Sunday, January 16, 2022

A Stirling week

There were no contests that I'd like to mention this week, but this offers a good chance to look back at the first full week of 2022 :) The new competitive programming year started with Hello 2022 on Codeforces (problems, results, top 5 on the left, analysis). The new year started in just the same way it ended: with a victory for tourist, who has been in amazing form recently. Congratulations!

My round was unfortunately ruined by getting bogged in minor details of the problems. I've started by forgetting to reset the segment tree back to the initial state after processing each group in problem E, but that was still a relatively minor setback. The next problem F was nastier, and I spent quite a lot of time and energy to deal with the first and last groups of zeros properly. Finally, in problem G I just could not untangle myself in time, and my bug turned out to be simply forgetting to ignore smaller numbers to the right of a right-maximum when processing it.

TopCoder SRM 821 on Saturday started the race to qualify for the TCO22 regional events (problems, results, top 5 on the left, analysis). tourist has continued his impressive run of form by becoming the first person ever to win 5 SRMs in a row, also achieving a ridiculous rating of 4034 in the process. Very impressive!

The round featured a 250-pointer that required a math olympiad-style solution. Sure, after knowing that such solution is expected finding it is not so hard, but I somehow tried to think more algorithmically and failed. Having seen that others are submitting the problem for very high scores, I've assumed that the greedy works and submitted it, which at the very least allowed me to get unstuck and move on to the other two problems. Having submitted them, I've implemented a stress-test for the 250 and quickly learned that the greedy does not in fact work, but still could not come up with a working solution, hence no resubmit. Therefore I guess I was lucky to believe in the greedy, as otherwise I wouldn't have had enough time left for the two remaining problems even had I figured the trick out.

I've managed to make up the points I knew I was going to lose by challenging others who submitted greedy and similar solutions for this problem, but even +450 challenge points were not enough as tourist has found +425 for himself. Nevertheless, I enjoyed the excitement of the challenge phase, and the need to make tradeoffs between confidence that your challenge might work with the need to find many challenges in 15 minutes. Props to the TopCoder format for making such exciting rounds possible!

Finally, CodeChef SnackDown 2021 Final Round on Sunday wrapped up the contests of the previous year (problems, results, top 5 on the left, analysis). It was the first online edition of SnackDown, and the first one without the two-person team format (even though the organizers mentioned that they're planning to come back to the team format when the pandemic allows). Mingyang Deng has made steady progress while others stumbled, and therefore had a much better penalty on 6 problems and got the first prize. Well done!

My round went wrong when I quickly reduced the permutation problem to finding the Stirling numbers of the second kind S(n, k) for all k from 1 to a, but then spent at least an hour, and probably more, trying to figure out how to actually do that, including opening a physical copy of Concrete Mathematics searching for some identities that would help me :) However, this was a moment where no magical trick was needed, and one just needed to apply the technique that I described in my blog 5 years ago. I guess my statement back then was wrong: "However, C(nk) seems to be the only practically relevant function with this property".

This huge delay has set the stage for my last-hour comeback, as I managed first to deal with the quite tricky problem where one needs to take advantage of the fact that some probabilities are too low to significantly reduce the number of dynamic programming states, then to solve the problem about incremental maximums, and then finally see the FFT trick with about 5 minutes remaining, managing to finish coding and submit my solution with about 5 seconds left on the contest clock, after it failed to compile locally, and it passed!

The compilation issues were due to the fact I'm currently using Visual Studio as the backend for CLion, which provides for much faster compilation, but unfortunately lacks int128 support. I had int128 used in my library to apply CRT when doing FFT on three prime moduli, back from the times I used gcc and clang. JHelper's inlining only works on file level, therefore this version of FFT was being included even though it was not actually used.

Here is the incremental maximums problem: you are given a permutation of length 200000, and need to tell if it's possible to split it into two disjoint subsequences in such a way that both parts have the same number of incremental maximums (numbers that are bigger than all previous numbers in the part).

This is one of those problems that look quite ugly after reading the statement, but as you work towards the solution you develop a good understanding of the problem structure, and the solution is therefore quite rewarding when you see it in the end. So I encourage you to take a crack at it, even though of course it is very hard.

This problem has appeared before, but I somehow did not remember it at all even though I solved the relevant round. I guess I focused on other problems then, did not do much upsolving, and did not find it interesting enough for the blog. Well, at least that allowed me to enjoy solving it this time!

Thanks for reading, and check back next week.

Saturday, December 18, 2021

A pseudobubble week

Last week's contests were all concentrated on the weekend. Codeforces Round 758 took place early on Saturday (problems, results, top 5 on the left, analysis). The round seems to have been downvoted into oblivion because of weak pretests, and I think it's quite unfortunate — I think there's definitely place for weak pretests in the Codeforces format, and proving/testing one's solution is a part of the game. Congratulations to maroonrk who was able to solve six problems despite the weak pretests and win the round!

Facebook Hacker Cup 2021 Finals followed a few hours later (problems, results, top 5 on the left, analysis). Once again the points+ICPC penalty system has allowed for many different strategies, and this time A+D+F was the winning choice. Congratulations to ecnerwala and Um_nik for the successful execution!

I was trying to solve the same 3 problems, but my solution for F did not manage to finish in 6 minutes because it used too much memory and started swapping. After the round I've asked Egor to run it on his machine with 64G RAM, and it finished in 3 minutes there, but it turned out it also gave wrong answer on one testcase. It was surprisingly satisfying to learn that I was not so close, and my (relatively) weak laptop was not the only issue :)

My real issue in this round is that I've treated it as a Facebook Hacker Cup round. I've convinced myself that the Hacker Cup always has tedious implementation problems, and therefore I've started implementing the first solution that came to mind instead of thinking further to find a much simpler approach. In problem D, this ended up being segment trees with lazy propagation on a structure with 6 fields, applied on the previsit orders on the components of the centroid decomposition. The reference solution is a relatively simple dynamic programming on a tree after applying the trick from Aliens. In problem F, I've used a persistent segment tree containing persistent treaps in its nodes to build a graph of size O(n*log^2(n)) to find strongly connected components in, and this was already too much — it has taken me ages to implement, used too much memory and also probably had a bug (or was fully incorrect — I still haven't debugged the wrong answer).

Well, I hope this will be a good reminder to look for simpler solutions next time!

Finally, Open Cup 2021-22 Grand Prix of Nanjing wrapped up the week and the 2021 Open Cup rounds on Sunday (problems, results, top 5 on the left, onsite results). Team USA1 was once again in their own league, getting to 12 problems with 1.5 hours remaining, and being the only team to solve everything in the end. Well done!

Problem D was relatively standard, but as part of the story it contained the following algorithm:

for i in 1..n:
  for j in 1..n:
    if a[i] < a[j]:
      swap(a[i], a[j])

At first sight, this looks to be a standard sorting algorithm. At second sight, though, it seems to be buggy. But it turns out that this algorithm does in fact sort the array in increasing order :)

Thanks for reading, and check back next week!

Sunday, October 10, 2021

A 2020 week

ICPC 2019-20 World Finals was the first event of this week (problems, results, top 12 on the left, official broadcast, our stream). As this is the main event of the year for many competitors, there are quite a few reports to check out from various participants (1, 2, 3, 4, please post more links in comments!). I did not watch the contest myself as we were solving it on stream instead, so let me just congratulate all medalists, and especially the NNSU team on the victory!

It was amazing to meet so many old and new friends in person after the long break, so big thanks to the organizers on making this event possible! At the same time, I think that the 3-computer rule and the fact that it was only announced a few weeks before the event were quite unfortunate. I hope that we will return to the 1-computer format next year!

On the heels of the World Finals where JetBrains was one of the main sponsors, Codeforces hosted Kotlin Heroes Episode 8 (problems, results, top 5 on the left, analysis). Judging by the standings, there was a nice difficulty curve that allowed everyone to have something interesting to solve. 3 contestants were able to solve one of the two hardest problems, SpyCheese finishing ahead of Heltion and eatmore thanks to being much faster on the relatively easier first 8. Congratulations!

TopCoder SRM 815 took place on Friday (problems, results, top 5 on the left). I was solving the round with a cold and a headache, however somehow this did not affect the results much, maybe because the 250 and 500 were very standard, which allowed me to go through them more or less on autopilot.

Solving the hard problem, on the other hand, required some gut feeling for what will be fast in practice. Given two positive integers n<=50 and m such that mn<=1018, what is the largest possible value of the least common multiple of n positive integers, each up to m?

Facebook Hacker Cup 2021 Round 3 wrapped up the week on Saturday (problems, results, top 5 on the left, analysis). The combination of point values with the ICPC penalty time system created an interesting strategic dilemma: solving tasks D2+D3 (which was not really much more difficult than solving one of them) required more time than solving problem B or problem C, but since they were counted as two for penalty time purposes, it could pay off in terms of penalty time. From the top two finishers, tourist went for D2+D3 first, and Radewoosh went for B+C first, and the strategies turned out to be really close in practice. Well done to both, and congratulations to the top 25 who will compete in the online finals!

Going with D2+D3 before problem C had another drawback: in case one did not have enough time to solve everything, since D2+D3 were worth less points than C, one would lose to the alternative approach. The set of people on 68 points in the scoreboard has quite a few favorites who ended up not qualifying despite solving the most difficult problem. So in effect one had to decide between a strategy that would deliver the smallest penalty time when solving everything, and a strategy that would give more points when solving all problems except one (if we treat D2+D3 as one problem, as they really were).

I would also like to bring two important competitive programming websites to your attention. clist.by has the information about virtually all competitive programming contests, and allows everyone to assemble a custom calendar with regexp-based filters to see only the contests one is interested in. It also has Telegram notifications for this customized calendar and also imports the standings from most of those contests which allows for contestant-vs-contestant comparisons across platforms. It obtains the links between handles on different platforms through crowdsourcing, you can link all your handles together after registering on clist.by.

The recently launched cphof.org is similar in spirit, as it tries to create a unified profile for the contestants through various platforms and one-off competitions. However, it focuses only on big tournaments and not on regular rounds, and this limited scope allowed it to do the linking without relying on crowdsourcing. As a result, it has really good coverage for the top competitors.

Huge thanks to the maintainers of those websites, I use them every week! And thanks to all for reading, please check back next week.

Monday, October 4, 2021

An invitational week

TopCoder SRM 814 was the first round of last week (problems, results, top 5 on the left). ecnerwal was about 5 minutes faster than the rest of the field in solving the 1000-pointer, and therefore got enough margin to maintain the first place through the challenge phase and deny me qualification to Round 4 of TCO22 :) Well done!

Codeforces Round 745 followed later on the same day (problems, results, top 5 on the left, analysis). The last two problems were a lot harder than the rest, and only medium_waxberry and maroonrk were able to crack them. Congratulations to both!

ACM ICPC 2019-20 World Finals finally takes place tomorrow in Moscow. In case you haven't been following, quite a few teams could not make it and participated in a different online contest instead, and for those who could, the format was changed to 3 computers instead of 1. From the teams that made it to Moscow (list, but you have to exclude the online teams), the strongest three are the MIT, the Nizhny Novgorod SU and the Seoul NU. It seems that the battle tomorrow will be quite exciting!

I hope there will be an online mirror, and we'll solve it together with Gennady Korotkevich and Mikhail Tikhomirov for the fourth time in a row (2017, 2018, 2019), streaming the process online. We will use only 1 computer and stream to Gennady's Twitch channel this time. Tune in tomorrow around 9:15am Moscow time and ask your questions in chat! There is also an amazing official stream with commentary.

Sunday, September 26, 2021

A 202 week

Facebook Hacker Cup 2021 Round 2 narrowed the field to 500 qualifiers on Saturday (problems, results, top 5 on the left, analysis). maroonrk was a lot faster than everybody else, and he also bravely skipped writing a dedicated solution for C1, submitting problem C2 first and then quickly adapting its solution for C1. Congratulations on the clear first place!

This contest ended up being quite a struggle for me. First, I did run into the stack overflow issues that are a recurring topic for the Hacker Cup. It seems that I have changed my setup compared to the last year, and now even about 15 minutes of googling during the round and another half an hour after the round did not help me find a way to increase the stack size in the new setup. In the end I just gave up and submitted by compiling with g++ in the command line (in hindsight, this decision should have come earlier).

However, I'm still curious how to increase the stack size in my setup, which is CLion that uses WSL as the toolchain, compiling with clang++. I've tried various compiler/linker flags such as -Wl,-stack_size and  -Wl,--stack, but none of them worked. I've tried editing /etc/security/limits.conf inside WSL, but that did not seem to have any effect. I've found answers suggesting that the stack size can only be changed in WSL 2, and I double-checked that I do in fact have WSL 2. I've tried calling setrlimit from within my code, but it did not help either. I've found mentions that calling sudo prlimit --stack=unlimited --pid $$; ulimit -s unlimited would help, but it's not clear how to do it when I'm executing my program from within CLion, as I could not find a way to specify that some commands need to be run before my program. Putting that command in .profile did not help.

So, I'm turning to the readers of my blog for help. Do you know how to increase the stack size in the CLion+WSL+clang setup?

EDIT: this has been resolved in comments, it turns out that I did not double-check enough and I had WSL1. With WSL2 /etc/security/limits.conf does the trick.

This was not the end of my fight with C++, however. After implementing the solution for C2, I've tried to check it before submitting by feeding it the input from C1 augmented with many random changes. I've noticed that it did agree with my solution for C1 in debug mode, but it produced a different answer in release mode! I went on to debug this for around half an hour, could not find any issues, and finally noticed that compiling from the command line with g++ with optimization on (the release mode mentioned above was compiling with clang++ from within CLion) produces the correct answers, so I've used that to make my submission which turned out to be correct.

After the contest I discovered what the problem was using binary search on clang++ flags: in release mode, CLion (which inherits this behavior from CMake as I understand) passes -DNDEBUG flag, which makes assert(f()) statements not execute f() at all, and I've used assertions to verify that the elements that I try to erase from sets are actually found: assert(set.erase(x)). I guess everyone has to learn this lesson once :)

Finally, I was quite excited to come up with a cool solution for problem D inspired by the algorithm described in this post from 9 years ago (see the second solution to problem F). However, I missed the fact that even though I only discarded at most 18 numbers in my solution, some of those discarded numbers are differences of, or differences of differences of original numbers (and so on), so we can discard much more than 18 original numbers. Therefore my solution has failed, but luckily the first four problems were enough to advance. Funnily enough, if I'm reading the feedback page correctly, this solution has almost passed, only failing one testcase where it discarded 31 numbers (it was only allowed to discard 25).

Open Cup 2021-22 Grand Prix of XiAn followed on Sunday (results, top 5 on the left, analysis). Team USA1 has returned to the winning ways, even though they were not the first team to reach 10 problems: Ildar Gainullin did this 20 minutes earlier, but his penalty time was ruined by the +24 in problem I. Congratulations to both teams!

In my previous summary, I have mentioned an Open Cup problem: you are given a positive integer k up to 106, and need to produce any set of at most 30 integers with absolute values not exceeding 1016 such that the number of its subsets with zero sum, including the empty subset, is exactly k.

Since we did not have anything else to code, I got to play with this problem using a computer during the round. First, I've tried a greedy approach where we start with a set containing just the number 1, and then we keep adding numbers that give us as many additional zero-sum subsets as possible (but not exceeding k). This approach found a solution of size ~100 for large values of k if I remember correctly. When looking at the solution, I've noticed that after starting with a few 1s and -1s, it starts using bigger and bigger numbers, which, while locally optimal, means that on later steps we have many different sums, and therefore not so many subsets with the same sum.

I've tried to improve the solution by prescribing it to use more 1s and -1s before falling back to the greedy approach, which brought the solution size to ~70, still quite far from the required 30.

Then I've tried to think how a more constructive solution to this problem look like. One idea that came to mind is to split the solution into two parts: in the first part, we use only positive numbers, and therefore do not get any zero-sum subsets. Then, we use only negative numbers, and we have some kind of a knapsack problem: for each negative number, we know the number of ways to build its negation as the sum of some positive numbers, and we need those numbers to add up to k. It's not exactly a knapsack though since there might also be zero-sum subsets with more than one negative number.

However, this line of thought allowed me to come back to my greedy approach with a new idea: instead of starting with 1s ands -1s, I've tried starting with 1s and 2s first (just iterating how many of each we have, up to a total of 30), and then greedily adding only negative numbers, each time maximizing the number of new zero-sum subsets, but not exceeding k. This has dramatically improved its performance, finding solutions with less than 30 numbers for k=106 and k=106-1. I'm not sure if this rationalization is correct, but I think the reason for such improvement might be the fact that now we do the "exponential blowup" only once instead of twice (using only positive numbers for it).

This solution did not pass, but it passed quite a few tests with 1000 cases in each one, so I've modified it to also try having some 3s, which was enough :)

Thanks for reading, and check back next week!

Saturday, September 25, 2021

An unwrapping week

TopCoder SRM 813 took place last Friday (problems, results, top 5 on the left). With this victory tourist has guaranteed himself the TCO qualification spot, so that he does not have to wake up for the last SRM which takes place in the middle of the night — well done! I have grossly overestimated the chances of other solutions being broken, so my challenges for solutuons that did not "look right" brought me -75 points, and that despite one of the challenges being successful :)

The second stage of the new Open Cup season, the Grand Prix of IMO, took place on Sunday (results, top 5 on the left, analysis). Team USA1 continued their streak with the 21st top-3 finish in a row, spanning from the end of the 2019-20 season, but they could not get their fifth victory in a row since team ext71 was 5 penalty minutes faster after getting their 12th problem dramatically at 4:58 from +3. Congratulations to both teams!

I've had some fun digging through the constructive problem K: you are given a positive integer k up to 106, and need to produce any set of at most 30 integers with absolute values not exceeding 1016 such that the number of its subsets with zero sum, including the empty subset, is exactly k. I doubt it's possible to solve this problem completely in one's head or on paper, so if you're interested, I encourage you to try some idea out with a computer :)

Finaly, CodeChef September Cook-Off wrapped up the week (problems, results, top 5 on the left, analysis). The two last problems were a lot more difficult than the rest, and only gennady.korotkevich was able to solve them both. Great job!

In my previous summary, I have mentioned a VK Cup problem: you have a tape that is infinite in both directions and is split into cells. You're going to write a given random permutation of size n<=15000 onto this tape in the following manner: you write the first number somewhere, then you either stay in the same cell or move one cell to the left or to the right, then write the second number there, then again either stay or move to an adjacent cell, write the third number there, and so on. Whenever you write a number into a cell that already has a number, the old number is overwritten and therefore discarded. After all n numbers have been written, we look at the resulting sequence written on the tape from left to right. What is the maximum possible size of an increasing subsequence of this sequence? Note that the given permutation is guaranteed to have been picked uniformly at random from all permutations of size n.

The first idea is relatively standard for problems of this type: overwriting is hard to deal with, so let us look at the process from the end. Then instead of overwriting we will just skip writing numbers to cells that already have a number, and therefore we will always have a segment of cells that have numbers that will not change anymore, and the rest of the cells free.

The second thing to get out of the way is the "increasing subsequence" complication. In the new setup where we go from the end, it's clear that there's no point at all writing numbers that will not end up in that increasing subsequence, as we can simply stay in the current position and skip writing the unnecessary number instead of moving left/right to write it. The only exception to this is the very first number we write (in reverse order, so the last number of the input permutation) which will always be written even if it's useless for the increasing subsequence. We can deal with this peculiar property of the last number separately, so for the remainder of this post I will forget about it for simplicity.

Now all numbers we write must form an increasing sequence. The property "increasing" is local, so all subsequent numbers will only be compared with the first or the last number written so far, which leads us to a dynamic programming solution. Having processed some prefix of the reversed permutation, only four things matter:

  1. what is the first number of the sequence we have so far
  2. what is the last number of the sequence we have so far
  3. what is the length of the sequence we have so far
  4. where in this sequence are we
However, these four things plus the length of the prefix we have processed lead to O(n5) states, so while polynomial, this solution does not really cut it for n<=15000. So we need a few tricks to reduce the state space.

The first trick is to notice that since the permutation is random, the length of the sequence will be O(n0.5) instead of O(n) (more info). Therefore our position inside the sequence also has just O(n0.5) options, so we have reduced the number of states to O(n4) just by staring at the solution for a bit and without changing it :)

The next relatively standard trick is to notice that we don't need the first two dimensions in such detail. Instead, we can store the smallest last number for each combination of (first number, length, our position), which brings us down to O(n3) states. Symmetrically, we could have also stored the biggest first number for each combination of (last number, length, our position).

Here comes another optimization: if we only consider states right after we have written a new number, as opposed to at all moments, then three dimensions of our dynamic programming collapse into one! For example, if we have just written a new first number, then "where in the sequence are we" has just one option — at that first number. Moreover, the prefix of the reversed permutation that we have processed is also fixed — it's the prefix up to that first number. This optimization without the previous optimization would have brought the number of states down to O(n2.5).

Can we apply both optimizations together? This is what I could not figure out during the contest, but in retrospect it's quite straightforward: yes, we can! The key thing to notice is that we can't apply just "smallest last number" or just "biggest first number" optimizations by themselves. We must use both: if we have just written a new first number, then we need to store the smallest last number for this (first number, length) pair, and if we have just written a new last number, then we need to store the biggest first number for this (last number, length) pair! This brings the number of states down to just O(n1.5), finally something manageable for n<=15000.

However, having reduced the number of states we have increased the number of transitions: instead of just considering what to do with the next number for a transition, we have to iterate over which will be the next number to be written, leading to O(n) transitions for each state and O(n2.5) running time.

However, we can use yet another relatively standard trick: instead of naively doing all transitions, we can use a segment tree that stores all "pending" transitions, and be able to find the dynamic programming value for a new state in O(log(n)), solving our problem in O(n1.5log(n)), which is fast enough. I hope you enjoyed this long process of unwrapping the present as much as I did :)

Thanks for reading, and check back for this week's summary!

Monday, July 19, 2021

A reversal week

TopCoder SRM 809 was the first event of the last week (problems, results, top 5 on the left, analysis). This was apparently the first SRM that counts towards the TCO22 qualification, and I have completely overlooked this fact and did not prioritize participating :( Congratulations to those who pay attention, and especially to tourist on solving the hard problem almost twice faster than everybody else and thus earning the clear first place!

The VK Cup 2021 Elimination Round for those who speak Russian and qualified, which was also available as Codeforces Round 733 for everybody else, followed on Saturday (problemsresults, top 5 on the left, parallel round results, top 5 on the right, analysis stream, analysis). Everything went pretty well for me in this round, I got the solution ideas quickly and only had to do heavy debugging once, in problem G. However, Um_nik chose a superior strategy and went for problem H in the end which gave him much more points, and won the round. Well done!

More worryingly for me, despite me feeling pretty good about my performance on the first six problems, some usual suspects in the parallel round were faster to solve those six problems and managed to solve both harder problems to boot! Congratulations to jiangly and ecnerwala. It seems that just having a good day is no longer enough for me to compete for the first place at Codeforces :)

The hardest problem H was a pretty nice example of DP optimization. You have a tape that is infinite in both directions and is split into cells. You're going to write a given random permutation of size n<=15000 onto this tape in the following manner: you write the first number somewhere, then you either stay in the same cell or move one cell to the left or to the right, then write the second number there, then again either stay or move to an adjacent cell, write the third number there, and so on. Whenever you write a number into a cell that already has a number, the old number is overwritten and therefore discarded. After all n numbers have been written, we look at the resulting sequence written on the tape from left to right. What is the maximum possible size of an increasing subsequence of this sequence? Note that the given permutation is guaranteed to have been picked uniformly at random from all permutations of size n (which conveniently makes preparing the testcases for this problem so much easier!)

Thanks for reading, and check back next week!