I have two lists, L and M, each containing thousands of 64-bit unsigned integers. I need to find out whether the sum of any two members of L is itself a member of M.
Is it possible to improve upon the performance of the following algorithm?
Sort(M)
for i = 0 to Length(L)
for j = i + 1 to Length(L)
BinarySearch(M, L[i] + L[j])
(I'm assuming your goal is to find all pairs in L that sum to something in M)
Forget hashtables!
Sort both lists.
Then do the outer loop of your algorithm: walk over every element i in L, then every larger element j in L. As you go, form the sum and check to see if it's in M.
But don't look using a binary search: simply do a linear scan from the last place you looked. Let's say you're working on some value i, and you have some value j, followed by some value j'. When searching for (i+j), you would have got to the point in M where that value is found, or the first largest value. You're now looking for (i+j'); since j' > j, you know that (i+j') > (i+j), and so it cannot be any earlier in M than the last place you got. If L and M are both smoothly distributed, there is an excellent chance that the point in M where you would find (i+j') is only a little way off.
If the arrays are not smoothly distributed, then better than a linear scan might be some sort of jumping scan - look forward N elements at a time, halving N if the jump goes too far.
I believe this algorithm is O(n^2), which is as fast as any proposed hash algorithm (which have an O(1) primitive operation, but still have to do O(n**2) of them. It also means that you don't have to worry about the O(n log n) to sort. It has much better data locality than the hash algorithms - it basically consists of paired streamed reads over the arrays, repeated n times.
EDIT: I have written implementations of Paul Baker's original algorithm, Nick Larsen's hashtable algorithm, and my algorithm, and a simple benchmarking framework. The implementations are simple (linear probing in the hashtable, no skipping in my linear search), and i had to make guesses at various sizing parameters. See http://urchin.earth.li/~twic/Code/SumTest/ for the code. I welcome corrections or suggestions, about any of the implementations, the framework, and the parameters.
For L and M containing 3438 items each, with values ranging from 1 to 34380, and with Larsen's hashtable having a load factor of 0.75, the median times for a run are:
Baker (binary search): 423 716 646 ns
Larsen (hashtable): 733 479 121 ns
Anderson (linear search): 62 077 597 ns
The difference is much bigger than i had expected (and, i admit, not in the direction i had expected). I suspect i have made one or more major mistakes in the implementation. If anyone spots one, i really would like to hear about it!
One thing is that i have allocated Larsen's hashtable inside the timed method. It is thus paying the cost of allocation and (some) garbage collection. I think this is fair, because it's a temporary structure only needed by the algorithm. If you think it's something that could be reused, it would be simple enough to move it into an instance field and allocate it only once (and Arrays.fill it with zero inside the timed method), and see how that affects performance.
The complexity of the example code in the question is O(m log m + l2 log m) where l=|L| and m=|M| as it runs binary search (O(log m)) for every pair of elements in L (O(l2)), and M is sorted first.
Replacing the binary search with a hash table reduces the complexity to O(l2) assuming that hash table insert and lookup are O(1) operations.
This is asymptotically optimal as long as you assume that you need to process every pair of numbers on the list L, as there are O(l2) such pairs. If there are a couple of thousands of numbers on L, and they are random 64-bit integers, then definitely you need to process all the pairs.
Instead of sorting M at a cost of n * log(n), you could create a hash set at the cost of n.
You could also store all sums in another hash set while iterating and add a check to make sure you don't perform the same search twice.
You can avoid binary search by using hashtable except sorted M array.
Alternatively, add all of the members of L to a hashset lSet, then iterate over M, performing these steps for each m in M:
add m to hashset mSet - if m is already in mSet, skip this iteration; if m is in hashset dSet, also skip this iteration.
subtract each member l of L less than m from m to give d, and test whether d is also in lSet;
if so, add (l, d) to some collection rSet; add d to hashset dSet.
This will require fewer iterations, at the cost of more memory. You will want to pre-allocate the memory for the structures, if this is to give you a speed increase.
Related
I have an application where I have a list of O(n) sets.
Each set Set(i) is an n-vector. Suppose n=4, for instance,
Set(1) could be [0|1|1|0]
Set(2) could be [1|1|1|0]
Set(3) could be [1|1|0|0]
Set(4) could be [1|1|1|0]
I'd like to process these sets so that as output, I only get the unique ones amongst them. So, in the example above, I would get as output:
Set(1), Set(2), Set(3). Note that Set(4) is discarded since it is same as Set(2).
A rather brute force way of figuring this gives me a worst-case bound of O(n^3):
Given: Input List of size O(n)
Output List L = Set(1)
for(j = 2 to Length of Input List){ // Loop Outer, check if Set(j) should be added to L
for(i = 1 to Length of L currently){ // Loop Inner
check if Set(i) is same as Set(j) //This step is O(n) since Set() has O(n) elements
if(they are same) exit inner loop
else
if( i is length of L currently) //so, Set(j) is unique thus far
Append Set(j) to L
}
}
There is no a priori bound on n: it can be arbitrarily large. This seems to preclude use of simple hash function which maps the binary set into decimal. I could be wrong.
Is there any other way this can be done in better worst-case running time other than O(n^3)?
O(n) sequences of length n makes an input of size O(n^2). You won't get complexity better than that, since you may at least be required to read all the input. All sequences might be the same, for example, but you'd have to read them all to know that.
A binary sequence of length n can be inserted into a trie or radix tree, while checking whether or not it already exists, in O(n) time. That's O(n^2) for all the sequences together, so simply using a trie or radix tree to find duplicates is optimal.
See: https://en.wikipedia.org/wiki/Trie
and: https://en.wikipedia.org/wiki/Radix_tree
You may consider implementing your set using a balanced binary tree. The cost of inserting a new node into such a tree is O(lgm), where m is the number of elements in the tree. Duplicates would implicitly be weeded out because if we detect that such a node already exists, then it would just not be added.
In your example, the total number of lookup/insertion operations would be n*n, since there are n sets, and each set has n values. So, the overall time might scale as O(n^2*lg(n^2)). This outperforms O(n^3) by some amount.
First of all, these are not sets but bitstrings.
Next, for every bitstring you can convert it to a number and put that number in a hashset (or simply store the original bitstrings, most hashset implementations can do that). Afterwards, your hashset contains all the unique items. O(N) time, O(N) space. If you need to maintain the original order of strings, then in the first loop check for each string if it is in the hashset already, and if not, output it and insert in the hashset.
If you can use O(n) extra space, you can try this:
First of all, let's assume the vectors are binary numbers, so 0110 becomes 6.
This is in case numbers in vectors are [0,1], else you can multiply by 10 instead of 2.
Converting all vectors into decimals would take O(4n).
For each converted number we'll map the vector by the decimal number. To implement this, we'll be using an n-sized hash-map.
HM <- n-sized hash-map
for each vector v:
num <- decimal number converted of v
map v into HM by num
loop over HM and take only one for each index
runtime by steps:
O(n)
O(n*(4+1)) , when 1 is the time for mapping, 4 is the vector length
O(n)
If I have an unsorted large set of n integers (say 2^20 of them) and would like to generate subsets with k elements each (where k is small, say 5) in increasing order of their sums, what is the most efficient way to do so?
Why I need to generate these subsets in this fashion is that I would like to find the k-element subset with the smallest sum satisfying a certain condition, and I thus would apply the condition on each of the k-element subsets generated.
Also, what would be the complexity of the algorithm?
There is a similar question here: Algorithm to get every possible subset of a list, in order of their product, without building and sorting the entire list (i.e Generators) about generating subsets in order of their product, but it wouldn't fit my needs due to the extremely large size of the set n
I intend to implement the algorithm in Mathematica, but could do it in C++ or Python too.
If your desired property of the small subsets (call it P) is fairly common, a probabilistic approach may work well:
Sort the n integers (for millions of integers i.e. 10s to 100s of MB of ram, this should not be a problem), and sum the k-1 smallest. Call this total offset.
Generate a random k-subset (say, by sampling k random numbers, mod n) and check it for P-ness.
On a match, note the sum-total of the subset. Subtract offset from this to find an upper bound on the largest element of any k-subset of equivalent sum-total.
Restrict your set of n integers to those less than or equal to this bound.
Repeat (goto 2) until no matches are found within some fixed number of iterations.
Note the initial sort is O(n log n). The binary search implicit in step 4 is O(log n).
Obviously, if P is so rare that random pot-shots are unlikely to get a match, this does you no good.
Even if only 1 in 1000 of the k-sized sets meets your condition, That's still far too many combinations to test. I believe runtime scales with nCk (n choose k), where n is the size of your unsorted list. The answer by Andrew Mao has a link to this value. 10^28/1000 is still 10^25. Even at 1000 tests per second, that's still 10^22 seconds. =10^14 years.
If you are allowed to, I think you need to eliminate duplicate numbers from your large set. Each duplicate you remove will drastically reduce the number of evaluations you need to perform. Sort the list, then kill the dupes.
Also, are you looking for the single best answer here? Who will verify the answer, and how long would that take? I suggest implementing a Genetic Algorithm and running a bunch of instances overnight (for as long as you have the time). This will yield a very good answer, in much less time than the duration of the universe.
Do you mean 20 integers, or 2^20? If it's really 2^20, then you may need to go through a significant amount of (2^20 choose 5) subsets before you find one that satisfies your condition. On a modern 100k MIPS CPU, assuming just 1 instruction can compute a set and evaluate that condition, going through that entire set would still take 3 quadrillion years. So if you even need to go through a fraction of that, it's not going to finish in your lifetime.
Even if the number of integers is smaller, this seems to be a rather brute force way to solve this problem. I conjecture that you may be able to express your condition as a constraint in a mixed integer program, in which case solving the following could be a much faster way to obtain the solution than brute force enumeration. Assuming your integers are w_i, i from 1 to N:
min sum(i) w_i*x_i
x_i binary
sum over x_i = k
subject to (some constraints on w_i*x_i)
If it turns out that the linear programming relaxation of your MIP is tight, then you would be in luck and have a very efficient way to solve the problem, even for 2^20 integers (Example: max-flow/min-cut problem.) Also, you can use the approach of column generation to find a solution since you may have a very large number of values that cannot be solved for at the same time.
If you post a bit more about the constraint you are interested in, I or someone else may be able to propose a more concrete solution for you that doesn't involve brute force enumeration.
Here's an approximate way to do what you're saying.
First, sort the list. Then, consider some length-5 index vector v, corresponding to the positions in the sorted list, where the maximum index is some number m, and some other index vector v', with some max index m' > m. The smallest sum for all such vectors v' is always greater than the smallest sum for all vectors v.
So, here's how you can loop through the elements with approximately increasing sum:
sort arr
for i = 1 to N
for v = 5-element subsets of (1, ..., i)
set = arr{v}
if condition(set) is satisfied
break_loop = true
compute sum(set), keep set if it is the best so far
break if break_loop
Basically, this means that you no longer need to check for 5-element combinations of (1, ..., n+1) if you find a satisfying assignment in (1, ..., n), since any satisfying assignment with max index n+1 will have a greater sum, and you can stop after that set. However, there is no easy way to loop through the 5-combinations of (1, ..., n) while guaranteeing that the sum is always increasing, but at least you can stop checking after you find a satisfying set at some n.
This looks to be a perfect candidate for map-reduce (http://en.wikipedia.org/wiki/MapReduce). If you know of any way of partitioning them smartly so that passing candidates are equally present in each node then you can probably get a great throughput.
Complete sort may not really be needed as the map stage can take care of it. Each node can then verify the condition against the k-tuples and output results into a file that can be aggregated / reduced later.
If you know of the probability of occurrence and don't need all of the results try looking at probabilistic algorithms to converge to an answer.
This is an interview question. I have K machines each of which is connected to 1 central machine. Each of the K machines have an array of 4 byte numbers in file. You can use any data structure to load those numbers into memory on those machines and they fit. Numbers are not unique across K machines. Find the K largest numbers in the union of the numbers across all K machines. What is the fastest I can do this?
(This is an interesting problem because it involves parallelism. As I haven't encountered parallel algorithm optimization before, it's quite amusing: you can get away with some ridiculously high-complexity steps, because you can make up for it later. Anyway, onto the answer...)
> "What is the fastest I can do this?"
The best you can do is O(K). Below I illustrate both a simple O(K log(K)) algorithm, and the more complex O(K) algorithm.
First step:
Each computer needs enough time to read every element. This means that unless the elements are already in memory, one of the two bounds on the time is O(largest array size). If for example your largest array size varies as O(K log(K)) or O(K^2) or something, no amount of algorithmic trickery will let you go faster than that. Thus the actual best running time is O(max(K, largestArraySize)) technically.
Let us say the arrays have a max length of N, which is <=K. With the above caveat, we're allowed to bound N<K since each computer has to look at each of its elements at least once (O(N) preprocessing per computer), each computer can pick the largest K elements (this is known as finding kth-order-statistics, see these linear-time algorithms). Furthermore, we can do so for free (since it's also O(N)).
Bounds and reasonable expectations:
Let's begin by thinking of some worst-case scenarios, and estimates for the minimum amount of work necessary.
One minimum-work-necessary estimate is O(K*N/K) = O(N), because we need to look at every element at the very least. But, if we're smart, we can distribute the work evenly across all K computers (hence the division by K).
Another minimum-work-necessary estimate is O(N): if one array is larger than all elements on all other computers, we return the set.
We must output all K elements; this is at least O(K) to print them out. We can avoid this if we are content merely knowing where the elements are, in which case the O(K) bound does not necessarily apply.
Can this bound of O(N) be achieved? Let's see...
Simple approach - O(NlogN + K) = O(KlogK):
For now let's come up with a simple approach, which achieves O(NlogN + K).
Consider the data arranged like so, where each column is a computer, and each row is a number in the array:
computer: A B C D E F G
10 (o) (o)
9 o (o) (o)
8 o (o)
7 x x (x)
6 x x (x)
5 x ..........
4 x x ..
3 x x x . .
2 x x . .
1 x x .
0 x x .
You can also imagine this as a sweep-line algorithm from computation geometry, or an efficient variant of the 'merge' step from mergesort. The elements with parentheses represent the elements with which we'll initialize our potential "candidate solution" (in some central server). The algorithm will converge on the correct o responses by dumping the (x) answers for the two unselected os.
Algorithm:
All computers start as 'active'.
Each computer sorts its elements. (parallel O(N logN))
Repeat until all computers are inactive:
Each active computer finds the next-highest element (O(1) since sorted) and gives it to the central server.
The server smartly combines the new elements with the old K elements, and removes an equal number of the lowest elements from the combined set. To perform this step efficiently, we have a global priority queue of fixed size K. We insert the new potentially-better elements, and bad elements fall out of the set. Whenever an element falls out of the set, we tell the computer which sent that element to never send another one. (Justification: This always raises the smallest element of the candidate set.)
(sidenote: Adding a callback hook to falling out of a priority queue is an O(1) operation.)
We can see graphically that this will perform at most 2K*(findNextHighest_time + queueInsert_time) operations, and as we do so, elements will naturally fall out of the priority queue. findNextHighest_time is O(1) since we sorted the arrays, so to minimize 2K*queueInsert_time, we choose a priority queue with an O(1) insertion time (e.g. a Fibonacci-heap based priority queue). This gives us an O(log(queue_size)) extraction time (we cannot have O(1) insertion and extraction); however, we never need to use the extract operation! Once we are done, we merely dump the priority queue as an unordered set, which takes O(queue_size)=O(K) time.
We'd thus have O(N log(N) + K) total running time (parallel sorting, followed by O(K)*O(1) priority queue insertions). In the worst case of N=K, this is O(K log(K)).
The better approach - O(N+K) = O(K):
However I have come up with a better approach, which achieves O(K). It is based on the median-of-median selection algorithm, but parallelized. It goes like this:
We can eliminate a set of numbers if we know for sure that there are at least K (not strictly) larger numbers somewhere among all the computers.
Algorithm:
Each computer finds the sqrt(N)th highest element of its set, and splits the set into elements < and > it. This takes O(N) time in parallel.
The computers collaborate to combine those statistics into a new set, and find the K/sqrt(N)th highest element of that set (let's call it the 'superstatistic'), and note which computers have statistics < and > the superstatistic. This takes O(K) time.
Now consider all elements less than their computer's statistics, on computers whose statistic is less than the superstatistic. Those elements can be eliminated. This is because the elements greater than their computer's statistic, on computers whose statistic is larger than the superstatistic, are a set of K elements which are larger. (See the visual here).
Now, the computers with the uneliminated elements evenly redistribute their data to the computers who lost data.
Recurse: you still have K computers, but the value of N has decreased. Once N is less than a predetermined constant, use the previous algorithm I mentioned in "simple approach - O(NlogN + K)"; except in this case, it is now O(K). =)
It turns out that the reductions are O(N) total (amazingly not order K), except perhaps the final step which might by O(K). Thus this algorithm is O(N+K) = O(K) total.
Analysis and simulation of O(K) running time below. The statistics allow us to divide the world into four unordered sets, represented here as a rectangle divided into four subboxes:
------N-----
N^.5
________________
| | s | <- computer
| | #=K s REDIST. | <- computer
| | s | <- computer
| K/N^.5|-----S----------| <- computer
| | s | <- computer
K | s | <- computer
| | s ELIMIN. | <- computer
| | s | <- computer
| | s | <- computer
| |_____s__________| <- computer
LEGEND:
s=statistic, S=superstatistic
#=K -- set of K largest elements
(I'd draw the relation between the unordered sets of rows and s-column here, but it would clutter things up; see the addendum right now quickly.)
For this analysis, we will consider N as it decreases.
At a given step, we are able to eliminate the elements labelled ELIMIN; this has removed area from the rectangle representation above, reducing the problem size from K*N to , which hilariously simplifies to
Now, the computers with the uneliminated elements redistribute their data (REDIST rectangle above) to the computers with eliminated elements (ELIMIN). This is done in parallel, where the bandwidth bottleneck corresponds to the length of the short size of REDIST (because they are outnumbered by the ELIMIN computers which are waiting for their data). Therefore the data will take as long to transfer as the long length of the REDIST rectangle (another way of thinking about it: K/√N * (N-√N) is the area, divided by K/√N data-per-time, resulting in O(N-√N) time).
Thus at each step of size N, we are able to reduce the problem size to K(2√N-1), at the cost of performing N + 3K + (N-√N) work. We now recurse. The recurrence relation which will tell us our performance is:
T(N) = 2N+3K-√N + T(2√N-1)
The decimation of the subproblem size is much faster than the normal geometric series (being √N rather than something like N/2 which you'd normally get from common divide-and-conquers). Unfortunately neither the Master Theorem nor the powerful Akra-Bazzi theorem work, but we can at least convince ourselves it is linear via a simulation:
>>> def T(n,k=None):
... return 1 if n<10 else sqrt(n)*(2*sqrt(n)-1)+3*k+T(2*sqrt(n)-1, k=k)
>>> f = (lambda x: x)
>>> (lambda n: T((10**5)*n,k=(10**5)*n)/f((10**5)*n) - T(n,k=n)/f(n))(10**30)
-3.552713678800501e-15
The function T(N) is, at large scales, a multiple of the linear function x, hence linear (doubling the input doubles the output). This method, therefore, almost certainly achieves the bound of O(N) we conjecture. Though see the addendum for an interesting possibility.
...
Addendum
One pitfall is accidentally sorting. If we do anything which accidentally sorts our elements, we will incur a log(N) penalty at the least. Thus it is better to think of the arrays as sets, to avoid the pitfall of thinking that they are sorted.
Also we might initially think that with the constant amount of work at each step of 3K, so we would have to do work 3Klog(log(N)) work. But the -1 has a powerful role to play in the decimation of the problem size. It is very slightly possible that the running time is actually something above linear, but definitely much smaller than even Nlog(log(log(log(N)))). For example it might be something like O(N*InverseAckermann(N)), but I hit the recursion limit when testing.
The O(K) is probably only due to the fact that we have to print them out; if we are content merely knowing where the data is, we might even be able to pull off an O(N) (e.g. if the arrays are of length O(log(K)) we might be able to achieve O(log(K)))... but that's another story.
The relation between the unordered sets is as follows. Would have cluttered things up in explanation.
.
_
/ \
(.....) > s > (.....)
s
(.....) > s > (.....)
s
(.....) > s > (.....)
\_/
v
S
v
/ \
(.....) > s > (.....)
s
(.....) > s > (.....)
s
(.....) > s > (.....)
\_/
Find the k largest numbers on each machine. O(n*log(k))
Combine the results (on a centralized server, if k is not huge, otherwise you can merge them in a tree-hierarchy accross the server cluster).
Update: to make it clear, the combine step is not a sort. You just pick the top k numbers from the results. There are many ways to do this efficiently. You can use a heap for example, pushing the head of each list. Then you can remove the head from the heap and push the head from the list the element belonged to. Doing this k times gives you the result. All this is O(k*log(k)).
Maintain a min heap of size 'k' in the centralized server.
Initially insert first k elements into the min heap.
For the remaining elements
Check(peek) for the min element in the heap (O(1))
If the min element is lesser than the current element, then remove the min element from heap and insert the current element.
Finally min heap will have 'k' largest elements
This would require n(log k) time.
I would suggest something like this:
take the k largest numbers on each machine in sorted order O(Nk) where N is the number of element on each machine
sort each of these arrays of k elements by largest element (you will get k arrays of k elements sorted by largest element : a square matrix kxk)
take the "upper triangle" of the matrix made of these k arrays of k elements, (the k largest element will be in this upper triangle)
the central machine can now find the k largest element of these k(k+1)/2 elements
Let the machines find the out k largest elements copy it into a
datastructure (stack), sort it and pass it on to the Central
machine.
At the central machine receive the stacks from all the machine. Find
the greatest of the elements at the top of the stacks.
Pop out the greatest element form its stack and copy it to the 'TopK list'.
Leave the other stacks intact.
Repeat step 3, k times to get Top K numbers.
1) sort the items on every machine
2) use a k - binary heap on the central machine
a) populate the heap with first (max) element from each machine
b) extract the first element, and put back in the heap the first element from the machine that you extracted the element. (of course heapify your heap, after the element is added).
Sort will be O(Nlog(N)) where N is the max array on the machines.
O(k) - to build the heap
O(klog(k)) to extract and populate the heap k times.
Complexity is max(O(klog(k)),O(Nlog(N)))
I would think the MapReduce paradigm would be well suited to a task like this.
Every machine runs it's own independent map task to find the maximum value in its array (depends on the language used) and this will probably be O(N) complexity for N numbers on each machine.
The reduce task compares the result from the individual machines' outputs to give you the largest k numbers.
Is there a way to generate all of the subset sums s1, s2, ..., sk that fall in a range [A,B] faster than O((k+N)*2N/2), where k is the number of sums there are in [A,B]? Note that k is only known after we have enumerated all subset sums within [A,B].
I'm currently using a modified Horowitz-Sahni algorithm. For example, I first call it to for the smallest sum greater than or equal to A, giving me s1. Then I call it again for the next smallest sum greater than s1, giving me s2. Repeat this until we find a sum sk+1 greater than B. There is a lot of computation repeated between each iteration, even without rebuilding the initial two 2N/2 lists, so is there a way to do better?
In my problem, N is about 15, and the magnitude of the numbers is on the order of millions, so I haven't considered the dynamic programming route.
Check the subset sum on Wikipedia. As far as I know, it's the fastest known algorithm, which operates in O(2^(N/2)) time.
Edit:
If you're looking for multiple possible sums, instead of just 0, you can save the end arrays and just iterate through them again (which is roughly an O(2^(n/2) operation) and save re-computing them. The value of all the possible subsets is doesn't change with the target.
Edit again:
I'm not wholly sure what you want. Are we running K searches for one independent value each, or looking for any subset that has a value in a specific range that is K wide? Or are you trying to approximate the second by using the first?
Edit in response:
Yes, you do get a lot of duplicate work even without rebuilding the list. But if you don't rebuild the list, that's not O(k * N * 2^(N/2)). Building the list is O(N * 2^(N/2)).
If you know A and B right now, you could begin iteration, and then simply not stop when you find the right answer (the bottom bound), but keep going until it goes out of range. That should be roughly the same as solving subset sum for just one solution, involving only +k more ops, and when you're done, you can ditch the list.
More edit:
You have a range of sums, from A to B. First, you solve subset sum problem for A. Then, you just keep iterating and storing the results, until you find the solution for B, at which point you stop. Now you have every sum between A and B in a single run, and it will only cost you one subset sum problem solve plus K operations for K values in the range A to B, which is linear and nice and fast.
s = *i + *j; if s > B then ++i; else if s < A then ++j; else { print s; ... what_goes_here? ... }
No, no, no. I get the source of your confusion now (I misread something), but it's still not as complex as what you had originally. If you want to find ALL combinations within the range, instead of one, you will just have to iterate over all combinations of both lists, which isn't too bad.
Excuse my use of auto. C++0x compiler.
std::vector<int> sums;
std::vector<int> firstlist;
std::vector<int> secondlist;
// Fill in first/secondlist.
std::sort(firstlist.begin(), firstlist.end());
std::sort(secondlist.begin(), secondlist.end());
auto firstit = firstlist.begin();
auto secondit = secondlist.begin();
// Since we want all in a range, rather than just the first, we need to check all combinations. Horowitz/Sahni is only designed to find one.
for(; firstit != firstlist.end(); firstit++) {
for(; secondit = secondlist.end(); secondit++) {
int sum = *firstit + *secondit;
if (sum > A && sum < B)
sums.push_back(sum);
}
}
It's still not great. But it could be optimized if you know in advance that N is very large, for example, mapping or hashmapping sums to iterators, so that any given firstit can find any suitable partners in secondit, reducing the running time.
It is possible to do this in O(N*2^(N/2)), using ideas similar to Horowitz Sahni, but we try and do some optimizations to reduce the constants in the BigOh.
We do the following
Step 1: Split into sets of N/2, and generate all possible 2^(N/2) sets for each split. Call them S1 and S2. This we can do in O(2^(N/2)) (note: the N factor is missing here, due to an optimization we can do).
Step 2: Next sort the larger of S1 and S2 (say S1) in O(N*2^(N/2)) time (we optimize here by not sorting both).
Step 3: Find Subset sums in range [A,B] in S1 using binary search (as it is sorted).
Step 4: Next, for each sum in S2, find using binary search the sets in S1 whose union with this gives sum in range [A,B]. This is O(N*2^(N/2)). At the same time, find if that corresponding set in S2 is in the range [A,B]. The optimization here is to combine loops. Note: This gives you a representation of the sets (in terms of two indexes in S2), not the sets themselves. If you want all the sets, this becomes O(K + N*2^(N/2)), where K is the number of sets.
Further optimizations might be possible, for instance when sum from S2, is negative, we don't consider sums < A etc.
Since Steps 2,3,4 should be pretty clear, I will elaborate further on how to get Step 1 done in O(2^(N/2)) time.
For this, we use the concept of Gray Codes. Gray codes are a sequence of binary bit patterns in which each pattern differs from the previous pattern in exactly one bit.
Example: 00 -> 01 -> 11 -> 10 is a gray code with 2 bits.
There are gray codes which go through all possible N/2 bit numbers and these can be generated iteratively (see the wiki page I linked to), in O(1) time for each step (total O(2^(N/2)) steps), given the previous bit pattern, i.e. given current bit pattern, we can generate the next bit pattern in O(1) time.
This enables us to form all the subset sums, by using the previous sum and changing that by just adding or subtracting one number (corresponding to the differing bit position) to get the next sum.
If you modify the Horowitz-Sahni algorithm in the right way, then it's hardly slower than original Horowitz-Sahni. Recall that Horowitz-Sahni works two lists of subset sums: Sums of subsets in the left half of the original list, and sums of subsets in the right half. Call these two lists of sums L and R. To obtain subsets that sum to some fixed value A, you can sort R, and then look up a number in R that matches each number in L using a binary search. However, the algorithm is asymmetric only to save a constant factor in space and time. It's a good idea for this problem to sort both L and R.
In my code below I also reverse L. Then you can keep two pointers into R, updated for each entry in L: A pointer to the last entry in R that's too low, and a pointer to the first entry in R that's too high. When you advance to the next entry in L, each pointer might either move forward or stay put, but they won't have to move backwards. Thus, the second stage of the Horowitz-Sahni algorithm only takes linear time in the data generated in the first stage, plus linear time in the length of the output. Up to a constant factor, you can't do better than that (once you have committed to this meet-in-the-middle algorithm).
Here is a Python code with example input:
# Input
terms = [29371, 108810, 124019, 267363, 298330, 368607,
438140, 453243, 515250, 575143, 695146, 840979, 868052, 999760]
(A,B) = (500000,600000)
# Subset iterator stolen from Sage
def subsets(X):
yield []; pairs = []
for x in X:
pairs.append((2**len(pairs),x))
for w in xrange(2**(len(pairs)-1), 2**(len(pairs))):
yield [x for m, x in pairs if m & w]
# Modified Horowitz-Sahni with toolow and toohigh indices
L = sorted([(sum(S),S) for S in subsets(terms[:len(terms)/2])])
R = sorted([(sum(S),S) for S in subsets(terms[len(terms)/2:])])
(toolow,toohigh) = (-1,0)
for (Lsum,S) in reversed(L):
while R[toolow+1][0] < A-Lsum and toolow < len(R)-1: toolow += 1
while R[toohigh][0] <= B-Lsum and toohigh < len(R): toohigh += 1
for n in xrange(toolow+1,toohigh):
print '+'.join(map(str,S+R[n][1])),'=',sum(S+R[n][1])
"Moron" (I think he should change his user name) raises the reasonable issue of optimizing the algorithm a little further by skipping one of the sorts. Actually, because each list L and R is a list of sizes of subsets, you can do a combined generate and sort of each one in linear time! (That is, linear in the lengths of the lists.) L is the union of two lists of sums, those that include the first term, term[0], and those that don't. So actually you should just make one of these halves in sorted form, add a constant, and then do a merge of the two sorted lists. If you apply this idea recursively, you save a logarithmic factor in the time to make a sorted L, i.e., a factor of N in the original variable of the problem. This gives a good reason to sort both lists as you generate them. If you only sort one list, you have some binary searches that could reintroduce that factor of N; at best you have to optimize them somehow.
At first glance, a factor of O(N) could still be there for a different reason: If you want not just the subset sum, but the subset that makes the sum, then it looks like O(N) time and space to store each subset in L and in R. However, there is a data-sharing trick that also gets rid of that factor of O(N). The first step of the trick is to store each subset of the left or right half as a linked list of bits (1 if a term is included, 0 if it is not included). Then, when the list L is doubled in size as in the previous paragraph, the two linked lists for a subset and its partner can be shared, except at the head:
0
|
v
1 -> 1 -> 0 -> ...
Actually, this linked list trick is an artifact of the cost model and never truly helpful. Because, in order to have pointers in a RAM architecture with O(1) cost, you have to define data words with O(log(memory)) bits. But if you have data words of this size, you might as well store each word as a single bit vector rather than with this pointer structure. I.e., if you need less than a gigaword of memory, then you can store each subset in a 32-bit word. If you need more than a gigaword, then you have a 64-bit architecture or an emulation of it (or maybe 48 bits), and you can still store each subset in one word. If you patch the RAM cost model to take account of word size, then this factor of N was never really there anyway.
So, interestingly, the time complexity for the original Horowitz-Sahni algorithm isn't O(N*2^(N/2)), it's O(2^(N/2)). Likewise the time complexity for this problem is O(K+2^(N/2)), where K is the length of the output.
I have two arrays, N and M. they are both arbitrarily sized, though N is usually smaller than M. I want to find out what elements in N also exist in M, in the fastest way possible.
To give you an example of one possible instance of the program, N is an array 12 units in size, and M is an array 1,000 units in size. I want to find which elements in N also exist in M. (There may not be any matches.) The more parallel the solution, the better.
I used to use a hash map for this, but it's not quite as efficient as I'd like it to be.
Typing this out, I just thought of running a binary search of M on sizeof(N) independent threads. (Using CUDA) I'll see how this works, though other suggestions are welcome.
1000 is a very small number. Also, keep in mind that parallelizing a search will only give you speedup as the number of cores you have increases. If you have more threads than cores, your application will start to slow down again due to context switching and aggregating information.
A simple solution for your problem is to use a hash join. Build a hash table from M, then look up the elements of N in it (or vice versa; since both your arrays are small it doesn't matter much).
Edit: in response to your comment, my answer doesn't change too much. You can still speed up linearly only until your number of threads equals your number of processors, and not past that.
If you want to implement a parallel hash join, this would not be difficult. Start by building X-1 hash tables, where X is the number of threads/processors you have. Use a second hash function which returns a value modulo X-1 to determine which hash table each element should be in.
When performing the search, your main thread can apply the auxiliary hash function to each element to determine which thread to hand it off to for searching.
Just sort N. Then for each element of M, do a binary search for it over sorted N. Finding the M items in N is trivially parallel even if you do a linear search over an unsorted N of size 12.