How fast can 'finding the max in an array' possibly get? - algorithm

This question comes from a discussion that was touched off on this other question:
Parallelize already linear-time algorithm. It is not homework.
You are given an array of N numbers, and a machine with P processors and a shared CREW memory (Concurrent Read, Exclusive Write memory).
What is the tightest upper bound on the fastest algorithm to find the largest number in the array? [Obviously, also: What is the algorithm itself?]
I am not referring to the total amount of work performed [which can never be less than O(N)].

I think it's O(N/P') + O(Log2(P')), where P'=min{N,P}. P' processors search for max of N/P' elements each, followed by Log2 pairwise merges done in parallel. The first P'/2 merges are done by even-numbered processors, next 'P'/4' - by processors at locations divisible by 8, then by 16, and so on.
Edit P' is introduced to cover the case when you have significantly more processor nodes than the elements that you need to search.

Cook, Dwork, and Reischuk showed that any CREW algorithm for finding the maximum of n elements must run in Omega(lg n) time, even with an unlimited number of processors and unlimited memory. If I remember correctly, an algorithm with a matching upper bound appears in their paper:
Stephen Cook, Cynthia Dwork, and Rüdiger Reischuk. Upper and lower time bounds for parallel random access machines without simultaneous writes. SIAM Journal on Computing, 15(1):87-97, 1986.

The following is optimal bound:
If p <= n/log n you can do it in O(n/p) time; otherwise it's O(log n) i.e. when p>n/log n you gain nothing compared to p=n/log n.
Proof - lower bound:
Claim 1: You can never do faster than Ω(n/p), because p processors can give only speedup of p
Claim 2: You can never do faster than Ω(log n), because of CREW model (see unforgiven's paper); if you want to check if a 0-1 array has at least one 1, you need O(log n) time.
Proof - upper bound:
Claim 3: You can find maximum using n/log n processors and in O(log n) time
Proof: It is easy to find maximum using n processors and log n time; but in fact, in this algorithm most processors are dormant most of the time; by suitable dovetailing, (see e.g. Papadimitriou's complexity book) their number can be lowered to n/log n.
Now, given less than n/log n processors you can give work assigned to K processors to 1 processor, this divides processor requirement by K and multiplies required time by K.
Let K=(n/log n)/p; the previous algorithm runs in time O(K log n) = O(n/p), and requires n / (log n * K) = p processors.
Edited: I just realized that when p <= n/log n, dasblinkenlight's algorithm has the same asymptotic runtime:
n/p + log p <= n/p + log(n/log n) <= n/p + log n <= n/p + n/p <= 2n/p = O(n/p)
so you can use that algorithm, which has complexity O(n/p) when p <= n/log n and O(log n) otherwise.

I suspect this to be O(N/P)+O(P)
Sharing the work between P processors has a cost of O(P)
combining the work done by P processors has also costs of O(P)
A perfect parallell search of N items by P processors has time cost of O(N/P)
My naive algorithm would be to
write item 0 into a CREW cell labelled "result"
start P completly independent searches, each through 1/P th of the N items
upon completion of each search use CAS spinloop to replace "result" with result of partial search, if it is larger. (Depending on your definition of CREW you might not need the spinloop)

For P=N^2 it is O(1).
All initialize Boolean array CannotBeMax[i] = FALSE
Proc (i,j) sets CannotBeMax[i] = A[i] < A[j]
Max is A[CannotBeMax[i] == FALSE]
Note that all concurrent writes attempt to write identical info so the answer is consistent no matter which one succeeds.

Related

Parity of permutation with parallelism

I have an integer array of length N containing the values 0, 1, 2, .... (N-1), representing a permutation of integer indexes.
What's the most efficient way to determine if the permutation has odd or even parity, given I have parallel compute of O(N) as well?
For example, you can sum N numbers in log(N) with parallel computation. I expect to find the parity of permutations in log(N) as well, but cannot seem to find an algorithm. I also do not know how this "complexity order with parallel computation" is called.
The number in each array slot is the proper slot for that item. Think of it as a direct link from the "from" slot to the "to" slot. An array like this is very easy to sort in O(N) time with a single CPU just by following the links, so it would be a shame to have to use a generic sorting algorithm to solve this problem. Thankfully...
You can do this easily in O(log N) time with Ω(N) CPUs.
Let A be your array. Since each array slot has a single link out (the number in that slot) and a single link in (that slot's number is in some slot), the links break down into some number of cycles.
The parity of the permutation is the oddness of N-m, where N is the length of the array and m is the number of cycles, so we can get your answer by counting the cycles.
First, make an array S of length N, and set S[i] = i.
Then:
Repeat ceil(log_2(N)) times:
foreach i in [0,N), in parallel:
if S[i] < S[A[i]] then:
S[A[i]] = S[i]
A[i] = A[A[i]]
When this is finished, every S[i] will contain the smallest index in the cycle containing i. The first pass of the inner loop propagates the smallest S[i] to the next slot in the cycle by following the link in A[i]. Then each link is made twice as long, so the next pass will propagate it to 2 new slots, etc. It takes at most ceil(log_2(N)) passes to propagate the smallest S[i] around the cycle.
Let's call the smallest slot in each cycle the cycle's "leader". The number of leaders is the number of cycles. We can find the leaders just like this:
foreach i in [0,N), in parallel:
if (S[i] == i) then:
S[i] = 1 //leader
else
S[i] = 0 //not leader
Finally, we can just add up the elements of S to get the number of cycles in the permutation, from which we can easily calculate its parity.
You didn't specify a machine model, so I'll assume that we're working with an EREW PRAM. The complexity measure you care about is called "span", the number of rounds the computation takes. There is also "work" (number of operations, summed over all processors) and "cost" (span times number of processors).
From the point of view of theory, the obvious answer is to modify an O(log n)-depth sorting network (AKS or Goodrich's Zigzag Sort) to count swaps, then return (number of swaps) mod 2. The code is very complex, and the constant factors are quite large.
A more practical algorithm is to use Batcher's bitonic sorting network instead, which raises the span to O(log2 n) but has reasonable constant factors (such that people actually use it in practice to sort on GPUs).
I can't think of a practical deterministic algorithm with span O(log n), but here's a randomized algorithm with span O(log n) with high probability. Assume n processors and let the (modifiable) input be Perm. Let Coin be an array of n Booleans.
In each of O(log n) passes, the processors do the following in parallel, where i ∈ {0…n-1} identifies the processor, and swaps ← 0 initially. Lower case variables denote processor-local variables.
Coin[i] ← true with probability 1/2, false with probability 1/2
(barrier synchronization required in asynchronous models)
if Coin[i]
j ← Perm[i]
if not Coin[j]
Perm[i] ← Perm[j]
Perm[j] ← j
swaps ← swaps + 1
end if
end if
(barrier synchronization required in asynchronous models)
Afterwards, we sum up the local values of swaps and mod by 2.
Each pass reduces the number of i such that Perm[i] ≠ i by 1/4 of the current total in expectation. Thanks to the linearity of expectation, the expected total is at most n(3/4)r, so after r = 2 log4/3 n = O(log n) passes, the expected total is at most 1/n, which in turn bounds the probability that the algorithm has not converged to the identity permutation as required. On failure, we can just switch to the O(n)-span serial algorithm without blowing up the expected span, or just try again.

time complexity to find k elements in unsorted array using quick partition [duplicate]

According to Wikipedia, partition-based selection algorithms such as quickselect have runtime of O(n), but I am not convinced by it. Can anyone explain why it is O(n)?
In the normal quick-sort, the runtime is O(n log n). Every time we partition the branch into two branches (greater than the pivot and lesser than the pivot), we need to continue the process in both branches, whereas quickselect only needs to process one branch. I totally understand these points.
However, if you think in the Binary Search algorithm, after we chose the middle element, we are also searching only one side of the branch. So does that make the algorithm O(1)? No, of course, the Binary Search Algorithm is still O(log N) instead of O(1). This is also the same thing as the search element in a Binary Search Tree. We only search for one side, but we still consider O(log n) instead of O(1).
Can someone explain why in quickselect, if we continue the search in one side of pivot, it is considered O(1) instead of O(log n)? I consider the algorithm to be O(n log n), O(N) for the partitioning, and O(log n) for the number of times to continue finding.
There are several different selection algorithms, from the much simpler quickselect (expected O(n), worst-case O(n2)) to the more complex median-of-medians algorithm (Θ(n)). Both of these algorithms work by using a quicksort partitioning step (time O(n)) to rearrange the elements and position one element into its proper position. If that element is at the index in question, we're done and can just return that element. Otherwise, we determine which side to recurse on and recurse there.
Let's now make a very strong assumption - suppose that we're using quickselect (pick the pivot randomly) and on each iteration we manage to guess the exact middle of the array. In that case, our algorithm will work like this: we do a partition step, throw away half of the array, then recursively process one half of the array. This means that on each recursive call we end up doing work proportional to the length of the array at that level, but that length keeps decreasing by a factor of two on each iteration. If we work out the math (ignoring constant factors, etc.) we end up getting the following time:
Work at the first level: n
Work after one recursive call: n / 2
Work after two recursive calls: n / 4
Work after three recursive calls: n / 8
...
This means that the total work done is given by
n + n / 2 + n / 4 + n / 8 + n / 16 + ... = n (1 + 1/2 + 1/4 + 1/8 + ...)
Notice that this last term is n times the sum of 1, 1/2, 1/4, 1/8, etc. If you work out this infinite sum, despite the fact that there are infinitely many terms, the total sum is exactly 2. This means that the total work is
n + n / 2 + n / 4 + n / 8 + n / 16 + ... = n (1 + 1/2 + 1/4 + 1/8 + ...) = 2n
This may seem weird, but the idea is that if we do linear work on each level but keep cutting the array in half, we end up doing only roughly 2n work.
An important detail here is that there are indeed O(log n) different iterations here, but not all of them are doing an equal amount of work. Indeed, each iteration does half as much work as the previous iteration. If we ignore the fact that the work is decreasing, you can conclude that the work is O(n log n), which is correct but not a tight bound. This more precise analysis, which uses the fact that the work done keeps decreasing on each iteration, gives the O(n) runtime.
Of course, this is a very optimistic assumption - we almost never get a 50/50 split! - but using a more powerful version of this analysis, you can say that if you can guarantee any constant factor split, the total work done is only some constant multiple of n. If we pick a totally random element on each iteration (as we do in quickselect), then on expectation we only need to pick two elements before we end up picking some pivot element in the middle 50% of the array, which means that, on expectation, only two rounds of picking a pivot are required before we end up picking something that gives a 25/75 split. This is where the expected runtime of O(n) for quickselect comes from.
A formal analysis of the median-of-medians algorithm is much harder because the recurrence is difficult and not easy to analyze. Intuitively, the algorithm works by doing a small amount of work to guarantee a good pivot is chosen. However, because there are two different recursive calls made, an analysis like the above won't work correctly. You can either use an advanced result called the Akra-Bazzi theorem, or use the formal definition of big-O to explicitly prove that the runtime is O(n). For a more detailed analysis, check out "Introduction to Algorithms, Third Edition" by Cormen, Leisserson, Rivest, and Stein.
Let me try to explain the difference between selection & binary search.
Binary search algorithm in each step does O(1) operations. Totally there are log(N) steps and this makes it O(log(N))
Selection algorithm in each step performs O(n) operations. But this 'n' keeps on reducing by half each time. There are totally log(N) steps.
This makes it N + N/2 + N/4 + ... + 1 (log(N) times) = 2N = O(N)
For binary search it is 1 + 1 + ... (log(N) times) = O(logN)
In Quicksort, the recursion tree is lg(N) levels deep and each of these levels requires O(N) amount of work. So the total running time is O(NlgN).
In Quickselect, the recurision tree is lg(N) levels deep and each level requires only half the work of the level above it. This produces the following:
N * (1/1 + 1/2 + 1/4 + 1/8 + ...)
or
N * Summation(1/i^2)
1 < i <= lgN
The important thing to note here is that i goes from 1 to lgN, but not from 1 to N and also not from 1 to infinity.
The summation evaluates to 2. Hence Quickselect = O(2N).
Quicksort does not have a big-O of nlogn - it's worst case runtime is n^2.
I assume you're asking about Hoare's Selection Algorithm (or quickselect) not the naive selection algorithm that is O(kn). Like quicksort, quickselect has a worst case runtime of O(n^2) (if bad pivots are chosen), not O(n). It can run in expectation time n because it's only sorting one side, as you point out.
Because for selection, you're not sorting, necessarily. You can simply count how many items there are which have any given value. So an O(n) median can be performed by counting how many times each value comes up, and picking the value that has 50% of items above and below it. It's 1 pass through the array, simply incrementing a counter for each element in the array, so it's O(n).
For example, if you have an array "a" of 8 bit numbers, you can do the following:
int histogram [ 256 ];
for (i = 0; i < 256; i++)
{
histogram [ i ] = 0;
}
for (i = 0; i < numItems; i++)
{
histogram [ a [ i ] ]++;
}
i = 0;
sum = 0;
while (sum < (numItems / 2))
{
sum += histogram [ i ];
i++;
}
At the end, the variable "i" will contain the 8-bit value of the median. It was about 1.5 passes through the array "a". Once through the entire array to count the values, and half through it again to get the final value.

Exponential complexity of a program

I got a program which depends on two entries of sizes m and n respectively. If T(m,n) is the running time of the problem, it follows:
T(m,n)=T(m-1,n-1)+T(m-1,n)+T(m,n-1)+C
for a given constant C.
I could prove that the time complexity is in Omega(2^min(m,n)). However, it seems that it is of complexity Omega(2^max(m,n)) (it was just confirmed to me) but I can't find a formal proof. Anyone has a trick?
Thanks in advance!
From the top of my head:
I assume the recursion of T(m,n) stops when you reach T(0,x) or T(x,0).
You have 3 factors contributing to the complexity:
Factor 1: T(m-1,n-1) decreases both m and n, so its lenght until m or n become 0 is min(m,n) steps (See note below)
Factor 2: T(m-1,n) decreases m only, so its length until m=0 is m steps.
Factor 3: T(m,n-1) same as above but until n=0 is n steps.
The overall complexity is the greater of all complexities, so it must be related to maximum of
max( min(m,n) steps, m steps, n steps) = max(m,n)
rather than min(m,n).
I guess you can fill in the details. The constant C does not contribute, or more precisely contributes with O(1), which is the lowest of all complexities here.
Note about item 1: This factor has also a branch for m-1 until 0 and n-1 until 0, so strictly speaking it s complexity will also be max(m,n)
First, you should define that, for all values of x, T(0, x) = T(x, 0) = 0 to have your recursion stop - or at least T(0, x) = T(x, 0) = C.
Second, "time complexity is in Omega(2^min(m,n))" must obviously be wrong. Set m=10000 and n=1. Now try to prove to me that the complexity is the same as m=1 and n=1. The T(m-1, n-1) and T(m, n-1) parts disappear quickly, but you still have to walk the T(m-1, n) part.
Third, this observation directly leads to 2^max(m, n). Try to find out the number of recursion steps for a few low values of m and n. Then try to make up a formula for the number of steps depending on m and n. (Hint: Fibonacci). When you have that formula, you're finished.

Given a set of intervals, why is the average case for computing pred[i] theta(n logn)?

So I was reading my notes and I don't really get this part:
Define f, s as the starting and finishing time of an interval.
Sort all intervals by finish time.
So suppose we have a set of intervals I_1, ..., I_n and define pred[i] = the largest index j such that f_j <= s_i. (So essentially the next closest interval before i that does not overlap i).
Now we want to solve all pred[i] for all I_i in n. Why is the runtime of this theta(n log n)? I would think that, in the worst case, all of I_1, ..., I_(i-1) all overlap with I_i; and thus would take n - 1 comparisons, then n - 2 comparisons and so on. And the best case would be that none of the intervals overlap and pred(i) takes 1 comparison for each interval. I'm not sure why the runtime for "sorting and computing pred[i] values" is theta(n logn).
The condition you named, max j such that f[j] < s[i], is not an overlap check.
It is, rather, the last interval that starts before interval i finishes
(they may or may not actually overlap). I suppose you can get intervals
sorted by start times which is theta(nlogn) and then use binary search to
locate f[j] for every s[i] (also theta(nlog n) )
P.S by theta I mean average case, not the worst case scenario you mentioned.

Why is the runtime of the selection algorithm O(n)?

According to Wikipedia, partition-based selection algorithms such as quickselect have runtime of O(n), but I am not convinced by it. Can anyone explain why it is O(n)?
In the normal quick-sort, the runtime is O(n log n). Every time we partition the branch into two branches (greater than the pivot and lesser than the pivot), we need to continue the process in both branches, whereas quickselect only needs to process one branch. I totally understand these points.
However, if you think in the Binary Search algorithm, after we chose the middle element, we are also searching only one side of the branch. So does that make the algorithm O(1)? No, of course, the Binary Search Algorithm is still O(log N) instead of O(1). This is also the same thing as the search element in a Binary Search Tree. We only search for one side, but we still consider O(log n) instead of O(1).
Can someone explain why in quickselect, if we continue the search in one side of pivot, it is considered O(1) instead of O(log n)? I consider the algorithm to be O(n log n), O(N) for the partitioning, and O(log n) for the number of times to continue finding.
There are several different selection algorithms, from the much simpler quickselect (expected O(n), worst-case O(n2)) to the more complex median-of-medians algorithm (Θ(n)). Both of these algorithms work by using a quicksort partitioning step (time O(n)) to rearrange the elements and position one element into its proper position. If that element is at the index in question, we're done and can just return that element. Otherwise, we determine which side to recurse on and recurse there.
Let's now make a very strong assumption - suppose that we're using quickselect (pick the pivot randomly) and on each iteration we manage to guess the exact middle of the array. In that case, our algorithm will work like this: we do a partition step, throw away half of the array, then recursively process one half of the array. This means that on each recursive call we end up doing work proportional to the length of the array at that level, but that length keeps decreasing by a factor of two on each iteration. If we work out the math (ignoring constant factors, etc.) we end up getting the following time:
Work at the first level: n
Work after one recursive call: n / 2
Work after two recursive calls: n / 4
Work after three recursive calls: n / 8
...
This means that the total work done is given by
n + n / 2 + n / 4 + n / 8 + n / 16 + ... = n (1 + 1/2 + 1/4 + 1/8 + ...)
Notice that this last term is n times the sum of 1, 1/2, 1/4, 1/8, etc. If you work out this infinite sum, despite the fact that there are infinitely many terms, the total sum is exactly 2. This means that the total work is
n + n / 2 + n / 4 + n / 8 + n / 16 + ... = n (1 + 1/2 + 1/4 + 1/8 + ...) = 2n
This may seem weird, but the idea is that if we do linear work on each level but keep cutting the array in half, we end up doing only roughly 2n work.
An important detail here is that there are indeed O(log n) different iterations here, but not all of them are doing an equal amount of work. Indeed, each iteration does half as much work as the previous iteration. If we ignore the fact that the work is decreasing, you can conclude that the work is O(n log n), which is correct but not a tight bound. This more precise analysis, which uses the fact that the work done keeps decreasing on each iteration, gives the O(n) runtime.
Of course, this is a very optimistic assumption - we almost never get a 50/50 split! - but using a more powerful version of this analysis, you can say that if you can guarantee any constant factor split, the total work done is only some constant multiple of n. If we pick a totally random element on each iteration (as we do in quickselect), then on expectation we only need to pick two elements before we end up picking some pivot element in the middle 50% of the array, which means that, on expectation, only two rounds of picking a pivot are required before we end up picking something that gives a 25/75 split. This is where the expected runtime of O(n) for quickselect comes from.
A formal analysis of the median-of-medians algorithm is much harder because the recurrence is difficult and not easy to analyze. Intuitively, the algorithm works by doing a small amount of work to guarantee a good pivot is chosen. However, because there are two different recursive calls made, an analysis like the above won't work correctly. You can either use an advanced result called the Akra-Bazzi theorem, or use the formal definition of big-O to explicitly prove that the runtime is O(n). For a more detailed analysis, check out "Introduction to Algorithms, Third Edition" by Cormen, Leisserson, Rivest, and Stein.
Let me try to explain the difference between selection & binary search.
Binary search algorithm in each step does O(1) operations. Totally there are log(N) steps and this makes it O(log(N))
Selection algorithm in each step performs O(n) operations. But this 'n' keeps on reducing by half each time. There are totally log(N) steps.
This makes it N + N/2 + N/4 + ... + 1 (log(N) times) = 2N = O(N)
For binary search it is 1 + 1 + ... (log(N) times) = O(logN)
In Quicksort, the recursion tree is lg(N) levels deep and each of these levels requires O(N) amount of work. So the total running time is O(NlgN).
In Quickselect, the recurision tree is lg(N) levels deep and each level requires only half the work of the level above it. This produces the following:
N * (1/1 + 1/2 + 1/4 + 1/8 + ...)
or
N * Summation(1/i^2)
1 < i <= lgN
The important thing to note here is that i goes from 1 to lgN, but not from 1 to N and also not from 1 to infinity.
The summation evaluates to 2. Hence Quickselect = O(2N).
Quicksort does not have a big-O of nlogn - it's worst case runtime is n^2.
I assume you're asking about Hoare's Selection Algorithm (or quickselect) not the naive selection algorithm that is O(kn). Like quicksort, quickselect has a worst case runtime of O(n^2) (if bad pivots are chosen), not O(n). It can run in expectation time n because it's only sorting one side, as you point out.
Because for selection, you're not sorting, necessarily. You can simply count how many items there are which have any given value. So an O(n) median can be performed by counting how many times each value comes up, and picking the value that has 50% of items above and below it. It's 1 pass through the array, simply incrementing a counter for each element in the array, so it's O(n).
For example, if you have an array "a" of 8 bit numbers, you can do the following:
int histogram [ 256 ];
for (i = 0; i < 256; i++)
{
histogram [ i ] = 0;
}
for (i = 0; i < numItems; i++)
{
histogram [ a [ i ] ]++;
}
i = 0;
sum = 0;
while (sum < (numItems / 2))
{
sum += histogram [ i ];
i++;
}
At the end, the variable "i" will contain the 8-bit value of the median. It was about 1.5 passes through the array "a". Once through the entire array to count the values, and half through it again to get the final value.

Resources