# Re: [Rd] unique.matrix issue [Was: Anomaly with unique and match]

From: Petr Savicky <savicky_at_cs.cas.cz>
Date: Thu, 10 Mar 2011 09:29:10 +0100

On Wed, Mar 09, 2011 at 02:11:49PM -0500, Simon Urbanek wrote:
> match() is a red herring here -- it is really a very specific thing that has to do with the fact that you're running unique() on a matrix. Also it's much easier to reproduce:
>
> > x=c(1,1+0.2e-15)
> > x
> [1] 1 1
> > sprintf("%a",x)
> [1] "0x1p+0" "0x1.0000000000001p+0"
> > unique(x)
> [1] 1 1
> > sprintf("%a",unique(x))
> [1] "0x1p+0" "0x1.0000000000001p+0"
> > unique(matrix(x,2))
> [,1]
> [1,] 1
>
> and this comes from the fact that unique.matrix uses string representation since it has to take into account all values of a row/column so it pastes all values into one string, but for the two numbers that is the same:
> > as.character(x)
> [1] "1" "1"

I understand the use of match() in the original message by Terry Therneau as an example of a situation, where the behavior of unique.matrix() becomes visible even without looking at the last bits of the numbers.

Let me suggest to consider the following example.

x <- 1 + c(1.1, 1.3, 1.7, 1.9)*1e-14
a <- cbind(rep(x, each=2), 2)
rownames(a) <- 1:nrow(a)

The correct set of rows may be obtained using

unique(a - 1)

[,1] [,2]

```  1 1.110223e-14    1
3 1.310063e-14    1
5 1.709743e-14    1
7 1.909584e-14    1

```

However, due to the use of paste(), which uses as.character(), in unique.matrix(), we also have

unique(a)

[,1] [,2]
1 1 2
5 1 2

Let me suggest to consider a transformation of the numeric columns by rank() before the use of paste(). For example

unique.mat <- function(a)
{

```      temp <- apply(a, 2, rank, ties.method="max")
temp <- apply(temp, 1, function(x) paste(x, collapse = "\r"))
a[!duplicated(temp), , drop=FALSE]
```
}

unique.mat(a)

[,1] [,2]
1 1 2
3 1 2
5 1 2
7 1 2

Petr Savicky.

R-devel_at_r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-devel Received on Thu 10 Mar 2011 - 08:38:06 GMT

Archive maintained by Robert King, hosted by the discipline of statistics at the University of Newcastle, Australia.
Archive generated by hypermail 2.2.0, at Thu 10 Mar 2011 - 12:30:28 GMT.

Mailing list information is available at https://stat.ethz.ch/mailman/listinfo/r-devel. Please read the posting guide before posting to the list.