I would like to divide each row of a matrix by a fixed vector. For example
mat<-matrix(1,ncol=2,nrow=2,TRUE)
dev<-c(5,10)
Giving mat/dev divides each column by dev.
[,1] [,2]
[1,] 0.2 0.2
[2,] 0.1 0.1
However, I would like to have this as a result, i.e. do the operation row-wise :
rbind(mat[1,]/dev, mat[2,]/dev)
[,1] [,2]
[1,] 0.2 0.1
[2,] 0.2 0.1
Is there an explicit command to get there?
Here are a few ways in order of increasing code length:
t(t(mat) / dev)
mat / dev[col(mat)] # @DavidArenburg & @akrun
mat %*% diag(1 / dev)
sweep(mat, 2, dev, "/")
t(apply(mat, 1, "/", dev))
plyr::aaply(mat, 1, "/", dev)
mat / rep(dev, each = nrow(mat))
mat / t(replace(t(mat), TRUE, dev))
mapply("/", as.data.frame(mat), dev) # added later
mat / matrix(dev, nrow(mat), ncol(mat), byrow = TRUE) # added later
do.call(rbind, lapply(as.data.frame(t(mat)), "/", dev))
mat2 <- mat; for(i in seq_len(nrow(mat2))) mat2[i, ] <- mat2[i, ] / dev
All the solutions that begin with mat / also work if mat is a data frame and produce a data frame result. The same is also the case for the sweep solution and the last, i.e. mat2, solution. The mapply solutions works with data.frames but produces a matrix.
If mat is a plain vector rather than a matrix then either of these return a one column matrix
t(t(mat) / dev)
mat / t(replace(t(mat), TRUE, dev))
and this one returns a vector:
plyr::aaply(mat, 1, "/", dev)
The others give an error, warning or not the desired answer.
The brevity and clarity of the code may be more important than speed but for purposes of completeness here are some benchmarks using 10 repetitions and then 100 repetitions.
library(microbenchmark)
library(plyr)
set.seed(84789)
mat<-matrix(runif(1e6),nrow=1e5)
dev<-runif(10)
microbenchmark(times=10L,
"1" = t(t(mat) / dev),
"2" = mat %*% diag(1/dev),
"3" = sweep(mat, 2, dev, "/"),
"4" = t(apply(mat, 1, "/", dev)),
"5" = mat / rep(dev, each = nrow(mat)),
"6" = mat / t(replace(t(mat), TRUE, dev)),
"7" = aaply(mat, 1, "/", dev),
"8" = do.call(rbind, lapply(as.data.frame(t(mat)), "/", dev)),
"9" = {mat2 <- mat; for(i in seq_len(nrow(mat2))) mat2[i, ] <- mat2[i, ] / dev},
"10" = mat/dev[col(mat)])
giving:
Unit: milliseconds
expr min lq mean median uq max neval
1 7.957253 8.136799 44.13317 8.370418 8.597972 366.24246 10
2 4.678240 4.693771 10.11320 4.708153 4.720309 58.79537 10
3 15.594488 15.691104 16.38740 15.843637 16.559956 19.98246 10
4 96.616547 104.743737 124.94650 117.272493 134.852009 177.96882 10
5 17.631848 17.654821 18.98646 18.295586 20.120382 21.30338 10
6 19.097557 19.365944 27.78814 20.126037 43.322090 48.76881 10
7 8279.428898 8496.131747 8631.02530 8644.798642 8741.748155 9194.66980 10
8 509.528218 524.251103 570.81573 545.627522 568.929481 821.17562 10
9 161.240680 177.282664 188.30452 186.235811 193.250346 242.45495 10
10 7.713448 7.815545 11.86550 7.965811 8.807754 45.87518 10
Re-running the test on all those that took <20 milliseconds with 100 repetitions:
microbenchmark(times=100L,
"1" = t(t(mat) / dev),
"2" = mat %*% diag(1/dev),
"3" = sweep(mat, 2, dev, "/"),
"5" = mat / rep(dev, each = nrow(mat)),
"6" = mat / t(replace(t(mat), TRUE, dev)),
"10" = mat/dev[col(mat)])
giving:
Unit: milliseconds
expr min lq mean median uq max neval
1 8.010749 8.188459 13.972445 8.560578 10.197650 299.80328 100
2 4.672902 4.734321 5.802965 4.769501 4.985402 20.89999 100
3 15.224121 15.428518 18.707554 15.836116 17.064866 42.54882 100
5 17.625347 17.678850 21.464804 17.847698 18.209404 303.27342 100
6 19.158946 19.361413 22.907115 19.772479 21.142961 38.77585 100
10 7.754911 7.939305 9.971388 8.010871 8.324860 25.65829 100
So on both these tests #2 (using diag) is fastest. The reason may lie in its almost direct appeal to the BLAS, whereas #1 relies on the costlier t.
You're looking for the apply function, applied on the rows:
t(apply(mat, 1, function(x) x/dev))
Not only faster than the fastest base method (using matrix multiplication, see @G. Grothendieck's answer) but also more readable is the use of rray package that allows numpy-like broadcasting for division with its %b/% operator:
mat %b/% matrix(dev, nrow = 1)
#install.packages("rray")
library(rray)
set.seed(84789)
mat <- matrix(runif(1e6), nrow = 1e5)
dev <- runif(10)
bench::mark(rray = {ret <- mat %b/% matrix(dev, nrow = 1); attr(ret, "dimnames") <- NULL; ret},
matmult = mat %*% diag(1/dev)
)
#> # A tibble: 2 x 6
#> expression min median `itr/sec` mem_alloc `gc/sec`
#> <bch:expr> <bch:tm> <bch:tm> <dbl> <bch:byt> <dbl>
#> 1 rray 8.24ms 8.82ms 108. 7.67MB 46.4
#> 2 matmult 11.17ms 12.01ms 77.6 7.66MB 13.9
