Related
I would like to generate a sequence of n random integers in the interval [1,n] without duplicates, i.e. a permutation of the sequence [1,2,...,n] with O(log(n)) space complexity (or a polynomial function of log(n)).
One hint is that I can assume that I have a family of l-wise uniform hash functions h : [n] -> [k] (with l<=n) such that for any y_1, y_2,..., y_l and any distinct x_1, x_2,..., x_l :
P(h(x_1) = y_1 and h(x_2) = y_2 and ... and h(x_l) = y_l) = 1/(k^l)
My first idea was to use the hash function to generate the i-th element of the sequence, i.e. x_i = h(i) , check if x_i is already used (has already been returned by the hash function for some 0<j<i) and if it's the case increment x_i by 1 and check again until x_i is a new number. My problem is I can not have a vector of booleans of size n to check if the value x_i is already used. And if I do a recursive function to get the j-th value I will need at some point O(n log2(n)) bits...
I also found here that pseudorandom generator like Linear congruential generator can be used for this kind of problem with something like x_i+1 = (a*x_i + c)%n + 1 but I am not sure to understand how to choose a for any value of n to have a period of length n. In that case the hint is not really useful except for generating the first number of the sequence thus I don't think it's the right way.
Here's a fun super simple solution with constant space; when N is a power of 2 and your definition of "random" is incredibly loose (the resulting sequence will alternate between even and odd numbers).
N = power of 2
P = prime number larger than N.
S = random starting number between 0 and N-1
For i = 1 TO N
// add our prime to the starting random number
S += P
// S Modulus N
// Bitwise And N-1 works because N is a pow of 2
T = S & (N - 1)
//T is [0, (N-1)] => we want [1, N]
PRINT (T + 1)
Next I
JS
for(let N = 64, P = 73, S = N * Math.random(), i = 1; i <= N; i++) { S += P; console.log((S & (N - 1)) + 1); }
Another answer would probably be to consider all of the numbers [1, N] as leaf nodes in a tree and your Log(N) space is the size of a the path through the tree. Your solution would be a function that permutes all N paths through the tree. The way you permute the paths in a pseudo random way would basically be a Linear Feedback Shift Register type generator that has a period grater than N.
https://www.maximintegrated.com/en/design/technical-documents/app-notes/4/4400.html
Given an array arr of n integers, what is the highest score that a player can reach, playing the following game?
Choose an index 0 < i < n-1 in the array
Add arr[i-1] * arr[i+1] points to the score (initially the score is 0)
Shrink the array by removing element i (forall j >= i: arr[j] = arr[j+1]; then n = n - 1
Repeat steps 1-3 until n == 2.
Do the above until there are only 2 elements (which are the first and the last element because you can't remove them).
What is the highest score you can get ?
Example
arr = [1 2 3 4]
Choose i=2, get: 2*4 = 8 points, remove 3
Remaining: arr = [1 2 4]
Choose i=1, get 1*4 = 4 points, remove 2
Remaining: arr = [1 4].
The sum of points is 8 + 4 = 12, which is the highest possible score on this example.
I think it is related to Dynamic programming but I'm not sure how to solve it.
This problem has a dynamic programming approach similar to Matrix-chain multiplication problem. You can find further explanation in the book "Introduction to Algorithms", 3rd Edition (Cormen, page 370).
Let's find the optimal substructure property and then use it to construct an optimal solution to the problem from optimal solutions to subproblems.
Notation: Ci..j, where i ≤ j, stands for elements Ci,Ci+1,...,Cj.
Definition: A removal sequence for Ci..j is a permutation of i+1,i+2,...,j-1.
A removal sequence for Ci..j is optimal if the score achieved by removing the elements of Ci..j in that order is maximum among all possible removal sequences for Ci..j.
1. Characterize the structure of an optimal solution
If the problem is nontrivial, i.e. i + 1 < j, then any solution has a last removed element which corresponding index is k in the range
i < k < j. Such k split the problem into Ci..k and Ck..j. That is, for some value k, we first remove non extremal elements of Ci..k and Ck..j and then we remove element k. As removing non extremal elements of Ci..k doesn't affect score obtained by removing non extremal elements of Ck..j and an analogous reasoning for removing non extremal elements of Ck..j is also true we state that both subproblems are independent. Then, for a given removal sequence where kth-element is last, the score of Ci..j is equal to the sum of scores of Ci..k and Ck..j, plus the score of removing kth-element (C[i] * C[j]).
The optimal substructure of this problem is as follows. Suppose there is an optimal removal sequence O for Ci..j that ends at kth-element, then the ordering of removed elements from Ci..k must be optimal too. We can prove it by contradiction: If there was a removal sequence for Ci..k that scored higher than removal subsequence extracted from O for Ci..k then we can produce another removal sequence for Ci..j with higher score than optimal removal sequence (contradiction). A similar observation holds for the ordering of removed elements from Ck..j in the optimal removal sequence for Ci..j: it must be optimal too.
We can build an optimal solution for nontrivial instances of the problem by splitting the problem into two subproblems, finding optimal solutions to subproblem instances, and them combining these optimal subproblem solutions.
2. Recursively define the value of an optimal solution.
For this problem our subproblems are the maximum score obtained in Ci..j for 1 ≤ i ≤ j ≤ N. Let S[i, j] be the maximum score obtained in Ci..j; for the full problem, the highest score when evaluating the given rules is S[1, N].
We can define S[i, j] recursively as follows:
If j ≤ i + 1 then S[i, j] = 0
If i + 1 < j then S[i, j] = maxi < k < j{S[i, k] + S[k, j] + C[i] * C[j]}
We ensure that we search for the correct place to split because we consider all possible places, so that we are sure of having examined the optimal one.
3. Compute the value of an optimal solution
You can use your favorite method to compute S:
top-down approach (recursive)
bottom-up approach (iterative)\
I would use bottom-up for computing the solution since it would be < 5 lines long in almost any programming language.
Example in C++11:
for(int l = 2; l <= N; ++l) \\ increasing length intervals
for(int i = 1, j = i + l; j <= N; ++i, ++j)
for(int k = i + 1; k < j; ++k)
S[i, j] = max(S[i, j], S[i, k] + S[k, j] + C[i] * C[j])
4. Time Complexity and Space Complexity
There are nC2 + n = Θ(n2) subproblems and every subproblem do an operation which running time is Θ(l) where l is length of the subproblem so the math yield a running time of Θ(n3) for the algorithm (it's easy to spot the O(n3) part :-)). Also, the algorithm requires Θ(n2) space to store the S table.
How to check if n can be partitioned to sum of a sequence of consecutive prime numbers.
For example, 12 is equal to 5+7 which 5 and 7 are consecutive primes, but 20 is equal to 3+17 which 3 and 17 are not consecutive.
Note that, repetition is not allowed.
My idea is to find and list all primes below n, then use 2 loops to sum all primes. The first 2 numbers, second 2 numbers, third 2 numbers etc. and then first 3 numbers, second 3 numbers and so far. But it takes lot of time and memory.
Realize that a consecutive list of primes is defined only by two pieces of information, the starting and the ending prime number. You just have to find these two numbers.
I assume that you have all the primes at your disposal, sorted in the array called primes. Keep three variables in memory: sum which initially is 2 (the smallest prime), first_index and last_index which are initially 0 (index of the smallest prime in array primes).
Now you have to "tweak" these two indices, and "travel" the array along the way in the loop:
If sum == n then finish. You have found your sequence of primes.
If sum < n then enlarge the list by adding next available prime. Increment last_index by one, and then increment sum by the value of new prime, which is primes[last_index]. Repeat the loop. But if primes[last_index] is larger than n then there is no solution, and you must finish.
If sum > n then reduce the list by removing the smallest prime from the list. Decrement sum by that value, which is primes[first_index], and then increment first_index by one. Repeat the loop.
Dialecticus's algorithm is the classic O(m)-time, O(1)-space way to solve this type of problem (here I'll use m to represent the number of prime numbers less than n). It doesn't depend on any mysterious properties of prime numbers. (Interestingly, for the particular case of prime numbers, AlexAlvarez's algorithm is also linear time!) Dialecticus gives a clear and correct description, but seems at a loss to explain why it is correct, so I'll try to do this here. I really think it's valuable to take the time to understand this particular algorithm's proof of correctness: although I had to read a number of explanations before it finally "sank in", it was a real "Aha!" moment when it did! :) (Also, problems that can be efficiently solved in the same manner crop up quite a lot.)
The candidate solutions this algorithm tries can be represented as number ranges (i, j), where i and j are just the indexes of the first and last prime number in a list of prime numbers. The algorithm gets its efficiency by ruling out (that is, not considering) sets of number ranges in two different ways. To prove that it always gives the right answer, we need to show that it never rules out the only range with the right sum. To that end, it suffices to prove that it never rules out the first (leftmost) range with the right sum, which is what we'll do here.
The first rule it applies is that whenever we find a range (i, j) with sum(i, j) > n, we rule out all ranges (i, k) having k > j. It's easy to see why this is justified: the sum can only get bigger as we add more terms, and we have determined that it's already too big.
The second, trickier rule, crucial to the linear time complexity, is that whenever we advance the starting point of a range (i, j) from i to i+1, instead of "starting again" from (i+1, i+1), we start from (i+1, j) -- that is, we avoid considering (i+1, k) for all i+1 <= k < j. Why is it OK to do this? (To put the question the other way: Couldn't it be that doing this causes us to skip over some range with the right sum?)
[EDIT: The original version of the next paragraph glossed over a subtlety: we might have advanced the range end point to j on any previous step.]
To see that it never skips a valid range, we need to think about the range (i, j-1). For the algorithm to advance the starting point of the current range, so that it changes from (i, j) to (i+1, j), it must have been that sum(i, j) > n; and as we will see, to get to a program state in which the range (i, j) is being considered in the first place, it must have been that sum(i, j-1) < n. That second claim is subtle, because there are two different ways to arrive in such a program state: either we just incremented the end point, meaning that the previous range was (i, j-1) and this range was found to be too small (in which case our desired property sum(i, j-1) < n obviously holds); or we just incremented the start point after considering (i-1, j) and finding it to be too large (in which case it's not obvious that the property still holds).
What we do know, however, is that regardless of whether the end point was increased from j-1 to j on the previous step, it was definitely increased at some time before the current step -- so let's call the range that triggered this end point increase (k, j-1). Clearly sum(k, j-1) < n, since this was (by definition) the range that caused us to increase the end point from j-1 to j; and just as clearly k <= i, since we only process ranges in increasing order of their start points. Since i >= k, sum(i, j-1) is just the same as sum(k, j-1) but with zero or more terms removed from the left end, and all of these terms are positive, so it must be that sum(i, j-1) <= sum(k, j-1) < n.
So we have established that whenever we increase i to i+1, we know that sum(i, j-1) < n. To finish the analysis of this rule, what we (again) need to make use of is that dropping terms from either end of this sum can't make it any bigger. Removing the first term leaves us with sum(i+1, j-1) <= sum(i, j-1) < n. Starting from that sum and successively removing terms from the other end leaves us with sum(i+1, j-2), sum(i+1, j-3), ..., sum(i+1, i+1), all of which we know must be less than n -- that is, none of the ranges corresponding to these sums can be valid solutions. Therefore we can safely avoid considering them in the first place, and that's exactly what the algorithm does.
One final potential stumbling block is that it might seem that, since we are advancing two loop indexes, the time complexity should be O(m^2). But notice that every time through the loop body, we advance one of the indexes (i or j) by one, and we never move either of them backwards, so if we are still running after 2m loop iterations we must have i + j = 2m. Since neither index can ever exceed m, the only way for this to hold is if i = j = m, which means that we have reached the end: i.e. we are guaranteed to terminate after at most 2m iterations.
The fact that primes have to be consecutive allows to solve quite efficiently this problem in terms of n. Let me suppose that we have previously computed all the primes less or equal than n. Therefore, we can easily compute sum(i) as the sum of the first i primes.
Having this function precomputed, we can loop over the primes less or equal than n and see whether there exists a length such that starting with that prime we can sum up to n. But notice that for a fixed starting prime, the sequence of sums is monotone, so we can binary search over the length.
Thus, let k be the number of primes less or equal than n. Precomputing the sums has cost O(k) and the loop has cost O(klogk), dominating the cost. Using the Prime number theorem, we know that k = O(n/logn), and then the whole algorithm has cost O(n/logn log(n/logn)) = O(n).
Let me put a code in C++ to make it clearer, hope there are not bugs:
#include <iostream>
#include <vector>
using namespace std;
typedef long long ll;
int main() {
//Get the limit for the numbers
int MAX_N;
cin >> MAX_N;
//Compute the primes less or equal than MAX_N
vector<bool> is_prime(MAX_N + 1, true);
for (int i = 2; i*i <= MAX_N; ++i) {
if (is_prime[i]) {
for (int j = i*i; j <= MAX_N; j += i) is_prime[j] = false;
}
}
vector<int> prime;
for (int i = 2; i <= MAX_N; ++i) if (is_prime[i]) prime.push_back(i);
//Compute the prefixed sums
vector<ll> sum(prime.size() + 1, 0);
for (int i = 0; i < prime.size(); ++i) sum[i + 1] = sum[i] + prime[i];
//Get the number of queries
int n_queries;
cin >> n_queries;
for (int z = 1; z <= n_queries; ++z) {
int n;
cin >> n;
//Solve the query
bool found = false;
for (int i = 0; i < prime.size() and prime[i] <= n and not found; ++i) {
//Do binary search over the lenght of the sum:
//For all x < ini, [i, x] sums <= n
int ini = i, fin = int(prime.size()) - 1;
while (ini <= fin) {
int mid = (ini + fin)/2;
int value = sum[mid + 1] - sum[i];
if (value <= n) ini = mid + 1;
else fin = mid - 1;
}
//Check the candidate of the binary search
int candidate = ini - 1;
if (candidate >= i and sum[candidate + 1] - sum[i] == n) {
found = true;
cout << n << " =";
for (int j = i; j <= candidate; ++j) {
cout << " ";
if (j > i) cout << "+ ";
cout << prime[j];
}
cout << endl;
}
}
if (not found) cout << "No solution" << endl;
}
}
Sample input:
1000
5
12
20
28
17
29
Sample output:
12 = 5 + 7
No solution
28 = 2 + 3 + 5 + 7 + 11
17 = 2 + 3 + 5 + 7
29 = 29
I'd start by noting that for a pair of consecutive primes to sum to the number, one of the primes must be less than N/2, and the other prime must be greater than N/2. For them to be consecutive primes, they must be the primes closest to N/2, one smaller and the other larger.
If you're starting with a table of prime numbers, you basically do a binary search for N/2. Look at the primes immediately larger and smaller than that. Add those numbers together and see if they sum to your target number. If they don't, then it can't be the sum of two consecutive primes.
If you don't start with a table of primes, it works out pretty much the same way--you still start from N/2 and find the next larger prime (we'll call that prime1). Then you subtract N-prime1 to get a candidate for prime2. Check if that's prime, and if it is, search the range prime2...N/2 for other primes to see if there was a prime in between. If there's a prime in between your number is a sum of non-consecutive primes. If there's no other prime in that range, then it is a sum of consecutive primes.
The same basic idea applies for sequences of 3 or more primes, except that (of course) your search starts from N/3 (or whatever number of primes you want to sum to get to the number).
So, for three consecutive primes to sum to N, 2 of the three must be the first prime smaller than N/3 and the first prime larger than N/3. So, we start by finding those, then compute N-(prime1+prime2). That gives use our third candidate. We know these three numbers sum to N. We still need to prove that this third number is a prime. If it is prime, we need to verify that it's consecutive to the other two.
To give a concrete example, for 10 we'd start from 3.333. The next smaller prime is 3 and the next larger is 5. Those add to 8. 10-8 = 2. 2 is prime and consecutive to 3, so we've found the three consecutive primes that add to 10.
There are some other refinements you can make as well. The most obvious would be based on the fact that all primes (other than 2) are odd numbers. Therefore (assuming we can ignore 2), an even number can only be the sum of an even number of primes, and an odd number can only be a sum of an odd number of primes. So, given 123456789, we know immediately that it can't possibly be the sum of 2 (or 4, 6, 8, 10, ...) consecutive primes, so the only candidates to consider are 3, 5, 7, 9, ... primes. Of course, the opposite works as well: given, say, 12345678, the simple fact that it's even lets us immediately rule out the possibility that it could be the sum of 3, 5, 7 or 9 consecutive primes; we only need to consider sequences of 2, 4, 6, 8, ... primes. We violate this basic rule only when we get to a large enough number of primes that we could include 2 as part of the sequence.
I haven't worked through the math to figure out exactly how many that would be be for a given number, but I'm pretty sure it should be fairly easy and it's something we want to know anyway (because it's the upper limit on the number of consecutive primes to look for for a given number). If we use M for the number of primes, the limit should be approximately M <= sqrt(N), but that's definitely only an approximation.
I know that this question is a little old, but I cannot refrain from replying to the analysis made in the previous answers. Indeed, it has been emphasized that all the three proposed algorithms have a run-time that is essentially linear in n. But in fact, it is not difficult to produce an algorithm that runs at a strictly smaller power of n.
To see how, let us choose a parameter K between 1 and n and suppose that the primes we need are already tabulated (if they must be computed from scratch, see below). Then, here is what we are going to do, to search a representation of n as a sum of k consecutive primes:
First we search for k<K using the idea present in the answer of Jerry Coffin; that is, we search k primes located around n/k.
Then to explore the sums of k>=K primes we use the algorithm explained in the answer of Dialecticus; that is, we begin with a sum whose first element is 2, then we advance the first element one step at a time.
The first part, that concerns short sums of big primes, requires O(log n) operations to binary search one prime close to n/k and then O(k) operations to search for the other k primes (there are a few simple possible implementations). In total this makes a running time
R_1=O(K^2)+O(Klog n).
The second part, that is about long sums of small primes, requires us to consider sums of consecutive primes p_1<\dots<p_k where the first element is at most n/K.
Thus, it requires to visit at most n/K+K primes (one can actually save a log factor by a weak version of the prime number theorem). Since in the algorithm every prime is visited at most O(1) times, the running time is
R_2=O(n/K) + O(K).
Now, if log n < K < \sqrt n we have that the first part runs with O(K^2) operations and the second part runs in O(n/K). We optimize with the choice K=n^{1/3}, so that the overall running time is
R_1+R_2=O(n^{2/3}).
If the primes are not tabulated
If we also have to find the primes, here is how we do it.
First we use Erathostenes, that in C_2=O(T log log T) operations finds all the primes up to T, where T=O(n/K) is the upper bound for the small primes visited in the second part of the algorithm.
In order to perform the first part of the algorithm we need, for every k<K, to find O(k) primes located around n/k. The Riemann hypothesis implies that there are at least k primes in the interval [x,x+y] if y>c log x (k+\sqrt x) for some constant c>0. Therefore a priori we need to find the primes contained in an interval I_k centered at n/k with width |I_k|= O(k log n)+O(\sqrt {n/k} log n).
Using the sieve Eratosthenes to sieve the interval I_k requires O(|I_k|log log n) + O(\sqrt n) operations. If k<K<\sqrt n we get a time complexity C_1=O(\sqrt n log n log log n) for every k<K.
Summing up, the time complexity C_1+C_2+R_1+R_2 is maximized when
K = n^{1/4} / (log n \sqrt{log log n}).
With this choice have the sublinear time complexity
R_1+R_2+C_1+C_2 = O(n^{3/4}\sqrt{log log n}.
If we do not assume the Riemann Hypothesis we will have to search on larger intervals, but we still get at the end a sublinear time complexity. If instead we assume stronger conjectures on prime gaps, we may only need to search on intervals I_k with width |I_k|=k (log n)^A for some A>0. Then, instead of Erathostenes, we can use other deterministic primality tests. For example, suppose that you can test a single number for primality in O((log n)^B) operations, for some B>0.
Then you can search the interval I_k in O(k(log n)^{A+B}) operations. In this case the optimal K is still K\approx n^{1/3}, up to logarithmic factors, and so the total complexity is O(n^{2/3}(log n)^D for some D>0.
The question is Number of solutions to a1 x1+a2 x2+....+an xn=k with constraints: 1)ai>0 and ai<=15 2)n>0 and n<=15 3)xi>=0 I was able to formulate a Dynamic programming solution but it is running too long for n>10^10. Please guide me to get a more efficient soution.
The code
int dp[]=new int[16];
dp[0]=1;
BigInteger seen=new BigInteger("0");
while(true)
{
for(int i=0;i<arr[0];i++)
{
if(dp[0]==0)
break;
dp[arr[i+1]]=(dp[arr[i+1]]+dp[0])%1000000007;
}
for(int i=1;i<15;i++)
dp[i-1]=dp[i];
seen=seen.add(new BigInteger("1"));
if(seen.compareTo(n)==0)
break;
}
System.out.println(dp[0]);
arr is the array containing coefficients and answer should be mod 1000000007 as the number of ways donot fit into an int.
Update for real problem:
The actual problem is much simpler. However, it's hard to be helpful without spoiling it entirely.
Stripping it down to the bare essentials, the problem is
Given k distinct positive integers L1, ... , Lk and a nonnegative integer n, how many different finite sequences (a1, ..., ar) are there such that 1. for all i (1 <= i <= r), ai is one of the Lj, and 2. a1 + ... + ar = n. (In other words, the number of compositions of n using only the given Lj.)
For convenience, you are also told that all the Lj are <= 15 (and hence k <= 15), and n <= 10^18. And, so that the entire computation can be carried out using 64-bit integers (the number of sequences grows exponentially with n, you wouldn't have enough memory to store the exact number for large n), you should only calculate the remainder of the sequence count modulo 1000000007.
To solve such a problem, start by looking at the simplest cases first. The very simplest cases are when only one L is given, then evidently there is one admissible sequence if n is a multiple of L and no admissible sequence if n mod L != 0. That doesn't help yet. So consider the next simplest cases, two L values given. Suppose those are 1 and 2.
0 has one composition, the empty sequence: N(0) = 1
1 has one composition, (1): N(1) = 1
2 has two compositions, (1,1); (2): N(2) = 2
3 has three compositions, (1,1,1);(1,2);(2,1): N(3) = 3
4 has five compositions, (1,1,1,1);(1,1,2);(1,2,1);(2,1,1);(2,2): N(4) = 5
5 has eight compositions, (1,1,1,1,1);(1,1,1,2);(1,1,2,1);(1,2,1,1);(2,1,1,1);(1,2,2);(2,1,2);(2,2,1): N(5) = 8
You may see it now, or need a few more terms, but you'll notice that you get the Fibonacci sequence (shifted by one), N(n) = F(n+1), thus the sequence N(n) satisfies the recurrence relation
N(n) = N(n-1) + N(n-2) (for n >= 2; we have not yet proved that, so far it's a hypothesis based on pattern-spotting). Now, can we see that without calculating many values? Of course, there are two types of admissible sequences, those ending with 1 and those ending with 2. Since that partitioning of the admissible sequences restricts only the last element, the number of ad. seq. summing to n and ending with 1 is N(n-1) and the number of ad. seq. summing to n and ending with 2 is N(n-2).
That reasoning immediately generalises, given L1 < L2 < ... < Lk, for all n >= Lk, we have
N(n) = N(n-L1) + N(n-L2) + ... + N(n-Lk)
with the obvious interpretation if we're only interested in N(n) % m.
Umm, that linear recurrence still leaves calculating N(n) as an O(n) task?
Yes, but researching a few of the mentioned keywords quickly leads to an algorithm needing only O(log n) steps ;)
Algorithm for misinterpreted problem, no longer relevant, but may still be interesting:
The question looks a little SPOJish, so I won't give a complete algorithm (at least, not before I've googled around a bit to check if it's a contest question). I hope no restriction has been omitted in the description, such as that permutations of such representations should only contribute one to the count, that would considerably complicate the matter. So I count 1*3 + 2*4 = 11 and 2*4 + 1*3 = 11 as two different solutions.
Some notations first. For m-tuples of numbers, let < | > denote the canonical bilinear pairing, i.e.
<a|x> = a_1*x_1 + ... + a_m*x_m. For a positive integer B, let A_B = {1, 2, ..., B} be the set of positive integers not exceeding B. Let N denote the set of natural numbers, i.e. of nonnegative integers.
For 0 <= m, k and B > 0, let C(B,m,k) = card { (a,x) \in A_B^m × N^m : <a|x> = k }.
Your problem is then to find \sum_{m = 1}^15 C(15,m,k) (modulo 1000000007).
For completeness, let us mention that C(B,0,k) = if k == 0 then 1 else 0, which can be helpful in theoretical considerations. For the case of a positive number of summands, we easily find the recursion formula
C(B,m+1,k) = \sum_{j = 0}^k C(B,1,j) * C(B,m,k-j)
By induction, C(B,m,_) is the convolution¹ of m factors C(B,1,_). Calculating the convolution of two known functions up to k is O(k^2), so if C(B,1,_) is known, that gives an O(n*k^2) algorithm to compute C(B,m,k), 1 <= m <= n. Okay for small k, but our galaxy won't live to see you calculating C(15,15,10^18) that way. So, can we do better? Well, if you're familiar with the Laplace-transformation, you'll know that an analogous transformation will convert the convolution product to a pointwise product, which is much easier to calculate. However, although the transformation is in this case easy to compute, the inverse is not. Any other idea? Why, yes, let's take a closer look at C(B,1,_).
C(B,1,k) = card { a \in A_B : (k/a) is an integer }
In other words, C(B,1,k) is the number of divisors of k not exceeding B. Let us denote that by d_B(k). It is immediately clear that 1 <= d_B(k) <= B. For B = 2, evidently d_2(k) = 1 if k is odd, 2 if k is even. d_3(k) = 3 if and only if k is divisible by 2 and by 3, hence iff k is a multiple of 6, d_3(k) = 2 if and only if one of 2, 3 divides k but not the other, that is, iff k % 6 \in {2,3,4} and finally, d_3(k) = 1 iff neither 2 nor 3 divides k, i.e. iff gcd(k,6) = 1, iff k % 6 \in {1,5}. So we've seen that d_2 is periodic with period 2, d_3 is periodic with period 6. Generally, like reasoning shows that d_B is periodic for all B, and the minimal positive period divides B!.
Given any positive period P of C(B,1,_) = d_B, we can split the sum in the convolution (k = q*P+r, 0 <= r < P):
C(B,m+1, q*P+r) = \sum_{c = 0}^{q-1} (\sum_{j = 0}^{P-1} d_B(j)*C(B,m,(q-c)*P + (r-j)))
+ \sum_{j = 0}^r d_B(j)*C(B,m,r-j)
The functions C(B,m,_) are no longer periodic for m >= 2, but there are simple formulae to obtain C(B,m,q*P+r) from C(B,m,r). Thus, with C(B,1,_) = d_B and C(B,m,_) known up to P, calculating C(B,m+1,_) up to P is an O(P^2) task², getting the data necessary for calculating C(B,m+1,k) for arbitrarily large k, needs m such convolutions, hence that's O(m*P^2).
Then finding C(B,m,k) for 1 <= m <= n and arbitrarily large k is O(n^2*P^2), in time and O(n^2*P) in space.
For B = 15, we have 15! = 1.307674368 * 10^12, so using that for P isn't feasible. Fortunately, the smallest positive period of d_15 is much smaller, so you get something workable. From a rough estimate, I would still expect the calculation of C(15,15,k) to take time more appropriately measured in hours than seconds, but it's an improvement over O(k) which would take years (for k in the region of 10^18).
¹ The convolution used here is (f \ast g)(k) = \sum_{j = 0}^k f(j)*g(k-j).
² Assuming all arithmetic operations are O(1); if, as in the OP, only the residue modulo some M > 0 is desired, that holds if all intermediate calculations are done modulo M.
Assume that I have a vector, V, of positive integers. If the sum of the integers are larger than a positive integer N, I want to rescale the integers in V so that the sum is <= N. The elements in V must remain above zero. The length of V is guaranteed to be <= N.
Is there an algorithm to perform this rescaling in linear time?
This is not homework, BTW :). I need to rescale a map from symbols to symbol frequencies to use range encoding.
Some quick thinking and googling has not given a solution to the problem.
EDIT:
Ok, the question was somewhat unclear. "Rescale" means "normalize". That is, transform the integers in V, for example by multiplying them by a constant, to smaller positive integers so the criterion of sum(V) <= N is fulfilled. The better the ratios between the integers are preserved, the better the compression will be.
The problem is open-ended in that way, the method does not need to find the optimal (in, say, a least squares fit sense) way to preserve the ratios, but a "good" one. Setting the entire vector to 1, as suggested, is not acceptable (unless forced). "Good" enough would for example be finding the smallest divisor (defined below) that fulfills the sum criterion.
The following naive algorithm does not work.
Find the current sum(V), Sv
divisor := int(ceil(Sv/N))
Divide each integer in V by divisor, rounding down, but not to less than 1.
This fails on v = [1,1,1,10] with N = 5.
divisor = ceil(13 / 5) = 3.
V := [1,1,1, max(1, floor(10/3)) = 3]
Sv is now 6 > 5.
In this case, the correct normalization is [1,1,1,2]
One algorithm that would work is to do a binary search for divisor (defined above) until the smallest divisor in [1,N] fulfilling the sum criterion is found. Starting with the ceil(Sv/N) guess. This is however, not linear in number of operations, but proportional to len(V)*log(len(V)).
I am starting to think that it is impossible to do well, in linear time, in the general case. I might resort to some sort of heuristic.
Just divide all the integers by their Greatest Common Divisor. You can find the GCD efficiently with multiple applications of Euclid's Algorithm.
d = 0
for x in xs:
d = gcd(d, x)
xs = [x/d for x in xs]
The positive point is that you always have a small as possible representation this way, without throwing away any precision and without needing to choose a specific N. The downside is that if your frequencies are large coprime numbers you will have no choice but to sacrifice precision (and you didn't specify what should be done in this case).
How about this:
Find the current sum(V), Sv
divisor := int(ceil(Sv/(N - |V| + 1))
Divide each integer in V by divisor, rounding up
On v = [1,1,1,10] with N = 5:
divisor = ceil(13 / 2) = 7.
V := [1,1,1, ceil(10/7)) = 2]
I think you should just rescale the part above 1. So, subtract 1 from all values, and V.length from N. Then rescale normally, then add 1 back. You can even do slightly better if you keep running totals as you go along, instead of choosing just one factor, which will usually waste some "number space". Something like this:
public static void rescale(int[] data, int N) {
int sum = 0;
for (int d : data)
sum += d;
if (sum > N) {
int n = N - data.length;
sum -= data.length;
for (int a = 0; a < data.length; a++) {
int toScale = data[a] - 1;
int scaled = Math.round(toScale * (float) n / sum);
data[a] = scaled + 1;
n -= scaled;
sum -= toScale;
}
}
}
This is a problem of 'range normalization', but it's very easy. Suppose that S is the sum of the elements of the vector, and S>=N, then S=dN, for some d>=1. Therefore d=S/N. So just multiply every element of the vector by N/S (i.e. divide by d). The result is a vector with rescaled components which sum is exactly N. This procedure is clearly linear :)