#line 1 "library/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 FOR_subset(t, s) for (ll t = (s); t >= 0; t = (t == 0 ? -1 : (t - 1) & (s)))
#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_mod_2(int x) { return __builtin_parity(x); }
int popcnt_mod_2(u32 x) { return __builtin_parity(x); }
int popcnt_mod_2(ll x) { return __builtin_parityll(x); }
int popcnt_mod_2(u64 x) { return __builtin_parityll(x); }
// (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 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 "library/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); }
#line 3 "main.cpp"

#line 2 "library/convex/larsch.hpp"

// https://noshi91.github.io/Library/algorithm/larsch.cpp.html
template <class T>
class LARSCH {
  struct reduce_row;
  struct reduce_col;

  struct reduce_row {
    int n;
    std::function<T(int, int)> f;
    int cur_row;
    int state;
    std::unique_ptr<reduce_col> rec;

    reduce_row(int n_) : n(n_), f(), cur_row(0), state(0), rec() {
      const int m = n / 2;
      if (m != 0) { rec = std::make_unique<reduce_col>(m); }
    }

    void set_f(std::function<T(int, int)> f_) {
      f = f_;
      if (rec) {
        rec->set_f([&](int i, int j) -> T { return f(2 * i + 1, j); });
      }
    }

    int get_argmin() {
      const int cur_row_ = cur_row;
      cur_row += 1;
      if (cur_row_ % 2 == 0) {
        const int prev_argmin = state;
        const int next_argmin = [&]() {
          if (cur_row_ + 1 == n) {
            return n - 1;
          } else {
            return rec->get_argmin();
          }
        }();
        state = next_argmin;
        int ret = prev_argmin;
        for (int j = prev_argmin + 1; j <= next_argmin; j += 1) {
          if (f(cur_row_, ret) > f(cur_row_, j)) { ret = j; }
        }
        return ret;
      } else {
        if (f(cur_row_, state) <= f(cur_row_, cur_row_)) {
          return state;
        } else {
          return cur_row_;
        }
      }
    }
  };

  struct reduce_col {
    int n;
    std::function<T(int, int)> f;
    int cur_row;
    std::vector<int> cols;
    reduce_row rec;

    reduce_col(int n_) : n(n_), f(), cur_row(0), cols(), rec(n) {}

    void set_f(std::function<T(int, int)> f_) {
      f = f_;
      rec.set_f([&](int i, int j) -> T { return f(i, cols[j]); });
    }

    int get_argmin() {
      const int cur_row_ = cur_row;
      cur_row += 1;
      const auto cs = [&]() -> std::vector<int> {
        if (cur_row_ == 0) {
          return {{0}};
        } else {
          return {{2 * cur_row_ - 1, 2 * cur_row_}};
        }
      }();
      for (const int j: cs) {
        while ([&]() {
          const int size = cols.size();
          return size != cur_row_ && f(size - 1, cols.back()) > f(size - 1, j);
        }()) {
          cols.pop_back();
        }
        if (int(cols.size()) != n) { cols.push_back(j); }
      }
      return cols[rec.get_argmin()];
    }
  };

  std::unique_ptr<reduce_row> base;

public:
  LARSCH(int n, std::function<T(int, int)> f)
      : base(std::make_unique<reduce_row>(n)) {
    base->set_f(f);
  }

  int get_argmin() { return base->get_argmin(); }
};
#line 2 "library/convex/smawk.hpp"

// select(i,j,k) は (i,j) -> (i,k) を行うかどうか
// 残念ながら monotone minima より高速な場合が存在しない説がある
// https://codeforces.com/contest/1423/problem/M
template <typename F>
vc<int> smawk(int H, int W, F select) {
  auto dfs = [&](auto& dfs, vc<int> X, vc<int> Y) -> vc<int> {
    int N = len(X);
    if (N == 0) return {};
    vc<int> YY;
    for (auto&& y: Y) {
      while (len(YY)) {
        int py = YY.back(), x = X[len(YY) - 1];
        if (!select(x, py, y)) break;
        YY.pop_back();
      }
      if (len(YY) < len(X)) YY.eb(y);
    }
    vc<int> XX;
    FOR(i, 1, len(X), 2) XX.eb(X[i]);
    vc<int> II = dfs(dfs, XX, YY);
    vc<int> I(N);
    FOR(i, len(II)) I[i + i + 1] = II[i];
    int p = 0;
    FOR(i, 0, N, 2) {
      int LIM = (i + 1 == N ? Y.back() : I[i + 1]);
      int best = Y[p];
      while (Y[p] < LIM) {
        ++p;
        if (select(X[i], best, Y[p])) best = Y[p];
      }
      I[i] = best;
    }
    return I;
  };
  vc<int> X(H), Y(W);
  iota(all(X), 0), iota(all(Y), 0);
  return dfs(dfs, X, Y);
}
#line 1 "library/other/fibonacci_search.hpp"
// returns: {fx, x}
// [L, R) での極小値をひとつ求める、単峰は不要
template <typename T, bool MINIMIZE, typename F>
pair<T, ll> fibonacci_search(F f, ll L, ll R) {
  assert(L < R);
  --R;
  ll a = L, b = L + 1, c = L + 2, d = L + 3;
  int n = 0;
  while (d < R) { b = c, c = d, d = b + c - a, ++n; }
  auto get = [&](ll x) -> T {
    if (R < x) return infty<T>;
    return (MINIMIZE ? f(x) : -f(x));
  };
  T ya = get(a), yb = get(b), yc = get(c), yd = get(d);
  // この中で極小ならば全体でも極小、を維持する
  FOR(n) {
    if (yb <= yc) {
      d = c, c = b, b = a + d - c;
      yd = yc, yc = yb, yb = get(b);
    } else {
      a = b, b = c, c = a + d - b;
      ya = yb, yb = yc, yc = get(c);
    }
  }
  ll x = a;
  T y = ya;
  if (chmin(y, yb)) x = b;
  if (chmin(y, yc)) x = c;
  if (chmin(y, yd)) x = d;
  if (MINIMIZE) return {y, x};
  return {-y, x};
}
#line 4 "library/convex/monge.hpp"

// 定義域 [0, N] の範囲で f の monge 性を確認
template <typename T, typename F>
bool check_monge(int N, F f) {
  FOR(l, N + 1) FOR(k, l) FOR(j, k) FOR(i, j) {
    T lhs = f(i, l) + f(j, k);
    T rhs = f(i, k) + f(j, l);
    if (lhs < rhs) {
      print("monge ng");
      print(i, j, k, l, f(i, k), f(i, l), f(j, k), f(j, l), lhs, rhs);
      return false;
    }
  }
  print("monge ok");
  return true;
}

// newdp[j] = min (dp[i] + f(i,j))
template <typename T, typename F>
vc<T> monge_dp_update(int N, vc<T>& dp, F f) {
  assert(len(dp) == N + 1);
  auto select = [&](int i, int j, int k) -> int {
    if (i <= k) return j;
    return (dp[j] + f(j, i) > dp[k] + f(k, i) ? k : j);
  };
  vc<int> I = SMAWK(N + 1, N + 1, select);
  vc<T> newdp(N + 1, infty<T>);
  FOR(j, N + 1) {
    int i = I[j];
    chmin(newdp[j], dp[i] + f(i, j));
  }
  return newdp;
}

// 遷移回数を問わない場合
template <typename T, typename F>
vc<T> monge_shortest_path(int N, F f) {
  vc<T> dp(N + 1, infty<T>);
  dp[0] = 0;
  LARSCH<T> larsch(N, [&](int i, int j) -> T {
    ++i;
    if (i <= j) return infty<T>;
    return dp[j] + f(j, i);
  });
  FOR(r, 1, N + 1) {
    int l = larsch.get_argmin();
    dp[r] = dp[l] + f(l, r);
  }
  return dp;
}

// https://noshi91.github.io/algorithm-encyclopedia/d-edge-shortest-path-monge
// |f| の上限 f_lim も渡す
// ・larsch が結構重いので、自前で dp できるならその方がよい
// ・複数の d で計算するとき：同じ lambda
// に対する計算をメモ化しておくと定数倍高速？ 　・ABC305
template <typename T, typename F>
T monge_shortest_path_d_edge(int N, int d, T f_lim, F f) {
  assert(d <= N);
  auto calc_L = [&](T lambda) -> T {
    auto cost = [&](int frm, int to) -> T { return f(frm, to) + lambda; };
    vc<T> dp = monge_shortest_path<T>(N, cost);
    return dp[N] - lambda * d;
  };

  auto [x, fx] = fibonacci_search<T, false>(calc_L, -3 * f_lim, 3 * f_lim + 1);
  return fx;
}

// https://topcoder-g-hatena-ne-jp.jag-icpc.org/spaghetti_source/20120915/1347668163.html
// Prop 1
// 上三角 monge A, B
// C[i][j] = min_k (A[i][k] + B[k][j])
template <typename T, typename F1, typename F2>
vvc<T> monge_matrix_product(int N, F1 A, F2 B) {
  vv(T, C, N + 1, N + 1, infty<T>);
  vc<int> K(N + 1);
  FOR(i, N + 1) C[i][i] = A(i, i) + B(i, i), K[i] = i;
  FOR(s, 1, N + 1) {
    vc<int> newK(N + 1 - s);
    FOR(i, N + 1 - s) {
      int j = i + s;
      int p = K[i], q = K[i + 1];
      FOR(k, p, q + 1) if (chmin(C[i][j], A(i, k) + B(k, j))) newK[i] = k;
    }
    swap(K, newK);
  }
  return C;
}
#line 1 "library/other/golden_search.hpp"
// return : (fx, x)
// f が評価される回数：2 + iter
// 幅が 1/phi^{iter} 倍になる. iter = 44: 1e-9.
template <typename Re, bool MINIMIZE, typename F>
pair<Re, Re> golden_search(F f, Re lx, Re rx, int iter = 50) {
  assert(lx <= rx);
  Re inv_phi = (sqrtl(5) - 1.0) * 0.5;
  Re inv_phi_2 = inv_phi * inv_phi;
  Re x1 = lx, x4 = rx;
  Re x2 = x1 + inv_phi_2 * (x4 - x1);
  Re x3 = x1 + inv_phi * (x4 - x1);
  Re y2 = f(x2), y3 = f(x3);
  auto comp = [&](Re a, Re b) -> bool {
    if constexpr (MINIMIZE)
      return a < b;
    else
      return a > b;
  };

  FOR(iter) {
    if (comp(y2, y3)) {
      x4 = x3, x3 = x2, y3 = y2;
      x2 = x1 + inv_phi_2 * (x4 - x1);
      y2 = f(x2);
    } else {
      x1 = x2, x2 = x3, y2 = y3;
      x3 = x1 + inv_phi * (x4 - x1);
      y3 = f(x3);
    }
  }
  return (comp(y2, y3) ? pair<Re, Re>{y2, x2} : pair<Re, Re>{y3, x3});
}
#line 2 "library/random/base.hpp"

u64 RNG_64() {
  static u64 x_ = u64(chrono::duration_cast<chrono::nanoseconds>(chrono::high_resolution_clock::now().time_since_epoch()).count()) * 10150724397891781847ULL;
  x_ ^= x_ << 7;
  return x_ ^= x_ >> 9;
}

u64 RNG(u64 lim) { return RNG_64() % lim; }

ll RNG(ll l, ll r) { return l + RNG_64() % (r - l); }
#line 7 "main.cpp"

using Re = double;

template <typename T, typename F>
T monge_shortest_path_d_edge_real(int N, int d, T f_lim, F f) {
  assert(d <= N);
  auto calc_L = [&](T lambda) -> T {
    auto cost = [&](int frm, int to) -> T { return f(frm, to) + lambda; };
    vc<T> dp = monge_shortest_path<T>(N, cost);
    return dp[N] - lambda * d;
  };

  auto [fx, x] = golden_search<T, false>(calc_L, -3 * f_lim, 3 * f_lim + 1, 100);
  return fx;
}

void solve() {
  LL(N, M);
  vc<Re> dp(N + 1, infty<Re>);
  dp[0] = 0;
  VEC(ll, A, N);
  sort(all(A));

  auto Ac = cumsum<ll>(A);
  auto cost = [&](int i, int j) -> Re { return sqrtl((j - i) * (Ac[j] - Ac[i])); };
  auto ANS = monge_shortest_path_d_edge_real<Re>(N, M, 1e15, cost);
  print(ANS);
}

void test() {
  int N = RNG(1, 10);
  vi A(N);
  FOR(i, N) A[i] = RNG(1, 1000);
  sort(all(A));

  auto Ac = cumsum<ll>(A);
  auto cost = [&](int i, int j) -> Re { return sqrtl((j - i) * (Ac[j] - Ac[i])); };
  bool ok = check_monge<Re>(N, cost);
  assert(ok);
}

signed main() {
  // FOR(10000) test();
  // return 0;
  solve();
}