lsum: Properly Compute the Logarithm of a Sum (of Exponentials)

In DPQ: Density, Probability, Quantile ('DPQ') Computations

Properly Compute the Logarithm of a Sum (of Exponentials)

Description

Properly compute \log(x_1 + \ldots + x_n). for given log(x_1),..,log(x_n). Here, x_i > 0 for all i.

If the inputs are denoted l_i = log(x_i) for i = 1,2,..,n, we compute log(sum(exp(l[]))), numerically stably.

Simple vector version of copula:::lsum() (CRAN package copula).

Usage

lsum(lx, l.off = max(lx))

Arguments

lx	n-vector of values log(x_1),..,log(x_n).
l.off	the offset to substract and re-add; ideally in the order of the maximum of each column.

Value

log(x_1 + .. + x_n) = log(sum(x)) = log(sum(exp(log(x)))) = = log(exp(log(x_max))*sum(exp(log(x)-log(x_max)))) = = log(x_max) + log(sum(exp(log(x)-log(x_max))))) = = lx.max + log(sum(exp(lx-lx.max)))

Author(s)

Originally, via paired programming: Marius Hofert and Martin Maechler.

See Also

lssum() which computes a sum in log scale with specified (typically alternating) signs.

Examples

## The "naive" version :
lsum0 <- function(lx) log(sum(exp(lx)))

lx1 <- 10*(-80:70) # is easy
lx2 <- 600:750     # lsum0() not ok [could work with rescaling]
lx3 <- -(750:900)  # lsum0() = -Inf - not good enough
m3 <- cbind(lx1,lx2,lx3)
lx6 <- lx5 <- lx4 <- lx3
lx4[149:151] <- -Inf ## = log(0)
lx5[150] <- Inf
lx6[1] <- NA_real_
m6 <- cbind(m3,lx4,lx5,lx6)
stopifnot(exprs = {
  all.equal(lsum(lx1), lsum0(lx1))
  all.equal((ls1 <- lsum(lx1)),  700.000045400960403, tol=8e-16)
  all.equal((ls2 <- lsum(lx2)),  750.458675145387133, tol=8e-16)
  all.equal((ls3 <- lsum(lx3)), -749.541324854612867, tol=8e-16)
  ## identical: matrix-version <==> vector versions
  identical(lsum(lx4), ls3)
  identical(lsum(lx4), lsum(head(lx4, -3))) # the last three were -Inf
  identical(lsum(lx5), Inf)
  identical(lsum(lx6), lx6[1])
  identical((lm3 <- apply(m3, 2, lsum)), c(lx1=ls1, lx2=ls2, lx3=ls3))
  identical(apply(m6, 2, lsum), c(lm3, lx4=ls3, lx5=Inf, lx6=lx6[1]))
})

DPQ documentation built on Nov. 3, 2024, 3 a.m.