1734.md (23250B)
1 Source: https://codeforces.com/blog/entry/107231 2 3 Thank you for your participation! 4 5 ## [1734A — Select Three Sticks](https://codeforces.com/contest/1734/problem/A) 6 7 We first sort the array $a$ in non-decreasing order. 8 9 Denote the indices of the elements that we choose from $a$ to be $x$, $y$, and $z$, where $1 \le x \lt y \lt z \le n$, and the final value (after performing the operations) of the concerned elements to be $v$. 10 11 The minimum required number of operations is then $|a_x-v|+|a_y-v|+|a_z-v|$. It is well-known that such expression attains its minimum value when $v$ is the median of $a_x$, $a_y$, and $a_z$. Since the array $a$ has already been sorted, it is best to assign $v$ to be $a_y$. 12 13 Our expression then becomes $|a_x-a_y|+|a_y-a_y|+|a_z-a_y|=(a_y-a_x)+0+(a_z-a_y)=a_z-a_x$. We would like to minimize the value of $a_z$, which implies $z$ should be as small as possible since $a$ is sorted. It is clear that taking $z=y+1$ would minimize the value of the expression. Similarly, we can show that we can take $x=y-1$ to minimize the value of the expression. 14 15 Therefore, the only possible values of the triplets $(x,y,z)$ are of the form $(t,t+1,t+2)$ for positive integers $1 \le t \le n-2$, and we can iterate through all such triplets and find the best one. 16 17 The time complexity is $\Theta(n \log n)$ per case due to sorting. 18 19 **Code in C++** 20 21 ``` 22 #include <bits/stdc++.h> 23 using namespace std; 24 void solve(){ 25 int n; 26 cin>>n; 27 int a[n+1]; 28 for (int i=1; i<=n; i++){ 29 cin>>a[i]; 30 } 31 sort(a+1,a+n+1); 32 int ans=2e9; 33 for (int i=3; i<=n; i++){ 34 ans=min(ans,a[i]-a[i-2]); 35 } 36 cout<<ans<<'\n'; 37 } 38 int main(){ 39 int t; 40 cin>>t; 41 for (int i=1; i<=t; i++){ 42 solve(); 43 } 44 } 45 ``` 46 47 ## [1734B — Bright, Nice, Brilliant](https://codeforces.com/contest/1734/problem/B) 48 49 Note that the brightnesses of the rooms on the $i$-th floor is at most $i$. This is because in room $(i,1)$, only $i$ rooms, namely, $(1,1)$, $(2,1)$, $\ldots$, $(i,1)$ can reach to $(i,1)$ through some number of staircases. 50 51 It is also possible to find a configuration of torches in the pyramid such that the brightnesses of the rooms on the $i$-th floor is exactly $i$, i.e. it attains the upper bound. 52 53 The configuration is as follows: Room $(i,j)$ contains a torch if and only if it is the leftmost room ($i=1$) or the rightmost room ($i=j$) on the $i$-th floor. 54 55 This is valid because for all rooms $(i,j)$, it can be reached from $(1,1)$, $(2,1)$, $(3,1)$, $\ldots$, $(i-j+1,1)$ and $(2,2)$, $(3,3)$, $\ldots$, $(j,j)$. In other words, room $(i,j)$ has brightness $(i-j+1)+j-1=i$, so the pyramid is nice. 56 57 **Code in C++** 58 59 ``` 60 #include <bits/stdc++.h> 61 using namespace std; 62 int main() { 63 int t; 64 cin >> t; 65 while (t--) { 66 int n; 67 cin >> n; 68 for (int i = 1; i <= n; i++) { 69 for (int j = 1; j <= i; j++) { 70 cout << (j == 1 || j == i) << ' '; 71 } 72 cout << '\n'; 73 } 74 } 75 } 76 ``` 77 78 ## [1734C — Removing Smallest Multiples](https://codeforces.com/contest/1734/problem/C) 79 80 One operation should be used to remove every element not belonging to $T$. 81 82 Let $v$ be an element not belonging to $T$. Suppose a $x$-cost operation removes value $v$, then $v$ must be divisible by $x$. Furthermore, the multiples $x,2x,\cdots (k-1)x$ must have been already removed from $S$, where we write $v = kx$. 83 84 Since removed elements stay removed, the above is only possible if all of $x,2x,\cdots (k-1)x$ does not belong to $T$. 85 86 For each $v$, let $f(v)$ be the smallest integer $x$ satisfying the above condition. As we can always remove $v$ using a $v$-cost operation, $f(v) \leq v$ and in particular $f(v)$ exists. 87 88 The total cost must be at least $\sum_{i \not \in T} f(i)$. We claim that this cost can be achieved. 89 90 To do so, we should remove the required elements in ascending order. When removing $v$, we assume all $w \not\in T$ with $w \lt v$ have already been removed. At this state, an $f(v)$-cost operation would be able to remove $v$. 91 92 It remains to find the values $f(v)$. To do so efficiently, we can perform the above process in a bottom-up manner similar to the Sieve of Eratosthenes. Please refer to the code below for implementation details. The overall complexity is $n (1 +\frac{1}{2} + \frac{1}{3} + \cdots + \frac{1}{n}) = \Theta(n \log n)$. 93 94 **Code in C++** 95 96 ``` 97 #include <bits/stdc++.h> 98 using namespace std; 99 void solve() { 100 int n; 101 cin >> n; 102 bool a[n + 1]; 103 string str; 104 cin >> str; 105 for (int i = 1; i <= n; i++) { 106 a[i] = (str[i - 1] == '1'); 107 } 108 long long ans = 0; 109 int cost[n + 1]; 110 for (int i = n; i >= 1; i--) { 111 for (int j = i; j <= n; j += i) { 112 if (a[j]) break; 113 cost[j] = i; 114 } 115 } 116 for (int i = 1; i <= n; i++) { 117 if (!a[i]) ans += cost[i]; 118 } 119 cout << ans << '\n'; 120 } 121 int main() { 122 int t; 123 cin >> t; 124 while (t--) solve(); 125 } 126 ``` 127 128 ## [1734D — Slime Escape](https://codeforces.com/contest/1734/problem/D) 129 130 Let's call a group of slime **good** if their total health is at least $0$, or if defeating this group allows you to reach the exits. 131 132 We partition the slimes into good groups in a two-pointer like manner. To form the groups to the right, start from position $k$, then find the smallest position $r$ such that slimes from $k+1$ through $r$ form a good group. We do the same starting from $r+1$ again. Repeat this process until slimes to the right are partitioned into groups, which can be done by maintaining the sum of health. We partition the left slimes into groups in a similar way. 133 134 We can observe that in an optimal strategy, we may assume the player absorbs group-by-group. 135 136 **Proof** 137 138 Assuming there is a valid strategy $S$ to reach the exit. Let $X$ be the first group to the left, and $Y$ be the first group to the right. Without loss of generality, assume all slimes of $X$ are absorbed before all slimes of $Y$ are absorbed first. Suppose there are $x$ slimes in $X$ and $y$ slimes in $Y$. 139 140 Suppose we did not absorb all of $X$ at a time. For example, if there are $6$ slimes in $X$ and $5$ or more slimes in $Y$, the beginning of our strategy may look like this: 141 142 $LLRRLRLLRL....$ 143 144 Consider instead the strategy where we move all `L' to the front. We claim that the strategy stays valid. That is, the following strategy 145 146 $LLLLLLRRRR....$ is at least as good. 147 148 We check that at every moment of strategy $LLLLLLRRRR....$, there is some point in time $LLRRLRLLRL....$ which we would have less than or equal amount of health. 149 150 For the state $T_1$ that we take $a$ moves to the left in $LLLLLLRRRR....$, we compare it with any moment $T_2$ in $LLRRLRLLRL....$ that we have taken $a$ moves to the left. To reach $T_2$, we take $b$ moves to the right. We can check that under the assumption, $0 \leq b \lt y$. Since $Y$ is the smallest good group to the right, taking these $b$ extra right moves must have reduced our health. So, we have more health at $T_1$ than in $T_2$. 151 152 Now consider the state $T_3$ in $LLLLLLRRRR....$ where we take all $x$ left moves and $b$ more right moves. We should compare it with any moment $T_4$ in $LLRRLRLLRL....$ that we have taken $b$ right moves and some $a$ left moves. We can check that under the assumption $0 \leq a \lt x$. Since $X$ is the smallest good group, taking only $a$ moves to the left gives us less health compared to taking all $x$ moves to the left. Therefore, we have more health at $T_3$ than in $T_4$. 153 154 Therefore, if $LLRRLRLLRL....$ is valid, $LLLLLLRRRR....$ must be valid. So, we only need to consider strategies of the second form. 155 156 By applying this claim recursively, we notice that we only need to consider strategies that absorb whole groups at a time. 157 158 This works for all strategies, we just referred to the starting strategy as $LLRRLRLLRL....$, and the better strategy as $LLLLLLRRRR....$ to simplify notations. 159 160 For any good group, since the total health is positive, there is no drawback to absorbing a good group. In other words, whenever it is possible to absorb a good group, we will absorb it. 161 162 For each group $G$, we calculate the ``requirement'' of the group — the lowest health we can begin with, such that we can absorb the group while maintaining non-negative health at all times. The requirement of a group of slime with health $a_1,a_2 \cdots a_n$ can be expressed as 163 164 $ - \min_{k=0}^n (\sum_{i=1}^k a_i)$ 165 166 Finally, we can simply simulate the process. We repeatedly attempt to absorb good groups to the left or to the right. We keep track of the current health, initially equal to $a_k$. Whenever we consider whether to absorb a group or not, we absorb it if and only if the current health is at least as high as its requirement. Otherwise, we ignore it for now and attempt to do so for the group on the other side. If it is possible to reach a state where either all left groups or all right groups are absorbed, then we can win the game. If at some point, it is not possible to absorb the left group nor the right group, then we lose. 167 168 The overall complexity is $\Theta(n)$. 169 170 It is also possible to use a range $\max$/$\min$ segment tree form the groups instead of using two-pointers, in which case its complexity would be $\Theta(n \log n)$. 171 172 **Code in C++** 173 174 ``` 175 #include <bits/stdc++.h> 176 using namespace std; 177 void solve() { 178 long long n, k, P; 179 cin >> n >> k; 180 k--; 181 long long L[n]; 182 for (int i = 0; i < n; i++) cin >> L[i]; 183 P = L[k]; 184 L[k] = 0; 185 long long ps[n + 1]; 186 ps[0] = 0; 187 for (int i = 1; i <= n; i++) ps[i] = ps[i - 1] + L[i - 1]; 188 int l = k, r = k + 1; 189 vector<pair<long long, long long> > LG, RG; 190 for (int i = k - 1; i >= 0; i--) { 191 if (ps[i] <= ps[l] || i == 0) { 192 long long worst = 0, cur = 0; 193 for (int j = l - 1; j >= i; j--) { 194 cur += L[j]; 195 worst = min(worst, cur); 196 } 197 LG.push_back({cur, -worst}); 198 l = i; 199 } 200 } 201 for (int i = k + 2; i <= n; i++) { 202 if (ps[i] >= ps[r] || i == n) { 203 long long worst = 0, cur = 0; 204 for (int j = r; j <= i - 1; j++) { 205 cur += L[j]; 206 worst = min(worst, cur); 207 } 208 RG.push_back({cur, -worst}); 209 r = i; 210 } 211 } 212 reverse(LG.begin(), LG.end()); 213 reverse(RG.begin(), RG.end()); 214 long long curp = P; 215 while (true) { 216 bool acted = false; 217 if (!LG.empty() && curp >= LG.back().second) { 218 curp += LG.back().first; 219 LG.pop_back(); 220 acted = true; 221 } 222 if (!RG.empty() && curp >= RG.back().second) { 223 curp += RG.back().first; 224 RG.pop_back(); 225 acted = true; 226 } 227 if (LG.empty() || RG.empty()) { 228 cout << "YES\n"; 229 return; 230 } 231 if (!acted) { 232 cout << "NO\n"; 233 return; 234 } 235 } 236 } 237 int main() { 238 int t; 239 cin >> t; 240 while (t--) solve(); 241 } 242 ``` 243 244 ## [1734E — Rectangular Congruences](https://codeforces.com/contest/1734/problem/E) 245 246 We say a matrix to be good if it satisfies the congruence condition (the second condition). 247 248 When we have a good matrix, we can add any value $c$ to a whole row while maintaining the congruence relation. The same is true for adding the same value to a whole column. 249 250 Suppose we have any good matrix $A$, then by adding $b_i - a_{i,i}$ to the $i$-th row for each of $i=1,2,\cdots,n$, we obtain a good matrix that has the desired values on the diagonal. 251 252 In fact, there are a lot of possible constructions. We present a few of them here: 253 254 * $a_{i,j} = i \times j \pmod n$ 255 * $a_{i,j} = (i + j)^2 \pmod n$. This needs special handling when $n=2$. 256 * $a_{i,j} = \frac{(i+j)(i+j+1)}{2} \pmod n$. 257 * The coolest part is that all quadratic polynomials in the form $ai^2 + bij + cj^2 + di + ej + f$ are valid for all integers $a,b,c,d,e,f$ and $b \not\equiv 0 \pmod n$ 258 259 As a bonus, we prove that the general quadratic polynomial gives a good construction. 260 261 **Proof** 262 263 Since we can add values to a whole row or column, and $i^2, j^2 , i , j , 1$ are also constant on the rows and columns, adding them by $a_{i,j}$ has no effect. So we may just assume $a = c = d =e = f =0$. 264 265 So it suffices to show that $a_{i,j} = b i j \pmod n$ satisfies the condition. We can see directly that $a_{r_1,c_1} - a_{r_2,c_1} - a_{r_1,c_2} +a_{r_2,c_2} = b(r_1 - r_2)(c_1 - c_2)$. As $r_1 \neq r_2$, $c_1 \neq c_2$, $b \not\equiv 0 \pmod n$, and $n$ is a prime, this expression must be nonzero $\pmod n$. 266 267 Here are some extra observations that may enable one to find a good matrix more quickly. 268 269 **Some extra observations** 270 271 For each $j$, the values $a_{i,j} - a_{i-1,j}$ over all $0 \le i \lt n $ must be a permutation of ${0,1,\cdots,n-1}$. 272 273 A good matrix stays good if we permute the rows or permute the columns. Therefore, we can show that there exists some good matrix with $a_{1,j} = a_{i,1} = 0$, and $a_{2,j} = j-1$, $a_{i,2} = i-1$. Assuming this, it should not be difficult to discover $a_{i,j} = (i-1)(j-1)$ yields one of the good matrices. 274 275 Let $b$ be the two dimensional difference array of $a$, that is, $b_{i,j} = a_{i,j} - a_{i-1,j} - a_{i,j-1} + a_{i-1,j-1}$. Then, the condition becomes sum of any rectangles of $b$ must be nonzero $\pmod n$. It is easy to see $ b = 1$ is valid. This corresponds to the solution that $a_{i,j} = i \times j \pmod n$. 276 277 **Code in C++** 278 279 ``` 280 #include <bits/stdc++.h> 281 using namespace std; 282 283 int want[355]; 284 int board[355][355]; 285 int main() { 286 int n; 287 cin >> n; 288 for(int i = 0; i < n ; i ++){ 289 cin >> want[i]; 290 } 291 for (int i = 0; i < n; i++) { 292 for (int j = 0; j < n; j++) { 293 board[i][j] = (i * j) % n; 294 } 295 } 296 for(int i = 0; i < n; i ++){ 297 int extra = (want[i] + n - board[i][i] )% n ; 298 for (int j = 0; j < n; j++) { 299 board[i][j] += extra; 300 board[i][j] %= n; 301 } 302 } 303 for (int i = 0; i < n; i++) { 304 for (int j = 0; j < n; j++) { 305 cout << board[i][j] << ' '; 306 } 307 cout << endl; 308 } 309 } 310 ``` 311 312 ## [1734F — Zeros and Ones](https://codeforces.com/contest/1734/problem/F) 313 314 Observe that the $i$-th character is `1' if and only if $i$ has an odd number of set bits in its binary representation. Both solutions make use of this fact. 315 316 The constraints allows solutions of up to $\Theta(q \log^3 n)$. Yet, both of the model solution runs in $\Theta(\log n)$. 317 318 ### 1. Digit DP solution 319 320 The question can be reformulated as follows: How many integers $x$ between $0$ and $m-1$ inclusive have the property that the total number of set bits of $x$ and $x+n$ is an odd number? 321 322 This can be solved with digit DP. We process the bit position from $\lceil \log(\max) \rceil$ down to $0$. We maintain three states: 323 324 * $\text{ans}$, a boolean value; 325 * $\text{trailzeros}$, an integer between $0$ and $\lceil \log(\max) \rceil$ inclusive; and 326 * $\text{under}$, a boolean value. 327 328 We can thus conclude the following: the number of trailing zeros is all we need to decide the answer. 329 330 After processing each bit $k$, we should have the following: the number of integers $x$ between $0$ and $\lfloor \frac{m}{2^k} \rfloor$ inclusive which have the following property: 331 332 * the total number of set bits of $x$ and $x + \lfloor \frac{n}{2^k} \rfloor$ is equal to $\text{ans} \mod 2$; 333 * the number of trailing `1's of $x + \lfloor \frac{n}{2^k} \rfloor$ is equal to $\text{trailzeros}$; 334 * the boolean value $[x \lt \lfloor \frac{m}{2^k} \rfloor]$ (where $[]$ is the Iverson bracket). 335 336 Now onto the transitions. Suppose we are adding the $(k-1)$-th digit, and let $d$ be the new digit of $x$, and $z$ be the $(k-1)$-th digit of $n$. 337 338 * If $z+d = 0$, then $(\text{ans},\text{trailzeros})$ after digit $k$ will be transited to $(\text{ans},0)$ after digit $k-1$; 339 * if $z+d = 1$, then $(\text{ans},\text{trailzeros})$ after digit $k$ will be transited to $((\text{ans} + z +d) \mod 2, \text{trailzeros}+1)$ after digit $k-1$; 340 * if $z+d =2$, then $(\text{ans},\text{trailzeros})$ after digit $k$ will be transited to $((\text{ans} + z + \text{trail} + 1) \mod 2, 0)$ after digit $k-1$ 341 342 The final answer is the total number of values for which $\text{ans} =1 $ and $\text{under} = 1$. 343 344 The above solution runs in $\Theta(\log^2 (\max))$ per query. There is a simple way to optimize this to $\Theta(\log(\max))$: note that we only need to keep parity of $\text{trailzero}$. 345 346 There are many other digit DP approaches that give similar time complexity. The constraints should allow most of them pass. 347 348 **Code in Kotlin** 349 350 ``` 351 val Boolean.chi:Int get() = if(this) 1 else 0 //characteristic function 352 fun Int.has(i:Int):Boolean = (this and (1 shl i) != 0) 353 fun Long.has(i:Int):Boolean = (this and (1L shl i) != 0L) 354 fun main(){ 355 repeat(readLine()!!.toInt()){ 356 357 val (n,m) = (readLine()!!.split(" ").map { it.toLong() }) 358 var DP = Array(2){Array(2){LongArray(64)}} 359 360 DP[0][0][0] = 1 361 362 for(bit in 60 downTo 0 ){ 363 var new = Array(2){Array(2){LongArray(64)}} 364 for(under in 0 until 2){ 365 for(ans in 0 until 2){ 366 for(nd in 0 until 2){ 367 if(under == 0 && nd == 1 && !m.has(bit)) continue 368 val nunder = ((under == 1) || (nd == 0 && m.has(bit))).chi 369 370 val np = nd + n.has(bit).chi 371 372 for(trail in 0 until 62){ 373 var nans = ans 374 if(nd == 1){ 375 nans = 1 - nans 376 } 377 var newtrail:Int 378 379 if(np == 0){ 380 newtrail = 0 381 }else if(np == 1){ 382 nans = 1 - nans 383 newtrail = trail + 1 384 }else if(np == 2){ 385 nans = (nans + trail + 1) %2 386 newtrail = 0 387 }else{ 388 error("oh no") 389 } 390 391 new[nunder][nans][newtrail] += DP[under][ans][trail] 392 } 393 } 394 } 395 } 396 DP = new 397 } 398 399 var ret = 0L 400 for(trail in 0 until 64){ 401 ret += DP[1][1][trail] 402 } 403 println(ret) 404 405 } 406 } 407 ``` 408 409 ### 2. Recursive Solution, by [darkkcyan](/profile/darkkcyan "Гроссмейстер darkkcyan") 410 411 Define the function $f(x) := $ the parity of bit one in the number $x$. 412 413 We have thus reworded the statement into evaluating the follow expression: 414 415 $ T = \sum_{i = 0}^{k - 1} \left[f(i) \ne f(i + n)\right] $ 416 417 The formula can be further transformed as: 418 419 $ T = \sum_{i = 0}^{k - 1} f(i \oplus (i + n)) $ 420 421 since $\left[ f(a) \ne f(b) \right] = f(a \oplus b)$ holds true for all non-negative integers $a$ and $b$. 422 423 Imagine we construct a grid and assign the value at row $r$ and column $c$ to be $f(r \oplus c)$. Then, $T$ is sum of a diagonal of length $k$ which starts at either $(0, n)$ or $(n, 0)$. Without loss of generality, we use $(0, n)$ in this editorial. 424 425 The grid can be constructed similarly to the way we construct the string $S$. We start with a $1$-by-$1$ matrix $M_0=\begin{bmatrix} 0 \end{bmatrix}$. 426 427 Then, the matrix $M_i$ of size $2^i \times 2^i$ can be constructed as follows: 428 429 $ M_i = \begin{bmatrix} M_{i - 1} & \overline {M_{i - 1}} \\ \overline{M_{i - 1}} & M_{i - 1} \end{bmatrix} $ 430 431 where $\overline{M_{i - 1}}$ is the matrix $M_{i - 1}$ but with flipped bits. 432 433 Here is another way of constructing the grid: let $C_i$ be an infinite chess board with alternating colors, similar to a regular chess board, but with each of the cells being size $2^i \times 2^i$. For example, $C_0$, $C_1$ and $C_2$ in an $8$-by-$8$ grid is as follows: 434 435 **Illustration** 436 437 ``` 438 .*.*.*.* 439 *.*.*.*. 440 .*.*.*.* 441 *.*.*.*. 442 .*.*.*.* 443 *.*.*.*. 444 .*.*.*.* 445 *.*.*.*. 446 447 ..**..** 448 ..**..** 449 **..**.. 450 **..**.. 451 ..**..** 452 ..**..** 453 **..**.. 454 **..**.. 455 456 ....**** 457 ....**** 458 ....**** 459 ....**** 460 ****.... 461 ****.... 462 ****.... 463 ****.... 464 ``` 465 466 We claim that our grid is the $\text{xor}$ of all chess board $C_i$. The proof is easy: $C_i$ is constructed by $\text{xor}$-ing the $i$-th bit of the the row and column number. 467 468 We are therefore motivated to proceed in the following way: if we drop the least significant bit (by making it to be $0$), we are still solving a very similar problem to the original problem, because dropping the first bit is similar to removing $C_0$. And when we shift $C_i$ to $C_{i - 1}$, it is a recursion of the same problem! 469 470 Going back to the problem, where we are computing sum of a diagonal of length $k$. If $k$ is odd, we can make it odd by adding the last element to the result and decreasing $k$ by one. Now, $k$ is even, and we can utilize the recurrence as follows: 471 472 * remove $C_0$. 473 * scale the board down by 2 (including $n$ and $k$). By doing so, $C_i$ becomes $C_{i - 1}$. 474 * solve the new problem. 475 * scale the board up again and add $C_0$ back. 476 * from the result of the scaled down problem, some how calculate the result of the original problem 477 478 The result of the scaled down problem is the number of $2$-by-$2$ cells with value $1$. From the number of $2$-by-$2$ cells with value $1$, we compute the number of cells with value $0$ as well. It is not hard to observe that it crosses the $2$-by-$2$ cells at all places. The only thing that matters is the parity of $n$. 479 480 * If $n$ is even, then the diagonal crosses the diagonal of the $2$-by-$2$ cells. In the scaled-down version, the diagonal is still a single diagonal starting at $(0, \frac{n}{2})$; otherwise, 481 * if $n$ is odd, it crosses the corner of the $2$-by-$2$ cells. In the scaled-down version, the diagonal is actually $2$ neighboring diagonals starting at $(0, \frac{n-1}{2})$ and $(0, \frac{n+1}{2})$. 482 483 Also, the $2$-by-$2$ cells with values $0$ and $1$ respectively will also have the form: 484 485 **Illustration** 486 487 ``` 488 .* *. 489 *. .* 490 ``` 491 492 From here we have everything we need to compute the result of the original problem. 493 494 Overall, the number of states we have to visit is $\Theta(\log k)$. 495 496 **Code in Python** 497 498 ``` 499 import random 500 501 cache = {} 502 503 def popcount(n): 504 res = 0 505 while n: 506 res += 1 507 n &= n - 1 508 return res 509 510 def solve(n, k): 511 if k == 0: 512 return 0 513 if k == 1: 514 return popcount(n) & 1 515 if k % 2 == 1: 516 t = solve(n, k - 1) 517 x = popcount((k - 1) ^ (n + k - 1)) & 1 518 # print(t, x) 519 return t + x 520 if (n, k) in cache: 521 return cache[(n, k)] 522 if n % 2 == 0: 523 one_cell = 2 524 zero_cell = 0 525 cnt1 = solve(n // 2, k // 2) 526 cnt0 = k // 2 - cnt1 527 else: 528 one_cell = 0 529 zero_cell = 1 530 cnt1 = solve(n // 2, k // 2) + solve(n // 2 + 1, k // 2) 531 cnt0 = k - cnt1 532 res = one_cell * cnt1 + zero_cell * cnt0 533 # print(f"n = {n}, k = {k}, cnt1 = {cnt1}, cnt0 = {cnt0}, one_cell = {one_cell}, zero_cell = {zero_cell}") 534 cache[(n, k)] = res 535 return res 536 537 t = int(input()) 538 for _ in range(t): 539 cache.clear() 540 n, k = map(int, input().split()) 541 print(solve(n, k)) 542 ```