Блог пользователя sd0061

Автор sd0061, история, 7 лет назад, По-английски

Hi all,

Let's discuss the problems of this contest.

How to solve problem E, F, G and I?

  • Проголосовать: нравится
  • +5
  • Проголосовать: не нравится

»
7 лет назад, # |
  Проголосовать: нравится 0 Проголосовать: не нравится

Also C please

  • »
    »
    7 лет назад, # ^ |
    Rev. 2   Проголосовать: нравится 0 Проголосовать: не нравится

    For problem C: Using dynamic programming that f[x] means the maximum speed when the current mass of rocket is x. enumerating the type and the number of the new stage is O(NM^2). But this can be optimized to O(NM) by saving the prefix maximum of f[x]  -  Isp[i] * log(x) for every type of rocket i and every mass x.

    • »
      »
      »
      5 лет назад, # ^ |
        Проголосовать: нравится 0 Проголосовать: не нравится

      but how can we fix the order of the DP

      I just sort all the engines by Isp from large to small , and double the engines by copyiny the same to n + 1, 2n. then do DP on it. Though I get AC, but I don't think it is correct

  • »
    »
    7 лет назад, # ^ |
      Проголосовать: нравится 0 Проголосовать: не нравится

    Here is another way to look at DP optimization.

    When we add a stage with some drive type, we can put K >= Kmin drives of it. But instead of taking K drives at once, we can take exactly Kmin drives, and then add arbitrary number of same-type drives in special steps which follow. Then we need f[x] — maximal speed for complete rocket of mass x, and g[x, i] — maximal speed of rocket of mass x with last stage composed of drives of type i (which is not complete yet and can be extended).

»
7 лет назад, # |
  Проголосовать: нравится +5 Проголосовать: не нравится

Got: Disk write error (disk full?).

»
7 лет назад, # |
  Проголосовать: нравится 0 Проголосовать: не нравится

Hot to solve div-2. problem nr. 9 (Vim)?

»
7 лет назад, # |
  Проголосовать: нравится +18 Проголосовать: не нравится

F: First of all, you should be able to compute the volume of the region ax + by ≤ c, x2 + y2 ≤ 1, 0 ≤ z ≤ a1x + b1y + c1. I tried numerical integration during the contest and it wasn't precise enough. Got AC with closed formula in upsolving.

When a = cosθ and b = sinθ, by binary search you can find c such that the line ax + by = c equally divides the bread. Call it f(θ). Similarly, when a = cosθ and b = sinθ, by binary search you can find c such that the line ax + by = c equally divides (the bread + the butter). Call it g(θ). You want to find θ such that f(θ) = g(θ).

Notice that f(x) =  - f(x + π) and g(x) =  - g(x + π). Therefore, the signs of f(0) - g(0) and f(π) - g(π) are different. You can again do binary search to find a root in the interval [0, π].

  • »
    »
    7 лет назад, # ^ |
    Rev. 2   Проголосовать: нравится 0 Проголосовать: не нравится

    I tried to compute volume with numerical integration during contests, but my output error was too big. (Even it couldn't pass examples)

    I have a hypothesis(?) that planes which divide x2  +  y2 ≤ 1,  0 ≤ z ≤ a1x  +  b1y  +  c1z region in half and perpendicular to xy-plane always through same point, But I can't prove it.

    • »
      »
      »
      7 лет назад, # ^ |
        Проголосовать: нравится 0 Проголосовать: не нравится

      It's wrong, we checked it during the contest

      • »
        »
        »
        »
        7 лет назад, # ^ |
          Проголосовать: нравится 0 Проголосовать: не нравится

        Thanks for your advise.

        I draw lines with GeoGebra and zoom in, then lines don't through same points. I was mistaken that I wasn't zoom in, and It looks does.

  • »
    »
    7 лет назад, # ^ |
      Проголосовать: нравится 0 Проголосовать: не нравится

    Interesting, so basically the cutting plane can always be a vertical one, right?

  • »
    »
    7 лет назад, # ^ |
    Rev. 2   Проголосовать: нравится +25 Проголосовать: не нравится

    I want to share how to obtain a direction of the required plane in O(1) right after reading an input.

    First of all, assume we have a unit circle, a linear function over it (say, Ax + By + C, let it be positive for convenience) and a line ax + by + c = 0. We want to compute an integral of the linear function over one of the parts the line divides the circle. We choose a parameter t, say that an one-dimensional integral over the chord parallel to the line and passing through the point equals its length () multiplied by the value at the center (), and our integral equals smth like

    It equals something like

    the key observation to obtain it is to change t to .

    Our target is to find such a, b and c that this equals , or that the obtained linear function of A, B and C (the function above minus ) equals zero for (A1, B1, C1) (the bottom layer) and (A2, B2, C2) (the whole sandwich). This means that the obtained coefficients of this linear combination (, etc) represent a vector which is collinear to the cross product of (A1, B1, C1) and (A2, B2, C2). One can easily calculate this vector once he has read the input.

    Let the cross product equal (A * , B * , C * ). Then obviously , so we just said that a = A * , b = B *  and found c using one binary search, having handled cases when one or both planes are horizontal, this solution got accepted.

  • »
    »
    7 лет назад, # ^ |
    Rev. 2   Проголосовать: нравится 0 Проголосовать: не нравится

    We tried calculate this volume by Simpson's rule. Looks like it was very close by precision, but not enough. Do anybody know, how this precision could be predicted before writing code? Error could be estimated by . But how estimate this derivative?

    • »
      »
      »
      7 лет назад, # ^ |
        Проголосовать: нравится +10 Проголосовать: не нравится

      ruban (who created this problem) noticed that in this case order of convergence is different from classical, because the function being integrated has unbounded derivative at one end of the interval. More precisely, derivative of goes to infinity for .

      So the maximal value of derivative in your formula is infinite, hence the whole estimate is useless.

  • »
    »
    7 лет назад, # ^ |
      Проголосовать: нравится 0 Проголосовать: не нравится

    It's easy to calculate this volume if you notice that the integral of a linear function over some figure equals to the area of this figure multiplied by the value of this function in the center of mass of the figure.

»
7 лет назад, # |
Rev. 3   Проголосовать: нравится +10 Проголосовать: не нравится

5 (I think you meant this problem by E): what we really need to do is to find maximum number of music sheets that can stay in place. Sheets staying in place should form increasing subsequence in terms of pairs (number of composition, number of sheet in composition). Moreover, if one of sheets of same composition stays in place, then whole composition stays in place. Now let's do dynamic programming: dp[y] = max number of sheets staying in place among compositions in range [0, y] if composition y stays in place (here we number compositions according to their order in input, let's call this source number (and let's call original numbers of composition target number). It is recalculated more or less as follows: dp[y] = max (dp[x] + (size of composition y)) over all x that we have enough space to fit all intermediate compositions (intermediate in terms of target number) in segment between x and y (in terms of source number)`. This can be done with 2d segment tree that keeps values of dp[y] firstly by target number, then by second number: let's look at "free space" to the left of y (how many multifores to the left of y will remain empty if y stays in place) — it should be at least free space to the left of x for transition to be valid.

So we reduced problem to 2d prefix maximum and point update — this can be done with 2d segment tree in time and memory. There are also some nasty parts concerning reconstruction of the answer, but basic is here.

Code.

  • »
    »
    7 лет назад, # ^ |
      Проголосовать: нравится +30 Проголосовать: не нравится

    You can reduce the problem to the standard Longest Increasing Subsequence problem. Take all the compositions ordered by number, and write down for each of them: its position minus sum of all lengths of compositions with smaller number. Now you need to find longest nondecreasing subsequence in the resulting sequence (add zero before start and M-N after end as sentinels), and it can be done in O(N log N) with ordinary 1D segment tree or Fenwick tree.

»
7 лет назад, # |
  Проголосовать: нравится 0 Проголосовать: не нравится

How to solve K ?)

  • »
    »
    7 лет назад, # ^ |
    Rev. 3   Проголосовать: нравится +18 Проголосовать: не нравится

    11: There are two cases: UEF has no more than acquaintances, or no more than non-acquaintances.

    First case is just dp over subsets: for every mask of already listed acquaintances remember maximum possible number of new friends we can make and mask of people who stopped not knowing us already. Updating this when we list new acquaintances is kind of straightforward with bit operations.

    Second case: notice that nothing changes if list some acquaintance second time, or if we already made all people who can theoretically start knowing us know us). We will try to find longest path in following DAG: vertices are subsets of our original non-acquaintances that still don't know Uef yet, but will start knowing Uef after he lists all his acquaintances; our acquaintances naturally correspond to transitions in this graph (subsets only become smaller as we do transitions, it does not make sense to take transition that does not change the mask). We found some way to list all our acquaintances that makes all people who can theoretically know Uef actually know him and best among such paths. There can be some repeats and some acquaintances not listed, we just will delete repeats and add all not yet listed acquaintances to the end of list.

    This solution works in O(2kn) time and O(2k) memory, where .

    • »
      »
      »
      7 лет назад, # ^ |
        Проголосовать: нравится 0 Проголосовать: не нравится

      In fact the second case is also a DP over subsets, and a very similar to the first one. But we take as parameter of DP the subset of initially-unknown guys who we already know. It is important to notice that we don't need to store the subset of the guys we have already listed, since listing a guy many times does not change anything in this DP.

      So the solution with two DPs over subsets is the intended correct solution.

      However, I'm pretty sure someone managed to get recursive search with some pruning and dirty heuristics accepted. It was very hard to find tests which are not obvious for a search with pruning. Just push hard enough, and you'll be rewarded =)

      • »
        »
        »
        »
        7 лет назад, # ^ |
          Проголосовать: нравится +25 Проголосовать: не нравится

        Well, this DP is ok for both cases, if it's written as backtracking with memorization. All reachable states are unions of some subset of acquaintances, so they are limited by both 2#acquaintances and 2#non - acquaintances.

        By this argument most of well-written recursive searches are really works fast.

»
7 лет назад, # |
  Проголосовать: нравится +6 Проголосовать: не нравится

Contest is over but submission is still running.

Run ID 1069.

Sent at 4:57:02. In resulting table it wasnt counted neither as wrong nor accpeted.

»
7 лет назад, # |
  Проголосовать: нравится 0 Проголосовать: не нравится

For problem 4(Universities), Is O(T * K) fast enough to pass? I got TLE on test 9.

  • »
    »
    7 лет назад, # ^ |
      Проголосовать: нравится +5 Проголосовать: не нравится

    You can optimize solution to O(nk) — it does not make sense to make day an entrance day if there are no students who apply to NSU on this day (be careful with case where nobody actually applies to NSU). Also double-check speed of reading input, because input can be really large.

»
7 лет назад, # |
  Проголосовать: нравится 0 Проголосовать: не нравится

Can someone write how to solve problem 4(Universities)?

  • »
    »
    7 лет назад, # ^ |
    Rev. 2   Проголосовать: нравится +1 Проголосовать: не нравится

    Let us first calculate the expected gain of NSU for a given fixed entrance day t (1 ≤ t ≤ T).

    Suppose, S =  The set consisting of the students who applied to NSU.

    Where,

    P(i, t) = Probability that the i-th student will be available till the t-th day.

    Gi = Value of the i-th student.

    If you can calculate E(t) in O(K), then the total complexity of your solution will be O(TK) (as you can iterate over all possible valid t ). You can speed up the solution with this optimization.

    How to calculate P(i, t):

    Suppose,

    C =  The set consisting of the colleges where the i-th graduate has applied.

    rij =  The day on which the i-th graduate applied to the j-th college.

    G(c, l, r) =  Probability that the c-th college will have its entrance day between the l-th day and r-th day.

    Then:

»
7 лет назад, # |
  Проголосовать: нравится 0 Проголосовать: не нравится

And 10(J) please

  • »
    »
    7 лет назад, # ^ |
      Проголосовать: нравится +1 Проголосовать: не нравится

    This requires high school physics/dynamics formula, like: v = u + at, s = ut + 0.5*a*t^2, v = d/t etc.

    It is always better to have "higher acceleration". So sort the extinguishers by the acceleration parameters. Use one by one until you reach your goal, or run out of the extinguishers. In the second case, you will move the rest of the way in uniform speed (no acceleration)

    • »
      »
      »
      7 лет назад, # ^ |
        Проголосовать: нравится 0 Проголосовать: не нравится

      Yeah, but i constantly received the wa6. Is there a problem with the overflow or precision?

»
7 лет назад, # |
  Проголосовать: нравится 0 Проголосовать: не нравится

When upsolving will be opened?

»
7 лет назад, # |
  Проголосовать: нравится 0 Проголосовать: не нравится

9(I):

Use fi, j, k to represent the shortest distance to current_row , current_col and start_row.

It is easy to deal j = k . When j ≠ k, we can find the path is from one position along some row's end, and it happens only j or k key used. So we have to handle the up and down when start_row is useful. My approach is to divide every nodes to two, one represents it is going to jump to some row with small length. When use k key, it is possible for position (i, j) to go to the end of k row meets maxk ≤ t < i(length(rowt)) < j .Position (i, j) can go to the place where (i, j - 1) can reach, so we link (i, j) -  > (i, j - 1).And we can use monotonic stack to reduce the edge to O(n). And another node represent it jumps to nearest row with same j.

Finally, we can use shortest algorithm to get the distance and path.

  • »
    »
    5 лет назад, # ^ |
      Проголосовать: нравится 0 Проголосовать: не нравится

    So you use dijkstra to compute shortest path? and instead of the heap, we can use a link-list to maintain the current min because edge weight is small? I still worry that it will be too slow to pass, so I'll appreciate it if you could send me your code. thanks!