21

I have seen this question being asked multiple times on the R mailing list, but still could not find a satisfactory answer.

Suppose I a matrix m 

m <- matrix(rnorm(10000000), ncol=10)

I can get the mean of each row by:

system.time(rowMeans(m))  
   user  system elapsed   
  0.100   0.000   0.097

But obtaining the minimum value of each row by 

system.time(apply(m,1,min))  
   user  system elapsed   
 16.157   0.400  17.029

takes more than 100 times as long, is there a way to speed this up?

Marek
  • 49,472
  • 15
  • 99
  • 121
johannes
  • 14,043
  • 5
  • 40
  • 51

6 Answers6

22

You could use pmin, but you would have to get each column of your matrix into a separate vector. One way to do that is to convert it to a data.frame then call pmin via do.call (since data.frames are lists).

system.time(do.call(pmin, as.data.frame(m)))
#    user  system elapsed 
#   0.940   0.000   0.949 
system.time(apply(m,1,min))
#    user  system elapsed 
#   16.84    0.00   16.95
Joshua Ulrich
  • 173,410
  • 32
  • 338
  • 418
  • I like the use of `do.call`. I thought of `pmin`, but didn't think of a slick way to incorporate it. All the cool kids seem to be able to use `do.call` to achieve their goals...I need to do some reading on this. – Chase Jun 14 '11 at 03:08
  • `do.call` comes in handy when you want to be able to create function arguments dynamically (generally when the number of arguments passed via `...` isn't known). – Joshua Ulrich Jun 14 '11 at 03:17
  • 1
    Nice answer, thanks! with pmin.int() it was even a tiny bit faster – johannes Jun 14 '11 at 03:35
  • 1
    Hadley have nice vocabulary of [functions that you need to know](https://github.com/hadley/devtools/wiki/vocabulary). There is `pmin` too. – Marek Jun 14 '11 at 05:05
  • Some variation of your answer: `do.call(pmin, lapply(seq_len(ncol(m)), function(i) m[,i]))` – Marek Jun 14 '11 at 10:13
  • `@Marek`, how to handle NA values in your solution? – danas.zuokas May 18 '12 at 09:12
  • 1
    @danas.zuokas: `do.call(pmin, c(na.rm=TRUE, lapply(...)))` – Joshua Ulrich May 18 '12 at 11:31
18

Quite late to the party, but as the author of matrixStats and in case someone spots this, please note that matrixStats::rowMins() is very fast these days, e.g.

library(microbenchmark)
library(Biobase)     # rowMin()
library(matrixStats) # rowMins()
options(digits=3)

m <- matrix(rnorm(10000000), ncol=10) 

stats <- microbenchmark(
  rowMeans(m), ## A benchmark by OP
  rowMins(m),
  rowMin(m),
  do.call(pmin, as.data.frame(m)),
  apply(m, MARGIN=1L, FUN=min),
  times=10
)

> stats
Unit: milliseconds
                             expr    min     lq   mean median     uq    max
                      rowMeans(m)   77.7   82.7   85.7   84.4   90.3   98.2
                       rowMins(m)   72.9   74.1   88.0   79.0   90.2  147.4
                        rowMin(m)  341.1  347.1  395.9  383.4  395.1  607.7
  do.call(pmin, as.data.frame(m))  326.4  357.0  435.4  401.0  437.6  657.9
 apply(m, MARGIN = 1L, FUN = min) 3761.9 3963.8 4120.6 4109.8 4198.7 4567.4
HenrikB
  • 6,132
  • 31
  • 34
  • 1
    @HenirkB it would be great if matrixStats rowMins also worked on data.frames, (without the need of transform it to matrix first) – skan Jul 23 '15 at 19:51
  • 1
    @skan, unfortunately it's not obvious that this belongs to matrixStats for various reasons, please see https://github.com/HenrikBengtsson/matrixStats/issues/18 – HenrikB Jul 24 '15 at 20:51
9

If you want to stick to CRAN packages, then both the matrixStats and the fBasics packages have the function rowMins [note the s which is not in the Biobase function] and a variety of other row and column statistics.

Henry
  • 6,704
  • 2
  • 23
  • 39
5
library("sos")
findFn("rowMin")

gets a hit in the Biobase package, from Bioconductor ...

source("http://bioconductor.org/biocLite.R")
biocLite("Biobase")

m <- matrix(rnorm(10000000), ncol=10)
system.time(rowMeans(m))
##   user  system elapsed 
##  0.132   0.148   0.279 
system.time(apply(m,1,min))
##   user  system elapsed 
## 11.825   1.688  13.603
library(Biobase)
system.time(rowMin(m))
##    user  system elapsed 
##  0.688   0.172   0.864

Not as fast as rowMeans, but a lot faster than apply(...,1,min)

Ben Bolker
  • 211,554
  • 25
  • 370
  • 453
5

I've been meaning to try out the new compiler package in R 2.13.0. This essentially follows the post outlined by Dirk here.

library(compiler)
library(rbenchmark)
rowMin <- function(x, ind) apply(x, ind, min)
crowMin <- cmpfun(rowMin)

benchmark(
      rowMin(m,1)
    , crowMin(m,1)
    , columns=c("test", "replications","elapsed","relative")
    , order="relative"
    , replications=10)
)

And the results:

           test replications elapsed relative
2 crowMin(m, 1)           10 120.091   1.0000
1  rowMin(m, 1)           10 122.745   1.0221

Anticlimatic to say the least, though looks like you've gotten some other good options.

Chase
  • 67,710
  • 18
  • 144
  • 161
  • thanks for your answer, I will have to look deeper into your answer, that's new terrain for me :) – johannes Jun 14 '11 at 03:38
  • 1
    Compiler is better in optimization of explicit loops. Try e.g.: `rowMin <- function(x) {n <- nrow(x);r <- numeric(n);for (i in 1:n) r[i] <- min(x[i,]);r}` – Marek Jun 14 '11 at 05:22
2

Not particularly R-idiosyncratic, but surely the fastest method is just to use pmin and loop over columns: 

x <- m[,1]
for (i in 2:ncol(m)) x <- pmin(x, m[,i])

On my machine that takes just 3 times longer than rowMeans for the 1e+07x10 matrix, and is slightly faster than the do.call method via data.frame.

mdsumner
  • 29,099
  • 6
  • 83
  • 91