Related
As I understand, it is related to the partition problem.
But I would like to ask a slightly different problem which I don't care about the sum but the average. In this case, it needs to optimize 2 constraints (sum and number of items) at the same time. It seems to be a harder problem and I cannot see any solutions online.
Are there any solutions for this variant? Or how does it relate to the partition problem?
Example:
input X = [1,1,1,1,1,6]
output based on sum: A = [1,1,1,1,1], B=[6]
output based on average: A = [1], B=[1,1,1,1,6]
On some inputs, a modification of the dynamic program for the usual partition problem will give a speedup. We have to classify each partial solution by its count and sum instead of just sum, which slows things down a bit. Python 3 below (note that the use of dictionaries implicitly collapses functionally identical partial solutions):
def children(ab, x):
a, b = ab
yield a + [x], b
yield a, b + [x]
def proper(ab):
a, b = ab
return a and b
def avg(lst):
return sum(lst) / len(lst)
def abs_diff_avg(ab):
a, b = ab
return abs(avg(a) - avg(b))
def min_abs_diff_avg(lst):
solutions = {(0, 0): ([], [])}
for x in lst:
solutions = {
(sum(a), len(a)): (a, b)
for ab in solutions.values()
for (a, b) in children(ab, x)
}
return min(filter(proper, solutions.values()), key=abs_diff_avg)
print(min_abs_diff_avg([1, 1, 1, 1, 1, 6]))
let S_i the sum of a subset of v of size i
let S be the total sum of v, n the length of v
the err to minimize is
err_i = |avg(S_i) - avg(S-S_i)|
err_i = |S_i/i - (S-S_i)/(n-i)|
err_i = |(nS_i - iS)/(i(n-i))|
algorithm below does:
for all tuple sizes (1,...,n/2) as i
- for all tuples of size i-1 as t_{i-1}
- generate all possible tuple of size i from t_{i-1} by adjoining one elem from v
- track best tuple in regard of err_i
The only cut I found being:
for two tuples of size i having the same sum, keep the one whose last element's index is the smallest
e.g given tuples A, B (where X is some taken element from v)
A: [X,....,X....]
B: [.,X,.....,X..]
keep A because its right-most element has the minimal index
(idea being that at size 3, A will offer the same candidates as B plus some more)
function generateTuples (v, tuples) {
const nextTuples = new Map()
for (const [, t] of tuples) {
for (let l = t.l + 1; l < v.length; ++l) {
const s = t.s + v[l]
if (!nextTuples.has(s) || nextTuples.get(s).l > l) {
const nextTuple = { v: t.v.concat(l), s, l }
nextTuples.set(s, nextTuple)
}
}
}
return nextTuples
}
function processV (v) {
const fErr = (() => {
const n = v.length
const S = v.reduce((s, x) => s + x, 0)
return ({ s: S_i, v }) => {
const i = v.length
return Math.abs((n * S_i - i * S) / (i * (n - i)))
}
})()
let tuples = new Map([[0, { v: [], s: 0, l: -1 }]])
let best = null
let err = 9e3
for (let i = 0; i < Math.ceil(v.length / 2); ++i) {
const nextTuples = generateTuples(v, tuples)
for (const [, t] of nextTuples) {
if (fErr(t) <= err) {
best = t
err = fErr(t)
}
}
tuples = nextTuples
}
const s1Indices = new Set(best.v)
return {
sol: v.reduce(([v1, v2], x, i) => {
(s1Indices.has(i) ? v1 : v2).push(x)
return [v1, v2]
}, [[], []]),
err
}
}
console.log('best: ', processV([1, 1, 1, 1, 1, 6]))
console.log('best: ', processV([1, 2, 3, 4, 5]))
console.log('best: ', processV([1, 3, 5, 7, 7, 8]))
I am using a Vec to store a 2D (row major) matrix of values. I would like to iterate over this matrix with a sliding 2D sub-window to apply a filter (which is unfortunately non-separable).
I have seen in the slice documentation that a windows function exists, which is what I want but in 2 dimensions.
I thought about implementing this as:
fn main() {
// 4 rows 3 columns
let dim: (usize, usize) = (4, 3);
// Place-holder matrix
#[rustfmt::skip]
let mat = vec![0, 1, 2,
3, 4, 5,
6, 7, 8,
9, 10, 11];
// 2D index to 1D index
let linearize = |r, c| r * dim.1 + c;
// The dimensions of my sub-window
let win_size: usize = 2;
// Calculate the bounds for which the top left corner of each window may exist
let bounds: (usize, usize) = (dim.0 - win_size + 1, dim.1 - win_size + 1);
// Convert window 1D index into a 2D index
let split = |i| (i / win_size, i % win_size);
// Iterate over all the top left corners
let window_2d = (0..bounds.0 * bounds.1).map(|i| {
// Get the 2D index of the top left corner
let (r, c) = (i / bounds.1, i % bounds.1);
// Borrow the matrix, so our closure may own the reference
let bmat = &mat;
// Return an iterator for this window
return (0..win_size * win_size).map(move |x| {
let (wr, wc) = split(x);
return bmat[linearize(wr + r, wc + c)];
});
});
// Print the windows out
window_2d.for_each(|it| {
print!("[ ");
it.for_each(|x| print!("{} ", x));
println!("]");
});
}
Essentially creating an iterator over a range of indices and then mapping to the square bracket operator of the matrix.
As far as I know, this is going to have the overhead of a bounds check for each deref of the iterator.
I'm wondering if there's an alternative which would elide the bounds checks? Maybe using a combination of chunks, windows and zip, to chunk the matrix into rows, each with a sliding window, then zip the row's windows and flatten the result?
Thanks!
Edit:
I'm not looking to simply iterate over a 2D array, I want to slide a 2D window over the array, similar to how the std::slice::windows function works.
The best I've got for now is wrapping the matrix access in an unsafe block to elide the bounds check.
With some other misc changes, this is the full example now:
fn split_factory(cols: usize) -> impl Fn(usize) -> (usize, usize) {
// Declaring that cols must be positive allows more aggressive optimisation of div and mod.
if cols < 1 {
unreachable!()
}
move |i| (i / cols, i % cols)
}
fn main() {
// 4 rows 3 columns
let dim: (usize, usize) = (4, 3);
// Place-holder matrix
#[rustfmt::skip]
let mat = vec![0, 1, 2,
3, 4, 5,
6, 7, 8,
9, 10, 11];
// The dimensions of my sub-window
let win_dim = (3usize, 2usize);
// Calculate the bounds for which the top left corner of each window may exist
let bounds = (dim.0 - win_dim.0 + 1, dim.1 - win_dim.1 + 1);
// Iterate over all the top left corners
let convolution_iter = (0..bounds.0 * bounds.1)
.map(split_factory(bounds.1))
.map(|(r, c)| {
// Borrow the matrix, so our closure may own the reference
let bmat = &mat;
// Return an iterator for this window
return (0..win_dim.0 * win_dim.1)
.map(split_factory(win_dim.1))
.map(move |(wr, wc)| {
let px = (wr + r) * dim.1 + (wc + c);
(px, unsafe { *bmat.get_unchecked(px) })
});
});
// Print the windows out (badly...)
convolution_iter.for_each(|it| println!("{:?}", it.collect::<Vec<(usize, i32)>>()));
}
Would still be nicer to avoid unsafe and the indirection of looking up the matrix with indices.
I want to talk about how to which proceed.
1. Incorrect usage of Scala. I should try to more improve the code.
2. The efficiency of the algorithm is poor. I should think of an efficient algorithm.
Goal:
Can quickly calculate the max number from more than 1,000 Partition Numbers collections.
Partition Number:
e.g.,
5 -> (5), (1, 4), (2, 3), (1, 1, 3), (1, 2, 2), (1, 1, 1, 2), (1, 1, 1, 1, 1)
I ask that "I want to convert from Python to Scala that Partition Function using Vector", and I was taught to use Stream yesterday.
I fixed code, I can use 10, 50, and so on. But using big numbers(e. g., 100, 1,000 or 10,000) weren't calculate max number.
It calculate from Stream.last to Stream.head.
In my understanding that Stream type can add an element at the head only, so the order of the numbers is reversed form the fist code.
code
import scala.math.floor
class PartitionNumbers(startNum: Int, point: Int) {
var maxNum = 0
var tmpNum = 0
private def appendOnes(n: Int, s: Stream[Int] = Stream.empty[Int]): Stream[Int] = {
if (n == 0) s
else appendOnes(n - 1, 1 #:: s)
}
private def partition(n: Int, k: Int, tmpStream: Stream[Int] = Stream.empty): Int = {
if (n == 0) tmpNum = calculate(tmpStream)
else if (n == 1 | k == 1) tmpNum = calculate(appendOnes(n))
else {
if (n >= k) partition(n - k, k, k #:: tmpStream)
partition(n, k - 1, tmpStream)
}
if (maxNum < tmpNum) maxNum = tmpNum
maxNum
}
def searchMax(n: Int = point): Int = {
partition(n, n)
}
def calculate(usePointsStream: Stream[Int], num: Int = startNum): Int = {
if (usePointsStream.isEmpty) {
num
} else {
calculate(usePointsStream.init, floor(num * (100 + usePointsStream.last) / 100).toInt)
}
}
}
output example
val pn_1 = new PartitionNumbers(100, 10)
println(pn_1.searchMax()) // -> 110
val pn_2 = new PartitionNumbers(1000, 50)
println(pn_2.searchMax()) // -> 1630
val pn_3 = new PartitionNumbers(10000, 100)
println(pn_3.searchMax()) // Can't calculate within 3 minutes using Ryzen 7 2700X.
It is not a homework problem. I am just curious about this problem. And my approach is simple brute-force :-)
My brute-force C++ code:
int main()
{
ll l,r;
cin>>l>>r;
ll f=0;
ll i=l;
while(i<=r)
{
ll j=0;
string s;
ll c=0;
s=to_string(i);
// cout<<s<<" ";
ll x=s.length();
if(x==1)
{
c=0;
}
else
{
j=0;
//whil
while(j<=x-2)
{
string b,g;
b="1";
g="1";
b=s[j];
g=s[j+1];
ll k1,k2;
k1=stoi(b);
k2=stoi(g);
if(__gcd(k1,k2)==1)
{
c=1;
break;
}
j++;
}
}
ll d=0;
j=0;
while(j<=x-1)
{
if( s[j]=='2' || s[j]=='3' || s[j]=='5' || s[j]=='7')
{
string b;
b="1";
b=s[j];
ll k1=stoi(b);
if(i%k1==0)
{
//d=0;
}
else
{
d=1;
break;
}
}
j++;
}
if(c==1 || d==1)
{
// cout<<"NO";
}
else
{
f++;
// cout<<"PR";
}
// cout<<"\n";
i++;
}
cout<<f;
return 0;
}
You are given 2 integers 'L' and 'R' . You are required to find the count of all the PR numbers in the range 'L' to 'R' inclusively. PR number are the numbers which satisfy following properties:
No pair of adjacent digits are co-prime i.e. adjacent digits in a PR number will not be co-prime to each other.
PR number is divisible by all the single digit prime numbers which occur as a digit in the PR number.
Note: Two numbers 'a' and 'b' are co-prime, if gcd(a,b)=1.
Also, gcd(0,a)=a;
Example:
Input: [2,5].
Output: '4'.
(Note: '1' is not a prime-number, though its very common)
(All the integers: '2','3','4','5') satisfy the condition of PR numbers :-)
Constraints on 'L','R': 1 <= L, R <= 10^18
What can be the the most efficient algorithm to solve this ?
Note: This will solve only part 1 which is No pair of adjacent digits are co-prime i.e. adjacent digits in a PR number will not be co-prime to each other.
Here is a constructive approach in python: instead of going throught all numbers in range and filtering by conditions, we will just construct all numbers that satisfy the condition. Note that if we have a valid sequence of digits, for it to continue being valid only the rightmost digit matters in order to decide what the next digit will be.
def ways(max_number, prev_digit, current_number):
if current_number > max_number:
return 0
count = 1
if prev_digit == 0:
if current_number != 0:
count += ways(max_number, 0, current_number * 10)
for i in range(2, 10):
count += ways(max_number, i, current_number * 10 + i)
if prev_digit == 2 or prev_digit == 4 or prev_digit == 8:
for i in [0, 2, 4, 6, 8]:
count += ways(max_number, i, current_number * 10 + i)
if prev_digit == 3 or prev_digit == 9:
for i in [0, 3, 6, 9]:
count += ways(max_number, i, current_number * 10 + i)
if prev_digit == 5 or prev_digit == 7:
count += ways(max_number, 0, current_number * 10)
count += ways(max_number, prev_digit, current_number * 10 + prev_digit)
if prev_digit == 6:
for i in [0, 2, 3, 4, 6, 8, 9]:
count += ways(max_number, i, current_number * 10 + i)
return count
As we are generating all valid numbers up to max_number without any repeats, the complexity of this function is O(amount of numbers between 0 and max_number that satisfy condition 1). To calculate the range a to b, we just need to do ways(b) - ways(a - 1).
Takes less than 1 second to caculate these numbers from 0 to 1 million, as there are only 42935 numbers that satisfy the result. As there are few numbers that satisfy the condition, we can then check if they are multiple of its prime digits to satisfy also condition 2. I leave this part up to the reader as there are multiple ways to do it.
TL;DR: This is more commonly called "digit dynamic programming with bitmask"
In more competitive-programming-familiar terms, you'd compute dp[n_digit][mod_2357][is_less_than_r][digit_appeared][last_digit] = number of numbers with n_digit digits (including leading zeroes), less than the number formed by first n_digit digits of R and with the other properties match. Do it twice with R and L-1 then take the difference. The number of operations required would be about 19 (number of digits) * 210 (mod) * 2 * 24 (it's only necessary to check for appearance of single-digit primes) * 10 * 10, which is obviously manageable by today computers.
Think about how you'd check whether a number is valid.
Not the normal way. Using a finite state automaton that take the input from left to right, digit by digit.
For simplicity, assume the input has a fixed number of digits (so that comparison with L/R is easier. This is possible because the number has at most as many digits as R).
It's necessary for each state to keep track of:
which digit appeared in the number (use a bit mask, there are 4 1-digit primes)
is the number in range [L..R] (either this is guaranteed to be true/false by the prefix, otherwise the prefix matches with that of L/R)
what is the value of the prefix mod each single digit prime
the most recent digit (to check whether all pairs of consecutive digits are coprime)
After the finite state automaton is constructed, the rest is simple. Just use dynamic programming to count the number of path to any accepted state from the starting state.
Remark: This method can be used to count the number of any type of object that can be verified using a finite state automaton (roughly speaking, you can check whether the property is satisfied using a program with constant memory usage, and takes the object piece-by-piece in some order)
We need a table where we can look up the count of suffixes that would match a prefix to construct valid numbers. Given a prefix's
right digit
prime combination
mod combination
and a suffix length, we'd like the count of suffixes that have searchable:
left digit
length
prime combination
mod combination
I started coding in Python, then switched to JavaScript to be able to offer a snippet. Comments in the code describe each lookup table. There are a few of them to allow for faster enumeration. There are samples of prefix-suffix calculations to illustrate how one can build an arbitrary upper-bound using the table, although at least some, maybe all of the prefix construction and aggregation could be made during the tabulation.
function gcd(a,b){
if (!b)
return a
else
return gcd(b, a % b)
}
// (Started writing in Python,
// then switched to JavaScript...
// 'xrange(4)' -> [0, 1, 2, 3]
// 'xrange(2, 4)' -> [2, 3]
function xrange(){
let l = 0
let r = arguments[1] || arguments[0]
if (arguments.length > 1)
l = arguments[0]
return new Array(r - l).fill(0).map((_, i) => i + l)
}
// A lookup table and its reverse,
// mapping each of the 210 mod combinations,
// [n % 2, n % 3, n % 5, n % 7], to a key
// from 0 to 209.
// 'mod_combs[0]' -> [0, 0, 0, 0]
// 'mod_combs[209]' -> [1, 2, 4, 6]
// 'mod_keys[[0,0,0,0]]' -> 0
// 'mod_keys[[1,2,4,6]]' -> 209
let mod_combs = {}
let mod_keys = {}
let mod_key_count = 0
for (let m2 of xrange(2)){
for (let m3 of xrange(3)){
for (let m5 of xrange(5)){
for (let m7 of xrange(7)){
mod_keys[[m2, m3, m5, m7]] = mod_key_count
mod_combs[mod_key_count] = [m2, m3, m5, m7]
mod_key_count += 1
}
}
}
}
// The main lookup table built using the
// dynamic program
// [mod_key 210][l_digit 10][suffix length 20][prime_comb 16]
let table = new Array(210)
for (let mk of xrange(210)){
table[mk] = new Array(10)
for (let l_digit of xrange(10)){
table[mk][l_digit] = new Array(20)
for (let sl of xrange(20)){
table[mk][l_digit][sl] = new Array(16).fill(0)
}
}
}
// We build prime combinations from 0 (no primes) to
// 15 (all four primes), using a bitmask of up to four bits.
let prime_set = [0, 0, 1<<0, 1<<1, 0, 1<<2, 0, 1<<3, 0, 0]
// The possible digits that could
// follow a digit
function get_valid_digits(digit){
if (digit == 0)
return [0, 2, 3, 4, 5, 6, 7, 8, 9]
else if ([2, 4, 8].includes(digit))
return [0, 2, 4, 6, 8]
else if ([3, 9].includes(digit))
return [0, 3, 6, 9]
else if (digit == 6)
return [0, 2, 3, 4, 6, 8, 9]
else if (digit == 5)
return [0, 5]
else if (digit == 7)
return [0, 7]
}
// Build the table bottom-up
// Single digits
for (let i of xrange(10)){
let mod_key = mod_keys[[i % 2, i % 3, i % 5, i % 7]]
let length = 1
let l_digit = i
let prime_comb = prime_set[i]
table[mod_key][l_digit][length][prime_comb] = 1
}
// Everything else
// For demonstration, we just table up to 6 digits
// since either JavaScript, this program, or both seem
// to be too slow for a full demo.
for (let length of xrange(2, 6)){
// We're appending a new left digit
for (let new_l_digit of xrange(0, 10)){
// The digit 1 is never valid
if (new_l_digit == 1)
continue
// The possible digits that could
// be to the right of our new left digit
let ds = get_valid_digits(new_l_digit)
// For each possible digit to the right
// of our new left digit, iterate over all
// the combinations of primes and remainder combinations.
// The ones that are populated are valid paths, the
// sum of which can be aggregated for each resulting
// new combination of primes and remainders.
for (let l_digit of ds){
for (let p_comb of xrange(16)){
for (let m_key of xrange(210)){
new_prime_comb = prime_set[new_l_digit] | p_comb
// suffix's remainder combination
let [m2, m3, m5, m7] = mod_combs[m_key]
// new remainder combination
let m = Math.pow(10, length - 1) * new_l_digit
let new_mod_key = mod_keys[[(m + m2) % 2, (m + m3) % 3, (m + m5) % 5, (m + m7) % 7]]
// Aggregate any populated entries into the new
// table entry
table[new_mod_key][new_l_digit][length][new_prime_comb] += table[m_key][l_digit][length - 1][p_comb]
}
}
}
}
}
// If we need only a subset of the mods set to
// zero, we need to check all instances where
// this subset is zero. For example,
// for the prime combination, [2, 3], we need to
// check all mod combinations where the first two
// are zero since we don't care about the remainders
// for 5 and 7: [0,0,0,0], [0,0,0,1],... [0,0,4,6]
// Return all needed combinations given some
// predetermined, indexed remainders.
function prime_comb_to_mod_keys(remainders){
let mod_map = [2, 3, 5, 7]
let mods = []
for (let i of xrange(4))
mods.push(!remainders.hasOwnProperty(i) ? mod_map[i] - 1 : 0)
function f(ms, i){
if (i == ms.length){
for (let idx in remainders)
ms[idx] = remainders[idx]
return [mod_keys[ms]]
}
let result = []
for (let m=ms[i] - 1; m>=0; m--){
let _ms = ms.slice()
_ms[i] = m
result = result.concat(f(_ms, i + 1))
}
return result.concat(f(ms, i + 1))
}
return f(mods, 0)
}
function get_matching_mods(prefix, len_suffix, prime_comb){
let ps = [2, 3, 5, 7]
let actual_prefix = Math.pow(10, len_suffix) * prefix
let remainders = {}
for (let i in xrange(4)){
if (prime_comb & (1 << i))
remainders[i] = (ps[i] - (actual_prefix % ps[i])) % ps[i]
}
return prime_comb_to_mod_keys(remainders)
}
// A brute-force function to check the
// table is working. Returns a list of
// valid numbers of 'length' digits
// given a prefix.
function confirm(prefix, length){
let result = [0, []]
let ps = [0, 0, 2, 3, 0, 5, 0, 7, 0, 0]
let p_len = String(prefix).length
function check(suffix){
let num = Math.pow(10, length - p_len) * prefix + suffix
let temp = num
prev = 0
while (temp){
let d = temp % 10
if (d == 1 || gcd(prev, d) == 1 || (ps[d] && num % d))
return [0, []]
prev = d
temp = ~~(temp / 10)
}
return [1, [num]]
}
for (suffix of xrange(Math.pow(10, length - p_len))){
let [a, b] = check(suffix)
result[0] += a
result[1] = result[1].concat(b)
}
return result
}
function get_prime_comb(prefix){
let prime_comb = 0
while (prefix){
let d = prefix % 10
prime_comb |= prime_set[d]
prefix = ~~(prefix / 10)
}
return prime_comb
}
// A function to test the table
// against the brute-force method.
// To match a prefix with the number
// of valid suffixes of a chosen length
// in the table, we want to aggregate all
// prime combinations for all valid digits,
// where the remainders for each combined
// prime combination (prefix with suffix)
// sum to zero (with the appropriate mod).
function test(prefix, length, show=false){
let r_digit = prefix % 10
let len_suffix = length - String(prefix).length
let prefix_prime_comb = get_prime_comb(prefix)
let ds = get_valid_digits(r_digit)
let count = 0
for (let l_digit of ds){
for (let prime_comb of xrange(16)){
for (let i of get_matching_mods(prefix, len_suffix, prefix_prime_comb | prime_comb)){
let v = table[i][l_digit][len_suffix][prime_comb]
count += v
}
}
}
let c = confirm(prefix, length)
return `${ count }, ${ c[0] }${ show ? ': ' + c[1] : '' }`
}
// Arbitrary prefixes
for (let length of [3, 4]){
for (let prefix of [2, 30]){
console.log(`prefix, length: ${ prefix }, ${ length }`)
console.log(`tabled, brute-force: ${ test(prefix, length, true) }\n\n`)
}
}
let length = 6
for (let l_digit=2; l_digit<10; l_digit++){
console.log(`prefix, length: ${ l_digit }, ${ length }`)
console.log(`tabled, brute-force: ${ test(l_digit, length) }\n\n`)
}
I have one value like 24, and I have four textboxes. How can I dynamically generate four values that add up to 24?
All the values must be integers and can't be negative, and the result cannot be 6, 6, 6, 6; they must be different like: 8, 2, 10, 4. (But 5, 6, 6, 7 would be okay.)
For your stated problem, it is possible to generate an array of all possible solutions and then pick one randomly. There are in fact 1,770 possible solutions.
var solutions = [[Int]]()
for i in 1...21 {
for j in 1...21 {
for k in 1...21 {
let l = 24 - (i + j + k)
if l > 0 && !(i == 6 && j == 6 && k == 6) {
solutions.append([i, j, k, l])
}
}
}
}
// Now generate 20 solutions
for _ in 1...20 {
let rval = Int(arc4random_uniform(UInt32(solutions.count)))
println(solutions[rval])
}
This avoids any bias at the cost of initial setup time and storage.
This could be improved by:
Reducing storage space by only storing the first 3 numbers. The 4th one is always 24 - (sum of first 3)
Reducing storage space by storing each solution as a single integer: (i * 10000 + j * 100 + k)
Speeding up the generation of solutions by realizing that each loop doesn't need to go to 21.
Here is the solution that stores each solution as a single integer and optimizes the loops:
var solutions = [Int]()
for i in 1...21 {
for j in 1...22-i {
for k in 1...23-i-j {
if !(i == 6 && j == 6 && k == 6) {
solutions.append(i * 10000 + j * 100 + k)
}
}
}
}
// Now generate 20 solutions
for _ in 1...20 {
let rval = Int(arc4random_uniform(UInt32(solutions.count)))
let solution = solutions[rval]
// unpack the values
let i = solution / 10000
let j = (solution % 10000) / 100
let k = solution % 100
let l = 24 - (i + j + k)
// print the solution
println("\([i, j, k, l])")
}
Here is a Swift implementation of the algorithm given in https://stackoverflow.com/a/8064754/1187415, with a slight
modification because all numbers are required to be positive.
The method to producing N positive random integers with sum M is
Build an array containing the number 0, followed by N-1 different
random numbers in the range 1 .. M-1, and finally the number M.
Compute the differences of subsequent array elements.
In the first step, we need a random subset of N-1 elements out of
the set { 1, ..., M-1 }. This can be achieved by iterating over this
set and choosing each element with probability n/m, where
m is the remaining number of elements we can choose from and
n is the remaining number of elements to choose.
Instead of storing the chosen random numbers in an array, the
difference to the previously chosen number is computed immediately
and stored.
This gives the following function:
func randomNumbers(#count : Int, withSum sum : Int) -> [Int] {
precondition(sum >= count, "`sum` must not be less than `count`")
var diffs : [Int] = []
var last = 0 // last number chosen
var m = UInt32(sum - 1) // remaining # of elements to choose from
var n = UInt32(count - 1) // remaining # of elements to choose
for i in 1 ..< sum {
// Choose this number `i` with probability n/m:
if arc4random_uniform(m) < n {
diffs.append(i - last)
last = i
n--
}
m--
}
diffs.append(sum - last)
return diffs
}
println(randomNumbers(count: 4, withSum: 24))
If a solution with all elements equal (e.g 6+6+6+6=24) is not
allowed, you can repeat the method until a valid solution is found:
func differentRandomNumbers(#count : Int, withSum sum : Int) -> [Int] {
precondition(count >= 2, "`count` must be at least 2")
var v : [Int]
do {
v = randomNumbers(count: count, withSum: sum)
} while (!contains(v, { $0 != v[0]} ))
return v
}
Here is a simple test. It computes 1,000,000 random representations
of 7 as the sum of 3 positive integers, and counts the distribution
of the results.
let set = NSCountedSet()
for i in 1 ... 1_000_000 {
let v = randomNumbers(count: 3, withSum: 7)
set.addObject(v)
}
for (_, v) in enumerate(set) {
let count = set.countForObject(v)
println("\(v as! [Int]) \(count)")
}
Result:
[1, 4, 2] 66786
[1, 5, 1] 67082
[3, 1, 3] 66273
[2, 2, 3] 66808
[2, 3, 2] 66966
[5, 1, 1] 66545
[2, 1, 4] 66381
[1, 3, 3] 67153
[3, 3, 1] 67034
[4, 1, 2] 66423
[3, 2, 2] 66674
[2, 4, 1] 66418
[4, 2, 1] 66292
[1, 1, 5] 66414
[1, 2, 4] 66751
Update for Swift 3:
func randomNumbers(count : Int, withSum sum : Int) -> [Int] {
precondition(sum >= count, "`sum` must not be less than `count`")
var diffs : [Int] = []
var last = 0 // last number chosen
var m = UInt32(sum - 1) // remaining # of elements to choose from
var n = UInt32(count - 1) // remaining # of elements to choose
for i in 1 ..< sum {
// Choose this number `i` with probability n/m:
if arc4random_uniform(m) < n {
diffs.append(i - last)
last = i
n -= 1
}
m -= 1
}
diffs.append(sum - last)
return diffs
}
print(randomNumbers(count: 4, withSum: 24))
Update for Swift 4.2 (and later), using the unified random API:
func randomNumbers(count : Int, withSum sum : Int) -> [Int] {
precondition(sum >= count, "`sum` must not be less than `count`")
var diffs : [Int] = []
var last = 0 // last number chosen
var m = sum - 1 // remaining # of elements to choose from
var n = count - 1 // remaining # of elements to choose
for i in 1 ..< sum {
// Choose this number `i` with probability n/m:
if Int.random(in: 0..<m) < n {
diffs.append(i - last)
last = i
n -= 1
}
m -= 1
}
diffs.append(sum - last)
return diffs
}
func getRandomValues(amountOfValues:Int, totalAmount:Int) -> [Int]?{
if amountOfValues < 1{
return nil
}
if totalAmount < 1{
return nil
}
if totalAmount < amountOfValues{
return nil
}
var values:[Int] = []
var valueLeft = totalAmount
for i in 0..<amountOfValues{
if i == amountOfValues - 1{
values.append(valueLeft)
break
}
var value = Int(arc4random_uniform(UInt32(valueLeft - (amountOfValues - i))) + 1)
valueLeft -= value
values.append(value)
}
var shuffledArray:[Int] = []
for i in 0..<values.count {
var rnd = Int(arc4random_uniform(UInt32(values.count)))
shuffledArray.append(values[rnd])
values.removeAtIndex(rnd)
}
return shuffledArray
}
getRandomValues(4, 24)
This is not a final answer, but it should be a (good) starting point.
How it works: It takes 2 parameters. The amount of random values (4 in your case) and the total amount (24 in your case).
It takes a random value between the total Amount and 0, stores this in an array and it subtracts this from a variable which stores the amount that is left and stores the new value.
Than it takes a new random value between the amount that is left and 0, stores this in an array and it again subtracts this from the amount that is left and stores the new value.
When it is the last number needed, it sees what amount is left and adds that to the array
EDIT:
Adding a +1 to the random value removes the problem of having 0 in your array.
EDIT 2:
Shuffling the array does remove the increased chance of having a high value as the first value.
One solution that is unfortunatly non-deterministic but completely random is as follows:
For a total of 24 in 4 numbers:
pick four random numbers between 1 and 21
repeat until the total of the numbers equals 24 and they are not all 6.
This will, on average, loop about 100 times before finding a solution.
Here's a solution which should have significantly* less bias than some of the other methods. It works by generating the requested number of random floating point numbers, multiplying or dividing all of them until they add up to the target total, and then rounding them into integers. The rounding process changes the total, so we need to correct for that by adding or subtracting from random terms until they add up to the right amount.
func getRandomDoubles(#count: Int, #total: Double) -> [Double] {
var nonNormalized = [Double]()
nonNormalized.reserveCapacity(count)
for i in 0..<count {
nonNormalized.append(Double(arc4random()) / 0xFFFFFFFF)
}
let nonNormalizedSum = reduce(nonNormalized, 0) { $0 + $1 }
let normalized = nonNormalized.map { $0 * total / nonNormalizedSum }
return normalized
}
func getRandomInts(#count: Int, #total: Int) -> [Int] {
let doubles = getRandomDoubles(count: count, total: Double(total))
var ints = [Int]()
ints.reserveCapacity(count)
for double in doubles {
if double < 1 || double % 1 >= 0.5 {
// round up
ints.append(Int(ceil(double)))
} else {
// round down
ints.append(Int(floor(double)))
}
}
let roundingErrors = total - (reduce(ints, 0) { $0 + $1 })
let directionToAdjust: Int = roundingErrors > 0 ? 1 : -1
var corrections = abs(roundingErrors)
while corrections > 0 {
let index = Int(arc4random_uniform(UInt32(count)))
if directionToAdjust == -1 && ints[index] <= 1 { continue }
ints[index] += directionToAdjust
corrections--
}
return ints
}
*EDIT: Martin R has correctly pointed out that this is not nearly as uniform as one might expect, and is in fact highly biased towards numbers in the middle of the 1-24 range. I would not recommend using this solution, but I'm leaving it up so that others can know not to make the same mistake.
As a recursive function the algorithm is very nice:
func getRandomValues(amount: Int, total: Int) -> [Int] {
if amount == 1 { return [total] }
if amount == total { return Array(count: amount, repeatedValue: 1) }
let number = Int(arc4random()) % (total - amount + 1) + 1
return [number] + getRandomValues(amount - 1, total - number)
}
And with safety check:
func getRandomValues(amount: Int, total: Int) -> [Int]? {
if !(1...total ~= amount) { return nil }
if amount == 1 { return [total] }
if amount == total { return Array(count: amount, repeatedValue: 1) }
let number = Int(arc4random()) % (total - amount + 1) + 1
return [number] + getRandomValues(amount - 1, total - number)!
}
As #MartinR pointed out the code above is extremely biased. So in order to have a uniform distribution of the output values you should use this piece of code:
func getRandomValues(amount: Int, total: Int) -> [Int] {
var numberSet = Set<Int>()
// add splitting points to numberSet
for _ in 1...amount - 1 {
var number = Int(arc4random()) % (total - 1) + 1
while numberSet.contains(number) {
number = Int(arc4random()) % (total - 1) + 1
}
numberSet.insert(number)
}
// sort numberSet and return the differences between the splitting points
let sortedArray = (Array(numberSet) + [0, total]).sort()
return sortedArray.enumerate().flatMap{
indexElement in
if indexElement.index == amount { return nil }
return sortedArray[indexElement.index + 1] - indexElement.element
}
}
A javascript implementation for those who may be looking for such case:
const numbersSumTo = (length, value) => {
const fourRandomNumbers = Array.from({ length: length }, () => Math.floor(Math.random() * 6) + 1);
const res = fourRandomNumbers.map(num => (num / fourRandomNumbers.reduce((a, b) => a + b, 0)) * value).map(num => Math.trunc(num));
res[0] += Math.abs(res.reduce((a, b) => a + b, 0) - value);
return res;
}
// Gets an array with 4 items which sum to 100
const res = numbersSumTo(4, 100);
const resSum = res.reduce((a, b) => a + b, 0);
console.log({
res,
resSum
});
Also plenty of different methods of approach can be found here on this question: https://math.stackexchange.com/questions/1276206/method-of-generating-random-numbers-that-sum-to-100-is-this-truly-random