My goal is to sample k integers from 0, ... n-1 without duplication. The order of sampled integers doesn't matter. At every each call (which occurs very often), n and k will slightly vary but not much (n is about 250,000 and k is about 2,000). I've come up with the following amortized O(k) algorithm:
Prepare an array A with items 0, 1, 2, ... , n-1. This takes O(n) but since n is relatively stable, the cost can be made amortized constant.
Sample a random number r from [0:i] where i = n - 1. Here the cost is in fact related to n, but as n is not VERY BIG, this dependency is not critical.
Swap the rth item and the ith item in the array A.
Decrease i by 1.
Repeat k times the steps 2~4; now we have a random permutation of length k at the tail of A. Copy this.
We should roll back A to its initial state (0, ... , n-1) to keep the cost of the step 1 constant. This can be done by push r to a stack of length k at each pass of step 2. Preparation of the stack requires amortized constant cost.
I think uniform sampling of permutation/combination should be an exhaustively studied problem, so either (1) there is a much better solution, or at least (2) my solution is a (minor modification of) a well-known solution. Thus,
In case (1), I want to know that better solution.
In case (2), I want to find a reference.
Please help me. Thanks.
If k is much less than n -- say, less than half of n -- then the most efficient solution is to keep the numbers generated in a hash table (actually, a hash set, since there is no value associated with a key). If the random number happens to already be in the hash table, reject it and generate another one in its place. With the actual values of k and n suggested (k ∼ 2000; n ∼ 250,000) the expected number of rejections to generate k unique samples is less than 10, so it will hardly be noticeable. The size of the hash table is O(k), and it can simply be deleted at the end of the sample generation.
It is also possible to simulate the FYK shuffle algorithm using a hash table instead of a vector of n values, thereby avoiding having to reject generated random numbers. If you were using a vector A, you would start by initializing A[i] to i, for every 0 ≤ i < k. With the hash table H, you start with an empty hash table, and use the convention that H[i] is considered to be i if the key i is not in the hash table. Step 3 in your algorithm -- "swap A[r] with A[i]" -- becomes "add H[r] as the next element of the sample and set H[r] to H[i]". Note that it is unnecessary to set H[i] because that element will never be referred to again: all subsequent random numbers r are generate from a range which does not include i.
Because the hash table in this case contains both keys and values, it is larger than the hash set used in alternative 1, above, and the increased size (and consequent increase in memory cache misses) is likely to cause more overhead than is saved by eliminating rejections. However, it has the advantage of working even if k is occasionally close to n.
Finally, in your proposed algorithm, it is actually quite easy to restore A in O(k) time. A value A[j] will have been modified by the algorithm only if:
a. n − k ≤ j < n, or
b. there is some i such that n − k ≤ i < n and A[i] = j.
Consequently, you can restore the vector A by looking at each A[i] for n − k ≤ i < n: first, if A[i] < n−k, set A[A[i]] to A[i]; then, unconditionally set A[i] to i.
Related
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.
Given the following pseudo-code, the question is how many times on average is the variable m being updated.
A[1...n]: array with n random elements
m = a[1]
for I = 2 to n do
if a[I] < m then m = a[I]
end for
One might answer that since all elements are random, then the variable will be updated on average on half the number of iterations of the for loop plus one for the initialization.
However, I suspect that there must be a better (and possibly the only correct) way to prove it using binomial distribution with p = 1/2. This way, the average number of updates on m would be
M = 1 + Σi=1 to n-1[k.Cn,k.pk.(1-p)(n-k)]
where Cn,k is the binomial coefficient. I have tried to solve this but I have stuck some steps after since I do not know how to continue.
Could someone explain me which of the two answers is correct and if it is the second one, show me how to calculate M?
Thank you for your time
Assuming the elements of the array are distinct, the expected number of updates of m is the nth harmonic number, Hn, which is the sum of 1/k for k ranging from 1 to n.
The summation formula can also be represented by the recursion:
H1 = 1
Hn = Hn−1+1/n (n > 1)
It's easy to see that the recursion corresponds to the problem.
Consider all permutations of n−1 numbers, and assume that the expected number of assignments is Hn−1. Now, every permutation of n numbers consists of a permutation of n−1 numbers, with a new smallest number inserted in one of n possible insertion points: either at the beginning, or after one of the n−1 existing values. Since it is smaller than every number in the existing series, it will only be assigned to m in the case that it was inserted at the beginning. That has a probability of 1/n, and so the expected number of assignments of a permutation of n numbers is Hn−1 + 1/n.
Since the expected number of assignments for a vector of length one is obviously 1, which is H1, we have an inductive proof of the recursion.
Hn is asymptotically equal to ln n + γ where γ is the Euler-Mascheroni constant, approximately 0.577. So it increases without limit, but quite slowly.
The values for which m is updated are called left-to-right maxima, and you'll probably find more information about them by searching for that term.
I liked #rici answer so I decided to elaborate its central argument a little bit more so to make it clearer to me.
Let H[k] be the expected number of assignments needed to compute the min m of an array of length k, as indicated in the algorithm under consideration. We know that
H[1] = 1.
Now assume we have an array of length n > 1. The min can be in the last position of the array or not. It is in the last position with probability 1/n. It is not with probability 1 - 1/n. In the first case the expected number of assignments is H[n-1] + 1. In the second, H[n-1].
If we multiply the expected number of assignments of each case by their probabilities and sum, we get
H[n] = (H[n-1] + 1)*1/n + H[n-1]*(1 - 1/n)
= H[n-1]*1/n + 1/n + H[n-1] - H[n-1]*1/n
= 1/n + H[n-1]
which shows the recursion.
Note that the argument is valid if the min is either in the last position or in any the first n-1, not in both places. Thus we are using that all the elements of the array are different.
I have a number n and a set of numbers S ∈ [1..n]* with size s (which is substantially smaller than n). I want to sample a number k ∈ [1..n] with equal probability, but the number is not allowed to be in the set S.
I am trying to solve the problem in at worst O(log n + s). I am not sure whether it's possible.
A naive approach is creating an array of numbers from 1 to n excluding all numbers in S and then pick one array element. This will run in O(n) and is not an option.
Another approach may be just generating random numbers ∈[1..n] and rejecting them if they are contained in S. This has no theoretical bound as any number could be sampled multiple times even if it is in the set. But on average this might be a practical solution if s is substantially smaller than n.
Say s is sorted. Generate a random number between 1 and n-s, call it k. We've chosen the k'th element of {1,...,n} - s. Now we need to find it.
Use binary search on s to find the count of the elements of s <= k. This takes O(log |s|). Add this to k. In doing so, we may have passed or arrived at additional elements of s. We can adjust for this by incrementing our answer for each such element that we pass, which we find by checking the next larger element of s from the point we found in our binary search.
E.g., n = 100, s = {1,4,5,22}, and our random number is 3. So our approach should return the third element of [2,3,6,7,...,21,23,24,...,100] which is 6. Binary search finds that 1 element is at most 3, so we increment to 4. Now we compare to the next larger element of s which is 4 so increment to 5. Repeating this finds 5 in so we increment to 6. We check s once more, see that 6 isn't in it, so we stop.
E.g., n = 100, s = {1,4,5,22}, and our random number is 4. So our approach should return the fourth element of [2,3,6,7,...,21,23,24,...,100] which is 7. Binary search finds that 2 elements are at most 4, so we increment to 6. Now we compare to the next larger element of s which is 5 so increment to 7. We check s once more, see that the next number is > 7, so we stop.
If we assume that "s is substantially smaller than n" means |s| <= log(n), then we will increment at most log(n) times, and in any case at most s times.
If s is not sorted then we can do the following. Create an array of bits of size s. Generate k. Parse s and do two things: 1) count the number of elements < k, call this r. At the same time, set the i'th bit to 1 if k+i is in s (0 indexed so if k is in s then the first bit is set).
Now, increment k a number of times equal to r plus the number of set bits is the array with an index <= the number of times incremented.
E.g., n = 100, s = {1,4,5,22}, and our random number is 4. So our approach should return the fourth element of [2,3,6,7,...,21,23,24,...,100] which is 7. We parse s and 1) note that 1 element is below 4 (r=1), and 2) set our array to [1, 1, 0, 0]. We increment once for r=1 and an additional two times for the two set bits, ending up at 7.
This is O(s) time, O(s) space.
This is an O(1) solution with O(s) initial setup that works by mapping each non-allowed number > s to an allowed number <= s.
Let S be the set of non-allowed values, S(i), where i = [1 .. s] and s = |S|.
Here's a two part algorithm. The first part constructs a hash table based only on S in O(s) time, the second part finds the random value k ∈ {1..n}, k ∉ S in O(1) time, assuming we can generate a uniform random number in a contiguous range in constant time. The hash table can be reused for new random values and also for new n (assuming S ⊂ { 1 .. n } still holds of course).
To construct the hash, H. First set j = 1. Then iterate over S(i), the elements of S. They do not need to be sorted. If S(i) > s, add the key-value pair (S(i), j) to the hash table, unless j ∈ S, in which case increment j until it is not. Finally, increment j.
To find a random value k, first generate a uniform random value in the range s + 1 to n, inclusive. If k is a key in H, then k = H(k). I.e., we do at most one hash lookup to insure k is not in S.
Python code to generate the hash:
def substitute(S):
H = dict()
j = 1
for s in S:
if s > len(S):
while j in S: j += 1
H[s] = j
j += 1
return H
For the actual implementation to be O(s), one might need to convert S into something like a frozenset to insure the test for membership is O(1) and also move the len(S) loop invariant out of the loop. Assuming the j in S test and the insertion into the hash (H[s] = j) are constant time, this should have complexity O(s).
The generation of a random value is simply:
def myrand(n, s, H):
k = random.randint(s + 1, n)
return (H[k] if k in H else k)
If one is only interested in a single random value per S, then the algorithm can be optimized to improve the common case, while the worst case remains the same. This still requires S be in a hash table that allows for a constant time "element of" test.
def rand_not_in(n, S):
k = random.randint(len(S) + 1, n);
if k not in S: return k
j = 1
for s in S:
if s > len(S):
while j in S: j += 1
if s == k: return j
j += 1
Optimizations are: Only generate the mapping if the random value is in S. Don't save the mapping to a hash table. Short-circuit the mapping generation when the random value is found.
Actually, the rejection method seems like the practical approach.
Generate a number in 1...n and check whether it is forbidden; regenerate until the generated number is not forbidden.
The probability of a single rejection is p = s/n.
Thus the expected number of random number generations is 1 + p + p^2 + p^3 + ... which is 1/(1-p), which in turn is equal to n/(n-s).
Now, if s is much less than n, or even more up to s = n/2, this expected number is at most 2.
It would take s almost equal to n to make it infeasible in practice.
Multiply the expected time by log s if you use a tree-set to check whether the number is in the set, or by just 1 (expected value again) if it is a hash-set. So the average time is O(1) or O(log s) depending on the set implementation. There is also O(s) memory for storing the set, but unless the set is given in some special way, implicitly and concisely, I don't see how it can be avoided.
(Edit: As per comments, you do this only once for a given set.
If, additionally, we are out of luck, and the set is given as a plain array or list, not some fancier data structure, we get O(s) expected time with this approach, which still fits into the O(log n + s) requirement.)
If attacks against the unbounded algorithm are a concern (and only if they truly are), the method can include a fall-back algorithm for the cases when a certain fixed number of iterations didn't provide the answer.
Similarly to how IntroSort is QuickSort but falls back to HeapSort if the recursion depth gets too high (which is almost certainly a result of an attack resulting in quadratic QuickSort behavior).
Find all numbers that are in a forbidden set and less or equal then n-s. Call it array A.
Find all numbers that are not in a forbidden set and greater then n-s. Call it array B. It may be done in O(s) if set is sorted.
Note that lengths of A and B are equal, and create mapping map[A[i]] = B[i]
Generate number t up to n-s. If there is map[t] return it, otherwise return t
It will work in O(s) insertions to a map + 1 lookup which is either O(s) in average or O(s log s)
Imagine a standard permute function that takes an integer and returns a vector of the first N natural numbers in a random permutation. If you only need k (<= N) of them, but don't know k beforehand, do you still have to perform a O(N) generation of the permutation? Is there a better algorithm than:
for x in permute(N):
if f(x):
break
I'm imagining an API such as:
p = permuter(N)
for x = p.next():
if f(x):
break
where the initialization is O(1) (including memory allocation).
This question is often viewed as a choice between two competing algorithms:
Strategy FY: A variation on the Fisher-Yates shuffle where one shuffle step is performed for each desired number, and
Strategy HT: Keep all generated numbers in a hash table. At each step, random numbers are produced until a number which is not in the hash table is found.
The choice is performed depending on the relationship between k and N: if k is sufficiently large, the strategy FY is used; otherwise, strategy HT. The argument is that if k is small relative to n, maintaining an array of size n is a waste of space, as well as producing a large initialization cost. On the other hand, as k approaches n more and more random numbers need to be discarded, and towards the end producing new values will be extremely slow.
Of course, you might not know in advance the number of samples which will be requested. In that case, you might pessimistically opt for FY, or optimistically opt for HT, and hope for the best.
In fact, there is no real need for trade-off, because the FY algorithm can be implemented efficiently with a hash table. There is no need to initialize an array of N integers. Instead, the hash-table is used to store only the elements of the array whose values do not correspond with their indices.
(The following description uses 1-based indexing; that seemed to be what the question was looking for. Hopefully it is not full of off-by-one errors. So it generates numbers in the range [1, N]. From here on, I use k for the number of samples which have been requested to date, rather than the number which will eventually be requested.)
At each point in the incremental FY algorithm a single index r is chosen at random from the range [k, N]. Then the values at indices k and r are swapped, after which k is incremented for the next iteration.
As an efficiency point, note that we don't really need to do the swap: we simply yield the value at r and then set the value at r to be the value at k. We'll never again look at the value at index k so there is no point updating it.
Initially, we simulate the array with a hash table. To look up the value at index i in the (virtual) array, we see if i is present in the hash table: if so, that's the value at index i. Otherwise the value at index i is i itself. We start with an empty hash table (which saves initialization costs), which represents an array whose value at every index is the index itself.
To do the FY iteration, for each sample index k we generate a random index r as above, yield the value at that index, and then set the value at index r to the value at index k. That's exactly the procedure described above for FY, except for the way we look up values.
This requires exactly two hash-table lookups, one insertion (at an already looked-up index, which in theory can be done more quickly), and one random number generation for each iteration. That's one more lookup than strategy HT's best case, but we have a bit of a saving because we never need to loop to produce a value. (There is another small potential saving when we rehash because we can drop any keys smaller than the current value of k.)
As the algorithm proceeds, the hash table will grow; a standard exponential rehashing strategy is used. At some point, the hash table will reach the size of a vector of N-k integers. (Because of hash table overhead, this point will be reached at a value of k much less than N, but even if there were no overhead this threshold would be reached at N/2.) At that point, instead of rehashing, the hash is used to create the tail of the now non-virtual array, a procedure which takes less time than a rehash and never needs to be repeated; remaining samples will be selected using the standard incremental FY algorithm.
This solution is slightly slower than FY if k eventually reaches the threshold point, and it is slightly slower than HT if k never gets big enough for random numbers to be rejected. But it is not much slower in either case, and if never suffers from pathological slowdown when k has an awkward value.
In case that was not clear, here is a rough Python implementation:
from random import randint
def sampler(N):
k = 1
# First phase: Use the hash
diffs = {}
# Only do this until the hash table is smallish (See note)
while k < N // 4:
r = randint(k, N)
yield diffs[r] if r in diffs else r
diffs[r] = diffs[k] if k in diffs else k
k += 1
# Second phase: Create the vector, ignoring keys less than k
vbase = k
v = list(range(vbase, N+1))
for i, s in diffs.items():
if i >= vbase:
v[i - vbase] = s
del diffs
# Now we can generate samples until we hit N
while k <= N:
r = randint(k, N)
rv = v[r - vbase]
v[r - vbase] = v[k - vbase]
yield rv
k += 1
Note: N // 4 is probably pessimistic; computing the correct value would require knowing too much about hash-table implementation. If I really cared about speed, I'd write my own hash table implementation in a compiled language, and then I'd know :)
Given a vector V of n integers and an integer k, k <= n, you want a subvector (a sequence of consecutive elements of the vector ) of maximum length containing at most k distinct elements.
The technique that I use for the resolution of the problem is dynamic programming.
The complexity of this algorithm must be O(n*k).
The main problem is how to count distinct elements of the vector. as you would resolve it ?
How to write the EQUATION OF RECURRENCE ?
Thanks you!!!.
I don't know why you would insist on O(n*k), this can be solved in O(n) with 'sliding window' approach.
Maintain current 'window' [left..right]
At each step, if we can increase right by 1 (without violating 'at most k disctint elements' requirement), do it
Otherwise, increase left by 1
Check whether current window is the longest and go back to #2
Checking whether we can increase right in #2 is a little tricky. We can use hashtable storing for each element inside window how many times it occurred there.
So, the condition to allow right increase would look like
hash.size < k || hash.contains(V[right + 1])
And each time left or right is increased, we'll need to update hash (decrease or increase number of occurrences of the given element).
I'm pretty sure, any DP solution here would be longer and more complicated.
the main problem is how to count distinct elements of the vector. as you would resolve it?
If you allowed to use hashing, you could do the following
init Hashtable h
distinct_count := 0
for each element v of the vector V
if h does not contain v (O(1) time in average)
insert v into h (O(1) time in average)
distinct_count := distinct_count + 1
return distinct_count
This is in average O(n) time.
If not here is an O(n log n) solution - this time worst case
sort V (O(n log n) comparisons)
Then it should be easy to determine the number of different elements in O(n) time ;-)
I could also tell you an algorithm to sort V in O(n*b) where b is the bit count of the integers - if this helps you.
Here is the algorithm:
sort(vector, begin_index, end_index, currentBit)
reorder the vector[begin_index to end_index] so that the elements that have a 1 at bit currentBit are after those that have a 0 there (O(end_index-begin_index) time)
Let c be the count of elements that have a 0 at bit currentBit (O(end_index-begin_index) time; can be got from the step before)
if (currentBit is not 0)
call sort(vector, begin_index, begin_index+c)
call sort(vector, begin_index+c+1, end_index)
Call it with
vector = V
begin_index = 0
end_index = n-1
currentBit = bit count of the integers (=: b)-1.
This even uses dynamic programming as requested.
As you can determine very easily this is O(n*b) time with a recursion depth of b.