如何有效地生成对称矩阵的下三角指数

时间:2014-01-03 07:24:57

标签: r matrix indices triangular

我需要生成下三角矩阵索引(行和列对)。当对称矩阵变大(超过50K行)时,当前的实现是低效的(记忆明智的)。还有更好的方法吗?

rows <- 2e+01
id <- which(lower.tri(matrix(, rows, rows)) == TRUE, arr.ind=T)
head(id)

#      row col
# [1,]   2   1
# [2,]   3   1
# [3,]   4   1
# [4,]   5   1
# [5,]   6   1
# [6,]   7   1

2 个答案:

答案 0 :(得分:6)

这是另一种方法:

z <- sequence(rows)
cbind(
  row = unlist(lapply(2:rows, function(x) x:rows), use.names = FALSE),
  col = rep(z[-length(z)], times = rev(tail(z, -1))-1))

数据较大的基准:

library(microbenchmark)

rows <- 1000
m <- matrix(, rows, rows)

## Your current approach
fun1 <- function() which(lower.tri(m) == TRUE, arr.ind=TRUE)

## An improvement of your current approach
fun2 <- function() which(lower.tri(m), arr.ind = TRUE)

## The approach shared in this answer
fun3 <- function() {
  z <- sequence(rows)
  cbind(
    row = unlist(lapply(2:rows, function(x) x:rows), use.names = FALSE),
    col = rep(z[-length(z)], times = rev(tail(z, -1))-1))
}

## Sven's answer
fun4 <- function() {
  row <- rev(abs(sequence(seq.int(rows - 1)) - rows) + 1)
  col <- rep.int(seq.int(rows - 1), rev(seq.int(rows - 1)))
  cbind(row, col)
}

microbenchmark(fun1(), fun2(), fun3(), fun4())
# Unit: milliseconds
#    expr       min        lq   median       uq       max neval
#  fun1() 77.813577 85.343356 90.60689 95.71648 130.40059   100
#  fun2() 73.812204 82.103600 85.87555 90.59235 138.66547   100
#  fun3()  9.016237  9.382506 10.63291 13.20085  55.42137   100
#  fun4() 20.591863 24.999702 28.82232 31.90663  65.05169   100

答案 1 :(得分:2)

您的方法非常慢,因为必须创建多个矩阵。您可以使用matrix创建第一个矩阵。函数lower.tri在内部创建3个矩阵。结果与TRUE的比较创建了第五个矩阵。顺便说一句:与TRUE的比较是不必要的。

以下方法不会创建任何矩阵,而是计算索引:

rows <- 2e+01 # number of rows and columns (20)

x <- rev(abs(sequence(seq.int(rows - 1)) - rows) + 1)
y <- rep.int(seq.int(rows - 1), rev(seq.int(rows - 1)))

idx <- cbind(x, y)

(如果您想要稍微快一些的方法,可以将seq.int(rows - 1)的结果分配给变量,而不是使用此命令三次。)

与原始解决方案比较:

id <- which(lower.tri(matrix(, rows, rows)) == TRUE, arr.ind=T)

all(id == idx)
# TRUE