Checksum for an integer array? - arrays

I have an array that is of size 4,9,16 or 25 (according to the input) and the numbers in the array are the same but less by one (if the array size is 9 then the biggest element in the array would be 8) the numbers start with 0
and I would like to do some algorithm to generate some sort of a checksum for the array so that I can compare that 2 arrays are equal without looping through the whole array and checking each element one by one.
Where can I get this sort of information? I need something that is as simple as possible. Thank you.
edit: just to be clear on what I want:
-All the numbers in the array are distinct, so [0,1,1,2] is not valid because there is a repeated element (1)
-The position of the numbers matter, so [0,1,2,3] is not the same as [3,2,1,0]
-The array will contain the number 0, so this should also be taken into consideration.
EDIT:
Okay I tried to implement the Fletcher's algorithm here:
http://en.wikipedia.org/wiki/Fletcher%27s_checksum#Straightforward
int fletcher(int array[], int size){
int i;
int sum1=0;
int sum2=0;
for(i=0;i<size;i++){
sum1=(sum1+array[i])%255;
sum2=(sum2+sum1)%255;
}
return (sum2 << 8) | sum1;
}
to be honest I have no idea what does the return line do but unfortunately, the algorithm does not work.
For arrays [2,1,3,0] and [1,3,2,0] I get the same checksum.
EDIT2:
okay here's another one, the Adler checksum
http://en.wikipedia.org/wiki/Adler-32#Example_implementation
#define MOD 65521;
unsigned long adler(int array[], int size){
int i;
unsigned long a=1;
unsigned long b=0;
for(i=0;i<size;i++){
a=(a+array[i])%MOD;
b=(b+a)%MOD;
}
return (b <<16) | a;
}
This also does not work.
Arrays [2,0,3,1] and [1,3,0,2] generate same checksum.
I'm losing hope here, any ideas?

Let's take the case of your array of 25 integers. You explain that it can contains any permutations of the unique integers 0 to 24. According to this page, there is 25! (25 factorial) possible permutations, that is 15511210043330985984000000. Far more than a 32bit integer can contains.
The conclusion is that you will have collision, no matter how hard you try.
Now, here is a simple algorithm that account for position:
int checksum(int[] array, int size) {
int c = 0;
for(int i = 0; i < size; i++) {
c += array[i];
c = c << 3 | c >> (32 - 3); // rotate a little
c ^= 0xFFFFFFFF; // invert just for fun
}
return c;
}

I think what you want is in the answer of the following thread:
Fast permutation -> number -> permutation mapping algorithms
You just take the number your permutation is mapped to and take that as your Checksum. As there is exactly one Checksum per permutation there can't be a smaller Checksum that is collision free.

How about the checksum of weighted sum? Let's take an example for [0,1,2,3]. First pick a seed and limit, let's pick a seed as 7 and limit as 10000007.
a[4] = {0, 1, 2, 3}
limit = 10000007, seed = 7
result = 0
result = ((result + a[0]) * seed) % limit = ((0 + 0) * 7)) % 10000007 = 0
result = ((result + a[1]) * seed) % limit = ((0 + 1) * 7)) % 10000007 = 7
result = ((result + a[2]) * seed) % limit = ((7 + 2) * 7)) % 10000007 = 63
result = ((result + a[3]) * seed) % limit = ((63 + 3) * 7)) % 10000007 = 462
Your checksum is 462 for that [0, 1, 2, 3].
The reference is http://www.codeabbey.com/index/wiki/checksum

For an array of N unique integers from 1 to N, just adding up the elements will always be N*(N+1)/2. Therefore the only difference is in the ordering. If by "checksum" you imply that you tolerate some collisions, then one way is to sum the differences between consecutive numbers. So for example, the delta checksum for {1,2,3,4} is 1+1+1=3, but the delta checksum for {4,3,2,1} is -1+-1+-1=-3.
No requirements were given for collision rates or computational complexity, but if the above doesn't suit, then I recommend a position dependent checksum

From what I understand your array contains a permutation of numbers from 0 to N-1. One check-sum which will be useful is the rank of the array in its lexicographic ordering. What does it means ? Given 0, 1, 2
You have the possible permutations
1: 0, 1, 2
2: 0, 2, 1
3: 1, 0, 2
4: 1, 2, 0
5: 2, 0, 1
6: 2, 1, 0
The check-sum will be the first number, and computed when you create the array. There are solutions proposed in
Find the index of a given permutation in the list of permutations in lexicographic order
which can be helpful, although it seems the best algorithm was of quadratic complexity. To improve it to linear complexity you should cache the values of the factorials before hand.
The advantage? ZERO collision.
EDIT: Computation
The value is like the evaluation of a polynomial where factorial is used for the monomial instead of power. So the function is
f(x0,....,xn-1) = X0 * (0!) + X1 * (1!) + X2 * (2!) +...+ Xn-1 * (n-1!)
The idea is to use each values to get a sub-range of permutations, and with enough values you pinpoint an unique permutation.
Now for the implementation (like the one of a polynomial):
pre compute 0!.... to n-1! at the beginning of the program
Each time you set an array you use f(elements) to compute its checksum
you compare in O(1) using this checksum

Related

What is the formula for the range of a random float and int in C? [duplicate]

This is a follow on from a previously posted question:
How to generate a random number in C?
I wish to be able to generate a random number from within a particular range, such as 1 to 6 to mimic the sides of a die.
How would I go about doing this?
All the answers so far are mathematically wrong. Returning rand() % N does not uniformly give a number in the range [0, N) unless N divides the length of the interval into which rand() returns (i.e. is a power of 2). Furthermore, one has no idea whether the moduli of rand() are independent: it's possible that they go 0, 1, 2, ..., which is uniform but not very random. The only assumption it seems reasonable to make is that rand() puts out a Poisson distribution: any two nonoverlapping subintervals of the same size are equally likely and independent. For a finite set of values, this implies a uniform distribution and also ensures that the values of rand() are nicely scattered.
This means that the only correct way of changing the range of rand() is to divide it into boxes; for example, if RAND_MAX == 11 and you want a range of 1..6, you should assign {0,1} to 1, {2,3} to 2, and so on. These are disjoint, equally-sized intervals and thus are uniformly and independently distributed.
The suggestion to use floating-point division is mathematically plausible but suffers from rounding issues in principle. Perhaps double is high-enough precision to make it work; perhaps not. I don't know and I don't want to have to figure it out; in any case, the answer is system-dependent.
The correct way is to use integer arithmetic. That is, you want something like the following:
#include <stdlib.h> // For random(), RAND_MAX
// Assumes 0 <= max <= RAND_MAX
// Returns in the closed interval [0, max]
long random_at_most(long max) {
unsigned long
// max <= RAND_MAX < ULONG_MAX, so this is okay.
num_bins = (unsigned long) max + 1,
num_rand = (unsigned long) RAND_MAX + 1,
bin_size = num_rand / num_bins,
defect = num_rand % num_bins;
long x;
do {
x = random();
}
// This is carefully written not to overflow
while (num_rand - defect <= (unsigned long)x);
// Truncated division is intentional
return x/bin_size;
}
The loop is necessary to get a perfectly uniform distribution. For example, if you are given random numbers from 0 to 2 and you want only ones from 0 to 1, you just keep pulling until you don't get a 2; it's not hard to check that this gives 0 or 1 with equal probability. This method is also described in the link that nos gave in their answer, though coded differently. I'm using random() rather than rand() as it has a better distribution (as noted by the man page for rand()).
If you want to get random values outside the default range [0, RAND_MAX], then you have to do something tricky. Perhaps the most expedient is to define a function random_extended() that pulls n bits (using random_at_most()) and returns in [0, 2**n), and then apply random_at_most() with random_extended() in place of random() (and 2**n - 1 in place of RAND_MAX) to pull a random value less than 2**n, assuming you have a numerical type that can hold such a value. Finally, of course, you can get values in [min, max] using min + random_at_most(max - min), including negative values.
Following on from #Ryan Reich's answer, I thought I'd offer my cleaned up version. The first bounds check isn't required given the second bounds check, and I've made it iterative rather than recursive. It returns values in the range [min, max], where max >= min and 1+max-min < RAND_MAX.
unsigned int rand_interval(unsigned int min, unsigned int max)
{
int r;
const unsigned int range = 1 + max - min;
const unsigned int buckets = RAND_MAX / range;
const unsigned int limit = buckets * range;
/* Create equal size buckets all in a row, then fire randomly towards
* the buckets until you land in one of them. All buckets are equally
* likely. If you land off the end of the line of buckets, try again. */
do
{
r = rand();
} while (r >= limit);
return min + (r / buckets);
}
Here is a formula if you know the max and min values of a range, and you want to generate numbers inclusive in between the range:
r = (rand() % (max + 1 - min)) + min
unsigned int
randr(unsigned int min, unsigned int max)
{
double scaled = (double)rand()/RAND_MAX;
return (max - min +1)*scaled + min;
}
See here for other options.
Wouldn't you just do:
srand(time(NULL));
int r = ( rand() % 6 ) + 1;
% is the modulus operator. Essentially it will just divide by 6 and return the remainder... from 0 - 5
For those who understand the bias problem but can't stand the unpredictable run-time of rejection-based methods, this series produces a progressively less biased random integer in the [0, n-1] interval:
r = n / 2;
r = (rand() * n + r) / (RAND_MAX + 1);
r = (rand() * n + r) / (RAND_MAX + 1);
r = (rand() * n + r) / (RAND_MAX + 1);
...
It does so by synthesising a high-precision fixed-point random number of i * log_2(RAND_MAX + 1) bits (where i is the number of iterations) and performing a long multiplication by n.
When the number of bits is sufficiently large compared to n, the bias becomes immeasurably small.
It does not matter if RAND_MAX + 1 is less than n (as in this question), or if it is not a power of two, but care must be taken to avoid integer overflow if RAND_MAX * n is large.
Here is a slight simpler algorithm than Ryan Reich's solution:
/// Begin and end are *inclusive*; => [begin, end]
uint32_t getRandInterval(uint32_t begin, uint32_t end) {
uint32_t range = (end - begin) + 1;
uint32_t limit = ((uint64_t)RAND_MAX + 1) - (((uint64_t)RAND_MAX + 1) % range);
/* Imagine range-sized buckets all in a row, then fire randomly towards
* the buckets until you land in one of them. All buckets are equally
* likely. If you land off the end of the line of buckets, try again. */
uint32_t randVal = rand();
while (randVal >= limit) randVal = rand();
/// Return the position you hit in the bucket + begin as random number
return (randVal % range) + begin;
}
Example (RAND_MAX := 16, begin := 2, end := 7)
=> range := 6 (1 + end - begin)
=> limit := 12 (RAND_MAX + 1) - ((RAND_MAX + 1) % range)
The limit is always a multiple of the range,
so we can split it into range-sized buckets:
Possible-rand-output: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Buckets: [0, 1, 2, 3, 4, 5][0, 1, 2, 3, 4, 5][X, X, X, X, X]
Buckets + begin: [2, 3, 4, 5, 6, 7][2, 3, 4, 5, 6, 7][X, X, X, X, X]
1st call to rand() => 13
→ 13 is not in the bucket-range anymore (>= limit), while-condition is true
→ retry...
2nd call to rand() => 7
→ 7 is in the bucket-range (< limit), while-condition is false
→ Get the corresponding bucket-value 1 (randVal % range) and add begin
=> 3
In order to avoid the modulo bias (suggested in other answers) you can always use:
arc4random_uniform(MAX-MIN)+MIN
Where "MAX" is the upper bound and "MIN" is lower bound. For example, for numbers between 10 and 20:
arc4random_uniform(20-10)+10
arc4random_uniform(10)+10
Simple solution and better than using "rand() % N".
While Ryan is correct, the solution can be much simpler based on what is known about the source of the randomness. To re-state the problem:
There is a source of randomness, outputting integer numbers in range [0, MAX) with uniform distribution.
The goal is to produce uniformly distributed random integer numbers in range [rmin, rmax] where 0 <= rmin < rmax < MAX.
In my experience, if the number of bins (or "boxes") is significantly smaller than the range of the original numbers, and the original source is cryptographically strong - there is no need to go through all that rigamarole, and simple modulo division would suffice (like output = rnd.next() % (rmax+1), if rmin == 0), and produce random numbers that are distributed uniformly "enough", and without any loss of speed. The key factor is the randomness source (i.e., kids, don't try this at home with rand()).
Here's an example/proof of how it works in practice. I wanted to generate random numbers from 1 to 22, having a cryptographically strong source that produced random bytes (based on Intel RDRAND). The results are:
Rnd distribution test (22 boxes, numbers of entries in each box):
1: 409443 4.55%
2: 408736 4.54%
3: 408557 4.54%
4: 409125 4.55%
5: 408812 4.54%
6: 409418 4.55%
7: 408365 4.54%
8: 407992 4.53%
9: 409262 4.55%
10: 408112 4.53%
11: 409995 4.56%
12: 409810 4.55%
13: 409638 4.55%
14: 408905 4.54%
15: 408484 4.54%
16: 408211 4.54%
17: 409773 4.55%
18: 409597 4.55%
19: 409727 4.55%
20: 409062 4.55%
21: 409634 4.55%
22: 409342 4.55%
total: 100.00%
This is as close to uniform as I need for my purpose (fair dice throw, generating cryptographically strong codebooks for WWII cipher machines such as http://users.telenet.be/d.rijmenants/en/kl-7sim.htm, etc). The output does not show any appreciable bias.
Here's the source of cryptographically strong (true) random number generator:
Intel Digital Random Number Generator
and a sample code that produces 64-bit (unsigned) random numbers.
int rdrand64_step(unsigned long long int *therand)
{
unsigned long long int foo;
int cf_error_status;
asm("rdrand %%rax; \
mov $1,%%edx; \
cmovae %%rax,%%rdx; \
mov %%edx,%1; \
mov %%rax, %0;":"=r"(foo),"=r"(cf_error_status)::"%rax","%rdx");
*therand = foo;
return cf_error_status;
}
I compiled it on Mac OS X with clang-6.0.1 (straight), and with gcc-4.8.3 using "-Wa,q" flag (because GAS does not support these new instructions).
As said before modulo isn't sufficient because it skews the distribution. Heres my code which masks off bits and uses them to ensure the distribution isn't skewed.
static uint32_t randomInRange(uint32_t a,uint32_t b) {
uint32_t v;
uint32_t range;
uint32_t upper;
uint32_t lower;
uint32_t mask;
if(a == b) {
return a;
}
if(a > b) {
upper = a;
lower = b;
} else {
upper = b;
lower = a;
}
range = upper - lower;
mask = 0;
//XXX calculate range with log and mask? nah, too lazy :).
while(1) {
if(mask >= range) {
break;
}
mask = (mask << 1) | 1;
}
while(1) {
v = rand() & mask;
if(v <= range) {
return lower + v;
}
}
}
The following simple code lets you look at the distribution:
int main() {
unsigned long long int i;
unsigned int n = 10;
unsigned int numbers[n];
for (i = 0; i < n; i++) {
numbers[i] = 0;
}
for (i = 0 ; i < 10000000 ; i++){
uint32_t rand = random_in_range(0,n - 1);
if(rand >= n){
printf("bug: rand out of range %u\n",(unsigned int)rand);
return 1;
}
numbers[rand] += 1;
}
for(i = 0; i < n; i++) {
printf("%u: %u\n",i,numbers[i]);
}
}
Will return a floating point number in the range [0,1]:
#define rand01() (((double)random())/((double)(RAND_MAX)))

Given an array A[] of N numbers. Now, you need to find and print the Summation of the bitwise OR of all possible subsets of this array

For [1, 2, 3], all possible subsets are {1}, {2}, {3}, {1,2}, {1,3}, {2,3}, {1,2,3}
The sum of OR of these subsets are, 1 + 2 + 3 + 3 + 3 + 3 + 3 = 18.
My Approach is to generate all possible subset and find their OR and sum it but time complexity is O(2^n) , but I need a solution with O(nlogn) or less.
As you having 3 alements so 2^3=8 subsets will be created and you need to or all subset and print the sum of all subsets, By following logic you can get the solution you required
public class AndOfSubSetsOfSet {
public static void main(String[] args) {
findSubsets(new int[]{1, 2,3});
}
private static void findSubsets(int array[]) {
int numOfSubsets = 1 << array.length;
int a = 0;
for (int i = 0; i < numOfSubsets; i++) {
int pos = array.length - 1;
int bitmask = i;
int temp = 0;
int count = 0;
while (bitmask > 0) {
if ((bitmask & 1) == 1) {
if (count == 0) {
temp = array[pos];
} else
temp = array[pos] | temp;
count++;
}
//this will shift this number to left so one bit will be remove
bitmask >>= 1;
pos--;
}
count = 0;
a += temp;
temp = 0;
}
System.out.println(a);
}
}
`
one best approach you can use 3 loops outer loop would select number of elements of pair we have to make 2,3,4....upto n. and inner two loops would select elements according to outer loop. in the inner loop you can use bitwise OR so get the answer.
here time complexicity is better than exponential.
if any problem i would gave you code .
please vote if like.
Let's find the solution by calculating bitwise values. Consider the following points first. We will formulate the algorithm based on these points
For N numbers, there can be 2^N-1 such subsets.
For N numbers, where the maximum number of bits can be k, what can be the maximum output? Obviously when every subset sum is all 1's (i.e., for every combination there will be 1 in every bit of k positions). So calculate this MAX. In your example k = 2 and N = 3. So the MAX is when all the subset sum will be 11 (i.e.,3). SO MAX = (2^N-1)*(2^k-1) = 21.
Note that, the value of a bit of subset sum will only be 0 when the bits of every element of that subset is 0. So For every bit first calculate how many subsets can have 0 value in that bit. Then multiply that number with the corresponding value (2^bit_position) and deduct from MAX. In your case, for the leftmost position (i.e., position 0), there is only one 0 (in 2). So in 2^1-1 = 1 subset, the subsets sum's 0 position will be 0. So deduct 1*1 from MAX. Similarly for position 1, there can be only 1 subset with 0 at position 1 of subset sum ({2}). so deduct 1*2 from MAX. For every bit, calculate this value and keep deducting. the final MAX will be the result. If you consider 16 bit integer and you don't know about max k, then calculate using k = 16.
Let's consider another example with N = {1,4}. The subsets are {1},{4},{1,4}, and the result is = 1+4+5 = 10
here k = 3, N = 2. SO MAX = (2^K-1)*(2^N-1) = 21.
For 0 bit, there is only single 0 (in 4). so deduct 1*1 from MAX. So new MAX = 21 -1 = 20.
For 1 bit, both 1 and 4 has 0. so deduct (2^2-1)*2 from MAX. So new MAX = 20 -6 = 14.
For 2 bit, there is only single 0 (in 1). so deduct 1*4 from MAX. So new MAX = 14 -4 = 10.
As we have calculated for every bit position, thus the final result is 10.
Time Complexity
First and second steps can be calculated in constant time
In third step, the main thing is to find the number of 0 bit of each position. So for N number it takes O(k*N) in total. as k will be constant so the overall complexity will be O(N).

How to count the number of 1 bits set in 0, 1, 2, ..., n in O(n) time?

This was an interview question. The original questions asks:
Given a positive integer N, count the number of 1's in each integer from 0 to N, and return the count in an array of size N+1. Do it in O(n) time.
An Example would be:
Given 7, then return [0, 1, 1, 2, 1, 2, 2, 3]
Of course, the easiest way is to create a loop counting 1's for every integer, but that would be O(kn) time, with k as the size of the integers in bits. So either there's a way to count the number of 1's of an integer in O(1) time, or there's a way to directly generate the count going from 0 to N. I'm sure both methods exist, but can't figure out either.
Here's a nice little observation that you can use to do this in time O(n). Imagine you want to know how many 1 bits are set in the number k, and that you already know how many 1 bits are set in the numbers 0, 1, 2, ..., k - 1. If you can find a way to clear any of the 1 bits that are set in the number k, you'd get some smaller number (let's call it m), and the number of bits set in k would then be equal to one plus the number of bits set in m. So provided that we can find a way to clear any 1 bit from the number k, we can use this pattern to solve the problem:
result[0] = 0 // No bits set in 0
for k = 1 to n:
let m = k, with some 1 bit cleared
result[k] = result[m] + 1
There's a famous bit twiddling trick where
k & (k - 1)
yields the number formed by clearing the lowest 1 bit that's set in the number k, and does so in time O(1), assuming that the machine can do bitwise operations in constant time (which is usually a reasonable assumption). That means that this pseudocode should do it:
result[0] = 0
for k = 1 to n:
result[k] = result[k & (k - 1)] + 1
This does O(1) work per number O(n) total times, so the total work done is O(n).
Here's a different way to do this. Imagine, for example, that you know the counts of the bits in the numbers 0, 1, 2, and 3. You can use this to generate the counts of the bits of the numbers 4, 5, 6, and 7 by noticing that those numbers have bitwise representations which are formed by taking the bitwise representations of 0, 1, 2, and 3 and then prepending a 1. Similarly, if you then knew the bit counts of 0, 1, 2, 3, 4, 5, 6, and 7, you could generate the bit counts of 8, 9, 10, 11, 12, 13, 14, and 15 by noting that they too are formed by prepending a 1 bit to each of the lower numbers. That gives rise to this pseudocode, which for simplicity assumes that n is of the form 2k - 1 but could easily be adapted for a general n:
result[0] = 0;
for (int powerOfTwo = 1; powerOfTwo < n; powerOfTwo *= 2) {
for (int i = 0; i < powerOfTwo; i++) {
result[powerOfTwo + i] = result[i] + 1;
}
}
This also runs in time O(n). To see this, notice that across all iterations of all the loops here, each entry in the array is written to exactly once, with O(1) work done to determine which value is supposed to be put into the array at that slot.
Much simpler answer:
Let x be an array of length k+1 and x[i] has the number of set bits in i.
x[0] = 0
for i=1:k
x[i] = x[i>>1] + i&1
We start by manually evaluating the bit count for some small numbers. We note a recurrence relation between the bit count of n and previous results:
n: | count(n): | recurrence:
==============================
0 | 0 |
1 | 1 |
------------------------------
10 | 1 | = count(0) + 1
11 | 2 | = count(1) + 1
------------------------------
100 | 1 | = count(0) + 1
101 | 2 | = count(1) + 1
110 | 2 | = count(10) + 1
111 | 3 | = count(11) + 1
...
Given all bit counts up to 2 = 2¹, we can compute the bit counts up to 4 = 2² by adding 1. Given the bit counts up to 4 = 2², we can compute the bit counts up to 8 = 2³ by adding 1. We generalize this to the k-th power of two and can come up with following exemplary implementation:
// Counts and returns number of enabled bits for integers 0 - n:
function count_bits(n) {
let count = new Array(n);
count[0] = 0;
for (let i = 1, j = 0, k = 1; i <= n; ++i, ++j) {
if (i == 2**k) { // k-th bit of i has been enabled
k++;
j = 0;
}
count[i] = count[j] + 1;
}
return count;
}
// Example:
console.log(count_bits(17).join());
We note that all operations involved are incrementing by one, random array access, copying array elements and checking the k-th bit of loop increment i via i == 2**k which could be rewritten as i & 1 << k or - for arbitrary precision of i - as a random array access.
Assuming that all primitive operations listed above are in O(1) on our machine, the total runtime complexity is in O(n).
This would equally hold for arbitrary precision integers - where incrementing has an average runtime complexity of O(1) - unless copying count[i] = count[j] + 1 takes more than constant time. As this is unfortunately the case for arbitrarily large integers, our runtime complexity is in O(n log(log(n))) as we need O(log(log(n))) space to store the bit count of n.

Random Number within a range but exclusive (...] [duplicate]

This is a follow on from a previously posted question:
How to generate a random number in C?
I wish to be able to generate a random number from within a particular range, such as 1 to 6 to mimic the sides of a die.
How would I go about doing this?
All the answers so far are mathematically wrong. Returning rand() % N does not uniformly give a number in the range [0, N) unless N divides the length of the interval into which rand() returns (i.e. is a power of 2). Furthermore, one has no idea whether the moduli of rand() are independent: it's possible that they go 0, 1, 2, ..., which is uniform but not very random. The only assumption it seems reasonable to make is that rand() puts out a Poisson distribution: any two nonoverlapping subintervals of the same size are equally likely and independent. For a finite set of values, this implies a uniform distribution and also ensures that the values of rand() are nicely scattered.
This means that the only correct way of changing the range of rand() is to divide it into boxes; for example, if RAND_MAX == 11 and you want a range of 1..6, you should assign {0,1} to 1, {2,3} to 2, and so on. These are disjoint, equally-sized intervals and thus are uniformly and independently distributed.
The suggestion to use floating-point division is mathematically plausible but suffers from rounding issues in principle. Perhaps double is high-enough precision to make it work; perhaps not. I don't know and I don't want to have to figure it out; in any case, the answer is system-dependent.
The correct way is to use integer arithmetic. That is, you want something like the following:
#include <stdlib.h> // For random(), RAND_MAX
// Assumes 0 <= max <= RAND_MAX
// Returns in the closed interval [0, max]
long random_at_most(long max) {
unsigned long
// max <= RAND_MAX < ULONG_MAX, so this is okay.
num_bins = (unsigned long) max + 1,
num_rand = (unsigned long) RAND_MAX + 1,
bin_size = num_rand / num_bins,
defect = num_rand % num_bins;
long x;
do {
x = random();
}
// This is carefully written not to overflow
while (num_rand - defect <= (unsigned long)x);
// Truncated division is intentional
return x/bin_size;
}
The loop is necessary to get a perfectly uniform distribution. For example, if you are given random numbers from 0 to 2 and you want only ones from 0 to 1, you just keep pulling until you don't get a 2; it's not hard to check that this gives 0 or 1 with equal probability. This method is also described in the link that nos gave in their answer, though coded differently. I'm using random() rather than rand() as it has a better distribution (as noted by the man page for rand()).
If you want to get random values outside the default range [0, RAND_MAX], then you have to do something tricky. Perhaps the most expedient is to define a function random_extended() that pulls n bits (using random_at_most()) and returns in [0, 2**n), and then apply random_at_most() with random_extended() in place of random() (and 2**n - 1 in place of RAND_MAX) to pull a random value less than 2**n, assuming you have a numerical type that can hold such a value. Finally, of course, you can get values in [min, max] using min + random_at_most(max - min), including negative values.
Following on from #Ryan Reich's answer, I thought I'd offer my cleaned up version. The first bounds check isn't required given the second bounds check, and I've made it iterative rather than recursive. It returns values in the range [min, max], where max >= min and 1+max-min < RAND_MAX.
unsigned int rand_interval(unsigned int min, unsigned int max)
{
int r;
const unsigned int range = 1 + max - min;
const unsigned int buckets = RAND_MAX / range;
const unsigned int limit = buckets * range;
/* Create equal size buckets all in a row, then fire randomly towards
* the buckets until you land in one of them. All buckets are equally
* likely. If you land off the end of the line of buckets, try again. */
do
{
r = rand();
} while (r >= limit);
return min + (r / buckets);
}
Here is a formula if you know the max and min values of a range, and you want to generate numbers inclusive in between the range:
r = (rand() % (max + 1 - min)) + min
unsigned int
randr(unsigned int min, unsigned int max)
{
double scaled = (double)rand()/RAND_MAX;
return (max - min +1)*scaled + min;
}
See here for other options.
Wouldn't you just do:
srand(time(NULL));
int r = ( rand() % 6 ) + 1;
% is the modulus operator. Essentially it will just divide by 6 and return the remainder... from 0 - 5
For those who understand the bias problem but can't stand the unpredictable run-time of rejection-based methods, this series produces a progressively less biased random integer in the [0, n-1] interval:
r = n / 2;
r = (rand() * n + r) / (RAND_MAX + 1);
r = (rand() * n + r) / (RAND_MAX + 1);
r = (rand() * n + r) / (RAND_MAX + 1);
...
It does so by synthesising a high-precision fixed-point random number of i * log_2(RAND_MAX + 1) bits (where i is the number of iterations) and performing a long multiplication by n.
When the number of bits is sufficiently large compared to n, the bias becomes immeasurably small.
It does not matter if RAND_MAX + 1 is less than n (as in this question), or if it is not a power of two, but care must be taken to avoid integer overflow if RAND_MAX * n is large.
Here is a slight simpler algorithm than Ryan Reich's solution:
/// Begin and end are *inclusive*; => [begin, end]
uint32_t getRandInterval(uint32_t begin, uint32_t end) {
uint32_t range = (end - begin) + 1;
uint32_t limit = ((uint64_t)RAND_MAX + 1) - (((uint64_t)RAND_MAX + 1) % range);
/* Imagine range-sized buckets all in a row, then fire randomly towards
* the buckets until you land in one of them. All buckets are equally
* likely. If you land off the end of the line of buckets, try again. */
uint32_t randVal = rand();
while (randVal >= limit) randVal = rand();
/// Return the position you hit in the bucket + begin as random number
return (randVal % range) + begin;
}
Example (RAND_MAX := 16, begin := 2, end := 7)
=> range := 6 (1 + end - begin)
=> limit := 12 (RAND_MAX + 1) - ((RAND_MAX + 1) % range)
The limit is always a multiple of the range,
so we can split it into range-sized buckets:
Possible-rand-output: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Buckets: [0, 1, 2, 3, 4, 5][0, 1, 2, 3, 4, 5][X, X, X, X, X]
Buckets + begin: [2, 3, 4, 5, 6, 7][2, 3, 4, 5, 6, 7][X, X, X, X, X]
1st call to rand() => 13
→ 13 is not in the bucket-range anymore (>= limit), while-condition is true
→ retry...
2nd call to rand() => 7
→ 7 is in the bucket-range (< limit), while-condition is false
→ Get the corresponding bucket-value 1 (randVal % range) and add begin
=> 3
In order to avoid the modulo bias (suggested in other answers) you can always use:
arc4random_uniform(MAX-MIN)+MIN
Where "MAX" is the upper bound and "MIN" is lower bound. For example, for numbers between 10 and 20:
arc4random_uniform(20-10)+10
arc4random_uniform(10)+10
Simple solution and better than using "rand() % N".
While Ryan is correct, the solution can be much simpler based on what is known about the source of the randomness. To re-state the problem:
There is a source of randomness, outputting integer numbers in range [0, MAX) with uniform distribution.
The goal is to produce uniformly distributed random integer numbers in range [rmin, rmax] where 0 <= rmin < rmax < MAX.
In my experience, if the number of bins (or "boxes") is significantly smaller than the range of the original numbers, and the original source is cryptographically strong - there is no need to go through all that rigamarole, and simple modulo division would suffice (like output = rnd.next() % (rmax+1), if rmin == 0), and produce random numbers that are distributed uniformly "enough", and without any loss of speed. The key factor is the randomness source (i.e., kids, don't try this at home with rand()).
Here's an example/proof of how it works in practice. I wanted to generate random numbers from 1 to 22, having a cryptographically strong source that produced random bytes (based on Intel RDRAND). The results are:
Rnd distribution test (22 boxes, numbers of entries in each box):
1: 409443 4.55%
2: 408736 4.54%
3: 408557 4.54%
4: 409125 4.55%
5: 408812 4.54%
6: 409418 4.55%
7: 408365 4.54%
8: 407992 4.53%
9: 409262 4.55%
10: 408112 4.53%
11: 409995 4.56%
12: 409810 4.55%
13: 409638 4.55%
14: 408905 4.54%
15: 408484 4.54%
16: 408211 4.54%
17: 409773 4.55%
18: 409597 4.55%
19: 409727 4.55%
20: 409062 4.55%
21: 409634 4.55%
22: 409342 4.55%
total: 100.00%
This is as close to uniform as I need for my purpose (fair dice throw, generating cryptographically strong codebooks for WWII cipher machines such as http://users.telenet.be/d.rijmenants/en/kl-7sim.htm, etc). The output does not show any appreciable bias.
Here's the source of cryptographically strong (true) random number generator:
Intel Digital Random Number Generator
and a sample code that produces 64-bit (unsigned) random numbers.
int rdrand64_step(unsigned long long int *therand)
{
unsigned long long int foo;
int cf_error_status;
asm("rdrand %%rax; \
mov $1,%%edx; \
cmovae %%rax,%%rdx; \
mov %%edx,%1; \
mov %%rax, %0;":"=r"(foo),"=r"(cf_error_status)::"%rax","%rdx");
*therand = foo;
return cf_error_status;
}
I compiled it on Mac OS X with clang-6.0.1 (straight), and with gcc-4.8.3 using "-Wa,q" flag (because GAS does not support these new instructions).
As said before modulo isn't sufficient because it skews the distribution. Heres my code which masks off bits and uses them to ensure the distribution isn't skewed.
static uint32_t randomInRange(uint32_t a,uint32_t b) {
uint32_t v;
uint32_t range;
uint32_t upper;
uint32_t lower;
uint32_t mask;
if(a == b) {
return a;
}
if(a > b) {
upper = a;
lower = b;
} else {
upper = b;
lower = a;
}
range = upper - lower;
mask = 0;
//XXX calculate range with log and mask? nah, too lazy :).
while(1) {
if(mask >= range) {
break;
}
mask = (mask << 1) | 1;
}
while(1) {
v = rand() & mask;
if(v <= range) {
return lower + v;
}
}
}
The following simple code lets you look at the distribution:
int main() {
unsigned long long int i;
unsigned int n = 10;
unsigned int numbers[n];
for (i = 0; i < n; i++) {
numbers[i] = 0;
}
for (i = 0 ; i < 10000000 ; i++){
uint32_t rand = random_in_range(0,n - 1);
if(rand >= n){
printf("bug: rand out of range %u\n",(unsigned int)rand);
return 1;
}
numbers[rand] += 1;
}
for(i = 0; i < n; i++) {
printf("%u: %u\n",i,numbers[i]);
}
}
Will return a floating point number in the range [0,1]:
#define rand01() (((double)random())/((double)(RAND_MAX)))

speeding up "base conversion" for large integers

I am using a base-conversion algorithm to generate a permutation from a large integer (split into 32-bit words).
I use a relatively standard algorithm for this:
/* N = count,K is permutation index (0..N!-1) A[N] contains 0..N-1 */
i = 0;
while (N > 1) {
swap A[i] and A[i+(k%N)]
k = k / N
N = N - 1
i = i + 1
}
Unfortunately, the divide and modulo each iteration adds up, especially moving to large integers - But, it seems I could just use multiply!
/* As before, N is count, K is index, A[N] contains 0..N-1 */
/* Split is arbitrarily 128 (bits), for my current choice of N */
/* "Adjust" is precalculated: (1 << Split)/(N!) */
a = k*Adjust; /* a can be treated as a fixed point fraction */
i = 0;
while (N > 1) {
a = a*N;
index = a >> Split;
a = a & ((1 << Split) - 1); /* actually, just zeroing a register */
swap A[i] and A[i+index]
N = N - 1
i = i + 1
}
This is nicer, but doing large integer multiplies is still sluggish.
Question 1:
Is there a way of doing this faster?
Eg. Since I know that N*(N-1) is less than 2^32, could I pull out those numbers from one word, and merge in the 'leftovers'?
Or, is there a way to modify an arithetic decoder to pull out the indicies one at a time?
Question 2:
For the sake of curiosity - if I use multiplication to convert a number to base 10 without the adjustment, then the result is multiplied by (10^digits/2^shift). Is there a tricky way to remove this factor working with the decimal digits? Even with the adjustment factor, this seems like it would be faster -- why wouldn't standard libraries use this vs divide and mod?
Seeing that you are talking about numbers like 2^128/(N!), it seems that in your problem N is going to be rather small (N < 35 according to my calculations).
I suggest taking the original algorithm as a starting point; first switch the direction of the loop:
i = 2;
while (i < N) {
swap A[N - 1 - i] and A[N - i + k % i]
k = k / i
i = i + 1
}
Now change the loop to do several permutations per iteration. I guess the speed of division is the same regardless of the number i, as long as i < 2^32.
Split the range 2...N-1 into sub-ranges so that the product of the numbers in each sub-range is less than 2^32:
2, 3, 4, ..., 12: product is 479001600
13, 14, ..., 19: product is 253955520
20, 21, ..., 26: product is 3315312000
27, 28, ..., 32: product is 652458240
33, 34, 35: product is 39270
Then, divide the long number k by the products instead of dividing by i. Each iteration will yield a remainder (less than 2^32) and a smaller number k. When you have the remainder, you can work with it in an inner loop using the original algorithm; which will now be faster because it doesn't involve long division.
Here is some code:
static const int rangeCount = 5;
static const int rangeLimit[rangeCount] = {13, 20, 27, 33, 36};
static uint32_t rangeProduct[rangeCount] = {
479001600,
253955520,
3315312000,
652458240,
39270
};
for (int rangeIndex = 0; rangeIndex < rangeCount; ++rangeIndex)
{
// The following two lines involve long division;
// math libraries probably calculate both quotient and remainder
// in one function call
uint32_t rangeRemainder = k % rangeProduct[rangeIndex];
k /= rangeProduct[rangeIndex];
// A range starts where the previous range ended
int rangeStart = (rangeIndex == 0) ? 2 : rangeLimit[rangeIndex - 1];
// Iterate over range
for (int i = rangeStart; i < rangeLimit[rangeIndex] && i < n; ++i)
{
// The following two lines involve a 32-bit division;
// it produces both quotient and remainder in one Pentium instruction
int remainder = rangeRemainder % i;
rangeRemainder /= i;
std::swap(permutation[n - 1 - i], permutation[n - i + remainder]);
}
}
Of course, this code can be extended into more than 128 bits.
Another optimization could involve extraction of powers of 2 from the products of ranges; this might add a slight speedup by making the ranges longer. Not sure whether this is worthwhile (maybe for large values of N, like N=1000).
Dont know about algorithms, but the ones you use seems pretty simple, so i dont really see how you can optimize the algorithm.
You may use alternative approaches:
use ASM (assembler) - from my experience, after a long time trying to figure out how should a certain algorithm would be written in ASM, it ended up being slower than the version generated by the compiler:) Probably because the compiler also knows how to layout the code so the CPU cache would be more efficient, and/or what instructions are actually faster and what situations(this was on GCC/linux).
use multi-processing:
make your algorithm multithreaded, and make sure you run with the same number of threads as the number of available cpu cores(most cpu's nowdays do have multiple cores/multithreading)
make you algorithm capable of running on multiple machines on a network, and devise a way of sending these numbers to machines in a network, so you may use their CPU power.

Resources