Kolmogorov

Hy, i'm trying to perform the Kolmogorov-Smirnov test by using the function h = kstest(x,CDF). I have to use this test to verify the good agreement of my data set (matrix (20,6545)) to the Cumulative GEV distribution but i can't build the CDF matrix. This is the program that i wrote:

E=patterns_EOBS(:,6545); %dataset

M=6545;

N=20;

x=[-5:3:55]';

parmhat_EOBS=zeros(M,3);

P=zeros(N,M);

CDF=zeros(N,2);

h=zeros(M);

p=zeros(M);

kstat=zeros(M);

cv=zeros(M);

for j=1:M

    parmhat_EOBS(j,:)=gevfit(E(:,j));

    for i=1:N

        P(i,j)=gevcdf(x(i),parmhat_EOBS(j,1),parmhat_EOBS(j,2),parmhat_EOBS(j,3));

    end

    CDF(:,j)=horzcat(E(:,j),P(:,j));

    [h(j),p(j),kstat(j),cv(j)]=kstest(E(:,j),CDF(:,j));

    end

And this is the error that mathlab returns to me:

??? Subscripted assignment dimension mismatch.

Error in ==> KolmoEOBS at 22 CDF(:,j)=horzcat(E(:,j),P(:,j));

CAn you help me?

MArghe

[Andrew Newell]

Since you don't know the parameters of your distribution in advance, you should be using lillietest.

Answer 1

[Ivan van der Kroon]

Matlab is very clear here; your dimensions mitmatch: size(CDF(:,j))=[20,1] while size(horzcat(E(:,j),P(:,j)))=[20,2].

Another problem that will occur is that j will be greater than 2 at some point while size(CDF)=[20,2].

From this I think you are looking for

CDF=zeros(N,2,M);
CDF(:,:,j)=horzcat(E(:,j),P(:,j));

Answer 2

[MArghe]

First of all, thank you for answering me.

To perform the test, the matrix CDF has to be a two columns matrix:

the first column comprises the data set that i want to analyze (that is every column of the matrix E [size (20,6545)]), the second column comprises the Cumulative Density Function chosen (In this case the GEV). That is, i want a two column matrix at each cycle, to perform the test. I've tried to execute your suggest, but now the error is of the type OUT OF MEMORY at the line 8 of the program i wrote above.

Thanks for your help

MArgheirta

[Oleg Komarov]

Line 8: zeros(M) = zeros(M,M) --> zeros(6545) ~ 2.6 Gb = (6545^2) * 8 bytes

Answer 3

[Ivan van der Kroon]

I know that CDF has to be of size Nx2, but you want to have M of those in your for-loop. So you have to use CDF(:,:,j), which is a Nx2 again. If you are not interested in it, you should overwrite it every iteration, i.e. no preallocation fod CDF and no indexes:

      CDF=horzcat(E(:,j),P(:,j));

    [h(j),p(j),kstat(j),cv(j)]=kstest(E(:,j),CDF);

This way you save considerable memory. But working with a 6545x6545 system is hard as Oleg commented. It checked and it worked for me (on my 64bits version).