Correlation in C with matrix

PB0711 · 01-23-2012, 11:50 AM

Hello all,

I'm a higher level programmer, i.e. R, perl, matlab etc. I want to get started with some C coding but I'm finding it pretty difficult. I would like to interface my R program with some C code, but first I need the C code.

I have a matrix 6783x102 (rowXcol) and I want to get the correlation of the two rows if some criteria are met. So in pseudo code

Code:

for(int i=0: i > numRows; i++){
  for(int j=0; j> numRows; j++){
    if(VectorCriteria[i] > kon+VectorCriteria[j]){
      do correlation mat[i][ ] vs mat[j][ ]
    }
  }
}

So my question is with C how is the best way to represent a matrix or ow is the best way to do this. For the correlation I think I can probably find a library.

Thanks for the help,

Paul

johnsfine · 01-23-2012, 12:03 PM

Quote:

Originally Posted by PB0711

with C how is the best way to represent a matrix

If the matrix size is known at compile time, it is best (performance and simplicity) to represent the matrix as an array of arrays.

Code:

void do_correlation(double* x, double*y);
...
enum {rows=6783, columns=102};
double mat[rows][columns];
...
for (int i=0; i<rows; i++)
   for (int j=0; j<rows; j++)
      if (...)
         do_correlation( mat[i], mat[j] );

Notes:
I used enum for defining integer compile time constants. If you don't like that, use some other method, but enum has advantages.
In C, arrays are passed as pointers, which can be confusing. I pre declared do_correlation as taking its inputs by double* to make the true parameter passing method is more explicit.

If the number of rows is not known at compile time, that makes the above approach a bit trickier. If the number of columns is not known at compile time, that is a much bigger problem for the above approach. You probably want a more complicated way to represent the matrix if the number of columns is not known at compile time.

BTW, why C rather than C++? C++ makes things like this easier and is not necessarily any less efficient than C.

ta0kira · 01-24-2012, 09:01 PM

Quote:

Originally Posted by PB0711

Code:

for(int i=0: i > numRows; i++){
  for(int j=0; j> numRows; j++){
    if(VectorCriteria[i] > kon+VectorCriteria[j]){
      do correlation mat[i][ ] vs mat[j][ ]
    }
  }
}

This will be very inefficient. Using R code, the correlation matrix, where element (m,n) is the correlation coefficient between row m and row n of the original matrix, can be computed as follows:

Code:

vals <- #the matrix you described
vals.covar <- (function(x) x %*% t(x) / ncol(x))(vals) #the covariance
vals.corr  <- (function(x) diag( 1/sqrt(diag(x)) ) %*% x %*% diag( 1/sqrt(diag(x)) ))(vals.covar) #the correlation matrix

You can break this down into matrix multiplication (%*% R operator) in C, and since vals.covar and vals.corr are by definition symmetric you only need to compute half the elements. The computation of vals.corr is just a normalization of each row and column using the square-root of the corresponding diagonal element, leaving the diagonal elements equal to 1.

Quote:

Originally Posted by PB0711

I would like to interface my R program with some C code, but first I need the C code.

If you're going to write an R extension in C you'll need to write a wrapper in R, anyway. You might as well design it so you only compute in C what can't be adequately computed in R, e.g. compute the correlation in R and pass it to a C function that expects it to already be computed.
Kevin Barry