# "Distribution" of numbers $0\leq a\leq b\leq c\leq d\leq 1$

A friend came up with the next problem: Consider $0\leq a\leq b\leq c\leq d\leq 1$ numbers such that $a+b+c+d=1$. Are there numbers $a_{0}$, $b_{0}$, $c_{0}$, $d_{0}$ that minimize $$|a-a_{0}|+|b-b_{0}|+|c-c_{0}|+|d-d_{0}|$$ most of the time? I mean, if we repeat the process of choosing $a$, $b$, $c$ and $d$ randomly, is there a expected value for $a$, $b$, $c$ and $d$? And what it is?

I tried to start with an easiest problem, with just $a$ and $b$, such that $0\leq a\leq b\leq 1$ and $a+b=1$ but I had no idea where to start, so I decided to try with some numerical sampling, and I did the next in Python:

Choose a number $x\in [0,1]$ uniformely, and define $a=\min\{x,1-x\}$ and $b=\max\{x,1-x\}$. Clearly $0\leq a\leq b\leq 1$ and $a+b=1$. Doing this, and repeating a lot of times I got that the "expected value" for $a$ was $0.25$ and for $b$ was $0.75$, so the ratio is $1:3$.

Next, I tried the same in Python but with two values: Choose $x,y\in [0,1]$ uniformely, and let $x_{1}=\min\{x,y\}$ and $x_{2}=\max\{x,y\}$, now define $$A=\{x_{1},x_{2}-x_{1},1-x_{2}\}$$ and let $a_{1}$, $a_{2}$, $a_{3}$ be the elements of $A$ in increasing order. Clearly $0\leq a_{1}\leq a_{2}\leq a_{3}\leq 1$, and $a_{1}+a_{2}+a_{3}=1$, and repeting a lot of times I got that the "expected values" were of the ratio $2:5:11$.

Repeating the same method but now with one more variable I got that the "expected values" were on ratio $3:7:13:25$.

Lastly, with one more variable the ratios were $12:27:47:77:137$.

All this calculations were found just by trial and error, and are clearly non mathematically justified, but seems like, at least, a good conjecture.

Is there any hidden pattern behind these ratios? Is there any reason for this numbers to came up?

Any help would be appreciated.

• Not an answer yet (maybe I have more time later), but you have to integrate over the density of solutions to get your ratios. Easy to see for two numbers: $a$ will be between 0 and 0.5, $b$ will be between 0.5 and 1, and both solutions are evenly distributed over their solution space. So your expected value for $a$ is $\int_0^{\frac{1}{2}}dx = \frac{1}{4}$ and for $b$ it is $\int_{\frac{1}{2}}^1dx = \frac{3}{4}$, and the resulting ratio is 1:3. For more values, it gets more complicated, but the principle should be the same. Jul 2, 2018 at 8:20
• But for 3 or more variables which are the ranges to integrate?
– Nah
Jul 2, 2018 at 8:20
• The ranges are not difficult for $3$: $a$ must be between $0$ and $\frac{1}{3}$, $b$ between $0$ and $\frac{1}{2}$, and $c$ between $\frac{1}{3}$ and $1$. Analogously, you can extend this to $n$ getting the ranges $a_1: [0, \frac{1}{n}], a_2: [0, \frac{1}{n-1}], ..., a_{n-1}: [0, \frac{1}{2}], a_n: [\frac{1}{n},1]$. The problem is the weight for each integral, because it is not evenly distributed anymore. Jul 2, 2018 at 8:25
• I'd like to add that I am unsure if the patterns you get are really unique. They are obviously a consequence of the solution space, which you can influence by the method you have chosen to generate $a$, $b$, and $c$. For two values, there seems to be only one obvious solution, but for more than two values I suppose there is more than one way to come to a solution. Jul 2, 2018 at 8:46
• Also, can you specifiy "Repeating the same method but now with one more variable"? It is not completely clear how this is done. What do you use for the three values? Min of all, Max of all, Min of all except the least? Jul 2, 2018 at 10:43

Without the sorting, if you just choose $n$ random numbers $x_i$ such that $0 \leq x_i \leq 1$ and $\sum x_i = 1,$ the sample space can be represented as an $(n-1)$-dimensional polytope (specifically, a simplex) with vertices at $(1,0,0,\ldots,0)$, $(0,1,0,\ldots,0),$ $(0,0,1,0,\ldots,0), \ldots,$ $(0,0,\ldots,0,1).$ (That is, at each vertex exactly one coordinate is $1$ and the others are $0$.) An obvious choice of distribution over this region would be a uniform density.

I believe your unsorted differences achieve this distribution. Certainly they give the correct expected value for each variable.

Now you apply the restriction that $x_i \leq x_j$ if $i < j.$ The sample space for these variables is a smaller $(n-1)$-dimensional simplex that you can obtain by "cutting away" the parts of the original simplex in which $x_1 > x_2,$ or $x_2 > x_3,$ and so forth until you cut away the part in which $x_n > x_{n-1}.$

Generating the unsorted differences and then sorting them converts a uniform distribution over the larger simplex to a uniform distribution over the smaller one.

As an example, for $n = 3$ the final sample space has vertices at $(0,0,1),$ $\left(0,\frac12,\frac12\right),$ and $\left(\frac13,\frac13,\frac13\right).$ The centroid of this simplex is $\left(\frac19,\frac{5}{18},\frac{11}{18}\right),$ with coordinates in the ratio $2:5:11.$

For $n=4$ the vertices are $(0,0,0,1),$ $\left(0,0,\frac12,\frac12\right),$ $\left(0,\frac13,\frac13,\frac13\right),$ and $\left(\frac14,\frac14,\frac14,\frac14\right).$ For $n=5$ the vertices are $(0,0,0,0,1),$ $\left(0,0,0,\frac12,\frac12\right),$ $\left(0,0,\frac13,\frac13,\frac13\right),$ $\left(0,\frac14,\frac14,\frac14,\frac14\right),$ and $\left(\frac15,\frac15,\frac15,\frac15,\frac15\right),$ and so forth for larger $n.$

In short, each vertex is obtained by maximizing the value of $x_i$ for a different $i.$ It should be clear that for $n$ variables, the $i$th coordinate of the centroid of the final simplex, $i = 1, 2, \ldots, n,$ is $$\frac1n\left(\sum_{j=1}^{n} \frac1j - \sum_{j=1}^{n-i} \frac1j\right).$$

Since summing up terms of the harmonic series can be a bit painful, here's a simpler way to generate the ratios. Denote the difference between the expected values of $x_{i+1}$ and $x_i$ by $\Delta_i = E(x_{i+1}) - E(x_i).$ Notice that if we take the expected value of $x_1,$ and follow it by the increasing sequence of these differences, we get a sequence of numbers in the ratio $$E(x_i) : \Delta_1 : \Delta_2 : \cdots : \Delta_{n-3} : \Delta_{n-2} : \Delta_{n-1} = \frac1n : \frac{1}{n-1} :\frac{1}{n-2} : \cdots: \frac13 : \frac12 : 1. \tag{*}$$

In order to make the ratio on the right-hand side a ratio of integers, we merely need multiply each part by the least common multiple of $\{1,2,3,\ldots, n\}.$

For example, for $n = 5,$ we find that the least common multiple of $\{1,2,3,4,5\}$ is $60.$ The ratio $(^*)$ is therefore $$\frac15 : \frac14 :\frac13 :\frac12 : 1 = \frac{60}{5} : \frac{60}{4} :\frac{60}{3} :\frac{60}{2} : 60 = 12 : 15 : 20 : 30 : 60.$$

Now, recalling that except for the first part of this ratio, each part is proportional to the difference between two consecutive expected values, and observing that $E(x_k) = E(x_1) + \sum_{i=1}^{k-1} \Delta_i,$ we replace the ratio with a ratio of partial sums: $$12 : 15 : 20 : 30 : 60 \to 12 : 27 : 47 : 77 : 137,$$ and there's the ratio you found.

• Yes, of course, that was what I overlooked. I assumed the distribution to be non-evenly, but it has to be evenly due to the original values all randomly picked from 0 to 1. Then the rest follows by taking the mean value of all coordinates. Nice! Jul 2, 2018 at 14:50

It seems that your process is:

• choose $n-1$ values independently and uniformly on $[0,1]$
• sort these values into order
• find the gaps between successive sorted values, and between $0$ and the smallest, and between the largest and $1$, so you calculate $n$ non-negative gaps adding up to $1$
• sort these gaps into order
• consider the expected values of these sorted gaps

For what it is worth, I believe the $n$ gaps (before sorting) have identical but not independent distributions with density $f(x)=n(1-x)^{n-1}$ when $0 \le x \le 1$ and so mean $\frac{1}{n}$

You asked for an explanation of the ratios pattern you observed. Empirically it seems that the expected length of the $k$th sorted gap out of $n$ is related to the difference of two harmonic numbers $$\frac{1}{n}\left(\sum_{i=1}^{n} \frac1{i} - \sum_{j=1}^{n-k} \frac1{j} \right) \\ = \sum_{j=n-k+1}^{n} \frac1{nj}$$

So in your initial example with $n=4$, you get for different values of $k$:

1. : $\frac1{4\times4} = \frac1{16} = \frac{3}{48}$
2. : $\frac1{4\times3}+\frac1{4\times4} = \frac{7}{48}$
3. : $\frac1{4\times2}+\frac1{4\times3}+\frac1{4\times4} = \frac{13}{48}$
4. : $\frac1{4\times1}+\frac1{4\times2}+\frac1{4\times3}+\frac1{4\times4} = \frac{25}{48}$

reproducing your $3:7:13:25$ ratios. This happens with other values for $n$ too

I believe that these expected values do not minimise $\mathbb E [ |a-a_{0}|+|b-b_{0}|+|c-c_{0}|+|d-d_{0}| ]$ but they do minimise $\mathbb E[(a-a_{0})^2+(b-b_{0})^2+(c-c_{0})^2+(d-d_{0})^2]$

To minimise the expected absolute sums I think you would do better with the medians of the distributions of the ordered gaps, and I suspect these medians may not add up to $1$ when $n \gt 2$. So for example with $n=4$ the means are about $0.0625, 0.1458, 0.2708, 0.5208$ but the medians seem experimentally to be closer to something like $0.052, 0.145, 0.276, 0.500$