Log-Likelihood Visualization for Archimedean Copulas

In copula: Multivariate Dependence with Copulas

require(copula)
doExtras <- FALSE

Intro

This vignette visualizes (log) likelihood functions of Archimedean copulas, some of which are numerically challenging to compute. Because of this computational challenge, we also check for equivalence of some of the several computational methods, testing for numerical near-equality using all.equal(L1, L2).

Auxiliary functions

We start by defining the following auxiliary functions.

##' @title [m]inus Log-Likelihood for Archimedean Copulas ("fast version")
##' @param theta parameter (length 1 for our current families)
##' @param acop Archimedean copula (of class "acopula")
##' @param u data matrix n x d
##' @param n.MC if > 0 MC is applied with sample size equal to n.MC; otherwise,
##'        the exact formula is used
##' @param ... potential further arguments, passed to <acop> @dacopula()
##' @return negative log-likelihood
##' @author Martin Maechler (Marius originally)
mLogL <- function(theta, acop, u, n.MC=0, ...) { # -(log-likelihood)
    -sum(acop@dacopula(u, theta, n.MC=n.MC, log=TRUE, ...))
}

##' @title Plotting the Negative Log-Likelihood for Archimedean Copulas
##' @param cop an outer_nacopula (currently with no children)
##' @param u n x d  data matrix
##' @param xlim x-range for curve() plotting
##' @param main title for curve()
##' @param XtrArgs a list of further arguments for mLogL()
##' @param ... further arguments for curve()
##' @return invisible()
##' @author Martin Maechler
curveLogL <- function(cop, u, xlim, main, XtrArgs=list(), ...) {
    unam <- deparse(substitute(u))
    stopifnot(is(cop, "outer_nacopula"), is.list(XtrArgs),
              (d <- ncol(u)) >= 2, d == dim(cop),
              length(cop@childCops) == 0# not yet *nested* A.copulas
              )
    acop <- cop@copula
    th. <- acop@theta # the true theta
    acop <- setTheta(acop, NA) # so it's clear, the true theta is not used below
    if(missing(main)) {
        tau. <- cop@copula@tau(th.)
        main <- substitute("Neg. Log Lik."~ -italic(l)(theta, UU) ~ TXT ~~
               FUN(theta['*'] == Th) %=>% tau['*'] == Tau,
               list(UU = unam,
                TXT= sprintf("; n=%d, d=%d;  A.cop",
                         nrow(u), d),
                FUN = acop@name,
                Th = format(th.,digits=3),
                Tau = format(tau., digits=3)))
    }
    r <- curve(do.call(Vectorize(mLogL, "theta"), c(list(x, acop, u), XtrArgs)),
               xlim=xlim, main=main,
               xlab = expression(theta),
               ylab = substitute(- log(L(theta, u, ~~ COP)), list(COP=acop@name)),
               ...)
    if(is.finite(th.))
        axis(1, at = th., labels=expression(theta["*"]),
             lwd=2, col="dark gray", tck = -1/30)
    else warning("non-finite cop@copula@theta = ", th.)
    axis(1, at = initOpt(acop@name),
         labels = FALSE, lwd = 2, col = 2, tck = 1/20)
    invisible(r)
}

Ensure that we are told about it, if the numerical algorithms choose methods using Rmpfr (R package interfacing to multi precision arithmetic MPFR):

op <- options("copula:verboseUsingRmpfr"=TRUE) # see when "Rmpfr" methods are chosen automatically

Joe's family

Easy case ($\tau=0.2$)

n <- 200
d <- 100
tau <- 0.2
theta <- copJoe@iTau(tau)
cop <- onacopulaL("Joe", list(theta,1:d))
theta

Here, the three different methods work "the same":

set.seed(1)
U1 <- rnacopula(n,cop)
enacopula(U1, cop, "mle") # 1.432885 --  fine
th4 <- 1 + (1:4)/4
mL.tr <- c(-3558.5, -3734.4, -3299.5, -2505.)
mLt1 <- sapply(th4, function(th) mLogL(th, cop@copula, U1, method="log.poly")) # default
mLt2 <- sapply(th4, function(th) mLogL(th, cop@copula, U1, method="log1p"))
mLt3 <- sapply(th4, function(th) mLogL(th, cop@copula, U1, method="poly"))
stopifnot(all.equal(mLt1, mL.tr, tolerance=5e-5),
          all.equal(mLt2, mL.tr, tolerance=5e-5),
          all.equal(mLt3, mL.tr, tolerance=5e-5))
system.time(r1l  <- curveLogL(cop, U1, c(1, 2.5), X=list(method="log.poly")))
mtext("all three polyJ() methods on top of each other")
system.time({
    r1J <- curveLogL(cop, U1, c(1, 2.5), X=list(method="poly"),
                     add=TRUE, col=adjustcolor("red", .4))
    r1m  <- curveLogL(cop, U1, c(1, 2.5), X=list(method="log1p"),
                      add=TRUE, col=adjustcolor("blue",.5))
})

U2 <- rnacopula(n,cop)
summary(dCopula(U2, cop)) # => density for the *correct* parameter looks okay
## hmm: max = 5.5e177
if(doExtras)
    system.time(r2 <- curveLogL(cop, U2, c(1, 2.5)))
stopifnot(all.equal(enacopula(U2, cop, "mle"), 1.43992755, tolerance=1e-5),
          all.equal(mLogL(1.8, cop@copula, U2), -4070.1953,tolerance=1e-5)) # (was -Inf)

U3 <- rnacopula(n,cop)
(th. <- enacopula(U3, cop, "mle")) # 1.4495
system.time(r3 <- curveLogL(cop, U3, c(1, 2.5)))
axis(1, at = th., label = quote(hat(theta)))

U4 <- rnacopula(n,cop)
enacopula(U4, cop, "mle") # 1.4519  (prev. was 2.351 : "completely wrong")
summary(dCopula(U4, cop)) # ok (had one Inf)
if(doExtras)
    system.time(r4 <- curveLogL(cop, U4, c(1, 2.5)))
mLogL(2.2351, cop@copula, U4)
mLogL(1.5,    cop@copula, U4)
mLogL(1.2,    cop@copula, U4)
if(doExtras) # each curve takes almost 2 sec
    system.time({
        curveLogL(cop, U4, c(1, 1.01))
        curveLogL(cop, U4, c(1, 1.0001))
        curveLogL(cop, U4, c(1, 1.000001))
    })
## --> limit goes *VERY* steeply up to  0
## --> theta 1.164 is about the boundary:
stopifnot(identical(setTheta(cop, 1.164), onacopula(cop@copula, C(1.164, 1:100))),
      all.equal(600.59577,
            cop@copula@dacopula(U4[118,,drop=FALSE],
                    theta=1.164, log = TRUE), tolerance=1e-5)) # was "Inf"

Harder case ($d=150$, $\tau=0.3$)

n <- 200
d <- 150
tau <- 0.3
(theta <- copJoe@iTau(tau))
cop <- onacopulaL("Joe",list(theta,1:d))

set.seed(47)
U. <- rnacopula(n,cop)
enacopula(U., cop, "mle") # 1.784578
system.time(r. <- curveLogL(cop, U., c(1.1, 3)))
## => still looks very good

Even harder case ($d=180$, $\tau=0.4$)

d <- 180
tau <- 0.4
(theta <- copJoe@iTau(tau))
cop <- onacopulaL("Joe",list(theta,1:d))

U. <- rnacopula(n,cop)
enacopula(U., cop, "mle") # 2.217582
if(doExtras)
system.time(r. <- curveLogL(cop, U., c(1.1, 4)))
## => still looks very good

Gumbel's family

Easy case ($\tau=0.2$)

n <- 200
d <- 50 # smaller 'd' -- so as to not need 'Rmpfr' here
tau <- 0.2
(theta <- copGumbel@iTau(tau))
cop <- onacopulaL("Gumbel",list(theta,1:d))

set.seed(1)
U1 <- rnacopula(n,cop)
if(doExtras) {
    U2 <- rnacopula(n,cop)
    U3 <- rnacopula(n,cop)
}
enacopula(U1, cop, "mle") # 1.227659 (was 1.241927)
##--> Plots with "many" likelihood evaluations
system.time(r1 <- curveLogL(cop, U1, c(1, 2.1)))
if(doExtras) system.time({
    mtext("and two other generated samples")
    r2 <- curveLogL(cop, U2, c(1, 2.1), add=TRUE)
    r3 <- curveLogL(cop, U3, c(1, 2.1), add=TRUE)
})

Harder case ($d=150$, $\tau=0.6$)

d <- 150
tau <- 0.6
(theta <- copGumbel@iTau(tau))
cG.5 <- onacopulaL("Gumbel",list(theta,1:d))

set.seed(17)
U4 <- rnacopula(n,cG.5)
U5 <- rnacopula(n,cG.5)
U6 <- rnacopula(n,cG.5)
if(doExtras) { ## "Rmpfr" is used {2012-06-21}: -- therefore about 18 seconds!
 tol <- if(interactive()) 1e-12 else 1e-8
 print(system.time(
 ee. <- c(enacopula(U4, cG.5, "mle", tol=tol),
          enacopula(U5, cG.5, "mle", tol=tol),
          enacopula(U6, cG.5, "mle", tol=tol))))
dput(ee.)# in case the following fails
## tol=1e-12 Linux nb-mm3 3.2.0-25-generic x86_64 (2012-06-23):
##   c(2.47567251789004, 2.48424484287686, 2.50410767129408)
##   c(2.475672518,      2.484244763,      2.504107671),
stopifnot(all.equal(ee., c(2.475672518, 2.484244763, 2.504107671),
            tolerance= max(1e-7, 16*tol)))
}
## --> Plots with "many" likelihood evaluations
th. <- seq(1, 3, by= 1/4)
if(doExtras) # "default2012" (polyG default) partly uses Rmpfr here:
system.time(r4   <- sapply(th., mLogL, acop=cG.5@copula, u=U4))## 25.6 sec
## whereas this (polyG method) is very fast {and still ok}:
system.time(r4.p <- sapply(th., mLogL, acop=cG.5@copula, u=U4, method="pois"))
r4. <- c(0, -18375.33, -21948.033, -24294.995, -25775.502,
         -26562.609, -26772.767, -26490.809, -25781.224)
stopifnot(!doExtras ||
          all.equal(r4,   r4., tolerance = 8e-8),
          all.equal(r4.p, r4., tolerance = 8e-8))
## --> use fast method here as well:
system.time(r5.p <- sapply(th., mLogL, acop=cG.5@copula, u=U5, method="pois"))
system.time(r6.p <- sapply(th., mLogL, acop=cG.5@copula, u=U6, method="pois"))
if(doExtras) {
    if(FALSE) # for speed analysis, etc
        debug(copula:::polyG)
    mLogL(1.65, cG.5@copula, U4) # -23472.96
}
dd <- dCopula(U4, setTheta(cG.5, 1.64), log = TRUE,
              method = if(doExtras)"default" else "pois")
summary(dd)
stopifnot(!is.na(dd), # no NaN's anymore
      40 < range(dd), range(dd) < 710)

Frank's family (an already hard case)

n <- 64
d <- 5
tau <- 0.8
(theta <- copFrank@iTau(tau))
cop <- onacopulaL("Frank", list(theta,1:d))

set.seed(11) # these seeds give no problems: 101, 41, 21
U. <- rnacopula(n,cop)
cop@copula <- setTheta(cop@copula, NA) # forget the true theta
system.time(f.ML <- emle(U., cop)); f.ML # --> fine: theta = 18.033, Log-lik = 314.01
if(doExtras)
    system.time(f.mlMC <- emle(U., cop, n.MC = 1e4)) # with MC
stopifnot(all.equal(unname(coef(f.ML)), 18.03331, tolerance= 1e-6),
      all.equal(f.ML@min, -314.0143, tolerance=1e-6),
      !doExtras || ## Simulate MLE (= SMLE) is "extra" random,  hmm...
      all.equal(unname(coef(f.mlMC)), 17.8, tolerance= 0.01)
      ##           64-bit ubuntu: 17.817523
      ##         ? 64-bit Mac:    17.741
     )

cop@copula <- setTheta(cop@copula, theta)
r. <- curveLogL(cop, U., c(1, 200)) # => now looks fine
tail(as.data.frame(r.), 15)
stopifnot( is.finite( r.$y ),
      ## and is convex (everywhere):
      diff(r.$y, d=2) > 0)
options(op) # revert to previous state

Session information

print(sessionInfo(), locale=FALSE)

copula documentation built on Sept. 11, 2024, 7:48 p.m.