Can Principal Component Analysis Solve the Max Min Distribution Problem?

[EngWiPy]

Hello,

I have this probability:

\text{Pr}\left\{\underset{i,j}{\max\,}\underset{n}{\min\,}X_i(n)+X_j(n)<a\right\}

where X_i(n) and X_j(n) are i.i.d. for all i,j, and n. Can I find the distribution of

X_i(n_{\text{min}})

where:

\underset{n}{\min\,}X_i(n)+X_j(n)=X_i(n_{\text{min}})+X_j(n_{\text{min}})

??

Thanks in advance

 

[Stephen Tashi]

where X_i(n) and X_j(n) are i.i.d. for all i,j, and n. Can I find the distribution of

I find your notation mysterious. Why is the index n in parenthesis versus being a subscript like i and j are?

 

[EngWiPy]

I find your notation mysterious. Why is the index n in parenthesis versus being a subscript like i and j are?

Basically, I have a set of random variables X_i(n) for i=1,...,K, and for n=1,..., N. So, X_i(n) means the nth random variable of X_i. It is hard to explain. It is easier using communication systems.

 

[Stephen Tashi]

I have this probability:

\text{Pr}\left\{\underset{i,j}{\max\,}\underset{n}{\min\,}X_i(n)+X_j(n)<a\right\}

where X_i(n) and X_j(n) are i.i.d. for all i,j, and n.

Are you saying that you only know the above probability and do not know the common distribution of the X_i(k) ?

 

[EngWiPy]

Are you saying that you only know the above probability and do not know the common distribution of the X_i(k) ?

I know the distribution of x_i(n), but I do not know what is the distribution of X_i(n_min), because the minimization is done for X_i(n)+X_j(n).

 

[haruspex]

Can I find the distribution of
X_i(n_{\text{min}})
where:
\underset{n}{\min\,}X_i(n)+X_j(n)=X_i(n_{\text{min}})+X_j(n_{\text{min}})

I don't understand how
\underset{n}{\min\,}X_i(n)+X_j(n)=X_i(n_{\text{min}})+X_j(n_{\text{min}})
serves as a definition of
n_{\text{min}}
Isn't n_min a function of i and j?

 

[EngWiPy]

I don't understand how
\underset{n}{\min\,}X_i(n)+X_j(n)=X_i(n_{\text{min}})+X_j(n_{\text{min}})
serves as a definition of
n_{\text{min}}
Isn't n_min a function of i and j?

OK, let me state the problem in another way: suppose I have N×K i.i.d. random variables X_{i,n} for i=1,...,K and n=1,...,N.

Define

X_{ij}=\underset{n}{\min\,}X_{i,n}+X_{j,n}

for

i\neq j

Now I can find the distribution X_{ij}, but I need the distribution of:

\underset{i,j}{\max\,}X_{ij}

Is that doable?

 

[chiro]

OK, let me state the problem in another way: suppose I have N×K i.i.d. random variables X_{i,n} for i=1,...,K and n=1,...,N.

Define

X_{ij}=\underset{n}{\min\,}X_{i,n}+X_{j,n}

for

i\neq j

Now I can find the distribution X_{ij}, but I need the distribution of:

\underset{i,j}{\max\,}X_{ij}

Is that doable?

This looks like a standard order statistics problem. Are the domain for i and j fixed?

 

[EngWiPy]

This looks like a standard order statistics problem. Are the domain for i and j fixed?

i=1,...,N and j=1,...,N and i does not equal j.

The problem is that X_ij are not independent.

 

[chiro]

i=1,...,N and j=1,...,N and i does not equal j.

The problem is that X_ij are not independent.

Perhaps you could create an uncorrelated basis and go from there. Are you aware of Principal Component Analysis?

 

[EngWiPy]

Perhaps you could create an uncorrelated basis and go from there. Are you aware of Principal Component Analysis?

Not really, what is that?

 

[chiro]

Not really, what is that?

It's the main idea of principal components.

http://en.wikipedia.org/wiki/Principal_component_analysis

The idea is to create an orthogonal (but not necessarily orthonormal in general) basis where each basis vector is a linear combination of your random variables. The basic idea is to solve an optimization problem where one constraint is to set your covariance matrix of your new basis to zero.

This will create an uncorrelated basis and from there you can use techniques that would otherwise assume to have un-correlated random variables.

This isn't enough to solve your problem, but I think it's worth looking into as one part of the solution especially since you are faced with the dependencies between the variables.

 

[EngWiPy]

It's the main idea of principal components.

http://en.wikipedia.org/wiki/Principal_component_analysis

The idea is to create an orthogonal (but not necessarily orthonormal in general) basis where each basis vector is a linear combination of your random variables. The basic idea is to solve an optimization problem where one constraint is to set your covariance matrix of your new basis to zero.

This will create an uncorrelated basis and from there you can use techniques that would otherwise assume to have un-correlated random variables.

This isn't enough to solve your problem, but I think it's worth looking into as one part of the solution especially since you are faced with the dependencies between the variables.

OK, I will have a look on it. Thanks for interacting