Related
I was trying to do my friends problem set from a few years ago to sharpen up my knowledge about data structures etc. I came across this problem, and I'm not really sure where to start. Hopefully someone could help me out!
We are given n unsorted arrays, each array has n elements. Ex.
3 1 2
7 6 9
4 9 12
Now, say we take one element from each array, and add them up. Lets just call the sum of these elements an "n-sum".
I need to devise an algorithm that gives us the n smallest "n-sums" (duplicates are allowed).
In our above ex, the answer would be:
11, 12, 12
# 11 comes from: 1 (first array) + 6 (second array) + 4 (third array)
# 12 comes from: 2 (first array) + 6 (second array) + 4 (third array)
# 12 comes from: 1 (first array) + 7 (second array) + 4 (third array)
One of the suggestions given were to use a priority queue.
Thanks!
The time is at least O (n^2): You must visit all array elements, because if all elements were equal to 1000 except on in each row being 0, you would have to look at the n elements equal to 0, or you couldn't find the smallest sum.
Sort each row, taking O (n^2 log n) steps. In each row, subtract the first element from all elements in the row, so the first element in each row is 0; after you found the smallest sums you can compensate for that. Your example turns into
3 1 2 -> 1 2 3 -> 0 1 2
7 6 9 -> 6 7 9 -> 0 1 3
4 9 12 -> 4 9 12-> 0 5 7
Now finding all sums ≤ K can be done in m steps if there are m sums: In the first row, pick all values in turn as long as they are ≤ K. In the second row, pick all values in turn as long as the sum from two rows is ≤ K and so on. Since each row starts with 0, no time is wasted.
For example, sums ≤ 5 are: 0+0+0, 0+0+5, 0+1+0, 0+3+0, 1+0+0, 1+1+0, 1+3+0, 2+0+0, 2+1+0, 2+3+0. Many more than the three that we needed. If we stop after finding 3 sums ≤ 5, we know very quickly "there are at least 3 sums ≤ 5". We need to have an early stop, because in the general case there could be n^n possible sums.
If you pick K = "largest element in the second column", then you know there are at least n+1 sums with a value ≤ K, because you can pick all 0's, or all 0's except one value from the second column. In your example, K = 5 (we know that worked). Let X be the value where there are n sums ≤ X but fewer than n sums ≤ X - 1. We find X with binary search between 0 and K, and then we find the sums. Example:
K = 5 is known to be big enough. We try K = 2, and find 4 sums (actually we stop at 3 sums). Too many. We try K = 1, and there are three solutions 0+0+0, 0+1+0 and 1+0+0. We try K = 0, but only one solution.
This part goes very quick, so we'd try to reduce the time used for sorting. We notice that in this case looking at the first two columns was enough. We can in each row find the two smallest items, and in this case that would be enough. If the two smallest items are not enough to determine the n smallest sums, find the third smallest item etc. where needed. For example, since the 2nd largest item of the last row is 5, we wouldn't need the third item of the row, because even the 5 is not element of a sum if K ≤ 4.
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.
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.
Given an array of positive integers a I want to output array of integers b so that b[i] is the closest number to a[i] that is smaller then a[i], and is in {a[0], ... a[i-1]}. If such number doesn't exist, then b[i] = -1.
Example:
a = 2 1 7 5 7 9
b = -1 -1 2 2 5 7
b[0] = -1 since there is no number that is smaller than 2
b[1] = -1 since there is no number that is smaller than 1 from {2}
b[2] = 2, closest number to 7 that is smaller than 7 from {2,1} is 2
b[3] = 2, closest number to 5 that is smaller than 5 from {2,1,7} is 2
b[4] = 5, closest number to 7 that is smaller than 7 from {2,1,7,5} is 5
I was thinking about implementing balanced binary tree, however it will require a lot of work. Is there an easier way of doing this?
Here is one approach:
for i ← 1 to i ← (length(A)-1) {
// A[i] is added in the sorted sequence A[0, .. i-1] save A[i] to make a hole at index j
item = A[i]
j = i
// keep moving the hole to next smaller index until A[j - 1] is <= item
while j > 0 and A[j - 1] > item {
A[j] = A[j - 1] // move hole to next smaller index
j = j - 1
}
A[j] = item // put item in the hole
// if there are elements to the left of A[j] in sorted sequence A[0, .. i-1], then store it in b
// TODO : run loop so that duplicate entries wont hamper results
if j > 1
b[i] = A[j-1]
else
b[1] = -1;
}
Dry run:
a = 2 1 7 5 7 9
a[1] = 2
its straight forward, set b[1] to -1
a[2] = 1
insert into subarray : [1 ,2]
any elements before 1 in sorted array ? no.
So set b[2] to -1 . b: [-1, -1]
a[3] = 7
insert into subarray : [1 ,2, 7]
any elements before 7 in sorted array ? yes. its 2
So set b[3] to 2. b: [-1, -1, 2]
a[4] = 5
insert into subarray : [1 ,2, 5, 7]
any elements before 5 in sorted array ? yes. its 2
So set b[4] to 2. b: [-1, -1, 2, 2]
and so on..
Here's a sketch of a (nearly) O(n log n) algorithm that's somewhere in between the difficulty of implementing an insertion sort and balanced binary tree: Do the problem backwards, use merge/quick sort, and use binary search.
Pseudocode:
let c be a copy of a
let b be an array sized the same as a
sort c using an O(n log n) algorithm
for i from a.length-1 to 1
binary search over c for key a[i] // O(log n) time
remove the item found // Could take O(n) time
if there exists an item to the left of that position, b[i] = that item
otherwise, b[i] = -1
b[0] = -1
return b
There's a few implementation details that can make this have poor runtime.
For instance, since you have to remove items, doing this on a regular array and shifting things around will make this algorithm still take O(n^2) time. So, you could store key-value pairs instead. One would be the key, and the other would be the number of those keys (kind of like a multiset implemented on an array). "Removing" one would just be subtracting the second item from the pair and so on.
Eventually you will be left with a bunch of 0-value keys. This would eventually make the if there exists an item to the left take roughly O(n) time, and therefore, the entire algorithm would degrade to a O(n^2) for that reason. So another optimization might be to batch remove all of them periodically. For instance, when 1/2 of them are 0-values, perform a pruning.
The ideal option might be to implement another data structure that has a much more favorable remove time. Something along the lines of a modified unrolled linked list with indices could work, but it would certainly increase the implementation complexity of this approach.
I've actually implemented this. I used the first two optimizations above (storing key-value pairs for compression, and pruning when 1/2 of them are 0s). Here's some benchmarks to compare using an insertion sort derivative to this one:
a.length This method Insert sort Method
100 0.0262ms 0.0204ms
1000 0.2300ms 0.8793ms
10000 2.7303ms 75.7155ms
100000 32.6601ms 7740.36 ms
300000 98.9956ms 69523.6 ms
1000000 333.501 ms ????? Not patient enough
So, as you can see, this algorithm grows much, much slower than the insertion sort method I posted before. However, it took 73 lines of code vs 26 lines of code for the insertion sort method. So in terms of simplicity, the insertion sort method might still be the way to go if you don't have time requirements/the input is small.
You could treat it like an insertion sort.
Pseudocode:
let arr be one array with enough space for every item in a
let b be another array with, again, enough space for all elements in a
For each item in a:
perform insertion sort on item into arr
After performing the insertion, if there exists a number to the left, append that to b.
Otherwise, append -1 to b
return b
The main thing you have to worry about is making sure that you don't make the mistake of reallocating arrays (because it would reallocate n times, which would be extremely costly). This will be an implementation detail of whatever language you use (std::vector's reserve for C++ ... arr.reserve(n) for D ... ArrayList's ensureCapacity in Java...)
A potential downfall with this approach compared to using a binary tree is that it's O(n^2) time. However, the constant factors using this method vs binary tree would make this faster for smaller sizes. If your n is smaller than 1000, this would be an appropriate solution. However, O(n log n) grows much slower than O(n^2), so if you expect a's size to be significantly higher and if there's a time limit that you are likely to breach, you might consider a more complicated O(n log n) algorithm.
There are ways to slightly improve the performance (such as using a binary insertion sort: using binary search to find the position to insert into), but generally they won't improve performance enough to matter in most cases since it's still O(n^2) time to shift elements to fit.
Consider this:
a = 2 1 7 5 7 9
b = -1 -1 2 2 5 7
c 0 1 2 3 4 5 6 7 8 9
0 - - - - - - - - - -
Where the index of C is value of a[i] such that 0,3,4,6,8 would have null values.
and the 1st dimension of C contains the highest to date closest value to a[i]
So in step by a[3] we have the following
c 0 1 2 3 4 5 6 7 8 9
0 - -1 -1 - - 2 - 2 - -
and by step a[5] we have the following
c 0 1 2 3 4 5 6 7 8 9
0 - -1 -1 - - 2 - 5 - 7
This way when we get to the 2nd 7 at a[4] we know that 2 is the largest value to date and all we need to do is loop back through a[i-1] until we encounter a 7 again comparing the a[i] value to that in c[7] if bigger, replace c[7]. Once a[i-1] = the 7 we put c[7] into b[i] and move on to next a[i].
The main downfalls to this approach that I can see are:
footprint size depending on how big the c[] needs to be dimensioned..
the fact that you have to revisit elements of a[] that you've already touched. If the distribution of data is such that there are significant spaces between the two 7's then keeping track of the highest value as you go would presumably be faster. Alternatively it might be better to gather statistics on the a[i] up front to know what distributions exist and then use a hybrid method maintaining the max until such time that no more instances of that number are in the statistics.
Lets say we have the following set of numbers representing values over time
1 2 3 10 1 20 40 60
Now I am looking for an algorithm to find the highest percentage increase from one time to another. In the above case, the answer would be the pair (1, 60), which has a 6000% increase.
So far, the best algorithm I can think of is a brute-force method. We consider all possible pairs using a series of iterations:
1st Iteration:
1-2 1-3 1-10 .. 1-60
2nd Iteration
2-3 2-10 2-1 ... 2-60
(etc.)
This has complexity O(n3).
I've also been thinking about another approach. Find all the strictly increasing sequences, and determine only the perecentage increase in those strictly increasing sequences.
Does any other idea strike you guys? Please do correct me if my ideas are wrong!
I may have misunderstood the problem, but it seems that all you want is the largest and smallest numbers, since those are the two numbers that matter.
while true:
indexOfMax = max(list)
indexOfMin = min(list)
list.remove(indexOfMax)
list.remove(indexOfMin)
if(indexOfmax < indexOfMin)
contine
else if(indexOfMax == indexOfMin)
return -1
else
SUCCESS
As I understand (you didn't correct me in your comment), you want to maximize a[i]/a[j] for all j <= i. If that's correct, then for each i we only need to know smallest value before it.
int current_min = INFINITY;
double max_increase = 0;
for (int i = 0; i < n; ++i) {
current_min = min(current_min, a[i]);
max_increase = max(max_increase, a[i] / min);
}
So you just want to compare each number pair-wise and see which pair has the highest ratio from the second number to the first number? Just iterating with two loops (one with i=0 to n, and an inner loop with j=i+1 to n) is going to give you O(n^2). I guess this is actually your original solution, but you incorrectly said the complexity was O(n^3). It's n^2.
You could get to O(n log n), though. Take your list, make it into a list where each element is a pair of (index, value). Then sort it by the second element of the pair. Then have two pointers into the list, one coming from the left (0 to n-1), and the other coming from the right (n-1 to 0). Find the first pair of elements such that the left element's original index is less than the right element's original index. Done.
1 2 3 10 1 20 40 60
becomes
(1,0) (2,1) (3,2) (10,3) (1, 4) (20, 5) (40, 6) (60,7)
becomes
(1,0) (1,4) (2,1) (3,2) (10,3) (20,5) (40,6) (60,7)
So your answer is 60/1, from index 0 to index 7.
If this isn't what you're looking for, it would help if you said what the right answer was for your example numbers.
If I understand your problem correctly, you are looking for two indices (i, j) in the array with i < j that has the highest ratio A[j]/A[i]. If so, then you can reduce it to this related problem, which asks you to find the indices (i, j) with i ≤ j such that A[j] - A[i] is as large as possible. That problem has a very fast O(n)-time, O(1)-space algorithm that can be adapted to this problem as well. The intuition is to solve the problem for the array consisting of just the first element of your array, then for the first two elements, then the first three, etc. Once you've solved the problem for the first n elements of the array, you have an overall solution to the problem.
Let's think about how to do this. Initially, when you consider just the first element of the array, the best percentage increase you can get is 0% by comparing the element with itself. Now, suppose (inductively) that you've solved the problem for the first k array elements and want to see what happens when you look at the next array element. When this happens, one of two conditions holds. First, the maximum percentage increase over the first k elements might also be the maximum percentage increase over the first (k + 1) elements as well. For example, if the (k+1)st array element is an extremely small number, then chances are you can't get a large percentage increase from something in the first k elements to that value. Second, the maximum percentage increase might be from one of the first k elements to the (k + 1)st element. If this is the case, the highest percentage increase would be from the smallest of the first k elements to the (k + 1)st element.
Combining these two cases, we get that the best percentage increase over the first k + 1 elements is the maximum of
The highest percentage increase of the first k elements, or
The percentage increase from the smallest of the first k elements to the (k + 1)st element.
You can implement this by iterating across the array elements keeping track of two values - the minimum value you've seen so far and the pair that maximizes the percent increase. As an example, for your original example array of
1 2 3 10 1 20 40 60
The algorithm would work like this:
1 2 3 10 1 20 40 60
min 1 1 1 1 1 1 1 1
best (1,1) (1, 2) (1, 3) (1, 10) (1, 10) (1, 20) (1, 40) (1, 60)
and you'd output (1, 60) as the highest percentage increase. On a different array, like this one:
3 1 4 2 5
then the algorithm would trace out like this:
3 1 4 2 5
min 3 1 1 1 1
best (3,3) (3,3) (1,4) (1,4) (1,5)
and you'd output (1, 5).
This whole algorithm uses only O(1) space and runs in O(n) time, which is an extremely good solution to the problem.
Alternatively, you can think about reducing this problem directly to the maximum single-sell profit problem by taking the logarithm of all of the values in your array. In that case, if you find a pair of values where log A[j] - log A[i] is maximized, this is equivalent (using properties of logarithms) to finding a pair of values where log (A[j] / A[i]) is maximized. Since the log function is monotonically increasing, this means that you have found a pair of values where A[j] / A[i] is maximized, as intended.