A fast algorithm to minimize a pseudo Diophantine equation - algorithm

We're looking for an algorithm to solve this problem in under O(N).
given two real numbers a and b (without loss of generality you can assume they are both between 0 and 1)
Find an integer n between -N and N that minimizes the expression:
|a n - b - round(a n - b)|
We have thought that the Euclidean Algorithm might work well for this, but can't figure it out. It certainly looks like there should be much faster ways to do this than via an exhaustive search over integers n.
Note: in our situation a and b could be changing often, so fixing a and b for a lookup table is possible, it gets kind of ugly as N can vary as well. Haven't looked in detail into the lookup table yet, to see how small we can get it as a function of N.

It sounds like you may be looking for something like continued fractions...
How are they related? Suppose you can substitute b with a rational number b1/b2. Now you are looking for integers n and m such that an-b1/b2 is approximately m. Put it otherwise, you are looking for n and m such that (m+(b1/b2))/n = (mb2+b1)/nb1, a rational number, is approximately a. Set a1 = mb2+b1 and a2 = nb1. Find values for a1 and a2 from a continued fractions approximation and solve for n and m.
Another approach could be this:
Find a good rational approximations for a and b: a ~ a1/a2 and b ~ b1/b2.
Solve n(a1/a2)-(b1/b2) = m for n and m.
I'm not too sure it would work though. The accuracy needed for a depends on n and b.

You are effectively searching for the integer N that makes the expression aN - b as close as possible to an integer. Are a and b fixed? If yes you can pre-compute a lookup table and have O(1) :-)
If not consider looking for the N that makes aN close to I + b for all integers I.

You can compute a continued fraction for the ratio a/b. You can stop when the denominator is greater than N, or when your approximation is good enough.
// Initialize:
double ratio = a / b;
int ak = (int)(ratio);
double remainder = ratio - ak;
int n0 = 1;
int d0 = 0;
int n1 = ak;
int d1 = 1;
do {
ratio = 1 / remainder;
ak = (int)ratio;
int n2 = ak * n1 + n0;
int d2 = ak * d1 + d0;
n0 = n1;
d0 = d1;
n1 = n2;
d1 = d2;
remainder = ratio - ak;
} while (d1 < N);
The value for n you're looking for is d0 (or d1 if it is still smaller than N).
This doesn't necessarily give you the minimum solution, but it will likely be a very good approximation.

First, let us consider a simpler case where b=0 and 0 < a < 1. F(a,n) = |an-round(an)|
Let step_size = 1
Step 1. Let v=a
Step 2. Let period size p = upper_round( 1/v ).
Step 3. Now, for n=1..p, there must be a number i such that F(v,i) < v.
Step 4. v = F(v,i), step_size = stepsize * i
Step 5. Go to step 2
As you can see you can reduce F(v, *) to any level you want. Final solution n = step_size.

Related

Calculate integer powers with a given loop invariant

I need to derive an algorithm in C++ to calculate integer powers m^n that uses the loop invariant r = y^n and the loop condition y != m.
I tried using the instruction y= y+1 to advance, but I don´t know how to obtain (y+1)^n from y^n, and it shouldn't be difficult to find . So, probably, this isn't the correct path to follow
Could you help me to derive the program?
EDIT: this is a problem from the subject Data Structures and Algorithms. The difficulty ( if there is at all) shouldn't be mathematic.
EDIT2: Just to clarify, the difficulty of the problem is using the invariant y^n and loop condition y != m. If I vary the n I'm not achieving that
Given w and P such that 2^w > m, P > 2^(wn), and 2^((P-1)/2) = -1 mod P,
then 2 is a generator mod P, and there will be some x such that 2^x = m mod P, so:
if (m<=1 || n==1)
return m;
if (n==0)
return 1;
let y = 2;
let r = 1<<n;
while(y!=m)
{
y = (y*2)%P;
r = (r*(1<<n))%P;
}
return r;
Unless your function needs to produce bignum results, you can just pick the largest P that fits into an integer in your language.
There is no useful relation between (y+1)^n and y^n (you can write (y+1)^n = (√(y^n)+1)^n or (y+1)^n = (1+1/y)^n y^n, but this leads you nowhere).
If y was factored, you could exploit (a.b)^n = (a^n).(b^n), but you would need a table of the nth powers of the primes.
I can't see an answer that makes sense.
You can also think of the Binomial theorem,
(y+1)^n = y^n + n y^(n-1) + n(n-1)/2 y^(n-2) + ... 1
but this is worse than anything: you need to compute n binomial coefficients, and update all powers of y from 0 to n. The total cost of the computation would be ridiculously high.

Find the value of f(T) for big value T

I am trying to solve a problem which is described below,
Given value of f(0) and k , which are integers.
I need to find value of f( T ). where T<=1010
Recursive function is,
f(n) = 2*f(n-1) , if 4*f(n-1) <=k
k - ( 2*f(n-1) ) , if 4*f(n-1) > k
My efforts,
#include<iostream>
using namespace std;
int main(){
long k,f0,i;
cin>>k>>f0;
long operation ;
cin>>operation;
long answer=f0;
for(i=1;i<=operation;i++){
answer=(4*answer <= k )?(2*answer):(k-(2*answer));
}
cout<<answer;
return 0;
}
My code gives me right answer. But, The code will run 1010 time in worst case that gives me Time Limit Exceed. I need more efficient solution for this problem. Please help me. I don't know the correct algorithm.
If 2f(0) < k then you can compute this function in O(log n) time (using exponentiation by squaring modulo k).
r = f(0) * 2^n mod k
return 2 * r >= k ? k - r : r
You can prove this by induction. The induction hypothesis is that 0 <= f(n) < k/2, and that the above code fragment computes f(n).
Here's a Python program which checks random test cases, comparing a naive implementation (f) with an optimized one (g).
def f(n, k, z):
r = z
for _ in xrange(n):
if 4*r <= k:
r = 2 * r
else:
r = k - 2 * r
return r
def g(n, k, z):
r = (z * pow(2, n, k)) % k
if 2 * r >= k:
r = k - r
return r
import random
errs = 0
while errs < 20:
k = random.randrange(100, 10000000)
n = random.randrange(100000)
z = random.randrange(k//2)
a1 = f(n, k, z)
a2 = g(n, k, z)
if a1 != a2:
print n, k, z, a1, a2
errs += 1
print '.',
Can you use methmetical solution before progamming and compulating?
Actually,
f(n) = f0*2^(n-1) , if f(n-1)*4 <= k
k - f0*2^(n-1) , if f(n-1)*4 > k
thus, your code will write like this:
condition = f0*pow(2, operation-2)
answer = condition*4 =< k? condition*2: k - condition*2
For a simple loop, your answer looks pretty tight; one could optimise a little bit using answer<<2 instead of 4*answer, and answer<<1 for 2*answer, but quite possibly your compiler is already doing that. If you're blowing the time with this, it might be necessary to reduce the loop itself somehow.
I can't figure out a mathematical pattern that #Shannon was going for, but I'm thinking we could exploit the fact that this function will sooner or later cycle. If the cycle is short enough, then we could short the loop by just getting the answer at the same point in the cycle.
So let's get some cycle detection equipment in the form of Brent's algorithm, and see if we can cut the loop to reasonable levels.
def brent(f, x0):
# main phase: search successive powers of two
power = lam = 1
tortoise = x0
hare = f(x0) # f(x0) is the element/node next to x0.
while tortoise != hare:
if power == lam: # time to start a new power of two?
tortoise = hare
power *= 2
lam = 0
hare = f(hare)
lam += 1
# Find the position of the first repetition of length λ
mu = 0
tortoise = hare = x0
for i in range(lam):
# range(lam) produces a list with the values 0, 1, ... , lam-1
hare = f(hare)
# The distance between the hare and tortoise is now λ.
# Next, the hare and tortoise move at same speed until they agree
while tortoise != hare:
tortoise = f(tortoise)
hare = f(hare)
mu += 1
return lam, mu
f0 = 2
k = 198779
t = 10000000000
def f(x):
if 4 * x <= k:
return 2 * x
else:
return k - 2 * x
lam, mu = brent(f, f0)
t2 = t
if t >= mu + lam: # if T is past the cycle's first loop,
t2 = (t - mu) % lam + mu # find the equivalent place in the first loop
x = f0
for i in range(t2):
x = f(x)
print("Cycle start: %d; length: %d" % (mu, lam))
print("Equivalent result at index: %d" % t2)
print("Loop iterations skipped: %d" % (t - t2))
print("Result: %d" % x)
As opposed to the other proposed answers, this approach actually could use a memo array to speed up the process, since the start of the function is actually calculated multiple times (in particular, inside brent), or it may be irrelevant, depending on how big the cycle happens to be.
The algorithm you proposed already has O(n).
To come up with more efficient algorithms, there is not that much direction we can go about. Some typical options we have
1.Decease the coefficients of the linear term( but I doubt it would make a difference in this case
2.Change to O(Logn)(typically use some sort of divide and conquer technique)
3.Change to O(1)
In this case, we can do the last one.
The recursion function is a piece-wise function
f(n) = 2*f(n-1) , if 4*f(n-1) <=k
k - ( 2*f(n-1) ) , if 4*f(n-1) > k
Let's tackle it by case:
case 1: if 4*f(n-1) <= k (1)(assuming the starting index is zero)
this is a obvious a geometry series
a_n = 2*a_n-1
Therefore, have the formula
Sn = 2^(n-1)f(0) ----()
Case 2: if 4*f(n-1) > k (2), we have
a_n = -2a_n-1 + k
Assuming, a_j is the element in the sequence which just satisfy condition (2)
Nestedly sub in an_1 to the formula, you will obtain the equation
an = k -2k +4k -8k... +(-2)^(n-j)* a_j
k -2k 4k -8... is another gemo series
Sn = k*(1-2^(n-j))/(1-2) ---gemo series sum formula with starting value k and ratio = -2
Therefore, we have a formula for an in the case 2
an = k * (1-2^(n-j))/(1-2) + (-2)^(n-j) * a_j ----(**)
All we left to do it to find aj which just dissatisfy condition (1) and satisfy (2)
This can be obtained in constant time again using the formula we have for case 1:
find n such that, 4*an = 4*Sn = 4*2^(n-1)*f(0)
solve for n: 4*2^(n-1)*f(0) = k, if n is not integer, take ceiling of n
In my first attempt to solve this question, I had wrong assumption that the value of the sequence is monotonically increasing but in fact the sequence might jump between case 1 and case 2. Therefore, there might not be constant algorithm to solve the problem.
However, we can use utilize the result above to skip iterative update complexity.
The overall algorithm will look something like:
start with T, K, and f(0)
compute n that make the condition switch using either (*) or (**)
update f(0) with f(n), update T - n
repeat
terminate when T-n = 0(the last iteration might over compute causing T-n<0, therefore, you need to go back a little bit if that happen)
Create a map that can store your results. Before finding f(n) check in that map, if solution is already existed or not.
If exists, use that solution.
Otherwise find it, store it for future use.
For C++:
Definition:
map<long,long>result;
Insertion:
result[key]=value
Accessing:
value=result[key];
Checking:
map<long,long>::iterator it=result.find(key);
if(it==result.end())
{
//key was not found, find the solution and insert into result
}
else
{
return result[key];
}
Use above technique for better solution.

How do you determine the average-case complexity of this algorithm?

It's usually easy to calculate the time complexity for the best case and the worst case, but when it comes to the average case especially when there's a probability p given, I don't know where to start.
Let's look at the following algorithm to compute the product of all the elements in a matrix:
int computeProduct(int[][] A, int m, int n) {
int product = 1;
for (int i = 0; i < m; i++ {
for (int j = 0; j < n; j++) {
if (A[i][j] == 0) return 0;
product = product * A[i][j];
}
}
return product;
}
Suppose p is the probability of A[i][j] being 0 (i.e. the algorithm terminates there, return 0); how do we derive the average case time complexity for this algorithm?
Let’s consider a related problem. Imagine you have a coin that flips heads with probability p. How many times, on expectation, do you need to flip the coin before it comes up heads? The answer is 1/p, since
There’s a p chance that you need one flip.
There’s a p(1-p) chance that you need two flips (the first flip has to go tails and the second has to go heads).
There’s a p(1-p)^2 chance that you need three flips (the first two flips need to go tails and the third has to go heads)
...
There’s a p(1-p)^(k-1) chance that you need k flips (the first k-1 flips need to go tails and the kth needs to go heads.)
So this means the expected value of the number of flips is
p + 2p(1 - p) + 3p(1 - p)^2 + 4p(1 - p)^3 + ...
= p(1(1 - p)^0 + 2(1 - p)^1 + 3(1 - p)^2 + ...)
So now we need to work out what this summation is. The general form is
p sum from k = 1 to infinity (k(1 - p)^k).
Rather than solving this particular summation, let's make this more general. Let x be some variable that, later, we'll set equal to 1 - p, but which for now we'll treat as a free value. Then we can rewrite the above summation as
p sum from k = 1 to infinity (kx^(k-1)).
Now for a cute trick: notice that the inside of this expression is the derivative of x^k with respect to x. Therefore, this sum is
p sum from k = 1 to infinity (d/dx x^k).
The derivative is a linear operator, so we can move it out to the front:
p d/dx sum from k = 1 to infinity (x^k)
That inner sum (x + x^2 + x^3 + ...) is the Taylor series for 1 / (1 - x) - 1, so we can simplify this to get
p d/dx (1 / (1 - x) - 1)
= p / (1 - x)^2
And since we picked x = 1 - p, this simplifies to
p / (1 - (1 - p))^2
= p / p^2
= 1 / p
Whew! That was a long derivation. But it shows that the expected number of coin tosses needed is 1/p.
Now, in your case, your algorithm can be thought of as tossing mn coins that come up heads with probability p and stopping if any of them come up heads. Surely, the expected number of coins you’d need to toss won’t be more than the case where you’re allowed to flip infinitely often, so your expected runtime is at most O(1 / p) (assuming p > 0).
If we assume that p is independent of m and n, then we can notice that at after some initial growth, each added term into our summation as we increase the number of flips is exponentially lower than the previous ones. More specifically, after adding in roughly logarithmically many terms into the sum we’ll be off from the total in the case of the infinite summation. Therefore, provided that mn is roughly larger than Θ(log p), the sum ends up being Θ(1 / p). So in a big-O sense, if mn is independent of p, the runtime is Θ(1 / p).

Time complexity for this relation - matrix chain multiplication

I think an (inefficient) recursive procedure for Matrix chain multiplication problem can be this (based on recurrence relation given in Cormen):
MATRIX-CHAIN(i,j)
if i == j
return 0
if i < j
q = INF
for k = i to j-1
q = min (q, MATRIX-CHAIN(i,k) + MATRIX-CHAIN(k+1, j) + c)
//c = cost of multiplying two sub-matrices.
return q
Time complexity for this will be:
T(n) = summation over k varying from i to j [T(k) + T(n-k)]
Here, n = number of matrices to be multiplied.
What will be the value of T(n) and how?
This is http://en.wikipedia.org/wiki/Catalan_number
You can view the recurrence relation as doing parenthesis. The wiki page describes in depth how to arrive to the formula.
This might help:
you only have to work out each matrix-chain once (and store its value).
start = anywhere between i and j
end = anywhere between start and j
k = anywhere between start and end
if we think of a number with all 0's apart from three 1's (which represent start, k, end)
this special number has j-i+1 digits.
e.g. if i = 3 and j = 6 we need 4 digits giving us the following options:
1101 (i=3, k=4, j=6)
1011 (i=3, k=5, j=6)
0111 (i=4, k=5, j=6)
1110 (i=3, k=4, j=5)
number of choices for i,j,k = Combinations(3, j-i+1)
this is n!/(k! * (n-k)!) = (j-i+1)! / (3! * (j-i+1-3)!)

Number of Positive Solutions to a1 x1+a2 x2+......+an xn=k (k<=10^18)

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.

Resources