File size: 10,569 Bytes
1fd0050
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
#include <bits/stdc++.h>
using namespace std;

struct TestCase {
    int n;
    long long k;
    vector<vector<long long>> A;
    long long answer;
};

mt19937_64 rng_gen(42);

TestCase gen_matrix(int n, long long k, function<long long(int,int)> valfn) {
    TestCase tc;
    tc.n = n; tc.k = k;
    tc.A.assign(n+1, vector<long long>(n+1, 0));
    vector<long long> all;
    for (int i = 1; i <= n; i++)
        for (int j = 1; j <= n; j++) {
            tc.A[i][j] = valfn(i, j);
            all.push_back(tc.A[i][j]);
        }
    sort(all.begin(), all.end());
    tc.answer = all[k-1];
    return tc;
}

TestCase gen_multiplicative(int n, long long k) { return gen_matrix(n, k, [](int i, int j) -> long long { return (long long)i * j; }); }
TestCase gen_shifted(int n, long long k) { return gen_matrix(n, k, [n](int i, int j) -> long long { return (long long)(i + n) * (j + n); }); }
TestCase gen_additive(int n, long long k) { return gen_matrix(n, k, [](int i, int j) -> long long { return i + j; }); }
TestCase gen_random_sorted(int n, long long k) {
    TestCase tc;
    tc.n = n; tc.k = k;
    tc.A.assign(n+1, vector<long long>(n+1, 0));
    for (int i = 1; i <= n; i++)
        for (int j = 1; j <= n; j++)
            tc.A[i][j] = (long long)i * 1000000 + (long long)j * 1000 + (rng_gen() % 500);
    for (int i = 1; i <= n; i++)
        for (int j = 2; j <= n; j++)
            tc.A[i][j] = max(tc.A[i][j], tc.A[i][j-1]);
    for (int j = 1; j <= n; j++)
        for (int i = 2; i <= n; i++)
            tc.A[i][j] = max(tc.A[i][j], tc.A[i-1][j]);
    vector<long long> all;
    for (int i = 1; i <= n; i++)
        for (int j = 1; j <= n; j++)
            all.push_back(tc.A[i][j]);
    sort(all.begin(), all.end());
    tc.answer = all[k-1];
    return tc;
}

struct Solver {
    const TestCase& tc;
    int query_count;
    vector<long long> memo;
    int n;

    Solver(const TestCase& t) : tc(t), query_count(0), n(t.n) {
        memo.assign(2002 * 2002, -1);
    }

    long long do_query(int r, int c) {
        int key = r * 2001 + c;
        if (memo[key] != -1) return memo[key];
        query_count++;
        memo[key] = tc.A[r][c];
        return memo[key];
    }

    long long solve() {
        long long k = tc.k;
        long long N2 = (long long)n * n;

        if (n == 1) return do_query(1, 1);

        long long heap_k = min(k, N2 - k + 1);
        if (heap_k + n <= 24000) {
            if (k <= N2 - k + 1) {
                priority_queue<tuple<long long, int, int>, vector<tuple<long long, int, int>>, greater<>> pq;
                vector<vector<bool>> vis(n + 1, vector<bool>(n + 1, false));
                pq.emplace(do_query(1, 1), 1, 1);
                vis[1][1] = true;
                long long result = -1;
                for (long long i = 0; i < k; i++) {
                    auto [v, r, c] = pq.top(); pq.pop();
                    result = v;
                    if (r + 1 <= n && !vis[r + 1][c]) { vis[r + 1][c] = true; pq.emplace(do_query(r + 1, c), r + 1, c); }
                    if (c + 1 <= n && !vis[r][c + 1]) { vis[r][c + 1] = true; pq.emplace(do_query(r, c + 1), r, c + 1); }
                }
                return result;
            } else {
                long long kk = N2 - k + 1;
                priority_queue<tuple<long long, int, int>> pq;
                vector<vector<bool>> vis(n + 1, vector<bool>(n + 1, false));
                pq.emplace(do_query(n, n), n, n);
                vis[n][n] = true;
                long long result = -1;
                for (long long i = 0; i < kk; i++) {
                    auto [v, r, c] = pq.top(); pq.pop();
                    result = v;
                    if (r - 1 >= 1 && !vis[r - 1][c]) { vis[r - 1][c] = true; pq.emplace(do_query(r - 1, c), r - 1, c); }
                    if (c - 1 >= 1 && !vis[r][c - 1]) { vis[r][c - 1] = true; pq.emplace(do_query(r, c - 1), r, c - 1); }
                }
                return result;
            }
        }

        vector<int> L(n + 1, 1), R(n + 1, n);
        long long k_rem = k;

        // Store previous boundary values for adaptive pivot selection
        vector<long long> prev_boundary_vals;
        vector<int> prev_boundary_weights;

        for (int iter = 0; iter < 100; iter++) {
            vector<int> active;
            long long total_cand = 0;
            for (int i = 1; i <= n; i++) {
                if (L[i] <= R[i]) {
                    active.push_back(i);
                    total_cand += R[i] - L[i] + 1;
                }
            }
            int na = active.size();
            if (total_cand == 0) break;
            if (total_cand == 1) {
                for (int i : active) return do_query(i, L[i]);
                break;
            }

            long long budget = 49500 - query_count;
            if (k_rem + na <= budget) {
                priority_queue<tuple<long long, int, int>, vector<tuple<long long, int, int>>, greater<>> pq;
                for (int i : active) pq.emplace(do_query(i, L[i]), i, L[i]);
                for (long long t = 1; t < k_rem; t++) {
                    auto [v, r, c] = pq.top(); pq.pop();
                    if (c + 1 <= R[r]) pq.emplace(do_query(r, c + 1), r, c + 1);
                }
                return get<0>(pq.top());
            }
            long long rev_k = total_cand - k_rem + 1;
            if (rev_k + na <= budget) {
                priority_queue<tuple<long long, int, int>> pq;
                for (int i : active) pq.emplace(do_query(i, R[i]), i, R[i]);
                for (long long t = 1; t < rev_k; t++) {
                    auto [v, r, c] = pq.top(); pq.pop();
                    if (c - 1 >= L[r]) pq.emplace(do_query(r, c - 1), r, c - 1);
                }
                return get<0>(pq.top());
            }

            // Pivot selection: sample from each active row at target_frac position
            // But also use midpoint of each row (2 samples per row) for better coverage
            double target_frac = (double)(k_rem - 0.5) / total_cand;

            // Strategy: sample from a subset of rows, weighted by width
            // For rows with width > median_width, sample at both target_frac and 0.5 positions
            vector<pair<long long, int>> vw; // (value, weight=width)

            // Sample every active row at the target_frac position
            int sample_step = max(1, na / min(na, max(1, (int)ceil(sqrt((double)na) * 4))));
            for (int idx = 0; idx < na; idx += sample_step) {
                int i = active[idx];
                int width = R[i] - L[i] + 1;
                int col = L[i] + (int)(target_frac * (width - 1));
                col = max(L[i], min(R[i], col));
                vw.push_back({do_query(i, col), width});
            }

            // Sort by value and find weighted quantile
            sort(vw.begin(), vw.end());
            long long total_w = 0;
            for (auto& [v, w] : vw) total_w += w;
            long long target_w = (long long)(target_frac * total_w);
            long long cum = 0;
            long long pivot = vw[vw.size() / 2].first;
            for (auto& [v, w] : vw) {
                cum += w;
                if (cum >= target_w) {
                    pivot = v;
                    break;
                }
            }

            // Staircase walk (bottom-to-top)
            vector<int> p_le(n + 1, 0);
            {
                int j = 0;
                for (int idx = na - 1; idx >= 0; idx--) {
                    int i = active[idx];
                    j = max(j, L[i]);
                    while (j <= R[i] && do_query(i, j) <= pivot) j++;
                    p_le[i] = j - 1;
                }
            }

            long long cle = 0;
            for (int i : active) {
                int rl = min(p_le[i], R[i]);
                if (rl >= L[i]) cle += rl - L[i] + 1;
            }

            // Store boundary values for potential use in next iteration
            prev_boundary_vals.clear();
            prev_boundary_weights.clear();

            if (cle >= k_rem) {
                for (int i : active) {
                    if (p_le[i] >= L[i]) {
                        // Boundary: a[i][p_le[i]] is the max value in the kept region for row i
                        prev_boundary_vals.push_back(do_query(i, p_le[i])); // cached
                        prev_boundary_weights.push_back(min(p_le[i], R[i]) - L[i] + 1);
                    }
                    R[i] = min(R[i], p_le[i]);
                }
            } else {
                k_rem -= cle;
                for (int i : active) {
                    if (p_le[i] + 1 <= R[i]) {
                        prev_boundary_vals.push_back(do_query(i, p_le[i] + 1)); // boundary of > side, cached from walk
                        prev_boundary_weights.push_back(R[i] - max(L[i], p_le[i] + 1) + 1);
                    }
                    L[i] = max(L[i], p_le[i] + 1);
                }
            }
        }
        return -1;
    }
};

int main() {
    struct TestDef { string name; function<TestCase()> gen; };
    vector<TestDef> tests;
    tests.push_back({"additive n=100 k=5000", []{ return gen_additive(100, 5000); }});
    tests.push_back({"mult n=100 k=5000", []{ return gen_multiplicative(100, 5000); }});
    tests.push_back({"additive n=500 k=125000", []{ return gen_additive(500, 125000); }});
    tests.push_back({"mult n=500 k=125000", []{ return gen_multiplicative(500, 125000); }});
    tests.push_back({"random n=500 k=125000", []{ return gen_random_sorted(500, 125000); }});
    tests.push_back({"additive n=2000 k=2000000", []{ return gen_additive(2000, 2000000); }});
    tests.push_back({"mult n=2000 k=2000000", []{ return gen_multiplicative(2000, 2000000); }});
    tests.push_back({"shifted n=2000 k=2000000", []{ return gen_shifted(2000, 2000000); }});
    tests.push_back({"random n=2000 k=2000000", []{ return gen_random_sorted(2000, 2000000); }});
    tests.push_back({"mult n=2000 k=100000", []{ return gen_multiplicative(2000, 100000); }});
    tests.push_back({"mult n=2000 k=3900000", []{ return gen_multiplicative(2000, 3900000); }});

    for (auto& t : tests) {
        auto tc = t.gen();
        Solver s(tc);
        long long result = s.solve();
        bool correct = (result == tc.answer);
        int used = s.query_count;
        double score = !correct ? 0.0 : (used <= tc.n ? 1.0 : (used >= 50000 ? 0.0 : (50000.0 - used) / (50000.0 - tc.n)));
        printf("%-45s q=%6d %s score=%.4f\n", t.name.c_str(), used, correct ? "OK" : "WRONG", score);
    }
}