- poly/coef_of_rational_fps.hpp
- poly/convolution.hpp
- poly/convolution_all.hpp
- poly/convolution_karatsuba.hpp
- poly/convolution_naive.hpp
- poly/count_terms.hpp
- poly/fps_div.hpp
- poly/fps_inv.hpp
- poly/ntt.hpp
- poly/ntt_doubling.hpp
- poly/poly_divmod.hpp
- poly/poly_taylor_shift.hpp
Code
#define PROBLEM "https://yukicoder.me/problems/no/2587"
#include "my_template.hpp"
#include "other/io.hpp"
#include "mod/modint.hpp"
#include "graph/base.hpp"
#include "poly/poly_taylor_shift.hpp"
#include "poly/coef_of_rational_fps.hpp"
#include "graph/tree_walk_generating_function.hpp"
using mint = modint998;
using poly = vc<mint>;
void solve() {
LL(N, M, S, T);
--S, --T;
Graph<int, 0> G(N);
G.read_tree();
auto [g, f] = tree_walk_generating_function<false, mint>(
G, S, T, [&](int i, int j) -> mint { return 1; });
// W(x) = g(x)/f(x)
// ANS = sum_m binom(M,m)[x^m]W(x)
// EGF(W) = sum_m W_m 1/m! x^m
// [x^M]e^x EGF(W)
// 根が全部 -1 される
// ANS も C-rec で、f(x-1)
vc<mint> W = fps_div(g, f);
FOR(m, N) W[m] *= fact_inv<mint>(m);
vc<mint> tmp(N + 1);
FOR(m, N + 1) tmp[m] = fact_inv<mint>(m);
vc<mint> F = convolution(W, tmp);
F.resize(N);
FOR(i, N) F[i] *= fact<mint>(i);
reverse(all(f));
f = poly_taylor_shift<mint>(f, -1);
reverse(all(f));
g = convolution<mint>(F, f);
g.resize(N);
mint ANS = coef_of_rational_fps<mint>(g, f, M);
print(ANS);
}
signed main() {
int T = 1;
// INT(T);
FOR(T) solve();
return 0;
}#line 1 "test/3_yukicoder/2587_2.test.cpp"
#define PROBLEM "https://yukicoder.me/problems/no/2587"
#line 1 "my_template.hpp"
#if defined(LOCAL)
#include <my_template_compiled.hpp>
#else
// https://codeforces.com/blog/entry/96344
#pragma GCC optimize("Ofast,unroll-loops")
// いまの CF だとこれ入れると動かない？
// #pragma GCC target("avx2,popcnt")
#include <bits/stdc++.h>
using namespace std;
using ll = long long;
using u8 = uint8_t;
using u16 = uint16_t;
using u32 = uint32_t;
using u64 = uint64_t;
using i128 = __int128;
using u128 = unsigned __int128;
using f128 = __float128;
template <class T>
constexpr T infty = 0;
template <>
constexpr int infty<int> = 1'010'000'000;
template <>
constexpr ll infty<ll> = 2'020'000'000'000'000'000;
template <>
constexpr u32 infty<u32> = infty<int>;
template <>
constexpr u64 infty<u64> = infty<ll>;
template <>
constexpr i128 infty<i128> = i128(infty<ll>) * 2'000'000'000'000'000'000;
template <>
constexpr double infty<double> = infty<ll>;
template <>
constexpr long double infty<long double> = infty<ll>;
using pi = pair<ll, ll>;
using vi = vector<ll>;
template <class T>
using vc = vector<T>;
template <class T>
using vvc = vector<vc<T>>;
template <class T>
using vvvc = vector<vvc<T>>;
template <class T>
using vvvvc = vector<vvvc<T>>;
template <class T>
using vvvvvc = vector<vvvvc<T>>;
template <class T>
using pq = priority_queue<T>;
template <class T>
using pqg = priority_queue<T, vector<T>, greater<T>>;
#define vv(type, name, h, ...) vector<vector<type>> name(h, vector<type>(__VA_ARGS__))
#define vvv(type, name, h, w, ...) vector<vector<vector<type>>> name(h, vector<vector<type>>(w, vector<type>(__VA_ARGS__)))
#define vvvv(type, name, a, b, c, ...) \
vector<vector<vector<vector<type>>>> name(a, vector<vector<vector<type>>>(b, vector<vector<type>>(c, vector<type>(__VA_ARGS__))))
// https://trap.jp/post/1224/
#define FOR1(a) for (ll _ = 0; _ < ll(a); ++_)
#define FOR2(i, a) for (ll i = 0; i < ll(a); ++i)
#define FOR3(i, a, b) for (ll i = a; i < ll(b); ++i)
#define FOR4(i, a, b, c) for (ll i = a; i < ll(b); i += (c))
#define FOR1_R(a) for (ll i = (a)-1; i >= ll(0); --i)
#define FOR2_R(i, a) for (ll i = (a)-1; i >= ll(0); --i)
#define FOR3_R(i, a, b) for (ll i = (b)-1; i >= ll(a); --i)
#define overload4(a, b, c, d, e, ...) e
#define overload3(a, b, c, d, ...) d
#define FOR(...) overload4(__VA_ARGS__, FOR4, FOR3, FOR2, FOR1)(__VA_ARGS__)
#define FOR_R(...) overload3(__VA_ARGS__, FOR3_R, FOR2_R, FOR1_R)(__VA_ARGS__)
#define all(x) x.begin(), x.end()
#define len(x) ll(x.size())
#define elif else if
#define eb emplace_back
#define mp make_pair
#define mt make_tuple
#define fi first
#define se second
#define stoi stoll
int popcnt(int x) { return __builtin_popcount(x); }
int popcnt(u32 x) { return __builtin_popcount(x); }
int popcnt(ll x) { return __builtin_popcountll(x); }
int popcnt(u64 x) { return __builtin_popcountll(x); }
int popcnt_sgn(int x) { return (__builtin_parity(unsigned(x)) & 1 ? -1 : 1); }
int popcnt_sgn(u32 x) { return (__builtin_parity(x) & 1 ? -1 : 1); }
int popcnt_sgn(ll x) { return (__builtin_parityll(x) & 1 ? -1 : 1); }
int popcnt_sgn(u64 x) { return (__builtin_parityll(x) & 1 ? -1 : 1); }
// (0, 1, 2, 3, 4) -> (-1, 0, 1, 1, 2)
int topbit(int x) { return (x == 0 ? -1 : 31 - __builtin_clz(x)); }
int topbit(u32 x) { return (x == 0 ? -1 : 31 - __builtin_clz(x)); }
int topbit(ll x) { return (x == 0 ? -1 : 63 - __builtin_clzll(x)); }
int topbit(u64 x) { return (x == 0 ? -1 : 63 - __builtin_clzll(x)); }
// (0, 1, 2, 3, 4) -> (-1, 0, 1, 0, 2)
int lowbit(int x) { return (x == 0 ? -1 : __builtin_ctz(x)); }
int lowbit(u32 x) { return (x == 0 ? -1 : __builtin_ctz(x)); }
int lowbit(ll x) { return (x == 0 ? -1 : __builtin_ctzll(x)); }
int lowbit(u64 x) { return (x == 0 ? -1 : __builtin_ctzll(x)); }
template <typename T>
T kth_bit(int k) {
return T(1) << k;
}
template <typename T>
bool has_kth_bit(T x, int k) {
return x >> k & 1;
}
template <typename UINT>
struct all_bit {
struct iter {
UINT s;
iter(UINT s) : s(s) {}
int operator*() const { return lowbit(s); }
iter &operator++() {
s &= s - 1;
return *this;
}
bool operator!=(const iter) const { return s != 0; }
};
UINT s;
all_bit(UINT s) : s(s) {}
iter begin() const { return iter(s); }
iter end() const { return iter(0); }
};
template <typename UINT>
struct all_subset {
static_assert(is_unsigned<UINT>::value);
struct iter {
UINT s, t;
bool ed;
iter(UINT s) : s(s), t(s), ed(0) {}
int operator*() const { return s ^ t; }
iter &operator++() {
(t == 0 ? ed = 1 : t = (t - 1) & s);
return *this;
}
bool operator!=(const iter) const { return !ed; }
};
UINT s;
all_subset(UINT s) : s(s) {}
iter begin() const { return iter(s); }
iter end() const { return iter(0); }
};
template <typename T>
T floor(T a, T b) {
return a / b - (a % b && (a ^ b) < 0);
}
template <typename T>
T ceil(T x, T y) {
return floor(x + y - 1, y);
}
template <typename T>
T bmod(T x, T y) {
return x - y * floor(x, y);
}
template <typename T>
pair<T, T> divmod(T x, T y) {
T q = floor(x, y);
return {q, x - q * y};
}
template <typename T, typename U>
T SUM(const vector<U> &A) {
T sm = 0;
for (auto &&a: A) sm += a;
return sm;
}
#define MIN(v) *min_element(all(v))
#define MAX(v) *max_element(all(v))
#define LB(c, x) distance((c).begin(), lower_bound(all(c), (x)))
#define UB(c, x) distance((c).begin(), upper_bound(all(c), (x)))
#define UNIQUE(x) sort(all(x)), x.erase(unique(all(x)), x.end()), x.shrink_to_fit()
template <typename T>
T POP(deque<T> &que) {
T a = que.front();
que.pop_front();
return a;
}
template <typename T>
T POP(pq<T> &que) {
T a = que.top();
que.pop();
return a;
}
template <typename T>
T POP(pqg<T> &que) {
T a = que.top();
que.pop();
return a;
}
template <typename T>
T POP(vc<T> &que) {
T a = que.back();
que.pop_back();
return a;
}
template <typename F>
ll binary_search(F check, ll ok, ll ng, bool check_ok = true) {
if (check_ok) assert(check(ok));
while (abs(ok - ng) > 1) {
auto x = (ng + ok) / 2;
(check(x) ? ok : ng) = x;
}
return ok;
}
template <typename F>
double binary_search_real(F check, double ok, double ng, int iter = 100) {
FOR(iter) {
double x = (ok + ng) / 2;
(check(x) ? ok : ng) = x;
}
return (ok + ng) / 2;
}
template <class T, class S>
inline bool chmax(T &a, const S &b) {
return (a < b ? a = b, 1 : 0);
}
template <class T, class S>
inline bool chmin(T &a, const S &b) {
return (a > b ? a = b, 1 : 0);
}
// ? は -1
vc<int> s_to_vi(const string &S, char first_char) {
vc<int> A(S.size());
FOR(i, S.size()) { A[i] = (S[i] != '?' ? S[i] - first_char : -1); }
return A;
}
template <typename T, typename U>
vector<T> cumsum(vector<U> &A, int off = 1) {
int N = A.size();
vector<T> B(N + 1);
FOR(i, N) { B[i + 1] = B[i] + A[i]; }
if (off == 0) B.erase(B.begin());
return B;
}
// stable sort
template <typename T>
vector<int> argsort(const vector<T> &A) {
vector<int> ids(len(A));
iota(all(ids), 0);
sort(all(ids), [&](int i, int j) { return (A[i] == A[j] ? i < j : A[i] < A[j]); });
return ids;
}
// A[I[0]], A[I[1]], ...
template <typename T>
vc<T> rearrange(const vc<T> &A, const vc<int> &I) {
vc<T> B(len(I));
FOR(i, len(I)) B[i] = A[I[i]];
return B;
}
template <typename T, typename... Vectors>
void concat(vc<T> &first, const Vectors &... others) {
vc<T> &res = first;
(res.insert(res.end(), others.begin(), others.end()), ...);
}
#endif
#line 1 "other/io.hpp"
#define FASTIO
#include <unistd.h>
// https://judge.yosupo.jp/submission/21623
namespace fastio {
static constexpr uint32_t SZ = 1 << 17;
char ibuf[SZ];
char obuf[SZ];
char out[100];
// pointer of ibuf, obuf
uint32_t pil = 0, pir = 0, por = 0;
struct Pre {
char num[10000][4];
constexpr Pre() : num() {
for (int i = 0; i < 10000; i++) {
int n = i;
for (int j = 3; j >= 0; j--) {
num[i][j] = n % 10 | '0';
n /= 10;
}
}
}
} constexpr pre;
inline void load() {
memcpy(ibuf, ibuf + pil, pir - pil);
pir = pir - pil + fread(ibuf + pir - pil, 1, SZ - pir + pil, stdin);
pil = 0;
if (pir < SZ) ibuf[pir++] = '\n';
}
inline void flush() {
fwrite(obuf, 1, por, stdout);
por = 0;
}
void rd(char &c) {
do {
if (pil + 1 > pir) load();
c = ibuf[pil++];
} while (isspace(c));
}
void rd(string &x) {
x.clear();
char c;
do {
if (pil + 1 > pir) load();
c = ibuf[pil++];
} while (isspace(c));
do {
x += c;
if (pil == pir) load();
c = ibuf[pil++];
} while (!isspace(c));
}
template <typename T>
void rd_real(T &x) {
string s;
rd(s);
x = stod(s);
}
template <typename T>
void rd_integer(T &x) {
if (pil + 100 > pir) load();
char c;
do
c = ibuf[pil++];
while (c < '-');
bool minus = 0;
if constexpr (is_signed<T>::value || is_same_v<T, i128>) {
if (c == '-') { minus = 1, c = ibuf[pil++]; }
}
x = 0;
while ('0' <= c) { x = x * 10 + (c & 15), c = ibuf[pil++]; }
if constexpr (is_signed<T>::value || is_same_v<T, i128>) {
if (minus) x = -x;
}
}
void rd(int &x) { rd_integer(x); }
void rd(ll &x) { rd_integer(x); }
void rd(i128 &x) { rd_integer(x); }
void rd(u32 &x) { rd_integer(x); }
void rd(u64 &x) { rd_integer(x); }
void rd(u128 &x) { rd_integer(x); }
void rd(double &x) { rd_real(x); }
void rd(long double &x) { rd_real(x); }
void rd(f128 &x) { rd_real(x); }
template <class T, class U>
void rd(pair<T, U> &p) {
return rd(p.first), rd(p.second);
}
template <size_t N = 0, typename T>
void rd_tuple(T &t) {
if constexpr (N < std::tuple_size<T>::value) {
auto &x = std::get<N>(t);
rd(x);
rd_tuple<N + 1>(t);
}
}
template <class... T>
void rd(tuple<T...> &tpl) {
rd_tuple(tpl);
}
template <size_t N = 0, typename T>
void rd(array<T, N> &x) {
for (auto &d: x) rd(d);
}
template <class T>
void rd(vc<T> &x) {
for (auto &d: x) rd(d);
}
void read() {}
template <class H, class... T>
void read(H &h, T &... t) {
rd(h), read(t...);
}
void wt(const char c) {
if (por == SZ) flush();
obuf[por++] = c;
}
void wt(const string s) {
for (char c: s) wt(c);
}
void wt(const char *s) {
size_t len = strlen(s);
for (size_t i = 0; i < len; i++) wt(s[i]);
}
template <typename T>
void wt_integer(T x) {
if (por > SZ - 100) flush();
if (x < 0) { obuf[por++] = '-', x = -x; }
int outi;
for (outi = 96; x >= 10000; outi -= 4) {
memcpy(out + outi, pre.num[x % 10000], 4);
x /= 10000;
}
if (x >= 1000) {
memcpy(obuf + por, pre.num[x], 4);
por += 4;
} else if (x >= 100) {
memcpy(obuf + por, pre.num[x] + 1, 3);
por += 3;
} else if (x >= 10) {
int q = (x * 103) >> 10;
obuf[por] = q | '0';
obuf[por + 1] = (x - q * 10) | '0';
por += 2;
} else
obuf[por++] = x | '0';
memcpy(obuf + por, out + outi + 4, 96 - outi);
por += 96 - outi;
}
template <typename T>
void wt_real(T x) {
ostringstream oss;
oss << fixed << setprecision(15) << double(x);
string s = oss.str();
wt(s);
}
void wt(int x) { wt_integer(x); }
void wt(ll x) { wt_integer(x); }
void wt(i128 x) { wt_integer(x); }
void wt(u32 x) { wt_integer(x); }
void wt(u64 x) { wt_integer(x); }
void wt(u128 x) { wt_integer(x); }
void wt(double x) { wt_real(x); }
void wt(long double x) { wt_real(x); }
void wt(f128 x) { wt_real(x); }
template <class T, class U>
void wt(const pair<T, U> val) {
wt(val.first);
wt(' ');
wt(val.second);
}
template <size_t N = 0, typename T>
void wt_tuple(const T t) {
if constexpr (N < std::tuple_size<T>::value) {
if constexpr (N > 0) { wt(' '); }
const auto x = std::get<N>(t);
wt(x);
wt_tuple<N + 1>(t);
}
}
template <class... T>
void wt(tuple<T...> tpl) {
wt_tuple(tpl);
}
template <class T, size_t S>
void wt(const array<T, S> val) {
auto n = val.size();
for (size_t i = 0; i < n; i++) {
if (i) wt(' ');
wt(val[i]);
}
}
template <class T>
void wt(const vector<T> val) {
auto n = val.size();
for (size_t i = 0; i < n; i++) {
if (i) wt(' ');
wt(val[i]);
}
}
void print() { wt('\n'); }
template <class Head, class... Tail>
void print(Head &&head, Tail &&... tail) {
wt(head);
if (sizeof...(Tail)) wt(' ');
print(forward<Tail>(tail)...);
}
// gcc expansion. called automaticall after main.
void __attribute__((destructor)) _d() { flush(); }
} // namespace fastio
using fastio::read;
using fastio::print;
using fastio::flush;
#if defined(LOCAL)
#define SHOW(...) SHOW_IMPL(__VA_ARGS__, SHOW6, SHOW5, SHOW4, SHOW3, SHOW2, SHOW1)(__VA_ARGS__)
#define SHOW_IMPL(_1, _2, _3, _4, _5, _6, NAME, ...) NAME
#define SHOW1(x) print(#x, "=", (x)), flush()
#define SHOW2(x, y) print(#x, "=", (x), #y, "=", (y)), flush()
#define SHOW3(x, y, z) print(#x, "=", (x), #y, "=", (y), #z, "=", (z)), flush()
#define SHOW4(x, y, z, w) print(#x, "=", (x), #y, "=", (y), #z, "=", (z), #w, "=", (w)), flush()
#define SHOW5(x, y, z, w, v) print(#x, "=", (x), #y, "=", (y), #z, "=", (z), #w, "=", (w), #v, "=", (v)), flush()
#define SHOW6(x, y, z, w, v, u) print(#x, "=", (x), #y, "=", (y), #z, "=", (z), #w, "=", (w), #v, "=", (v), #u, "=", (u)), flush()
#else
#define SHOW(...)
#endif
#define INT(...) \
int __VA_ARGS__; \
read(__VA_ARGS__)
#define LL(...) \
ll __VA_ARGS__; \
read(__VA_ARGS__)
#define U32(...) \
u32 __VA_ARGS__; \
read(__VA_ARGS__)
#define U64(...) \
u64 __VA_ARGS__; \
read(__VA_ARGS__)
#define STR(...) \
string __VA_ARGS__; \
read(__VA_ARGS__)
#define CHAR(...) \
char __VA_ARGS__; \
read(__VA_ARGS__)
#define DBL(...) \
double __VA_ARGS__; \
read(__VA_ARGS__)
#define VEC(type, name, size) \
vector<type> name(size); \
read(name)
#define VV(type, name, h, w) \
vector<vector<type>> name(h, vector<type>(w)); \
read(name)
void YES(bool t = 1) { print(t ? "YES" : "NO"); }
void NO(bool t = 1) { YES(!t); }
void Yes(bool t = 1) { print(t ? "Yes" : "No"); }
void No(bool t = 1) { Yes(!t); }
void yes(bool t = 1) { print(t ? "yes" : "no"); }
void no(bool t = 1) { yes(!t); }
void YA(bool t = 1) { print(t ? "YA" : "TIDAK"); }
void TIDAK(bool t = 1) { YES(!t); }
#line 4 "test/3_yukicoder/2587_2.test.cpp"
#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 "ds/hashmap.hpp"
// u64 -> Val
template <typename Val>
struct HashMap {
// n は入れたいものの個数で ok
HashMap(u32 n = 0) { build(n); }
void build(u32 n) {
u32 k = 8;
while (k < n * 2) k *= 2;
cap = k / 2, mask = k - 1;
key.resize(k), val.resize(k), used.assign(k, 0);
}
// size を保ったまま. size=0 にするときは build すること.
void clear() {
used.assign(len(used), 0);
cap = (mask + 1) / 2;
}
int size() { return len(used) / 2 - cap; }
int index(const u64& k) {
int i = 0;
for (i = hash(k); used[i] && key[i] != k; i = (i + 1) & mask) {}
return i;
}
Val& operator[](const u64& k) {
if (cap == 0) extend();
int i = index(k);
if (!used[i]) { used[i] = 1, key[i] = k, val[i] = Val{}, --cap; }
return val[i];
}
Val get(const u64& k, Val default_value) {
int i = index(k);
return (used[i] ? val[i] : default_value);
}
bool count(const u64& k) {
int i = index(k);
return used[i] && key[i] == k;
}
// f(key, val)
template <typename F>
void enumerate_all(F f) {
FOR(i, len(used)) if (used[i]) f(key[i], val[i]);
}
private:
u32 cap, mask;
vc<u64> key;
vc<Val> val;
vc<bool> used;
u64 hash(u64 x) {
static const u64 FIXED_RANDOM = std::chrono::steady_clock::now().time_since_epoch().count();
x += FIXED_RANDOM;
x = (x ^ (x >> 30)) * 0xbf58476d1ce4e5b9;
x = (x ^ (x >> 27)) * 0x94d049bb133111eb;
return (x ^ (x >> 31)) & mask;
}
void extend() {
vc<pair<u64, Val>> dat;
dat.reserve(len(used) / 2 - cap);
FOR(i, len(used)) {
if (used[i]) dat.eb(key[i], val[i]);
}
build(2 * len(dat));
for (auto& [a, b]: dat) (*this)[a] = b;
}
};
#line 3 "graph/base.hpp"
template <typename T>
struct Edge {
int frm, to;
T cost;
int id;
};
template <typename T = int, bool directed = false>
struct Graph {
static constexpr bool is_directed = directed;
int N, M;
using cost_type = T;
using edge_type = Edge<T>;
vector<edge_type> edges;
vector<int> indptr;
vector<edge_type> csr_edges;
vc<int> vc_deg, vc_indeg, vc_outdeg;
bool prepared;
class OutgoingEdges {
public:
OutgoingEdges(const Graph* G, int l, int r) : G(G), l(l), r(r) {}
const edge_type* begin() const {
if (l == r) { return 0; }
return &G->csr_edges[l];
}
const edge_type* end() const {
if (l == r) { return 0; }
return &G->csr_edges[r];
}
private:
const Graph* G;
int l, r;
};
bool is_prepared() { return prepared; }
Graph() : N(0), M(0), prepared(0) {}
Graph(int N) : N(N), M(0), prepared(0) {}
void build(int n) {
N = n, M = 0;
prepared = 0;
edges.clear();
indptr.clear();
csr_edges.clear();
vc_deg.clear();
vc_indeg.clear();
vc_outdeg.clear();
}
void add(int frm, int to, T cost = 1, int i = -1) {
assert(!prepared);
assert(0 <= frm && 0 <= to && to < N);
if (i == -1) i = M;
auto e = edge_type({frm, to, cost, i});
edges.eb(e);
++M;
}
#ifdef FASTIO
// wt, off
void read_tree(bool wt = false, int off = 1) { read_graph(N - 1, wt, off); }
void read_graph(int M, bool wt = false, int off = 1) {
for (int m = 0; m < M; ++m) {
INT(a, b);
a -= off, b -= off;
if (!wt) {
add(a, b);
} else {
T c;
read(c);
add(a, b, c);
}
}
build();
}
#endif
void build() {
assert(!prepared);
prepared = true;
indptr.assign(N + 1, 0);
for (auto&& e: edges) {
indptr[e.frm + 1]++;
if (!directed) indptr[e.to + 1]++;
}
for (int v = 0; v < N; ++v) { indptr[v + 1] += indptr[v]; }
auto counter = indptr;
csr_edges.resize(indptr.back() + 1);
for (auto&& e: edges) {
csr_edges[counter[e.frm]++] = e;
if (!directed) csr_edges[counter[e.to]++] = edge_type({e.to, e.frm, e.cost, e.id});
}
}
OutgoingEdges operator[](int v) const {
assert(prepared);
return {this, indptr[v], indptr[v + 1]};
}
vc<int> deg_array() {
if (vc_deg.empty()) calc_deg();
return vc_deg;
}
pair<vc<int>, vc<int>> deg_array_inout() {
if (vc_indeg.empty()) calc_deg_inout();
return {vc_indeg, vc_outdeg};
}
int deg(int v) {
if (vc_deg.empty()) calc_deg();
return vc_deg[v];
}
int in_deg(int v) {
if (vc_indeg.empty()) calc_deg_inout();
return vc_indeg[v];
}
int out_deg(int v) {
if (vc_outdeg.empty()) calc_deg_inout();
return vc_outdeg[v];
}
#ifdef FASTIO
void debug() {
print("Graph");
if (!prepared) {
print("frm to cost id");
for (auto&& e: edges) print(e.frm, e.to, e.cost, e.id);
} else {
print("indptr", indptr);
print("frm to cost id");
FOR(v, N) for (auto&& e: (*this)[v]) print(e.frm, e.to, e.cost, e.id);
}
}
#endif
vc<int> new_idx;
vc<bool> used_e;
// G における頂点 V[i] が、新しいグラフで i になるようにする
// {G, es}
// sum(deg(v)) の計算量になっていて、
// 新しいグラフの n+m より大きい可能性があるので注意
Graph<T, directed> rearrange(vc<int> V, bool keep_eid = 0) {
if (len(new_idx) != N) new_idx.assign(N, -1);
int n = len(V);
FOR(i, n) new_idx[V[i]] = i;
Graph<T, directed> G(n);
vc<int> history;
FOR(i, n) {
for (auto&& e: (*this)[V[i]]) {
if (len(used_e) <= e.id) used_e.resize(e.id + 1);
if (used_e[e.id]) continue;
int a = e.frm, b = e.to;
if (new_idx[a] != -1 && new_idx[b] != -1) {
history.eb(e.id);
used_e[e.id] = 1;
int eid = (keep_eid ? e.id : -1);
G.add(new_idx[a], new_idx[b], e.cost, eid);
}
}
}
FOR(i, n) new_idx[V[i]] = -1;
for (auto&& eid: history) used_e[eid] = 0;
G.build();
return G;
}
Graph<T, true> to_directed_tree(int root = -1) {
if (root == -1) root = 0;
assert(!is_directed && prepared && M == N - 1);
Graph<T, true> G1(N);
vc<int> par(N, -1);
auto dfs = [&](auto& dfs, int v) -> void {
for (auto& e: (*this)[v]) {
if (e.to == par[v]) continue;
par[e.to] = v, dfs(dfs, e.to);
}
};
dfs(dfs, root);
for (auto& e: edges) {
int a = e.frm, b = e.to;
if (par[a] == b) swap(a, b);
assert(par[b] == a);
G1.add(a, b, e.cost);
}
G1.build();
return G1;
}
HashMap<int> MP_FOR_EID;
int get_eid(u64 a, u64 b) {
if (len(MP_FOR_EID) == 0) {
MP_FOR_EID.build(N - 1);
for (auto& e: edges) {
u64 a = e.frm, b = e.to;
u64 k = to_eid_key(a, b);
MP_FOR_EID[k] = e.id;
}
}
return MP_FOR_EID.get(to_eid_key(a, b), -1);
}
u64 to_eid_key(u64 a, u64 b) {
if (!directed && a > b) swap(a, b);
return N * a + b;
}
private:
void calc_deg() {
assert(vc_deg.empty());
vc_deg.resize(N);
for (auto&& e: edges) vc_deg[e.frm]++, vc_deg[e.to]++;
}
void calc_deg_inout() {
assert(vc_indeg.empty());
vc_indeg.resize(N);
vc_outdeg.resize(N);
for (auto&& e: edges) { vc_indeg[e.to]++, vc_outdeg[e.frm]++; }
}
};
#line 7 "test/3_yukicoder/2587_2.test.cpp"
#line 2 "poly/poly_taylor_shift.hpp"
#line 2 "nt/primetable.hpp"
template <typename T = int>
vc<T> primetable(int LIM) {
++LIM;
const int S = 32768;
static int done = 2;
static vc<T> primes = {2}, sieve(S + 1);
if (done < LIM) {
done = LIM;
primes = {2}, sieve.assign(S + 1, 0);
const int R = LIM / 2;
primes.reserve(int(LIM / log(LIM) * 1.1));
vc<pair<int, int>> cp;
for (int i = 3; i <= S; i += 2) {
if (!sieve[i]) {
cp.eb(i, i * i / 2);
for (int j = i * i; j <= S; j += 2 * i) sieve[j] = 1;
}
}
for (int L = 1; L <= R; L += S) {
array<bool, S> block{};
for (auto& [p, idx]: cp)
for (int i = idx; i < S + L; idx = (i += p)) block[i - L] = 1;
FOR(i, min(S, R - L)) if (!block[i]) primes.eb((L + i) * 2 + 1);
}
}
int k = LB(primes, LIM + 1);
return {primes.begin(), primes.begin() + k};
}
#line 3 "mod/powertable.hpp"
// a^0, ..., a^N
template <typename mint>
vc<mint> powertable_1(mint a, ll N) {
// table of a^i
vc<mint> f(N + 1, 1);
FOR(i, N) f[i + 1] = a * f[i];
return f;
}
// 0^e, ..., N^e
template <typename mint>
vc<mint> powertable_2(ll e, ll N) {
auto primes = primetable(N);
vc<mint> f(N + 1, 1);
f[0] = mint(0).pow(e);
for (auto&& p: primes) {
if (p > N) break;
mint xp = mint(p).pow(e);
ll pp = p;
while (pp <= N) {
ll i = pp;
while (i <= N) {
f[i] *= xp;
i += pp;
}
pp *= p;
}
}
return f;
}
#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 5 "poly/poly_taylor_shift.hpp"
// f(x) -> f(x+c)
template <typename mint>
vc<mint> poly_taylor_shift(vc<mint> f, mint c) {
if (c == mint(0)) return f;
ll N = len(f);
FOR(i, N) f[i] *= fact<mint>(i);
auto b = powertable_1<mint>(c, N);
FOR(i, N) b[i] *= fact_inv<mint>(i);
reverse(all(f));
f = convolution(f, b);
f.resize(N);
reverse(all(f));
FOR(i, N) f[i] *= fact_inv<mint>(i);
return f;
}
#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 2 "poly/ntt_doubling.hpp"
#line 4 "poly/ntt_doubling.hpp"
// 2^k 次多項式の長さ 2^k が与えられるので 2^k+1 にする
template <typename mint, bool transposed = false>
void ntt_doubling(vector<mint>& a) {
static array<mint, 30> root;
static bool prepared = 0;
if (!prepared) {
prepared = 1;
const int rank2 = mint::ntt_info().fi;
root[rank2] = mint::ntt_info().se;
FOR_R(i, rank2) { root[i] = root[i + 1] * root[i + 1]; }
}
if constexpr (!transposed) {
const int M = (int)a.size();
auto b = a;
ntt(b, 1);
mint r = 1, zeta = root[topbit(2 * M)];
FOR(i, M) b[i] *= r, r *= zeta;
ntt(b, 0);
copy(begin(b), end(b), back_inserter(a));
} else {
const int M = len(a) / 2;
vc<mint> tmp = {a.begin(), a.begin() + M};
a = {a.begin() + M, a.end()};
transposed_ntt(a, 0);
mint r = 1, zeta = root[topbit(2 * M)];
FOR(i, M) a[i] *= r, r *= zeta;
transposed_ntt(a, 1);
FOR(i, M) a[i] += tmp[i];
}
}
#line 2 "poly/poly_divmod.hpp"
#line 4 "poly/poly_divmod.hpp"
template <typename mint>
pair<vc<mint>, vc<mint>> poly_divmod(vc<mint> f, vc<mint> g) {
assert(g.back() != 0);
if (len(f) < len(g)) { return {{}, f}; }
auto rf = f, rg = g;
reverse(all(rf)), reverse(all(rg));
ll deg = len(rf) - len(rg) + 1;
rf.resize(deg), rg.resize(deg);
rg = fps_inv(rg);
auto q = convolution(rf, rg);
q.resize(deg);
reverse(all(q));
auto h = convolution(q, g);
FOR(i, len(f)) f[i] -= h[i];
while (len(f) > 0 && f.back() == 0) f.pop_back();
return {q, f};
}
#line 4 "poly/coef_of_rational_fps.hpp"
template <typename mint>
mint coef_of_rational_fps_small(vector<mint> P, vector<mint> Q, ll N) {
assert(0 <= len(P) && len(P) + 1 == len(Q) && len(Q) <= 16 && Q[0] == mint(1));
if (P.empty()) return 0;
int m = len(Q) - 1;
vc<u32> Q32(m + 1);
FOR(i, m + 1) Q32[i] = (-Q[i]).val;
using poly = vc<u64>;
auto dfs = [&](auto& dfs, const ll N) -> poly {
// x^N mod G
if (N == 0) {
poly f(m);
f[0] = 1;
return f;
}
poly f = dfs(dfs, N / 2);
poly g(len(f) * 2 - 1 + (N & 1));
FOR(i, len(f)) FOR(j, len(f)) { g[i + j + (N & 1)] += f[i] * f[j]; }
FOR(i, len(g)) g[i] = mint(g[i]).val;
FOR_R(i, len(g)) {
g[i] = mint(g[i]).val;
if (i >= m) FOR(j, 1, len(Q)) g[i - j] += Q32[j] * g[i];
}
g.resize(m);
return g;
};
poly f = dfs(dfs, N);
FOR(i, m) FOR(j, 1, i + 1) { P[i] -= Q[j] * P[i - j]; }
u64 res = 0;
FOR(i, m) res += f[i] * P[i].val;
return res;
}
template <typename mint>
mint coef_of_rational_fps_ntt(vector<mint> P, vector<mint> Q, ll N) {
assert(0 <= len(P) && len(P) + 1 == len(Q) && Q[0] == mint(1));
if (P.empty()) return 0;
int n = 1;
while (n < len(Q)) n += n;
vc<mint> W(n);
{
vc<int> btr(n);
int log = topbit(n);
FOR(i, n) { btr[i] = (btr[i >> 1] >> 1) + ((i & 1) << (log - 1)); }
int t = mint::ntt_info().fi;
mint r = mint::ntt_info().se;
mint dw = r.inverse().pow((1 << t) / (2 * n));
mint w = inv<mint>(2);
for (auto& i: btr) { W[i] = w, w *= dw; }
}
P.resize(2 * n), Q.resize(2 * n);
ntt(P, 0), ntt(Q, 0);
while (N >= n) {
if (N % 2 == 0) {
FOR(i, n) { P[i] = (P[2 * i] * Q[2 * i + 1] + P[2 * i + 1] * Q[2 * i]) * inv<mint>(2); }
} else {
FOR(i, n) { P[i] = (P[2 * i] * Q[2 * i + 1] - P[2 * i + 1] * Q[2 * i]) * W[i]; }
}
FOR(i, n) Q[i] = Q[2 * i] * Q[2 * i + 1];
P.resize(n), Q.resize(n);
N /= 2;
if (N < n) break;
ntt_doubling(P), ntt_doubling(Q);
}
ntt(P, 1), ntt(Q, 1);
Q = fps_inv<mint>(Q);
mint ans = 0;
FOR(i, N + 1) ans += P[i] * Q[N - i];
return ans;
}
template <typename mint>
mint coef_of_rational_fps_convolution(vector<mint> P, vector<mint> Q, ll N) {
assert(0 <= len(P) && len(P) + 1 == len(Q) && Q[0] == mint(1));
if (P.empty()) return 0;
while (N >= len(P)) {
vc<mint> Q1 = Q;
FOR(i, len(Q1)) if (i & 1) Q1[i] = -Q1[i];
P = convolution(P, Q1);
Q = convolution(Q, Q1);
FOR(i, len(Q1)) Q[i] = Q[2 * i];
FOR(i, len(Q1) - 1) P[i] = P[2 * i | (N & 1)];
P.resize(len(Q1) - 1);
Q.resize(len(Q1));
N /= 2;
}
return fps_div(P, Q)[N];
}
template <typename mint>
mint coef_of_rational_fps(vector<mint> P, vector<mint> Q, ll N) {
if (P.empty()) return 0;
assert(len(Q) > 0 && Q[0] != mint(0));
while (Q.back() == mint(0)) POP(Q);
mint c = mint(1) / Q[0];
for (auto& x: P) x *= c;
for (auto& x: Q) x *= c;
mint base = 0;
if (len(P) >= len(Q)) {
auto [f, g] = poly_divmod<mint>(P, Q);
base = (N < len(f) ? f[N] : mint(0));
P = g;
}
P.resize(len(Q) - 1);
int n = len(Q);
if (mint::ntt_info().fi != -1) {
if (n <= 10) return base + coef_of_rational_fps_small(P, Q, N);
if (n > 10) return base + coef_of_rational_fps_ntt(P, Q, N);
}
mint x = (n <= 16 ? coef_of_rational_fps_small(P, Q, N) : coef_of_rational_fps_convolution(P, Q, N));
return base + x;
}
#line 1 "graph/tree_walk_generating_function.hpp"
#line 2 "graph/tree.hpp"
#line 4 "graph/tree.hpp"
// HLD euler tour をとっていろいろ。
template <typename GT>
struct Tree {
using Graph_type = GT;
GT &G;
using WT = typename GT::cost_type;
int N;
vector<int> LID, RID, head, V, parent, VtoE;
vc<int> depth;
vc<WT> depth_weighted;
Tree(GT &G, int r = 0, bool hld = 1) : G(G) { build(r, hld); }
void build(int r = 0, bool hld = 1) {
if (r == -1) return; // build を遅延したいとき
N = G.N;
LID.assign(N, -1), RID.assign(N, -1), head.assign(N, r);
V.assign(N, -1), parent.assign(N, -1), VtoE.assign(N, -1);
depth.assign(N, -1), depth_weighted.assign(N, 0);
assert(G.is_prepared());
int t1 = 0;
dfs_sz(r, -1, hld);
dfs_hld(r, t1);
}
void dfs_sz(int v, int p, bool hld) {
auto &sz = RID;
parent[v] = p;
depth[v] = (p == -1 ? 0 : depth[p] + 1);
sz[v] = 1;
int l = G.indptr[v], r = G.indptr[v + 1];
auto &csr = G.csr_edges;
// 使う辺があれば先頭にする
for (int i = r - 2; i >= l; --i) {
if (hld && depth[csr[i + 1].to] == -1) swap(csr[i], csr[i + 1]);
}
int hld_sz = 0;
for (int i = l; i < r; ++i) {
auto e = csr[i];
if (depth[e.to] != -1) continue;
depth_weighted[e.to] = depth_weighted[v] + e.cost;
VtoE[e.to] = e.id;
dfs_sz(e.to, v, hld);
sz[v] += sz[e.to];
if (hld && chmax(hld_sz, sz[e.to]) && l < i) { swap(csr[l], csr[i]); }
}
}
void dfs_hld(int v, int ×) {
LID[v] = times++;
RID[v] += LID[v];
V[LID[v]] = v;
bool heavy = true;
for (auto &&e: G[v]) {
if (depth[e.to] <= depth[v]) continue;
head[e.to] = (heavy ? head[v] : e.to);
heavy = false;
dfs_hld(e.to, times);
}
}
vc<int> heavy_path_at(int v) {
vc<int> P = {v};
while (1) {
int a = P.back();
for (auto &&e: G[a]) {
if (e.to != parent[a] && head[e.to] == v) {
P.eb(e.to);
break;
}
}
if (P.back() == a) break;
}
return P;
}
int heavy_child(int v) {
int k = LID[v] + 1;
if (k == N) return -1;
int w = V[k];
return (parent[w] == v ? w : -1);
}
int e_to_v(int eid) {
auto e = G.edges[eid];
return (parent[e.frm] == e.to ? e.frm : e.to);
}
int v_to_e(int v) { return VtoE[v]; }
int get_eid(int u, int v) {
if (parent[u] != v) swap(u, v);
assert(parent[u] == v);
return VtoE[u];
}
int ELID(int v) { return 2 * LID[v] - depth[v]; }
int ERID(int v) { return 2 * RID[v] - depth[v] - 1; }
// 目標地点へ進む個数が k
int LA(int v, int k) {
assert(k <= depth[v]);
while (1) {
int u = head[v];
if (LID[v] - k >= LID[u]) return V[LID[v] - k];
k -= LID[v] - LID[u] + 1;
v = parent[u];
}
}
int la(int u, int v) { return LA(u, v); }
int LCA(int u, int v) {
for (;; v = parent[head[v]]) {
if (LID[u] > LID[v]) swap(u, v);
if (head[u] == head[v]) return u;
}
}
int meet(int a, int b, int c) { return LCA(a, b) ^ LCA(a, c) ^ LCA(b, c); }
int lca(int u, int v) { return LCA(u, v); }
int subtree_size(int v, int root = -1) {
if (root == -1) return RID[v] - LID[v];
if (v == root) return N;
int x = jump(v, root, 1);
if (in_subtree(v, x)) return RID[v] - LID[v];
return N - RID[x] + LID[x];
}
int dist(int a, int b) {
int c = LCA(a, b);
return depth[a] + depth[b] - 2 * depth[c];
}
WT dist_weighted(int a, int b) {
int c = LCA(a, b);
return depth_weighted[a] + depth_weighted[b] - WT(2) * depth_weighted[c];
}
// a is in b
bool in_subtree(int a, int b) { return LID[b] <= LID[a] && LID[a] < RID[b]; }
int jump(int a, int b, ll k) {
if (k == 1) {
if (a == b) return -1;
return (in_subtree(b, a) ? LA(b, depth[b] - depth[a] - 1) : parent[a]);
}
int c = LCA(a, b);
int d_ac = depth[a] - depth[c];
int d_bc = depth[b] - depth[c];
if (k > d_ac + d_bc) return -1;
if (k <= d_ac) return LA(a, k);
return LA(b, d_ac + d_bc - k);
}
vc<int> collect_child(int v) {
vc<int> res;
for (auto &&e: G[v])
if (e.to != parent[v]) res.eb(e.to);
return res;
}
vc<int> collect_light(int v) {
vc<int> res;
bool skip = true;
for (auto &&e: G[v])
if (e.to != parent[v]) {
if (!skip) res.eb(e.to);
skip = false;
}
return res;
}
vc<pair<int, int>> get_path_decomposition(int u, int v, bool edge) {
// [始点, 終点] の"閉"区間列。
vc<pair<int, int>> up, down;
while (1) {
if (head[u] == head[v]) break;
if (LID[u] < LID[v]) {
down.eb(LID[head[v]], LID[v]);
v = parent[head[v]];
} else {
up.eb(LID[u], LID[head[u]]);
u = parent[head[u]];
}
}
if (LID[u] < LID[v]) down.eb(LID[u] + edge, LID[v]);
elif (LID[v] + edge <= LID[u]) up.eb(LID[u], LID[v] + edge);
reverse(all(down));
up.insert(up.end(), all(down));
return up;
}
// 辺の列の情報 (frm,to,str)
// str = "heavy_up", "heavy_down", "light_up", "light_down"
vc<tuple<int, int, string>> get_path_decomposition_detail(int u, int v) {
vc<tuple<int, int, string>> up, down;
while (1) {
if (head[u] == head[v]) break;
if (LID[u] < LID[v]) {
if (v != head[v]) down.eb(head[v], v, "heavy_down"), v = head[v];
down.eb(parent[v], v, "light_down"), v = parent[v];
} else {
if (u != head[u]) up.eb(u, head[u], "heavy_up"), u = head[u];
up.eb(u, parent[u], "light_up"), u = parent[u];
}
}
if (LID[u] < LID[v]) down.eb(u, v, "heavy_down");
elif (LID[v] < LID[u]) up.eb(u, v, "heavy_up");
reverse(all(down));
concat(up, down);
return up;
}
vc<int> restore_path(int u, int v) {
vc<int> P;
for (auto &&[a, b]: get_path_decomposition(u, v, 0)) {
if (a <= b) {
FOR(i, a, b + 1) P.eb(V[i]);
} else {
FOR_R(i, b, a + 1) P.eb(V[i]);
}
}
return P;
}
// path [a,b] と [c,d] の交わり. 空ならば {-1,-1}.
// https://codeforces.com/problemset/problem/500/G
pair<int, int> path_intersection(int a, int b, int c, int d) {
int ab = lca(a, b), ac = lca(a, c), ad = lca(a, d);
int bc = lca(b, c), bd = lca(b, d), cd = lca(c, d);
int x = ab ^ ac ^ bc, y = ab ^ ad ^ bd; // meet(a,b,c), meet(a,b,d)
if (x != y) return {x, y};
int z = ac ^ ad ^ cd;
if (x != z) x = -1;
return {x, x};
}
// uv path 上で check(v) を満たす最後の v
// なければ （つまり check(v) が ng ）-1
template <class F>
int max_path(F check, int u, int v) {
if (!check(u)) return -1;
auto pd = get_path_decomposition(u, v, false);
for (auto [a, b]: pd) {
if (!check(V[a])) return u;
if (check(V[b])) {
u = V[b];
continue;
}
int c = binary_search([&](int c) -> bool { return check(V[c]); }, a, b, 0);
return V[c];
}
return u;
}
};
#line 2 "graph/ds/static_toptree.hpp"
/*
参考 joitour tatyam
クラスタは根が virtual なもののみであるような簡易版
N 個の (頂+辺) をマージしていって，木全体＋根から親への辺とする．
single(v) : v とその親辺を合わせたクラスタ
rake(L,R) : L の boundary を維持
compress(L,R) (top-down) 順に x,y
*/
template <typename TREE>
struct Static_TopTree {
int N;
TREE &tree;
vc<int> par, lch, rch, A, B; // A, B boundary (top-down)
vc<bool> is_compress;
Static_TopTree(TREE &tree) : tree(tree) { build(); }
void build() {
N = tree.N;
par.assign(N, -1), lch.assign(N, -1), rch.assign(N, -1), A.assign(N, -1), B.assign(N, -1), is_compress.assign(N, 0);
FOR(v, N) { A[v] = tree.parent[v], B[v] = v; }
build_dfs(tree.V[0]);
assert(len(par) == 2 * N - 1);
}
// 木全体での集約値を得る
// single(v) : v とその親辺を合わせたクラスタ
// rake(x, y, u, v) uv(top down) が boundary になるように rake (maybe v=-1)
// compress(x,y,a,b,c) (top-down) 順に (a,b] + (b,c]
template <typename TREE_DP, typename F>
typename TREE_DP::value_type tree_dp(F single) {
using Data = typename TREE_DP::value_type;
auto dfs = [&](auto &dfs, int k) -> Data {
if (0 <= k && k < N) return single(k);
Data x = dfs(dfs, lch[k]), y = dfs(dfs, rch[k]);
if (is_compress[k]) {
assert(B[lch[k]] == A[rch[k]]);
return TREE_DP::compress(x, y);
}
return TREE_DP::rake(x, y);
};
return dfs(dfs, 2 * N - 2);
}
private:
int new_node(int l, int r, int a, int b, bool c) {
int v = len(par);
par.eb(-1), lch.eb(l), rch.eb(r), A.eb(a), B.eb(b), is_compress.eb(c);
par[l] = par[r] = v;
return v;
}
// height, node idx
// compress 参考：https://atcoder.jp/contests/abc351/editorial/9910
// ただし heavy path の選び方までは考慮しない
pair<int, int> build_dfs(int v) {
assert(tree.head[v] == v);
auto path = tree.heavy_path_at(v);
vc<pair<int, int>> stack;
stack.eb(0, path[0]);
auto merge_last_two = [&]() -> void {
auto [h2, k2] = POP(stack);
auto [h1, k1] = POP(stack);
stack.eb(max(h1, h2) + 1, new_node(k1, k2, A[k1], B[k2], true));
};
FOR(i, 1, len(path)) {
pqg<pair<int, int>> que;
int k = path[i];
que.emplace(0, k);
for (auto &c: tree.collect_light(path[i - 1])) { que.emplace(build_dfs(c)); }
while (len(que) >= 2) {
auto [h1, i1] = POP(que);
auto [h2, i2] = POP(que);
if (i2 == k) swap(i1, i2);
int i3 = new_node(i1, i2, A[i1], B[i1], false);
if (k == i1) k = i3;
que.emplace(max(h1, h2) + 1, i3);
}
stack.eb(POP(que));
while (1) {
int n = len(stack);
if (n >= 3 && (stack[n - 3].fi == stack[n - 2].fi || stack[n - 3].fi <= stack[n - 1].fi)) {
auto [h3, k3] = POP(stack);
merge_last_two(), stack.eb(h3, k3);
}
elif (n >= 2 && stack[n - 2].fi <= stack[n - 1].fi) { merge_last_two(); }
else break;
}
}
while (len(stack) >= 2) { merge_last_two(); }
return POP(stack);
}
};
#line 3 "graph/shortest_path/bfs01.hpp"
template <typename T, typename GT>
pair<vc<T>, vc<int>> bfs01(GT& G, int v) {
assert(G.is_prepared());
int N = G.N;
vc<T> dist(N, infty<T>);
vc<int> par(N, -1);
deque<int> que;
dist[v] = 0;
que.push_front(v);
while (!que.empty()) {
auto v = que.front();
que.pop_front();
for (auto&& e: G[v]) {
if (dist[e.to] == infty<T> || dist[e.to] > dist[e.frm] + e.cost) {
dist[e.to] = dist[e.frm] + e.cost;
par[e.to] = e.frm;
if (e.cost == 0)
que.push_front(e.to);
else
que.push_back(e.to);
}
}
}
return {dist, par};
}
// 多点スタート。[dist, par, root]
template <typename T, typename GT>
tuple<vc<T>, vc<int>, vc<int>> bfs01(GT& G, vc<int> vs) {
assert(G.is_prepared());
int N = G.N;
vc<T> dist(N, infty<T>);
vc<int> par(N, -1);
vc<int> root(N, -1);
deque<int> que;
for (auto&& v: vs) {
dist[v] = 0;
root[v] = v;
que.push_front(v);
}
while (!que.empty()) {
auto v = que.front();
que.pop_front();
for (auto&& e: G[v]) {
if (dist[e.to] == infty<T> || dist[e.to] > dist[e.frm] + e.cost) {
dist[e.to] = dist[e.frm] + e.cost;
root[e.to] = root[e.frm];
par[e.to] = e.frm;
if (e.cost == 0)
que.push_front(e.to);
else
que.push_back(e.to);
}
}
}
return {dist, par, root};
}
#line 2 "ds/unionfind/unionfind.hpp"
struct UnionFind {
int n, n_comp;
vc<int> dat; // par or (-size)
UnionFind(int n = 0) { build(n); }
void build(int m) {
n = m, n_comp = m;
dat.assign(n, -1);
}
void reset() { build(n); }
int operator[](int x) {
while (dat[x] >= 0) {
int pp = dat[dat[x]];
if (pp < 0) { return dat[x]; }
x = dat[x] = pp;
}
return x;
}
ll size(int x) {
x = (*this)[x];
return -dat[x];
}
bool merge(int x, int y) {
x = (*this)[x], y = (*this)[y];
if (x == y) return false;
if (-dat[x] < -dat[y]) swap(x, y);
dat[x] += dat[y], dat[y] = x, n_comp--;
return true;
}
vc<int> get_all() {
vc<int> A(n);
FOR(i, n) A[i] = (*this)[i];
return A;
}
};
#line 5 "graph/characteristic_polynomial_of_tree_adjacency_matrix.hpp"
template <typename mint>
struct TREE_ADJ_MATRIX_DP {
using poly = vc<mint>;
using Data = array<array<poly, 2>, 2>;
using value_type = Data;
static void add(poly& f, poly g) {
if (len(f) < len(g)) f.resize(len(g));
FOR(i, len(g)) f[i] += g[i];
};
static Data rake(Data L, Data R) {
Data Z;
add(Z[0][0], convolution(L[0][0], R[0][1]));
add(Z[0][1], convolution(L[0][1], R[0][1]));
add(Z[1][0], convolution(L[0][0], R[1][1]));
add(Z[1][1], convolution(L[0][1], R[1][1]));
add(Z[1][0], convolution(L[1][0], R[0][1]));
add(Z[1][1], convolution(L[1][1], R[0][1]));
return Z;
}
static Data compress(Data L, Data R) {
Data Z;
FOR(p, 2) FOR(q, 2) FOR(r, 2) { add(Z[p][r], convolution<mint>(L[p][q], R[1 - q][r])); }
return Z;
}
};
// det(I-xA) の計算 (固有多項式の reverse になっている)
// weight(i,j)：A[i][j]
// 偶数次だけしか出てこないので loop ありより高速
template <typename mint, typename F>
vc<mint> characteristic_poly_of_tree_adjacency_matrix_not_allow_loop(Graph<int, 0>& G, F weight) {
using poly = vc<mint>;
Tree<Graph<int, 0>> tree(G);
Static_TopTree<decltype(tree)> STT(tree);
// u, v はもう計算したか
using Data = array<array<poly, 2>, 2>;
auto single = [&](int v) -> Data {
Data X;
int p = tree.parent[v];
mint wt = (p == -1 ? mint(0) : weight(p, v) * weight(v, p));
X[0][0] = poly{mint(1)};
X[0][1] = poly{mint(1)}; // loop
if (p != -1) X[1][1] = poly{mint(0), -wt}; // match
return X;
};
Data X = STT.tree_dp<TREE_ADJ_MATRIX_DP<mint>>(single);
vc<mint> ANS(G.N + 1);
FOR(i, len(X[0][1])) { ANS[2 * i] += X[0][1][i]; }
return ANS;
}
template <typename mint, typename F>
vc<mint> characteristic_poly_of_tree_adjacency_matrix_allow_loop(Graph<int, 0>& G, F weight) {
using poly = vc<mint>;
Tree<Graph<int, 0>> tree(G);
Static_TopTree<decltype(tree)> STT(tree);
using Data = array<array<poly, 2>, 2>;
auto single = [&](int v) -> Data {
Data X;
int p = tree.parent[v];
mint wt = (p == -1 ? mint(0) : weight(p, v) * weight(v, p));
X[0][0] = poly{mint(1)};
X[0][1] = poly{mint(1), -weight(v, v)}; // loop
if (p != -1) X[1][1] = poly{mint(0), mint(0), -wt}; // match
return X;
};
Data X = STT.tree_dp<TREE_ADJ_MATRIX_DP<mint>>(single);
vc<mint> ANS(G.N + 1);
FOR(i, len(X[0][1])) { ANS[i] += X[0][1][i]; }
return ANS;
}
// det(I-xA) の計算 (固有多項式の reverse になっている)
// weight(i,j)：A[i][j]
template <bool ALLOW_LOOP, typename mint, typename F>
vc<mint> characteristic_poly_of_tree_adjacency_matrix(Graph<int, 0>& G, F weight) {
if constexpr (ALLOW_LOOP) {
return characteristic_poly_of_tree_adjacency_matrix_allow_loop<mint>(G, weight);
} else {
return characteristic_poly_of_tree_adjacency_matrix_not_allow_loop<mint>(G, weight);
}
}
#line 2 "poly/convolution_all.hpp"
#line 5 "poly/convolution_all.hpp"
template <typename T>
vc<T> convolution_all(vc<vc<T>>& polys) {
if (len(polys) == 0) return {T(1)};
while (1) {
int n = len(polys);
if (n == 1) break;
int m = ceil(n, 2);
FOR(i, m) {
if (2 * i + 1 == n) {
polys[i] = polys[2 * i];
} else {
polys[i] = convolution(polys[2 * i], polys[2 * i + 1]);
}
}
polys.resize(m);
}
return polys[0];
}
// product of 1-A[i]x
template <typename mint>
vc<mint> convolution_all_1(vc<mint> A) {
if (!mint::can_ntt()) {
vvc<mint> polys;
for (auto& a: A) polys.eb(vc<mint>({mint(1), -a}));
return convolution_all(polys);
}
int D = 6;
using poly = vc<mint>;
int n = 1;
while (n < len(A)) n *= 2;
int k = topbit(n);
vc<mint> F(n), nxt_F(n);
FOR(i, len(A)) F[i] = -A[i];
FOR(d, k) {
int b = 1 << d;
if (d < D) {
fill(all(nxt_F), mint(0));
for (int L = 0; L < n; L += 2 * b) {
FOR(i, b) FOR(j, b) { nxt_F[L + i + j] += F[L + i] * F[L + b + j]; }
FOR(i, b) nxt_F[L + b + i] += F[L + i] + F[L + b + i];
}
}
elif (d == D) {
for (int L = 0; L < n; L += 2 * b) {
poly f1 = {F.begin() + L, F.begin() + L + b};
poly f2 = {F.begin() + L + b, F.begin() + L + 2 * b};
f1.resize(2 * b), f2.resize(2 * b), ntt(f1, 0), ntt(f2, 0);
FOR(i, b) nxt_F[L + i] = f1[i] * f2[i] + f1[i] + f2[i];
FOR(i, b, 2 * b) nxt_F[L + i] = f1[i] * f2[i] - f1[i] - f2[i];
}
}
else {
for (int L = 0; L < n; L += 2 * b) {
poly f1 = {F.begin() + L, F.begin() + L + b};
poly f2 = {F.begin() + L + b, F.begin() + L + 2 * b};
ntt_doubling(f1), ntt_doubling(f2);
FOR(i, b) nxt_F[L + i] = f1[i] * f2[i] + f1[i] + f2[i];
FOR(i, b, 2 * b) nxt_F[L + i] = f1[i] * f2[i] - f1[i] - f2[i];
}
}
swap(F, nxt_F);
}
if (k - 1 >= D) ntt(F, 1);
F.eb(1), reverse(all(F));
F.resize(len(A) + 1);
return F;
}
#line 4 "graph/tree_walk_generating_function.hpp"
// ループなし：1600ms(N=10^5)
// ループあり：3300ms(N=10^5)
template <bool ALLOW_LOOP, typename mint, typename F>
pair<vc<mint>, vc<mint>> tree_walk_generating_function(Graph<int, 0>& G, int s, int t, F weight) {
int N = G.N;
// 分母
auto f = characteristic_poly_of_tree_adjacency_matrix<ALLOW_LOOP, mint>(G, weight);
// 分子
// (s,t) パスに沿って成分をかけたものの符号調整 + 他の成分
using poly = vc<mint>;
vc<poly> polys;
pair<int, mint> path_poly = {0, mint(1)};
vc<bool> on_path(N);
auto [dist, par] = bfs01<int>(G, s);
on_path[t] = 1;
while (t != s) {
mint w = weight(par[t], t);
t = par[t], on_path[t] = 1;
path_poly.fi += 1, path_poly.se *= w; // +wx
}
UnionFind uf(N);
for (auto& e: G.edges) {
if (on_path[e.frm] || on_path[e.to]) continue;
uf.merge(e.frm, e.to);
}
vvc<int> comp(N);
FOR(v, N) comp[uf[v]].eb(v);
FOR(r, N) {
if (on_path[r] || uf[r] != r) continue;
vc<int>& V = comp[r];
Graph<int, 0> H = G.rearrange(V);
poly f = characteristic_poly_of_tree_adjacency_matrix<ALLOW_LOOP, mint>(H, [&](int i, int j) -> mint { return weight(V[i], V[j]); });
polys.eb(f);
}
poly B = convolution_all<mint>(polys);
int m = path_poly.fi;
poly g(len(B) + m);
FOR(i, len(B)) g[m + i] = path_poly.se * B[i];
return {g, f};
}
#line 11 "test/3_yukicoder/2587_2.test.cpp"
using mint = modint998;
using poly = vc<mint>;
void solve() {
LL(N, M, S, T);
--S, --T;
Graph<int, 0> G(N);
G.read_tree();
auto [g, f] = tree_walk_generating_function<false, mint>(
G, S, T, [&](int i, int j) -> mint { return 1; });
// W(x) = g(x)/f(x)
// ANS = sum_m binom(M,m)[x^m]W(x)
// EGF(W) = sum_m W_m 1/m! x^m
// [x^M]e^x EGF(W)
// 根が全部 -1 される
// ANS も C-rec で、f(x-1)
vc<mint> W = fps_div(g, f);
FOR(m, N) W[m] *= fact_inv<mint>(m);
vc<mint> tmp(N + 1);
FOR(m, N + 1) tmp[m] = fact_inv<mint>(m);
vc<mint> F = convolution(W, tmp);
F.resize(N);
FOR(i, N) F[i] *= fact<mint>(i);
reverse(all(f));
f = poly_taylor_shift<mint>(f, -1);
reverse(all(f));
g = convolution<mint>(F, f);
g.resize(N);
mint ANS = coef_of_rational_fps<mint>(g, f, M);
print(ANS);
}
signed main() {
int T = 1;
// INT(T);
FOR(T) solve();
return 0;
}