Generate random numbers summing to a predefined value

Generate 4 random numbers, compute their sum, divide each one by the sum and multiply by 40.

If you want Integers, then this will require a little non-randomness.

There are only 37^4 = 1,874,161 arrangements of four integers in the range [1,37] (with repeats allowed). Enumerate them, saving and counting the permutations that add up to 40.
(This will be a much smaller number, N).

Draw uniformly distributed random integers K in the interval [0, N-1] and return the K-th permutation. This can easily be seen to guarantee a uniform distribution over the space of possible outcomes, with each sequence position identically distributed. (Many of the answers I’m seeing will have the final choice biased lower than the first three!)

Here’s the standard solution. It’s similar to Laurence Gonsalves’ answer, but has two advantages over that answer.

1. It’s uniform: each combination of 4 positive integers adding up to 40 is equally likely to come up with this scheme.

and

2. it’s easy to adapt to other totals (7 numbers adding up to 100, etc.)

import random

def constrained_sum_sample_pos(n, total):
    """Return a randomly chosen list of n positive integers summing to total.
    Each such list is equally likely to occur."""

    dividers = sorted(random.sample(range(1, total), n - 1))
    return [a - b for a, b in zip(dividers + [total], [0] + dividers)]

Sample outputs:

>>> constrained_sum_sample_pos(4, 40)
[4, 4, 25, 7]
>>> constrained_sum_sample_pos(4, 40)
[9, 6, 5, 20]
>>> constrained_sum_sample_pos(4, 40)
[11, 2, 15, 12]
>>> constrained_sum_sample_pos(4, 40)
[24, 8, 3, 5]

Explanation: there’s a one-to-one correspondence between (1) 4-tuples (a, b, c, d) of positive integers such that a + b + c + d == 40, and (2) triples of integers (e, f, g) with 0 < e < f < g < 40, and it’s easy to produce the latter using random.sample. The correspondence is given by (e, f, g) = (a, a + b, a + b + c) in one direction, and (a, b, c, d) = (e, f - e, g - f, 40 - g) in the reverse direction.

If you want nonnegative integers (i.e., allowing 0) instead of positive ones, then there’s an easy transformation: if (a, b, c, d) are nonnegative integers summing to 40 then (a+1, b+1, c+1, d+1) are positive integers summing to 44, and vice versa. Using this idea, we have:

def constrained_sum_sample_nonneg(n, total):
    """Return a randomly chosen list of n nonnegative integers summing to total.
    Each such list is equally likely to occur."""

    return [x - 1 for x in constrained_sum_sample_pos(n, total + n)]

Graphical illustration of constrained_sum_sample_pos(4, 10), thanks to @FM. (Edited slightly.)

0 1 2 3 4 5 6 7 8 9 10  # The universe.
|                    |  # Place fixed dividers at 0, 10.
|   |     |       |  |  # Add 4 - 1 randomly chosen dividers in [1, 9]
  a    b      c    d    # Compute the 4 differences: 2 3 4 1

Use multinomial distribution

from numpy.random import multinomial
multinomial(40, [1/4.] * 4)

Each variable will be distributed as a binomial distribution with mean n * p equal to 40 * 1/4 = 10 in this example.

Building on @markdickonson by providing some control over distribution between the divisors. I introduce a variance/jiggle as a percentage of the uniform distance between each.

 def constrained_sum_sample(n, total, variance=50):
    """Return a random-ish list of n positive integers summing to total.

    variance: int; percentage of the gap between the uniform spacing to vary the result.
    """
    divisor = total/n
    jiggle = divisor * variance / 100 / 2
    dividers = [int((x+1)*divisor + random.random()*jiggle) for x in range(n-1)]
    result = [a - b for a, b in zip(dividers + [total], [0] + dividers)]
    return result

Sample output:

[12, 8, 10, 10]
[10, 11, 10, 9]
[11, 9, 11, 9]
[11, 9, 12, 8]

The idea remains to divide the population equally, then randomly move them left or right within the given range. Since each value is still bound to the uniform point we don’t have to worry about it drifting.

Good enough for my purposes, but not perfect. eg: the first number will always vary higher, and the last will always vary lower.

If you want true randomness then use:

import numpy as np
def randofsum_unbalanced(s, n):
    # Where s = sum (e.g. 40 in your case) and n is the output array length (e.g. 4 in your case)
    r = np.random.rand(n)
    a = np.array(np.round((r/np.sum(r))*s,0),dtype=int)
    while np.sum(a) > s:
        a[np.random.choice(n)] -= 1
    while np.sum(a) < s:
        a[np.random.choice(n)] += 1
    return a

If you want a greater level of uniformity then take advantage of the multinomial distribution:

def randofsum_balanced(s, n):
    return np.random.multinomial(s,np.ones(n)/n,size=1)[0]

Here’s a humble improvement to @Mark Dickinson‘s version to allow zeros in the generated integers (so that they are non-negative, rather than positive):

import random

def constrained_sum_sample_pos(n, total):
    """Return a randomly chosen list of n non-negative integers summing to total.
    Each such list is equally likely to occur."""

    dividers = sorted(random.choices(range(0, total), k=n-1))
    return [a - b for a, b in zip(dividers + [total], [0] + dividers)]

The function random.choices() samples with replacement, as opposed to random.sample() which samples without replacement. It’s new from Python 3.6.