Index values from a matrix using row, col indicies

Question

This is probably simple to solve. I have a 2D matrix mat with 500 rows × 335 columns, and a data.frame dat with 120425 rows. The data.frame dat has two columns I and J, which are integers to index the row, column from mat. I would like to add the values from mat to the rows of dat.

Here is my conceptual fail:

> dat$matval <- mat[dat$I, dat$J]
Error: cannot allocate vector of length 1617278737

(I am using R 2.13.1 on Win32). Digging a bit deeper, I see that I'm misusing matrix indexing, as it appears that I'm only getting a sub-matrix of mat, and not a single-dimension array of values as I expected, i.e.:

> str(mat[dat$I[1:100], dat$J[1:100]])
 int [1:100, 1:100] 20 1 1 1 20 1 1 1 1 1 ...

I was expecting something like int [1:100] 20 1 1 1 20 1 1 1 1 1 .... What is the correct way to index a 2D matrix using indices of row, column to get the values?

Answer 1

Here's a one-liner using apply's row-based operations

> dat <- as.data.frame(matrix(rep(seq(4),4),ncol=2))
> colnames(dat) <- c('I','J')
> dat
   I  J
1  1  1
2  2  2
3  3  3
4  4  4
5  1  1
6  2  2
7  3  3
8  4  4
> mat <- matrix(seq(16),ncol=4)
> mat
     [,1] [,2] [,3] [,4]
[1,]    1    5    9   13
[2,]    2    6   10   14
[3,]    3    7   11   15
[4,]    4    8   12   16

> dat$K <- apply( dat, 1, function(x,mat) mat[ x[1], x[2] ], mat=mat )
> dat
  I J  K
1 1 1  1
2 2 2  6
3 3 3 11
4 4 4 16
5 1 1  1
6 2 2  6
7 3 3 11
8 4 4 16

Answer 2

n <- 10
mat <- cor(matrix(rnorm(n*n),n,n))
ix <- matrix(NA,n*(n-1)/2,2)
k<-0
for (i in 1:(n-1)){
    for (j in (i+1):n){
    k <- k+1
    ix[k,1]<-i
    ix[k,2]<-j
    }
}
o <- rep(NA,nrow(ix))
o <- mat[ix]
out <- cbind(ix,o)

Answer 3

Almost. Needs to be offered to "[" as a two column matrix:

dat$matval <- mat[ cbind(dat$I, dat$J) ] # should do it.

There is a caveat: Although this also works for dataframes, they are first coerced to matrix-class and if any are non-numeric, the entire matrix becomes the "lowest denominator" class.

Answer 4

Using a matrix to index as DWin suggests is of course much cleaner, but for some strange reason doing it manually using 1-D indices is actually slightly faster:

# Huge sample data
mat <- matrix(sin(1:1e7), ncol=1000)
dat <- data.frame(I=sample.int(nrow(mat), 1e7, rep=T), 
                  J=sample.int(ncol(mat), 1e7, rep=T))

system.time( x <- mat[cbind(dat$I, dat$J)] )     # 0.51 seconds
system.time( mat[dat$I + (dat$J-1L)*nrow(mat)] ) # 0.44 seconds

The dat$I + (dat$J-1L)*nrow(m) part turns the 2-D indices into 1-D ones. The 1L is the way to specify an integer instead of a double value. This avoids some coercions.

...I also tried gsk3's apply-based solution. It's almost 500x slower though:

system.time( apply( dat, 1, function(x,mat) mat[ x[1], x[2] ], mat=mat ) ) # 212