randomx_123's blog

By randomx_123, history, 3 months ago, In English

So I gave the on-campus Google hiring test yesterday and when I saw the 2 problems I had to solve in an hour, I was really excited since they looked doable. However, things took a turn when I actually started to think about optimizations. Anyways, here goes :

Problem -1

We have an array A of N elements. We will be asked Q queries. each query is in the form of a single integer, X, and we have to tell whether there exists an index i in the array such that the bitwise AND of A[i] and X equals 0. If such an index exists print YES, otherwise print NO.

Constraints : 1<=N<=1e5, 1<=Q<=1e5, 1<=A[i]<=1e5

Problem 2

We have a binary string S of length N, and we have to find the number of substrings(that is, a contiguous range i to j) in which the frequency of 1 is strictly greater than the frequency of 0.

Constraints : 1<=N<=1e6;

I have spent a lot of time on them but could not come up with an optimal approach for either. I realize that the 2nd problem can be solved in O(NlogN) in a similar fashion in which we solve LIS in NlogN time, But other than that, I am clueless about both problems. Any help will be appreciated, Thanks!

 
 
 
 
  • Vote: I like it
  • +88
  • Vote: I do not like it

»
3 months ago, # |
  Vote: I like it 0 Vote: I do not like it

Auto comment: topic has been updated by randomx_123 (previous revision, new revision, compare).

»
3 months ago, # |
Rev. 3   Vote: I like it +76 Vote: I do not like it

A few hints for problem 2 in your post:

Hint 1
Hint Answer
Hint 2
Hint Answer
Hint 3
Hint Answer
  • »
    »
    3 months ago, # ^ |
      Vote: I like it +3 Vote: I do not like it

    I used segment tree to find no of prefixes with sum strictly less than current.

    • »
      »
      »
      3 months ago, # ^ |
        Vote: I like it 0 Vote: I do not like it

      Yeah, that's the way we do LIS in NlogN. But is there any simpler method?

      • »
        »
        »
        »
        3 months ago, # ^ |
        Rev. 4   Vote: I like it 0 Vote: I do not like it

        I simply made an array of size 2*N... First N indices to store frequency of negative sums and next N to store positive sums.

        Then at each index i called a query(0, N+cur-1) where cur is current prefix sum and updated +1 to present index.

        Could not think of anything else during challenge.

        • »
          »
          »
          »
          »
          3 months ago, # ^ |
            Vote: I like it 0 Vote: I do not like it

          Can you Please Share your code if you dont mind ? i need to know how excatly you did ? Sorry Sir but i am noob here .

      • »
        »
        »
        »
        3 months ago, # ^ |
          Vote: I like it +11 Vote: I do not like it

        Could you import ordered set on that compiler?

        • »
          »
          »
          »
          »
          3 months ago, # ^ |
            Vote: I like it 0 Vote: I do not like it

          Yes. you could have imported anything present in bits/stdc++ (almost everything) as it was pre-written in the code.

        • »
          »
          »
          »
          »
          3 months ago, # ^ |
            Vote: I like it +4 Vote: I do not like it

          Yes , I imported gnu_pbds

      • »
        »
        »
        »
        3 months ago, # ^ |
          Vote: I like it +8 Vote: I do not like it

        I think socho has explained this part.

        You have 2 indexes l and r.

        Now you have to find number of ordered pairs of (l, r).

        Such that l < r and a[l] < a[r].

        This can be changed into a known problem of counting-inversions.

        Totalways = (n choose 2) — Ways such that (l < r and a[l] >= a[r]).

        You can solve the latter using merge-sort.

      • »
        »
        »
        »
        3 months ago, # ^ |
          Vote: I like it +3 Vote: I do not like it

        You can count inversions using policy based data structure, ordered set in nlogn

      • »
        »
        »
        »
        3 months ago, # ^ |
          Vote: I like it +3 Vote: I do not like it
        1. Simplest & shortest of all — Use pbds(policy based data structure) to find how many smaller values are there
        2. Compress the prefix array since negative values can be there and use BIT(binary indexed tree) to find how many smaller values are there
        3. Use merge sort. While merge process we already know how many values in right half are smaller than current value in left half, use that to find how many smaller values are there

        You can find in detail explanations here

      • »
        »
        »
        »
        3 months ago, # ^ |
          Vote: I like it 0 Vote: I do not like it

        how did you did this question in LIS(Nlogn) approach ! can you please explain it

      • »
        »
        »
        »
        3 months ago, # ^ |
          Vote: I like it +14 Vote: I do not like it

        yes, there is. every next prefix will increase or decrease by 1, so you can use answer of previous element and add/subtract frequency of specific value. this can be done in O(n)

  • »
    »
    3 months ago, # ^ |
    Rev. 2   Vote: I like it +5 Vote: I do not like it

    I did the exact same thing, But then since the prefix sums are not increasing,how do we find the number of positive-sum subarrays in less than O(n^2) ? edit : thanks for taking up the time and explaining, Really appreciate.

  • »
    »
    3 months ago, # ^ |
    Rev. 2   Vote: I like it 0 Vote: I do not like it

    will it overflow when pre[i] will reach beyond 1e9 ?? how to deal then with fenwick tree ??

    Spoiler
    • »
      »
      »
      3 months ago, # ^ |
      Rev. 2   Vote: I like it 0 Vote: I do not like it
      Will it overflow?
      • »
        »
        »
        »
        3 months ago, # ^ |
          Vote: I like it 0 Vote: I do not like it

        Sir , in my implementation i have used prefix sum . So if all elements are 1 . Then sum of 1e6 numbers is nothing but sum of first 1e6 natural numbers . and it will overflow . so Am i doing correct in my implementation?? what do am i missing here Sir ??

        • »
          »
          »
          »
          »
          3 months ago, # ^ |
            Vote: I like it 0 Vote: I do not like it

          The sum of all elements in your modified (replace 0 with -1) array is at most $$$N$$$. You don't need the prefix sum on your prefix sum array, that's what you seem to be doing.

  • »
    »
    3 months ago, # ^ |
      Vote: I like it 0 Vote: I do not like it

    Continuing after Hint 1 — I think we do not have to count inversions. We could just use two pointer method to calculate number of subarrays with sum greater than or equal to X.

    • »
      »
      »
      3 months ago, # ^ |
        Vote: I like it 0 Vote: I do not like it

      bro, I was thinking the same but I got stuck on how to calculate numbers when u get start and end pointer.

»
3 months ago, # |
  Vote: I like it +62 Vote: I do not like it

Problem 1: a[i] & x == 0 iff a[i] is a submask of ~x. Rephrasing the problem, you're given a mask and you want to know whether there's its submask in the array. Use sum over submasks dp to pre-compute the answer for all ~x values at once and then answer each query in $$$\mathcal{O}(1)$$$ by accessing this pre-computed array.

  • »
    »
    3 months ago, # ^ |
      Vote: I like it +8 Vote: I do not like it

    Another approach would be to make a dp, where dp(i) denotes the smallest element in the set (of A) such that:
    i&dp(i) = 0, or -1, if no such element exists.
    Updates can be done as : dp[i] = dp[i^2j], where j is the highest active bit in i. Base case is that we will have to find dp(i) for all i which are powers to 2 separately, that can be easily.

  • »
    »
    3 months ago, # ^ |
      Vote: I like it +19 Vote: I do not like it
    Implementation
»
3 months ago, # |
  Vote: I like it -54 Vote: I do not like it

Both the problems are well known :) .

»
3 months ago, # |
  Vote: I like it 0 Vote: I do not like it

was this a hiring test for internships?

»
3 months ago, # |
  Vote: I like it 0 Vote: I do not like it

from where you got the link of test?

»
3 months ago, # |
  Vote: I like it -12 Vote: I do not like it

Problem-1 can be solved using Trie.

  • »
    »
    3 months ago, # ^ |
      Vote: I like it 0 Vote: I do not like it

    Can you explain more ? Split for every bit ? And then for each query in O(32) ?

  • »
    »
    3 months ago, # ^ |
    Rev. 2   Vote: I like it +30 Vote: I do not like it

    Not Really, you could have solved this using Trie if the operation was Bitwise-Xor instead of Bitwise-And. Because incase of And when your current bit is 0, you can go either way in the trie because it doesn't really matter, so you might end up traversing the whole trie for each query.

    The key observation to solve this problem is that A[i] <= 1e5 and then do something like sum of subsets dp to preprocess answer for every possible query

  • »
    »
    3 months ago, # ^ |
      Vote: I like it 0 Vote: I do not like it

    In reasonable time ~ n•log(n)? Doubt it. It’s not Xor.

»
3 months ago, # |
  Vote: I like it +1 Vote: I do not like it

Off topic,

I have seen another set in which the first question was some significantly easier stack question and the second question was same. So while selection is some kind of normalization done? Or is it the case that people of same college get same set?(which is kind of pointless)

  • »
    »
    3 months ago, # ^ |
      Vote: I like it 0 Vote: I do not like it

    I don't know about normalization is being done or not, but everybody gets a different set

»
3 months ago, # |
  Vote: I like it +8 Vote: I do not like it

Problem 1 is same as this one https://codeforces.com/contest/165/problem/E The only difference being the array itself forms the queries

»
3 months ago, # |
Rev. 4   Vote: I like it -18 Vote: I do not like it

the first one is a basic question of tries , you need to find the prefix of your choice if the number has a 1 find a 0 else if its 0 then look for 1 or 0.

the second one is also a standard question called inversion count , we can make a prefix. sum array of the string by converting the 0's to -1 and keeping 1 as 1, then it can be solved using merge sort, bit tree, avl tree, ordered set and idk if there are more ways

  • »
    »
    3 months ago, # ^ |
      Vote: I like it 0 Vote: I do not like it

    Wont't the time complexity for first one be O(n^2) in worst case in your solution?

  • »
    »
    3 months ago, # ^ |
      Vote: I like it +9 Vote: I do not like it

    Wouldn't looking for a 1 or 0 in the trie increase the time complexity? Because now instead of checking a single path in the trie, in the worst case (X is all zeroes) , you would have to traverse the whole trie, which takes O(n log(A)) per query.

»
3 months ago, # |
  Vote: I like it 0 Vote: I do not like it

I have a very easy approach for problem A which no-one has mentioned.

We'll create a vector of sets in which v[i] will contain set of possible elements if we consider rightmost i bits for every element

Now for X let j be the index of most significant bit of X (from right) So we will query in v[j] for complement of X and if it exists answer is yes otherwise no.

Proof :- if X&Y==0 then after removing those bits which are not present in X, Y will become compliment of X. If Y<X then we'll pad Y with zeros.

I think this approach will work, feel free to point out any mistakes you find.

»
3 months ago, # |
  Vote: I like it 0 Vote: I do not like it

I thought on a funny solution to problem 2 and it ended up being O(N) after some optimization (which is what i think is the complexity you were aiming for).

I'll define 0 to be -1 instead so it will make calculations easier. Make an array V of size 2*n+1, where we will be storing all the the number of sequences that end on the index we're iterating and have a certain sum. Initially we will assume the point in that array that represents sum 0 to be n. Let's say we filled this array from the start of the string to i-1. Consider these two cases:

If S[i] = 1: it's like the "zero point" of the sequences from j to i-1 "went down", because we will be adding 1 to all other sufixes that come before it. In addition, we do on V[newzeropoint+1].

If S[i] = 0: like the previous case, it's like the "zero point" went up, and in addition to that V[newzeropoint+1]++.

At any given index we're iterating over the initial string the number of subsequences that end on that point is the sum of all V[i] such that i > zeropoint. We can do that in O(logn) with a fenwick tree (aka BIT) but we just need to mantain a variable storing this sum and update it as we're moving the zero point up and down for a O(n) solution

»
3 months ago, # |
Rev. 2   Vote: I like it +3 Vote: I do not like it

Approach for the 2nd question:

1.replace each 0 with -1

2.calculate prefix sum of newly formed array after replacement ( call this array pref)

3.now we have to find the number of sub-arrays who's sum is > 0.

4.let's say we have to find the sub-arrays ending at index j.

5.then sub-array [i,j] has +ve sum iff, pref[j] — pref[i-1] > 0 =>pref[j] > pref[i-1]

6.it means we have to find number of elements pref[i] which are smaller than pref[j] such that i < j.

8.for this (in python ) we use SortedList a special type of container which keeps the list sorted and upon binary searching elements we can get the number of elements < pref[j] .

9.as SortedList is a special type of container it won't be allowed in the online rounds, hence here is the link for the source code for it, https://ideone.com/RyMNKu

10.we will traverse from left to right and keep pushing elements into the SortedList and add answer for each index to our final answer.

link for the code I have solved using SortedList: https://codeforces.com/contest/1536/submission/122169458

I have no idea how to solve this in c++ ,as set/multiset returns pointer, how to convert it to index ?

»
3 months ago, # |
  Vote: I like it +3 Vote: I do not like it

problem 2

Brute force (c++)
Pbds (c++)
Fenwick tree | BIT (c++)
»
3 months ago, # |
Rev. 2   Vote: I like it 0 Vote: I do not like it

Does anybody have an idea how to solve Problem-1? (Considering the Trie solution is not very optimal in the worst case)

Best I could think is maintaining a set for each i from 1 to 32. If a number is having set bit at position 'i', then we add it to that set. Now to answer for query X, we check the set bit positions and take intersection of all those sets. But again, taking intersection might give TLE. Can this be improved or any other way to solve this?

  • »
    »
    3 months ago, # ^ |
      Vote: I like it +3 Vote: I do not like it

    There is a concept known as SOS DP , which is helpful for the first question

    • »
      »
      »
      3 months ago, # ^ |
        Vote: I like it 0 Vote: I do not like it

      But won't it be of exponential complexity? It is given Q,X <= 7x10^3 Am I missing something obvious?

      • »
        »
        »
        »
        3 months ago, # ^ |
        Rev. 2   Vote: I like it +3 Vote: I do not like it

        a_i <= 10^ 5, which is something like 20 bits, so the complexity is gonna be O(20*2^20 +N) for precomputation and O(Q) for answering the queries.

»
3 months ago, # |
  Vote: I like it 0 Vote: I do not like it
»
3 months ago, # |
Rev. 2   Vote: I like it 0 Vote: I do not like it

I too got the same set.

I couldnt do 2nd one but I did 1st one For the first one I created an array of powers of 2 of length 20 and then for every element of this array I made sure whether it is in the keys array or not. Then for each query I traversed through the power of two's array and checked if that element exists in keys array and its and with the X = 0 then yes otherwise no.

PS: I had done 1st one partially...sorry for the mistake.

  • »
    »
    3 months ago, # ^ |
      Vote: I like it +17 Vote: I do not like it

    wait, what? you're saying you passed all test cases using this completely wrong method?

»
3 months ago, # |
Rev. 3   Vote: I like it -16 Vote: I do not like it

My implementation for 1st problem. please correct me if it is wrong.

const int N = 1e5 + 100;
int a[N];
int n,q,x;
struct trie{
    bool is;
    trie* left;
    trie* right;
};

struct trie* get(bool is){
    trie* r = new trie;
    r->left = nullptr;
    r->right = nullptr;
    r->is = is;
    return r;
}

void insert(trie* root,int n){
    trie* curr = root;
    while(n){
        int idx = 1 &mdash; n%2;
        if(idx == 0){
            if(!curr->left){
                curr->left = get(false);
            }
            curr = curr->left;
        }
        else{
            if(!curr->right)
                curr->right = get(false);
            curr = curr->right;
        }
        n/=2;
    }
    curr->is = true;
}

bool search(int n,trie* curr){
    if(!n) return true;
    int idx = n%2;
    if(idx == 0 && curr->left){
        if(curr->left->is){
            return true;
        }
        else return search(n/2,curr->left);
    }
    else if(idx == 1 && curr->right){
        if(curr->right->is)
            return true;
        else return search(n/2,curr->right);
    }
    return false;    
}

void test_case(int t) 
{
    cin>>n>>q;
    trie* root = get(false);
    for(int i=1;i<=n;i++){
        cin>>a[i];
        insert(root,a[i]);
    }
    while(q--){
        cin>>x;
        cout<<(search(x,root) ? "YES" : "NO")<<endl;
    }
}
»
3 months ago, # |
  Vote: I like it -6 Vote: I do not like it

I think there exists an o(n) solution for the 2nd question, we convert all 0 to -1 and want to find all l < r such that pref[r] > pref[l],

so the optimization basically boils down to : given an array arr, can you find all l < r such that arr[l] < arr[r] in O(n), if abs(arr[i + 1] — arr[i]) == 1; and this seems possible if we store just two arrays what is the current prefix and what have we not updated yet.

»
3 months ago, # |
  Vote: I like it +3 Vote: I do not like it

For the first, read about SOS DP