Closed. This question needs details or clarity. It is not currently accepting answers.
Want to improve this question? Add details and clarify the problem by editing this post.
Closed 8 years ago.
Improve this question
Problem:
Given a sorted array of integers find the most frequently occurring integer. If there are multiple integers that satisfy this condition, return any one of them.
My basic solution:
Scan through the array and keep track of how many times you've seen each integer. Since it's sorted, you know that once you see a different integer, you've gotten the frequency of the previous integer. Keep track of which integer had the highest frequency.
This is O(N) time, O(1) space solution.
I am wondering if there's a more efficient algorithm that uses some form of binary search. It will still be O(N) time, but it should be faster for the average case.
Asymptotically (big-oh wise), you cannot use binary search to improve the worst case, for the reasons the answers above mine have presented. However, here are some ideas that may or may not help you in practice.
For each integer, binary search for its last occurrence. Once you find it, you know how many times it appears in the array, and can update your counts accordingly. Then, continue your search from the position you found.
This is advantageous if you have only a few elements that repeat a lot of times, for example:
1 1 1 1 1 2 2 2 2 3 3 3 3 3 3 3 3 3 3
Because you will only do 3 binary searches. If, however, you have many distinct elements:
1 2 3 4 5 6
Then you will do O(n) binary searches, resulting in O(n log n) complexity, so worse.
This gives you a better best case and a worse worst case than your initial algorithm.
Can we do better? We could improve the worst case by finding the last occurrence of the number at position i like this: look at 2i, then at 4i etc. as long as the value at those positions are the same. If they are not, look at (i + 2i) / 2 etc.
For example, consider the array:
i
1 2 3 4 5 6 7 ...
1 1 1 1 1 2 2 2 2 3 3 3 3 3 3 3 3 3 3
We look at 2i = 2, it has the same value. We look at 4i = 4, same value. We look at 8i = 8, different value. We backtrack to (4 + 8) / 2 = 6. Different value. Backtrack to (4 + 6) / 2 = 5. Same value. Try (5 + 6) / 2 = 5, same value. We search no more, because our window has width 1, so we're done. Continue the search from position 6.
This should improve the best case, while keeping the worst case as fast as possible.
Asymptotically, nothing is improved. To see if it actually works better on average in practice, you'll have to test it.
Binary search, which eliminates half of the remaining candidates, probably wouldn't work. There are some techniques you could use to avoid reading every element in the array. Unless your array is extremely long or you're solving a problem for curiosity, the naive (linear scan) solution is probably good enough.
Here's why I think binary search wouldn't work: start with an array: given the value of the middle item, you do not have enough information to eliminate the lower or upper half from the search.
However, we can scan the array in multiple passes, each time checking twice as many elements. When we find two elements that are the same, make one final pass. If no other elements were repeated, you've found the longest element run (without even knowing how many of that element is in the sorted list).
Otherwise, investigate the two (or more) longer sequences to determine which is longest.
Consider a sorted list.
Index 0 1 2 3 4 5 6 7 8 9 a b c d e f
List 1 2 3 3 3 3 3 3 3 4 5 5 6 6 6 7
Pass1 1 . . . . . . 3 . . . . . . . 7
Pass2 1 . . 3 . . . 3 . . . 5 . . . 7
Pass3 1 2 . 3 . x . 3 . 4 . 5 . 6 . 7
After pass 3, we know that the run of 3's must be at least 5, while the longest run of any other number is at most 3. Therefore, 3 is the most frequently occurring number in the list.
Using the right data structures and algorithms (use binary-tree-style indexing), you can avoid reading values more than once. You can also avoid reading the 3 (marked as an x in pass 3) since you already know its value.
This solution has running time O(n/k) which degrades to O(n) for k=1 for a list with n elements and a longest run of k elements. For small k, the naive solution will perform better due to simpler logic, data structures, and higher RAM cache hits.
If you need to determine the frequency of the most common number, it would take O((n/k) log k) as indicated by David to find the first and last position of the longest run of numbers using binary search on up to n/k groups of size k.
The worst case cannot be better than O(n) time. Consider the case where each element exists once, except for one element which exists twice. In order to find that element, you'd need to look at every element in the array until you find it. This is because knowing the value of any array element does not give you any information regarding the location of the duplicate element, until it's actually found. This is in contrast to binary search, where the value of an array element allows you to rule out many other elements.
No, in the worst case we have to scan at least n - 2 elements, but see
below for an algorithm that exploits inputs with many duplicates.
Consider an adversary that, for the first n - 3 distinct probes into the
n-element array, returns m for the value at index m. Now the algorithm
knows that the array looks like
1 2 3 ... i-1 ??? i+1 ... j-1 ??? j+1 ... k-1 ??? k+1 ... n-2 n-1 n.
Depending on what the ???s are, the sole correct answer could be j-1
or j+1, so the algorithm isn’t done yet.
This example involved an array where there were very few duplicates. In
fact, we can design an algorithm that, if the most frequent element
occurs k times out of n, uses O((n/k) log k) probes into the array. For
j from ceil(log2(n)) - 1 down to 0, examine the subarray consisting of
every (2**j)th element. Stop if we find a duplicate. The cost so far
is O(n/k). Now, for each element in the subarray, use binary search to
find its extent (O(n/k) searches in subarrays of size O(k), for a total
of O((n/k) log k)).
It can be shown that all algorithms have a worst case of Omega((n/k) log
k), making this one optimal in the worst case up to constant factors.
Related
We have been given an array of size N 1 <= N <= 1e5, with Ai positive integers, such that
1 <= Ai <= 1e9.
we will be given Q queries. 1 <= Q <= 1e5
Every time in a query there will be two space separated integers b c , 1 <= b,c <= N
For every query we need to find that Is moving from index b of array to index c of array possible ?, and if it is then we have find a special sum, which i have explained below.
We can't just move in array simply from i to i+1 index, there is a restriction. If we want to move from i to j then A[j] should be strictly greater than A[i], i.e A[j] > A[i].
Note here one thing that : While moving we have to take the just next greater element than the current.
The sum what we need to find is sum of elements that came in the path taken to reach destination.
For Example
array : 3 2 5 4 6 6 7
query : 1 7
So, according to query we need to move from 1st element to last element if possible.
As, we can see we can take 3 --> 5 --> 6 --> 7 path to reach the destination and sum is 3+5+6+7 = 21
But if last element in array was 2
array : 3 2 5 4 6 6 2
query : 1 7
For this query we cant reach to destination as after 6 the destination element 2 is smaller than it. So for this query NO answer exist.
My approach
I know i can find the answer in O(n), by traversing the array simply from A(b) to A(c) and finding out that if answer exit or not as well as sum.
But the Problem is that There are a lot of queries so if i use O(n) solution the Time Complexity will be O(QN).
Time limit is only 1 sec, So i need to find a constant time O(c)solution for this.
One Thing more The becomes even tougher when Queries of second type appear.
Query type 2: In this query we need to update the value at an index with a given K.
query : b k , then A[b] = K.
Can anyone help me on this ??
The question is asking for N queries, the solution is most probably to do a pre-process to compute the possibilities and then query each of them in O(1) time.
Given an array {a0,a1,a2,a3,a4,.....an}.I need to find the minimum number of replacements of number to make the array strictly increasing.
I know that there exists dynamic programming solutions for the question.
https://www.geeksforgeeks.org/convert-to-strictly-increasing-integer-array-with-minimum-changes/
I have been trying out a greedy solution for the problem.
This is the Pseudo Code - (Greedy)
Start iterating on the array from a[1] (0-based indexing).
if a[i] > a[i-1] then continue
else -
if (a[i-2]+1 < a[i] && a[i-1] != a[i-2]+1) , then make a[i-1] = a[i-2] + 1
else make a[i] = a[i-1] + 1
Continue till end of array.
Basic approach is, that if an out of order element is found(a[i]) then first check if the previous element
(a[i-1]) can be decreased or not.
If it is possible to decrease it check if decreasing it to the minimum possible number changes the fact that a[i] < a[i-1] or not.
If it make no change to that fact then don't decrease a[i-1], instead increase a[i] to a[i-1] + 1.
eg - 3 2 5 3 8 4 4
2<3, so decrease 3 to 0. Array becomes - 0 2 5 3 8 4 4
5>2 so continue.
3<5, decreasing to 3 and 4 is possible but not effective. So increase 3 to 6. Array - 0 2 5 6 8 4 4
8>6 so continue
4<8, decreasing 8 to 7 is possible but not effective. Array - 0 2 5 6 8 9 4
4<9, decreasing 9 to 8 is not possible, so increase 4 to 10. Array - 0 2 5 6 8 9 10
Total changes = 4. Giving the answer in O(n).
I have been trying to prove the greedy strategy wrong but have been unable to do so. It is ignoring the LIS the same way that the DP approach ignores it. (i.e notice how 2,5,8 remain same and others change)
Is the solution correct ? If not can you please provide a counter-example.
Please also let me know your "thought" process that led you to think of the counter-example, if possible. Would be of great help in future questions.
Which leads me to the question i wanted to ask most -
How exactly do you figure out if your greedy solution is correct ? How do you realize that greedy wont work and you'll have to use dynamic programming ?
If the pseudo code is unclear for anyone,please let me know. I would be more than happy to submit the complete code.
This question already has answers here:
Sum of products of elements of all subarrays of length k
(2 answers)
Permutation of array
(13 answers)
Closed 7 years ago.
I will start with an example. Suppose we have an array of size 3 with elements a, b and c like: (where a, b and c are some numerical values)
|1 | 2| 3| |a | b| c|
(Assume index starts from 1 as shown in the example above)
Now all possible increasing sub-sequence of length 2 are:
12 23 13
so the sum of product of elements at those indexes is required, that is, ab+bc+ac
For length 3 we have only one increasing sub-sequence, that is, 123 so abc should be printed.
For length 4 we have no sequence so 0 is printed and the program terminates.
So output for the given array will be:
ab+bc+ac,abc,0
So for example if the elements a, b and c are 1, 2 and 3 respectively then the output should be 11,6,0
Similarly, for an array of size 4 with elements a,b,c,d the output will be:
ab+ac+ad+bc+bd+cd,abc+abd+acd+bcd,abcd,0
and so on...
Now obviously brute force will be too inefficient for large value of array size. I was wondering if there is an efficient algorithm to compute the output for an array of given size?
Edit 1: I tried finding a pattern. For example for an array of size 4:
The first value we need is :(ab+ac+bc)+d(a+b+c)= ab+ac+ad+bc+bd+cd (Take A=ab+ac+bd)
then the second value we need is:(abc) +d(A) = abc+abd+acd+bcd(B=abc)
then the third value we need is : (0) +d(B) = abcd(Let's take 0 as C)
then the fourth value we need is: +d(C) = 0
But it still requires a lot of computation and I can't figure out an efficient way to implement this.
Edit 2: My question is different then this since:
I don't need all possible permutations. I need all possible increasing sub-sequences from length 2 to n+1.
I also don't need to print all possible such sequences, I just need the value thus obtained (as explained above) and hence I am looking for some maths concept or/and some dynamic programming approach to solve this problem efficiently.
Note I am finding the set of all possible such increasing sub-sequences based on the index value and then computing based on the values at those index position as explained above.
As a post that seems to have disappeared pointed out one way is to get a recurrence relation. Let S(n,k) be the sum over increasing subsequences (of 1..n) of length k of the product of the array elements indexed by the sequence. Such a subsequence either ends in n or not; in the first case it's the concatenation of a subsequence of length k-1 of 1..n-1 and {n}; in the second case it's a subsequence of 1..n-1 of length k. Thus:
S(n,k) = S(n-1,k) + A[n] * S(n-1,k-1)
For this always to make sense we need to add:
S(n,0) = 1
S(n,m) = 0 for m>n
The classic 2sum question is simple and well-known:
You have an unsorted array, and you are given a value S. Find all pairs of elements in the array that add up to value S.
And it's always been said that this can be solved with the use of HashTable in O(N) time & space complexity or O(NlogN) time and O(1) space complexity by first sorting it and then moving from left and right,
well these two solution are obviously correct BUT I guess not for the following array :
{1,1,1,1,1,1,1,1}
Is it possible to print ALL pairs which add up to 2 in this array in O(N) or O(NlogN) time complexity ?
No, printing out all pairs (including duplicates) takes O(N2). The reason is because the output size is O(N2), thus the running time cannot be less than that (since it takes some constant amount of time to print each element in the output, thus to simply print the output would take CN2 = O(N2) time).
If all the elements are the same, e.g. {1,1,1,1,1}, every possible pair would be in the output:
1. 1 1
2. 1 1
3. 1 1
4. 1 1
5. 1 1
6. 1 1
7. 1 1
8. 1 1
9. 1 1
10. 1 1
This is N-1 + N-2 + ... + 2 + 1 (by taking each element with all elements to the right), which is
N(N-1)/2 = O(N2), which is more than O(N) or O(N log N).
However, you should be able to simply count the pairs in expected O(N) by:
Creating a hash-map map mapping each element to the count of how often it appears.
Looping through the hash-map and summing, for each element x up to S/2 (if we go up to S we'll include the pair x and S-x twice, let map[x] == 0 if x doesn't exist in the map):
map[x]*map[S-x] if x != S-x (which is the number of ways to pick x and S-x)
map[x]*(map[x]-1)/2 if x == S-x (from N(N-1)/2 above).
Of course you can also print the distinct pairs in O(N) by creating a hash-map similar to the above and looping through it, and only outputting x and S-x the value if map[S-x] exists.
Displaying or storing the results is O(N2) only.The worst case as highlighted by you clearly has N2 pairs and to write them onto the screen or storing them into a result array would clearly require at least that much time.In short, you are right!
No
You can pre-compute them in O(nlogn) using sorting but to print them you may need more than O(nlogn).In worst case It can be O(N^2).
Let's modify the algorithm to find all duplicate pairs.
As an example:
a[ ]={ 2 , 4 , 3 , 2 , 9 , 3 , 3 } and sum =6
After sorting:
a[ ] = { 2 , 2 , 3 , 3 , 3 , 4 , 9 }
Suppose you found pair {2,4}, now you have to find count of 2 and 4 and multiply them to get no of duplicate pairs.Here 2 occurs 2 times and 1 occurs 1 times.Hence {2,1} will appear 2*1 = 2 times in output.Now consider special case when both numbers are same then count no of occurrence and sq them .Here { 3,3 } sum to 6. occurrence of 3 in array is 3.Hence { 3,3 } will appear 9 times in output.
In your array {1,1,1,1,1} only pair {1,1} will sum to 2 and count of 1 is 5.hence there are going to 5^2=25 pairs of {1,1} in output.
Ideally I'm looking for a c# solution, but any help on the algorithm will do.
I have a 2-dimension array (x,y). The max columns (max x) varies between 2 and 10 but can be determined before the array is actually populated. Max rows (y) is fixed at 5, but each column can have a varying number of values, something like:
1 2 3 4 5 6 7...10
A 1 1 7 9 1 1
B 2 2 5 2 2
C 3 3
D 4
E 5
I need to come up with the total of all possible row-wise sums for the purpose of looking for a specific total. That is, a row-wise total could be the cells A1 + B2 + A3 + B5 + D6 + A7 (any combination of one value from each column).
This process will be repeated several hundred times with different cell values each time, so I'm looking for a somewhat elegant solution (better than what I've been able to come with). Thanks for your help.
The Problem Size
Let's first consider the worst case:
You have 10 columns and 5 (full) rows per column. It should be clear that you will be able to get (with the appropriate number population for each place) up to 5^10 ≅ 10^6 different results (solution space).
For example, the following matrix will give you the worst case for 3 columns:
| 1 10 100 |
| 2 20 200 |
| 3 30 300 |
| 4 40 400 |
| 5 50 500 |
resulting in 5^3=125 different results. Each result is in the form {a1 a2 a3} with ai ∈ {1,5}
It's quite easy to show that such a matrix will always exist for any number n of columns.
Now, to get each numerical result, you will need to do n-1 sums, adding up to a problem size of O(n 5^n). So, that's the worst case and I think nothing can be done about it, because to know the possible results you NEED to effectively perform the sums.
More benign incarnations:
The problem complexity may be cut off in two ways:
Less numbers (i.e. not all columns are full)
Repeated results (i.e. several partial sums give the same result, and you can join them in one thread). Much more in this later.
Let's see a simplified example of the later with two rows:
| 7 6 100 |
| 3 4 200 |
| 1 2 200 |
at first sight you will need to do 2 3^3 sums. But that's not the real case. As you add up the first column you don't get the expected 9 different results, but only 6 ({13,11,9,7,5,3}).
So you don't have to carry your nine results up to the third column, but only 6.
Of course, that is on the expense of deleting the repeating numbers from the list. The "Removal of Repeated Integer Elements" was posted before in SO and I'll not repeat the discussion here, but just cite that doing a mergesort O(m log m) in the list size (m) will remove the duplicates. If you want something easier, a double loop O(m^2) will do.
Anyway, I'll not try to calculate the size of the (mean) problem in this way for several reasons. One of them is that the "m" in the sort merge is not the size of the problem, but the size of the vector of results after adding up any two columns, and that operation is repeated (n-1) times ... and I really don't want to do the math :(.
The other reason is that as I implemented the algorithm, we will be able to use some experimental results and save us from my surely leaking theoretical considerations.
The Algorithm
With what we said before, it is clear that we should optimize for the benign cases, as the worst case is a lost one.
For doing so, we need to use lists (or variable dim vectors, or whatever can emulate those) for the columns and do a merge after every column add.
The merge may be replaced by several other algorithms (such as an insertion on a BTree) without modifying the results.
So the algorithm (procedural pseudocode) is something like:
Set result_vector to Column 1
For column i in (2 to n-1)
Remove repeated integers in the result_vector
Add every element of result_vector to every element of column i+1
giving a new result vector
Next column
Remove repeated integers in the result_vector
Or as you asked for it, a recursive version may work as follows:
function genResVector(a:list, b:list): returns list
local c:list
{
Set c = CartesianProduct (a x b)
Set c = Sum up each element {a[i],b[j]} of c </code>
Drop repeated elements of c
Return(c)
}
function ResursiveAdd(a:matrix, i integer): returns list
{
genResVector[Column i from a, RecursiveAdd[a, i-1]];
}
function ResursiveAdd(a:matrix, i==0 integer): returns list={0}
Algorithm Implementation (Recursive)
I choose a functional language, I guess it's no big deal to translate to any procedural one.
Our program has two functions:
genResVector, which sums two lists giving all possible results with repeated elements removed, and
recursiveAdd, which recurses on the matrix columns adding up all of them.
recursiveAdd, which recurses on the matrix columns adding up all of them.
The code is:
genResVector[x__, y__] := (* Header: A function that takes two lists as input *)
Union[ (* remove duplicates from resulting list *)
Apply (* distribute the following function on the lists *)
[Plus, (* "Add" is the function to be distributed *)
Tuples[{x, y}],2] (*generate all combinations of the two lists *)];
recursiveAdd[t_, i_] := genResVector[t[[i]], recursiveAdd[t, i - 1]];
(* Recursive add function *)
recursiveAdd[t_, 0] := {0}; (* With its stop pit *)
Test
If we take your example list
| 1 1 7 9 1 1 |
| 2 2 5 2 2 |
| 3 3 |
| 4 |
| 5 |
And run the program the result is:
{11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27}
The maximum and minimum are very easy to verify since they correspond to taking the Min or Max from each column.
Some interesting results
Let's consider what happens when the numbers on each position of the matrix is bounded. For that we will take a full (10 x 5 ) matrix and populate it with Random Integers.
In the extreme case where the integers are only zeros or ones, we may expect two things:
A very small result set
Fast execution, since there will be a lot of duplicate intermediate results
If we increase the Range of our Random Integers we may expect increasing result sets and execution times.
Experiment 1: 5x10 matrix populated with varying range random integers
It's clear enough that for a result set near the maximum result set size (5^10 ≅ 10^6 ) the Calculation time and the "Number of != results" have an asymptote. The fact that we see increasing functions just denote that we are still far from that point.
Morale: The smaller your elements are, the better chances you have to get it fast. This is because you are likely to have a lot of repetitions!
Note that our MAX calculation time is near 20 secs for the worst case tested
Experiment 2: Optimizations that aren't
Having a lot of memory available, we can calculate by brute force, not removing the repeated results.
The result is interesting ... 10.6 secs! ... Wait! What happened ? Our little "remove repeated integers" trick is eating up a lot of time, and when there are not a lot of results to remove there is no gain, but looses in trying to get rid of the repetitions.
But we may get a lot of benefits from the optimization when the Max numbers in the matrix are well under 5 10^5. Remember that I'm doing these tests with the 5x10 matrix fully loaded.
The Morale of this experiment is: The repeated integer removal algorithm is critical.
HTH!
PS: I have a few more experiments to post, if I get the time to edit them.