How to simulate a random signal sequence change

Specific Domains Statistics
1

I have got a simple problem to calculate Posterior after some signal disturbances.

Say the input signal is a sequence of 0 and 1 with 70% of chance is 1. And somehow we know that after the transmission, the output have 0 an 1 of equal chance. I want to calculate what the probability when we get 1 in output and it is truly 1 in the input.

I write the following code to implement the simulation. I want to randomly set seeds and see what happened to my simulation.

using Random, StatsBase
k = rand(1:2000);
Random.seed!(k);
t = sample(0:1,weights([0.3,0.7]),1000);
Random.seed!(k);
r = sample(0:1,weights([0.5,0.5]),1000);
Post = sum(t[r.==1])/sum(r[r.==1])
Prior =sum(r[t.==1])/sum(t[t.==1])

The result is not right, with Post always 1.0 and Prior around 0.7.

I am thinking of several potential problems:

  1. I was conceptually wrong to implement the simulation this way. Maybe r is not the correct realization of t after disturbance, and I should modify t base on my knowledge of the disturbance instead of generating r directly.
  2. I used sample() in a wrong way.
  3. I calculated Post and Prior wrongly.

Please help me out on this simulation problem.

2

I am not an expert at all, but my guess would be:

Prior = sum(t[t.==1])/length(t)
Post = sum(t[r.==1 .& t.==1])/sum(t[t.==1])

and both should read ~0.7.
I.e., the probability of the input being in state 1 given that the transmission was in state 1, should be 0.7 because the transmission is not changing the two-state probabilities. Seems analogous to the red-yellow cab problem well explained here Bayes and the cabs.

3

I think the problem is not fully specified. You don’t link the output with the input, you only talk about their individual distributions:

So you have a probability vector before transmission [0.7 0.3], and transmission procedure representable as 2x2 stochastic matrix Q and you tell us that [0.7 0.3]*Q = [0.5 0.5]. There are many Q which give you that.

julia> Q(λ) = [λ 1-λ; 5/3-7λ/3 1-( 5/3-7λ/3)]
Q (generic function with 1 method)

julia> Q(0.4)
2×2 Array{Float64,2}:
 0.4       0.6
 0.733333  0.266667

julia> [0.7 0.3]*Q(0.4)
1×2 Array{Float64,2}:
0.5 0.5

Here the entries of this particular Q = [ 0.4 0.6; 0.733 0.267] specify that a 1 gets corrupted to a 0 with probability 0.6 and a zero 0 gets corrupted to a 1 with probability 0.7333.

1 Like
4

PS:

n = 10000
p = [0.7,0.3]
t = sample(1:-1:0, weights(p), n)
Q(λ) = [λ 1-λ; 5/3-7λ/3 1-( 5/3-7λ/3)]
λ = 0.4
r = [sample(1:-1:0, weights(vec(Q(λ)[i==1 ? 1 : 2, :]))) for i in t]

Then the conditional probability of being 1 given that that 1 was observed is computed and estimated as follows

julia> prob = (p[1]*Q(λ)[1,1])/(p[1]*Q(λ)[1,1] + p[2]*Q(λ)[2, 1])
0.56

julia> prob_est =  dot(t,r)/n/mean(r)
0.5670391061452514
1 Like
5

Thank you, @mschauer, very clear explanation and helpful solution, too!

There are additional info on the disturbance process for us to model r.

It seems that independently generate r is not the correct solution. I will try as you suggested.