回顾以前的答案
在上一个答案中,我没有提到两种方法之间的区别。一般来说,如果我们选择最大似然推断,我建议使用MASS::fitdistr
,因为对于许多基本分布,它执行精确推断而不是数值优化。Doc of?fitdistr
说得很清楚:
For the Normal, log-Normal, geometric, exponential and Poisson
distributions the closed-form MLEs (and exact standard errors) are
used, and ‘start’ should not be supplied.
For all other distributions, direct optimization of the
log-likelihood is performed using ‘optim’. The estimated standard
errors are taken from the observed information matrix, calculated
by a numerical approximation. For one-dimensional problems the
Nelder-Mead method is used and for multi-dimensional problems the
BFGS method, unless arguments named ‘lower’ or ‘upper’ are
supplied (when ‘L-BFGS-B’ is used) or ‘method’ is supplied
explicitly.
另一方面fitdistrplus::fitdist
,即使存在精确推理,也始终以数字方式执行推理。当然,这样做的好处fitdist
是可以使用更多的推理原理:
Fit of univariate distributions to non-censored data by maximum
likelihood (mle), moment matching (mme), quantile matching (qme)
or maximizing goodness-of-fit estimation (mge).
本答案的目的
这个答案将探索正态分布的精确推断。它会有理论的味道,但没有可能性原理的证明;只给出结果。基于这些结果,我们编写了自己的 R 函数进行精确推理,可以与MASS::fitdistr
. 另一方面,为了与 比较fitdistrplus::fitdist
,我们使用optim
数值最小化负对数似然函数。
这是学习统计和相对高级使用optim
. 为方便起见,我将估计尺度参数:方差,而不是标准误差。
正态分布的精确推断
自己编写推理函数
下面的代码注释得很好。有一个开关exact
。如果设置FALSE
,则选择数值解。
## fitting a normal distribution
fitnormal <- function (x, exact = TRUE) {
if (exact) {
################################################
## Exact inference based on likelihood theory ##
################################################
## minimum negative log-likelihood (maximum log-likelihood) estimator of `mu` and `phi = sigma ^ 2`
n <- length(x)
mu <- sum(x) / n
phi <- crossprod(x - mu)[1L] / n # (a bised estimator, though)
## inverse of Fisher information matrix evaluated at MLE
invI <- matrix(c(phi, 0, 0, phi * phi), 2L,
dimnames = list(c("mu", "sigma2"), c("mu", "sigma2")))
## log-likelihood at MLE
loglik <- -(n / 2) * (log(2 * pi * phi) + 1)
## return
return(list(theta = c(mu = mu, sigma2 = phi), vcov = invI, loglik = loglik, n = n))
}
else {
##################################################################
## Numerical optimization by minimizing negative log-likelihood ##
##################################################################
## negative log-likelihood function
## define `theta = c(mu, phi)` in order to use `optim`
nllik <- function (theta, x) {
(length(x) / 2) * log(2 * pi * theta[2]) + crossprod(x - theta[1])[1] / (2 * theta[2])
}
## gradient function (remember to flip the sign when using partial derivative result of log-likelihood)
## define `theta = c(mu, phi)` in order to use `optim`
gradient <- function (theta, x) {
pl2pmu <- -sum(x - theta[1]) / theta[2]
pl2pphi <- -crossprod(x - theta[1])[1] / (2 * theta[2] ^ 2) + length(x) / (2 * theta[2])
c(pl2pmu, pl2pphi)
}
## ask `optim` to return Hessian matrix by `hessian = TRUE`
## use "..." part to pass `x` as additional / further argument to "fn" and "gn"
## note, we want `phi` as positive so box constraint is used, with "L-BFGS-B" method chosen
init <- c(sample(x, 1), sample(abs(x) + 0.1, 1)) ## arbitrary valid starting values
z <- optim(par = init, fn = nllik, gr = gradient, x = x, lower = c(-Inf, 0), method = "L-BFGS-B", hessian = TRUE)
## post processing ##
theta <- z$par
loglik <- -z$value ## flip the sign to get log-likelihood
n <- length(x)
## Fisher information matrix (don't flip the sign as this is the Hessian for negative log-likelihood)
I <- z$hessian / n ## remember to take average to get mean
invI <- solve(I, diag(2L)) ## numerical inverse
dimnames(invI) <- list(c("mu", "sigma2"), c("mu", "sigma2"))
## return
return(list(theta = theta, vcov = invI, loglik = loglik, n = n))
}
}
我们仍然使用之前的数据进行测试:
set.seed(0); x <- rnorm(500)
## exact inference
fit <- fitnormal(x)
#$theta
# mu sigma2
#-0.0002000485 0.9773790969
#
#$vcov
# mu sigma2
#mu 0.9773791 0.0000000
#sigma2 0.0000000 0.9552699
#
#$loglik
#[1] -703.7491
#
#$n
#[1] 500
hist(x, prob = TRUE)
curve(dnorm(x, fit$theta[1], sqrt(fit$theta[2])), add = TRUE, col = 2)
数值方法也相当准确,只是方差协方差在对角线上没有精确的 0:
fitnormal(x, FALSE)
#$theta
#[1] -0.0002235315 0.9773732277
#
#$vcov
# mu sigma2
#mu 9.773826e-01 5.359978e-06
#sigma2 5.359978e-06 1.910561e+00
#
#$loglik
#[1] -703.7491
#
#$n
#[1] 500