There is an arbitrary amount of distinct unsigned integer values within a known range.
The number of integer values is << the number of integers within the range.
I want to build a data structure which allows the following runtime complexities:
Insertion in O(1)
After insertion is done:
Deletion in O(1)
Get all values within a query range in O(k) with k being the number of result values (returned values do not have to be sorted)
Memory complexity is not restricted. However, an astronomically large amount of memory is not available ;-)
Here is an example:
range = [0, 1023]
insert 42
insert 350
insert 729
insert 64
insert 1
insert 680
insert 258
find values in [300;800] ; returns {350, 729, 680}
delete 350
delete 680
find values in [35;1000] ; returns {42, 258, 64, 729, 258}
delete 42
delete 258
find values in [0; 5] ; returns {1}
delete 1
Is such a data structure even possible? (with the aid of look-up tables etc)?
An approximation I thought about would be:
Bin the inserted values into buckets. 0..31 => bucket 0, 32..63 => bucket 1, 64..95 => bucket 2, 96..127 => bucket 3, ...
Insertion: find bucket id using simple shifting arithmetic, then insert it into an array per bucket
Find: find bucket id of start and endpoint using shifting arithmetic. Look through all values in the first and last bucket and check if they are within the range or outside the range. Add all values in all intermediate buckets to the search result
Delete: find bucket id using shifting. Swap value to delete with last value in bucket, then decrement count for this bucket.
Downside: if there are many queries which query a range which has a span of less than 32 values, the whole bucket will be searched every time.
Downside 2: if there are empty buckets within the range, they will also be visited during the search phase.
Theoretically speaking, a van Emde Boas tree is your best bet, with O(log log M)-time operations where M is the size of the range. The space usage is quite large, though there are more efficient variants.
Actually the theoretical state of the art is described in the paper On Range Reporting in One Dimension, by Mortensen, Pagh, and Patrascu.
I'm not sure if the existing lower bounds rule out O(1), but M won't be large enough to make the distinction matter. Instead of the vEB structure, I would just use a k-ary trie with k a power of two like 32 or 64.
EDIT: here's one way to do range search with a trie.
Let's assume each datum is a bit pattern (easy enough; that's how the CPU think of it). Each subtree consists of all of the nodes with a certain prefix. For example, {0000, 0011, 0101, 1001} is represented by the following 4-ary trie, where X denotes a null pointer.
+---+---+---+---+
|00\|01\|10\|11X|
+--|+--|+--|+---+
| | |
| | +----------------------------+
+--+ | |
| +------------+ |
| | |
v v v
+---+---+---+---+ +---+---+---+---+ +---+---+---+---+
|00\|01X|10X|11\| |00X|01\|10X|11X| |00X|01\|10X|11X|
+--|+---+---+--|+ +---+--|+---+---+ +---+--|+---+---+
| | | |
v v v v
0000 0011 0101 1001
A couple optimizations quickly become apparent. First, if all of the bit patterns are the same length, then we don't need to store them at the leaves—they can be reconstructed from the descent path. All we need is the bitmap, which if k is the number of bits in a machine word, fits nicely where the pointer from the previous level used to be.
+--------+--------+--------+--------+
|00(1001)|01(0100)|10(0100)|11(0000)|
+--------+--------+--------+--------+
In order to search the trie for a range like [0001, 1000], we start at the root, determine which subtrees might intersect the range and recurse on them. In this example, the relevant children of the root are 00, 01, and 10. The relevant children of 00 are the subtrees representing the prefixes 0001, 0010, and 0011.
For k fixed, reporting from a k-ary trie is O(log M + s), where M is the number of bit patterns and s is the number of hits. Don't be fooled though—when k is medium, each node occupies a couple cache lines but the trie isn't very high, so the number of cache misses is pretty small.
You could achieve your target (O(1),O(1) and O(k)) if the query operation required that it be told the value of at least one existing member that is already in the relevant range (the lower bound perhaps). Can you provide a guarantee that you will already know at least one member of the range? I guess not. I will expand if you can.
I'll now focus on the problem as specified. Each number in the data structure should form part of a linked list, such that each number knows the next highest number that is in the data structure. In C++
struct Number {
struct Number *next_highest;
int value;
};
Obviously, the largest value in the set will have next_highest==NULL, but otherwise this->value < this->next_highest->value
To add or remove or query, we need to be able to find the existing Numbers which are close to a particular lookup value.
set<Number *, specialized_comparator_to_compare_on_value_t >
Insertion and deletion would be O(log(N)), and query would be O(log(N)+k). N is the number of values currently in the set, which as you say will be much less than M (the number of possible values of the relevant datatype). Therefore log(N) < log(M). But in practice, other methods should also be considered, such as tries and such datastructures.
Related
Assume we have the following data, which consists of a consecutive 0's and 1's (the nature of data is that there are very very very few 1s.
data =
[0,0,0,0,0,0,0,0,1,1,1,0,0,0,0,0,0,0,1,1,0,0,0,0,0,0,0,0,0,0,0,1,1,1,1,0,0]
so a huge number of zeros, and then possibly some ones (which indicate that some sort of an event is happening).
You want to query this data many times. The query is that given two indices i and j what is sum(data[i:j]). For example, sum_query(i=12, j=25) = 2 in above example.
Note that you have all these queries in advance.
What sort of a data structure can help me evaluate all the queries as fast as possible?
My initial thoughts:
preprocess the data and obtain two shorter arrays: data_change and data_cumsum. The data_change will be filled up with the indices for when the sequence of 1s will start and when the next sequence of 0s will start, and so on. The data_cumsum will contain the corresponding cummulative sums up to indices represented in data_change, i.e. data_cumsum[k] = sum(data[0:data_change[k]])
In above example, the preprocessing results in: data_change=[8,11,18,20,31,35] and data_cumsum=[0,3,3,5,5,9]
Then if query comes for i=12 and j=25, I will do a binary search in this sorted data_change array to find the corresponding index for 12 and then for 25, which will result in the 0-based indices: bin_search(data_change, 12)=2 and bin_search(data_change, 25)=4.
Then I simply output the corresponding difference from the cumsum array: data_cumsum[4] - data_cumsum[2]. (I won't go into the detail of handling the situation where the any endpoint of the query range falls in the middle of the sequence of 1's, but those cases can be handled easily with an if-statement.
With linear space, linear preprocessing, constant query time, you can store an array of sums. The i'th position gets the sum of the first i elements. To get query(i,j) you take the difference of the sums (sums[j] - sums[i-1]).
I already gave an O(1) time, O(n) space answer. Here are some alternates that trade time for space.
1. Assuming that the number of 1s is O(log n) or better (say O(log n) for argument):
Store an array of ints representing the positions of the ones in the original array. so if the input is [1,0,0,0,1,0,1,1] then A = [0,4,6,7].
Given a query, use binary search on A for the start and end of the query in O(log(|A|)) = O(log(log(n)). If the element you're looking for isn't in A, find the smallest bigger index and the largest smaller index. E.g., for query (2,6) you'd return the indices for the 4 and the 6, which are (1,2). Then the answer is one more than the difference.
2. Take advantage of knowing all the queries up front (as mentioned by the OP in a comment to my other answer). Say Q = (Q1, Q2, ..., Qm) is the set of queries.
Process the queries, storing a map of start and end indices to the query. E.g., if Q1 = (12,92) then our map would include {92 => Q1, 12 => Q1}. This take O(m) time and O(m) space. Take note of the smallest start index and the largest end index.
Process the input data, starting with the smallest start index. Keep track of the running sum. For each index, check your map of queries. If the index is in the map, associate the current running sum with the appropriate query.
At the end, each query will have two sums associated with it. Add one to the difference to get the answer.
Worst case analysis:
O(n) + O(m) time, O(m) space. However, this is across all queries. The amortized time cost per query is O(n/m). This is the same as my constant time solution (which required O(n) preprocessing).
I would probably go with something like this:
# boilerplate testdata
from itertools import chain, permutations
data = [0,0,0,0,0,0,0,1,1,1]
chained = list(chain(*permutations(data,5))) # increase 5 to 10 if you dare
Preprozessing:
frSet = frozenset([i for i in range(len(chained)) if chained[i]==1])
"Counting":
# O(min(len(frSet), len(frozenset(range(200,500))))
summa = frSet.intersection(frozenset(range(200,500))) # use two sets for faster intersect
counted=len(summa)
"Sanity-Check"
print(sum([1 for x in frSet if x >= 200 and x<500]))
print(summa)
print(len(summa))
No edge cases needed, intersection will do all you need, slightly higher memory as you store each index not ranges of ones. Performance depends on intersection-Implementation.
This might be helpfull: https://wiki.python.org/moin/TimeComplexity#set
My data has large number of sets (few millions). Each of those set size is between few members to several tens of thousands integers. Many of those sets are subsets of larger sets (there are many of those super-sets). I'm trying to assign each subset to it's largest superset.
Please can anyone recommend algorithm for this type of task?
There are many algorithms for generating all possible sub-sets of a set, but this type of approach is time-prohibitive given my data size (e.g. this paper or SO question).
Example of my data-set:
A {1, 2, 3}
B {1, 3}
C {2, 4}
D {2, 4, 9}
E {3, 5}
F {1, 2, 3, 7}
Expected answer: B and A are subset of F (it's not important B is also subset of A); C is a subset of D; E remains unassigned.
Here's an idea that might work:
Build a table that maps number to a sorted list of sets, sorted first by size with largest first, and then, by size, arbitrarily but with some canonical order. (Say, alphabetically by set name.) So in your example, you'd have a table that maps 1 to [F, A, B], 2 to [F, A, D, C], 3 to [F, A, B, E] and so on. This can be implemented to take O(n log n) time where n is the total size of the input.
For each set in the input:
fetch the lists associated with each entry in that set. So for A, you'd get the lists associated with 1, 2, and 3. The total number of selects you'll issue in the runtime of the whole algorithm is O(n), so runtime so far is O(n log n + n) which is still O(n log n).
Now walk down each list simultaneously. If a set is the first entry in all three lists, then it's the largest set that contains the input set. Output that association and continue with the next input list. If not, then discard the smallest item among all the items in the input lists and try again. Implementing this last bit is tricky, but you can store the heads of all lists in a heap and get (IIRC) something like O(n log k) overall runtime where k is the maximum size of any individual set, so you can bound that at O(n log n) in the worst case.
So if I got everything straight, the runtime of the algorithm is overall O(n log n), which seems like probably as good as you're going to get for this problem.
Here is a python implementation of the algorithm:
from collections import defaultdict, deque
import heapq
def LargestSupersets(setlists):
'''Computes, for each item in the input, the largest superset in the same input.
setlists: A list of lists, each of which represents a set of items. Items must be hashable.
'''
# First, build a table that maps each element in any input setlist to a list of records
# of the form (-size of setlist, index of setlist), one for each setlist that contains
# the corresponding element
element_to_entries = defaultdict(list)
for idx, setlist in enumerate(setlists):
entry = (-len(setlist), idx) # cheesy way to make an entry that sorts properly -- largest first
for element in setlist:
element_to_entries[element].append(entry)
# Within each entry, sort so that larger items come first, with ties broken arbitrarily by
# the set's index
for entries in element_to_entries.values():
entries.sort()
# Now build up the output by going over each setlist and walking over the entries list for
# each element in the setlist. Since the entries list for each element is sorted largest to
# smallest, the first entry we find that is in every entry set we pulled will be the largest
# element of the input that contains each item in this setlist. We are guaranteed to eventually
# find such an element because, at the very least, the item we're iterating on itself is in
# each entries list.
output = []
for idx, setlist in enumerate(setlists):
num_elements = len(setlist)
buckets = [element_to_entries[element] for element in setlist]
# We implement the search for an item that appears in every list by maintaining a heap and
# a queue. We have the invariants that:
# 1. The queue contains the n smallest items across all the buckets, in order
# 2. The heap contains the smallest item from each bucket that has not already passed through
# the queue.
smallest_entries_heap = []
smallest_entries_deque = deque([], num_elements)
for bucket_idx, bucket in enumerate(buckets):
smallest_entries_heap.append((bucket[0], bucket_idx, 0))
heapq.heapify(smallest_entries_heap)
while (len(smallest_entries_deque) < num_elements or
smallest_entries_deque[0] != smallest_entries_deque[num_elements - 1]):
# First extract the next smallest entry in the queue ...
(smallest_entry, bucket_idx, element_within_bucket_idx) = heapq.heappop(smallest_entries_heap)
smallest_entries_deque.append(smallest_entry)
# ... then add the next-smallest item from the bucket that we just removed an element from
if element_within_bucket_idx + 1 < len(buckets[bucket_idx]):
new_element = buckets[bucket_idx][element_within_bucket_idx + 1]
heapq.heappush(smallest_entries_heap, (new_element, bucket_idx, element_within_bucket_idx + 1))
output.append((idx, smallest_entries_deque[0][1]))
return output
Note: don't trust my writeup too much here. I just thought of this algorithm right now, I haven't proved it correct or anything.
So you have millions of sets, with thousands of elements each. Just representing that dataset takes billions of integers. In your comparisons you'll quickly get to trillions of operations without even breaking a sweat.
Therefore I'll assume that you need a solution which will distribute across a lot of machines. Which means that I'll think in terms of https://en.wikipedia.org/wiki/MapReduce. A series of them.
Read the sets in, mapping them to k:v pairs of i: s where i is an element of the set s.
Receive a key of an integers, along with a list of sets. Map them off to pairs (s1, s2): i where s1 <= s2 are both sets that included to i. Do not omit to map each set to be paired with itself!
For each pair (s1, s2) count the size k of the intersection, and send off pairs s1: k, s2: k. (Only send the second if s1 and s2 are different.
For each set s receive the set of supersets. If it is maximal, send off s: s. Otherwise send off t: s for every t that is a strict superset of s.
For each set s, receive the set of subsets, with s in the list only if it is maximal. If s is maximal, send off t: s for every t that is a subset of s.
For each set we receive the set of maximal sets that it is a subset of. (There may be many.)
There are a lot of steps for this, but at its heart it requires repeated comparisons between pairs of sets with a common element for each common element. Potentially that is O(n * n * m) where n is the number of sets and m is the number of distinct elements that are in many sets.
Here is a simple suggestion for an algorithm that might give better results based on your numbers (n = 10^6 to 10^7 sets with m = 2 to 10^5 members, a lot of super/subsets). Of course it depends a lot on your data. Generally speaking complexity is much worse than for the other proposed algorithms. Maybe you could only process the sets with less than X, e.g. 1000 members that way and for the rest use the other proposed methods.
Sort the sets by their size.
Remove the first (smallest) set and start comparing it against the others from behind (largest set first).
Stop as soon as you found a superset and create a relation. Just remove if no superset was found.
Repeat 2. and 3. for all but the last set.
If you're using Excel, you could structure it as follows:
1) Create a cartesian plot as a two-way table that has all your data sets as titles on both the side and the top
2) In a seperate tab, create a row for each data set in the first column, along with a second column that will count the number of entries (ex: F has 4) and then just stack FIND(",") and MID formulas across the sheet to split out all the entries within each data set. Use the counter in the second column to do COUNTIF(">0"). Each variable you find can be your starting point in a subsequent FIND until it runs out of variables and just returns a blank.
3) Go back to your cartesian plot, and bring over the separate entries you just generated for your column titles (ex: F is 1,2,3,7). Use an AND statement to then check that each entry in your left hand column is in your top row data set using an OFFSET to your seperate area and utilizing your counter as the width for the OFFSET
There are many typical questions like https://softwareengineering.stackexchange.com/questions/150616/return-random-list-item-by-its-weight
Imagine more advanced problem.
You have N sources of pair (item_id, weight) information. Let's call them Shards. Shards contain lists of pairs (item_id, weight).
And you have central node, let's call it Central.
The problem is: on Central choose random item from The Big List (the list virtually merged from all lists on all shards) according to their weight through all weights.
For example, we have two shards:
+-------+---------+--------+
| shard | item_id | weight |
+-------+---------+--------+
| 1 | 1 | 7 |
| 1 | 2 | 4 |
| 1 | 3 | 2 |
| 2 | 4 | 5 |
| 2 | 5 | 1 |
+-------+---------+--------+
(Let item_id will be unique through all shards.)
First problem:
How to choose item_id randomly but weighted through all shards?
I.e. total_weight == 7+4+2+5+1 == 19, so 1 will be chosen with probability of 7/19, 2 - 4/19, 3 - 2/19 and so on.
Second problem:
How to range all items from all shards randomly, but weighted through all shards?
I.e. ideal ranging will be: 1, 4, 2, 3, 5 (according to their weights),
but there may be another ranging like 1, 2, 4, 3, 5, but slightly less frequently than previous,
...
and worst case 5, 3, 2, 4, 1 can also appear, but with very-very little probability.
Is there common problem in computer science for this?
I think your two questions are independent and should have been separate questions. Also I'm not sure that I have understood them correctly. But here we go:
Sharded weighted random query
If your reference to sharding has to do with distributing the item store over multiple network hosts and attempting to do some sort of network parallel random select, then you can use the modified reservoir sample algorithm which I outline at the end of this answer.
That algorithm was originally developed for use in a redundant network where the various storage hosts are not necessarily reachable directly from the central host and the connectivity is a graph, not a tree. In that case, you need to be able to deal with hosts which do not respond (which will bias an individual query, but if the network faults are infrequent and random will hopefully not bias a long series of queries). It's also necessary to deal with the possibility that a host is queried twice; in the outlined algorithm, I simply ignore the second and subsequent queries on the assumption that if a query reaches a host, then the response is likely to be returned to the querying host. That could be completely wrong, but it makes the problem much easier and again it's probably unbiased over a sufficiently large number of queries.
Without the complications, if the central host can reliably connect to every storage host, the algorithm is straight-forward. The central host queries all storage hosts in parallel, and each storage host returns a tuple of the total weight of objects it stores, and one object randomly selected according to those weights. (It uses some standard weighted random selection algorithm to do that. Which algorithm is used will depend on how often the objects and weights change.)
The central host maintains two variables: total, the sum of weights from servers which have responded (initially 0), and candidate, a random object which might be returned (initially some sentinel indicating "no object").
It handles the responses one at a time, in any order (such as the order it receives them). For each response <weight, object>, it does the following:
total ← total + weight
r ← a random integer in the range [0, total)
if r < weight: candidate ← object
When it decides that all remote servers have responded, it returns candidate.
Weighted random shuffle
(At least, I think you're asking for a weighted random shuffle).
I was going to suggest using the standard Fisher-Yates algorithm modified with weights, which I think will yield the sampling behaviour you expect. To do that, you start with the objects in any arbitrary order, and then for each value of i from 1 to n:
select j, the index of a (weighted) random element from the objects starting at i, and swap objects i and j.
To do that, you need to maintain the CDF of the successively smaller suffixes, which you can do in O(log N) by keeping the objects in a binary tree. (Or you can do it much more simply in O(N), if N is smallish.)
However, I did a quick search on SO before hitting the Post button, and concluded that this brilliant answer is actually better, because it achieves O(N log N) without any complicated data structure: For each object, you generate a random number from an exponential distribution whose rate is the corresponding weight. Then you sort the objects according to those random numbers, and the result is the random shuffle.
For your first problem, you can do something like
// Your pairs (item_id, weight).
ArrayList<Pair> theBigList;
// The total weight of your data. Get it updated.
int totalWeight;
public int weightedSearch() {
// Local copy of the value.
int total = totalWeight;
Random r = new Random();
// Random integer in the interval [1 - total]
int random = r.nextInt(total) + 1;
for (int i = 0; true; i++) {
if (theBigList.get(i).weigth < total)
total -= theBigList.get(i).weigth;
else
return theBigLits.get(i).itemId;
}
}
The random search, but weighted, is given by the random integer generated. In your example, if random is between 1 and 7 (7/19 prob.), your first element will be returned; if it is between 8 and 11 (4/19 prob.), second element is returned, etc.
Tip: Get your heavy items in the beginning, so you have more probability to return faster the weighted search (your loop ends before).
How can I generate a random number that is in the range (1,n) but not in a certain list (i,j)?
Example: range is (1,500), list is [1,3,4,45,199,212,344].
Note: The list may not be sorted
Rejection Sampling
One method is rejection sampling:
Generate a number x in the range (1, 500)
Is x in your list of disallowed values? (Can use a hash-set for this check.)
If yes, return to step 1
If no, x is your random value, done
This will work fine if your set of allowed values is significantly larger than your set of disallowed values:if there are G possible good values and B possible bad values, then the expected number of times you'll have to sample x from the G + B values until you get a good value is (G + B) / G (the expectation of the associated geometric distribution). (You can sense check this. As G goes to infinity, the expectation goes to 1. As B goes to infinity, the expectation goes to infinity.)
Sampling a List
Another method is to make a list L of all of your allowed values, then sample L[rand(L.count)].
The technique I usually use when the list is length 1 is to generate a random
integer r in [1,n-1], and if r is greater or equal to that single illegal
value then increment r.
This can be generalised for a list of length k for small k but requires
sorting that list (you can't do your compare-and-increment in random order). If the list is moderately long, then after the sort you can start with a bsearch, and add the number of values skipped to r, and then recurse into the remainder of the list.
For a list of length k, containing no value greater or equal to n-k, you
can do a more direct substitution: generate random r in [1,n-k], and
then iterate through the list testing if r is equal to list[i]. If it is
then set r to n-k+i (this assumes list is zero-based) and quit.
That second approach fails if some of the list elements are in [n-k,n].
I could try to invest something clever at this point, but what I have so far
seems sufficient for uniform distributions with values of k much less than
n...
Create two lists -- one of illegal values below n-k, and the other the rest (this can be done in place).
Generate random r in [1,n-k]
Apply the direct substitution approach for the first list (if r is list[i] then set r to n-k+i and go to step 5).
If r was not altered in step 3 then we're finished.
Sort the list of larger values and use the compare-and-increment method.
Observations:
If all values are in the lower list, there will be no sort because there is nothing to sort.
If all values are in the upper list, there will be no sort because there is no occasion on which r is moved into the hazardous area.
As k approaches n, the maximum size of the upper (sorted) list grows.
For a given k, if more value appear in the upper list (the bigger the sort), the chance of getting a hit in the lower list shrinks, reducing the likelihood of needing to do the sort.
Refinement:
Obviously things get very sorty for large k, but in such cases the list has comparatively few holes into which r is allowed to settle. This could surely be exploited.
I might suggest something different if many random values with the same
list and limits were needed. I hope that the list of illegal values is not the
list of results of previous calls to this function, because if it is then you
wouldn't want any of this -- instead you would want a Fisher-Yates shuffle.
Rejection sampling would be the simplest if possible as described already. However, if you didn't want use that, you could convert the range and disallowed values to sets and find the difference. Then, you could choose a random value out of there.
Assuming you wanted the range to be in [1,n] but not in [i,j] and that you wanted them uniformly distributed.
In Python
total = range(1,n+1)
disallowed = range(i,j+1)
allowed = list( set(total) - set(disallowed) )
return allowed[random.randrange(len(allowed))]
(Note that this is not EXACTLY uniform since in all likeliness, max_rand%len(allowed) != 0 but this will in most practical applications be very close)
I assume that you know how to generate a random number in [1, n) and also your list is ordered like in the example above.
Let's say that you have a list with k elements. Make a map(O(logn)) structure, which will ensure speed if k goes higher. Put all elements from list in map, where element value will be the key and "good" value will be the value. Later on I'll explain about "good" value. So when we have the map then just find a random number in [1, n - k - p)(Later on I'll explain what is p) and if this number is in map then replace it with "good" value.
"GOOD" value -> Let's start from k-th element. It's good value is its own value + 1, because the very next element is "good" for us. Now let's look at (k-1)th element. We assume that its good value is again its own value + 1. If this value is equal to k-th element then the "good" value for (k-1)th element is k-th "good" value + 1. Also you will have to store the largest "good" value. If the largest value exceed n then p(from above) will be p = largest - n.
Of course I recommend you this only if k is big number otherwise #Timothy Shields' method is perfect.
I have a collection of objects, each of which has a weight and a value. I want to pick the pair of objects with the highest total value subject to the restriction that their combined weight does not exceed some threshold. Additionally, I am given two arrays, one containing the objects sorted by weight and one containing the objects sorted by value.
I know how to do it in O(n2) but how can I do it in O(n)?
This is a combinatorial optimization problem, and the fact the values are sorted means you can easily try a branch and bound approach.
I think that I have a solution that works in O(n log n) time and O(n) extra space. This isn't quite the O(n) solution you wanted, but it's still better than the naive quadratic solution.
The intuition behind the algorithm is that we want to be able to efficiently determine, for any amount of weight, the maximum value we can get with a single item that uses at most that much weight. If we can do this, we have a simple algorithm for solving the problem: iterate across the array of elements sorted by value. For each element, see how much additional value we could get by pairing a single element with it (using the values we precomputed), then find which of these pairs is maximum. If we can do the preprocessing in O(n log n) time and can answer each of the above queries in O(log n) time, then the total time for the second step will be O(n log n) and we have our answer.
An important observation we need to do the preprocessing step is as follows. Our goal is to build up a structure that can answer the question "which element with weight less than x has maximum value?" Let's think about how we might do this by adding one element at a time. If we have an element (value, weight) and the structure is empty, then we want to say that the maximum value we can get using weight at most "weight" is "value". This means that everything in the range [0, max_weight - weight) should be set to value. Otherwise, suppose that the structure isn't empty when we try adding in (value, weight). In that case, we want to say that any portion of the range [0, weight) whose value is less than value should be replaced by value.
The problem here is that when we do these insertions, there might be, on iteration k, O(k) different subranges that need to be updated, leading to an O(n2) algorithm. However, we can use a very clever trick to avoid this. Suppose that we insert all of the elements into this data structure in descending order of value. In that case, when we add in (value, weight), because we add the elements in descending order of value, each existing value in the data structure must be higher than our value. This means that if the range [0, weight) intersects any range at all, those ranges will automatically be higher than value and so we don't need to update them. If we combine this with the fact that each range we add always spans from zero to some value, the only portion of the new range that could ever be added to the data structure is the range [weight, x), where x is the highest weight stored in the data structure so far.
To summarize, assuming that we visit the (value, weight) pairs in descending order of value, we can update our data structure as follows:
If the structure is empty, record that the range [0, value) has value "value."
Otherwise, if the highest weight recorded in the structure is greater than weight, skip this element.
Otherwise, if the highest weight recorded so far is x, record that the range [weight, x) has value "value."
Notice that this means that we are always splitting ranges at the front of the list of ranges we have encountered so far. Because of this, we can think about storing the list of ranges as a simple array, where each array element tracks the upper endpoint of some range and the value assigned to that range. For example, we might track the ranges [0, 3), [3, 9), and [9, 12) as the array
3, 9, 12
If we then needed to split the range [0, 3) into [0, 1) and [1, 3), we could do so by prepending 1 to he list:
1, 3, 9, 12
If we represent this array in reverse (actually storing the ranges from high to low instead of low to high), this step of creating the array runs in O(n) time because at each point we just do O(1) work to decide whether or not to add another element onto the end of the array.
Once we have the ranges stored like this, to determine which of the ranges a particular weight falls into, we can just use a binary search to find the largest element smaller than that weight. For example, to look up 6 in the above array we'd do a binary search to find 3.
Finally, once we have this data structure built up, we can just look at each of the objects one at a time. For each element, we see how much weight is left, use a binary search in the other structure to see what element it should be paired with to maximize the total value, and then find the maximum attainable value.
Let's trace through an example. Given maximum allowable weight 10 and the objects
Weight | Value
------+------
2 | 3
6 | 5
4 | 7
7 | 8
Let's see what the algorithm does. First, we need to build up our auxiliary structure for the ranges. We look at the objects in descending order of value, starting with the object of weight 7 and value 8. This means that if we ever have at least seven units of weight left, we can get 8 value. Our array now looks like this:
Weight: 7
Value: 8
Next, we look at the object of weight 4 and value 7. This means that with four or more units of weight left, we can get value 7:
Weight: 7 4
Value: 8 7
Repeating this for the next item (weight six, value five) does not change the array, since if the object has weight six, if we ever had six or more units of free space left, we would never choose this; we'd always take the seven-value item of weight four. We can tell this since there is already an object in the table whose range includes remaining weight four.
Finally, we look at the last item (value 3, weight 2). This means that if we ever have weight two or more free, we could get 3 units of value. The final array now looks like this:
Weight: 7 4 2
Value: 8 7 3
Finally, we just look at the objects in any order to see what the best option is. When looking at the object of weight 2 and value 3, since the maximum allowed weight is 10, we need tom see how much value we can get with at most 10 - 2 = 8 weight. A binary search over the array tells us that this value is 8, so one option would give us 11 weight. If we look at the object of weight 6 and value 5, a binary search tells us that with five remaining weight the best we can do would be to get 7 units of value, for a total of 12 value. Repeating this on the next two entries doesn't turn up anything new, so the optimum value found has value 12, which is indeed the correct answer.
Hope this helps!
Here is an O(n) time, O(1) space solution.
Let's call an object x better than an object y if and only if (x is no heavier than y) and (x is no less valuable) and (x is lighter or more valuable). Call an object x first-choice if no object is better than x. There exists an optimal solution consisting either of two first-choice objects, or a first-choice object x and an object y such that only x is better than y.
The main tool is to be able to iterate the first-choice objects from lightest to heaviest (= least valuable to most valuable) and from most valuable to least valuable (= heaviest to lightest). The iterator state is an index into the objects by weight (resp. value) and a max value (resp. min weight) so far.
Each of the following steps is O(n).
During a scan, whenever we encounter an object that is not first-choice, we know an object that's better than it. Scan once and consider these pairs of objects.
For each first-choice object from lightest to heaviest, determine the heaviest first-choice object that it can be paired with, and consider the pair. (All lighter objects are less valuable.) Since the latter object becomes lighter over time, each iteration of the loop is amortized O(1). (See also searching in a matrix whose rows and columns are sorted.)
Code for the unbelievers. Not heavily tested.
from collections import namedtuple
from operator import attrgetter
Item = namedtuple('Item', ('weight', 'value'))
sentinel = Item(float('inf'), float('-inf'))
def firstchoicefrombyweight(byweight):
bestsofar = sentinel
for x in byweight:
if x.value > bestsofar.value:
bestsofar = x
yield (x, bestsofar)
def firstchoicefrombyvalue(byvalue):
bestsofar = sentinel
for x in byvalue:
if x.weight < bestsofar.weight:
bestsofar = x
yield x
def optimize(items, maxweight):
byweight = sorted(items, key=attrgetter('weight'))
byvalue = sorted(items, key=attrgetter('value'), reverse=True)
maxvalue = float('-inf')
try:
i = firstchoicefrombyvalue(byvalue)
y = i.next()
for x, z in firstchoicefrombyweight(byweight):
if z is not x and x.weight + z.weight <= maxweight:
maxvalue = max(maxvalue, x.value + z.value)
while x.weight + y.weight > maxweight:
y = i.next()
if y is x:
break
maxvalue = max(maxvalue, x.value + y.value)
except StopIteration:
pass
return maxvalue
items = [Item(1, 1), Item(2, 2), Item(3, 5), Item(3, 7), Item(5, 8)]
for maxweight in xrange(3, 10):
print maxweight, optimize(items, maxweight)
This is similar to Knapsack problem. I will use naming from it (num - weight, val - value).
The essential part:
Start with a = 0 and b = n-1. Assuming 0 is the index of heaviest object and n-1 is the index of lightest object.
Increase a til objects a and b satisfy the limit.
Compare current solution with best solution.
Decrease b by one.
Go to 2.
Update:
It's the knapsack problem, except there is a limit of 2 items. You basically need to decide how much space you want for the first object and how much for the other. There is n significant ways to split available space, so the complexity is O(n). Picking the most valuable objects to fit in those spaces can be done without additional cost.