## Interesting probability problem

For the discussion of math. Duh.

Moderators: gmalivuk, Moderators General, Prelates

measure
Posts: 81
Joined: Sat Apr 04, 2015 4:31 pm UTC
Location: Time-traveling kayak

### Interesting probability problem

Here's a problem that was bugging me for a while. I was going to post it here for help, but I ended up solving it myself. I thought the solution looked kinda pretty, so I decided to post it here for y'all:

Suppose a value x is chosen at random uniformly on the interval [0,1]. Next two numbers a and b are chosen at random uniformly on the interval [0,x]. You are then told the absolute value of the difference of a and b δ=|a-b|. Given δ, what is the expected value of x?

Here is a python program I wrote to run a bunch (10^7) of trials and plot for each δ the average of the x values that resulted in that δ:
Spoiler:

Code: Select all

`import randomsteps = 1000trials = 10000000sd = []cd = []for i in range(steps):   sd.append(0)   cd.append(0)for i in range(trials):   x = random.uniform(0,1)   a = random.uniform(0,x)   b = random.uniform(0,x)   d = abs(a-b)   pd = int(steps*d)   sd[pd] += x   cd[pd] += 1for i in range(steps):   if cd[i] > 0:      print("> d = "+str(i/steps)+"\t("+str(cd[i])+")\tavg(x) = "+str(sd[i]/cd[i]))      #print(i/steps,sd[i]/cd[i])   else:      print("> d = "+str(i/steps)+"\t(0)")      #print(i/steps)`

My solution:
Spoiler:

Code: Select all

`_      δ*ln(δ) - δ + 1x(δ) = ―――――――――――――――        -ln(δ) + δ - 1`

Xanthir
My HERO!!!
Posts: 5186
Joined: Tue Feb 20, 2007 12:49 am UTC
Contact:

### Re: Interesting probability problem

Not looking at your solution yet, but I know from prior experience that the average distance between two uniformly-random points in an interval is 1/3 the length of the interval, so I believe I can just reverse that - E(x)=3δ.

Now I've looked at your trials, and: interesting!
(defun fibs (n &optional (a 1) (b 1)) (take n (unfold '+ a b)))

measure
Posts: 81
Joined: Sat Apr 04, 2015 4:31 pm UTC
Location: Time-traveling kayak

### Re: Interesting probability problem

Xanthir wrote:Not looking at your solution yet, but I know from prior experience that the average distance between two uniformly-random points in an interval is 1/3 the length of the interval, so I believe I can just reverse that - E(x)=3δ.

Yeah, that was my first thought as well. Remember though that x is capped at 1 and δ can be as large as 1 as well. (I think the average of δ/x does end up being 1/3, but that's just a first-order approximation, and I don't think in general the reciprocal of the average will be the average of the reciprocals.)

cyanyoshi
Posts: 356
Joined: Thu Sep 23, 2010 3:30 am UTC

### Re: Interesting probability problem

I got the same answer as measure. Seems like using Bayes' theorem and a lot of handwaving usually works!

measure
Posts: 81
Joined: Sat Apr 04, 2015 4:31 pm UTC
Location: Time-traveling kayak

### Re: Interesting probability problem

cyanyoshi wrote:I got the same answer as measure. Seems like using Bayes' theorem and a lot of handwaving usually works!

I tried to get it to work with Bayes' theorem, but I didn't trust my handwaving to apply it to continuous distributions. I ended up going with a geometric/calculus approach.

cyanyoshi
Posts: 356
Joined: Thu Sep 23, 2010 3:30 am UTC

### Re: Interesting probability problem

My intuition is that with random variables X, Y and Z, then you can get everything you need from the pdf f(x,y,z). By normalizing f(x,y0,z0), it seems that you should get f(x|y0,z0), the conditional probability distribution of X given that Y=y0 and Z=z0. Clearly, f(a,b|x0) should be constant over the square (a,b)∈[0,x0]2, so f(a,b|x0)=1/(x0)2. I am visualizing the domain of f(a,b,x) as an inverted pyramid with its vertex at (0,0,0) and its base is a square with vertices (0,0,1), (1,0,1), (1,1,1), and (0,1,1). Each horizontal slice should be a square with the same infinitesimal "mass" of dx. So then f(a,b,x) = 1/x2 within this pyramid and is zero elsewhere.

OK, now what about f(δ,x)? Fixing X=x0 again, f(a,b|x0) is just a uniform distribution over a square. My gut tells me that f(δ|x0) is proportional to the length of the level curve of δ(a,b) = |a-b| in this domain. This is just C*2*sqrt(2)*(x0-δ). Normalizing over the domain [0,x] gives f(δ,x)=2*(x-δ)/x2.

Now then, f(x|δ)=C*2*(x-δ)/x2. x can take any value in [δ,1], so this must integrate to 1 over that interval. Doing the integral makes C=2*δ-2-2*ln(δ). Now we got f(x|δ) = 1/(δ-1-ln(δ))*(x-δ)/x2. The expected value of x under this pdf is the integral of x*f(x|δ) on the interval of all possible x values, namely [δ,1]. Thus the expected value of x given δ is (1-δ+δ*ln(δ))(δ-1-ln(δ)).