R/npnormadw.R

In xiangjiexue/npfixedcompR: Non-parametric estimation of mixing distribution with possibily fixed components

npnormadw = R6::R6Class("npnormadw",
                       inherit = npnorm,
                       public = list(
                         beta = 1,
                         type = "npnorm",
                         w = NULL,
                         h = 10^-3,
                         initialize = function(data, mu0, pi0, beta, order = -3){
                           v = bin(data, order = order)
                           self$data = v$v
                           self$w = v$w
                           self$len = length(self$data)
                           self$h = 10^order
                           if (!missing(mu0)) self$mu0fixed = mu0
                           if (!missing(pi0)) self$pi0fixed = pi0
                           if (!missing(beta)) self$beta = beta
                           a1 = (2 * cumsum(self$w) - self$w) / sum(self$w)
                           private$precompute = list(a1 = a1 * self$w, a2 = (2 - a1) * self$w,
                                                     precompute1 = pnpdiscnorm(self$data, mu0 = self$mu0fixed, pi0 = self$pi0fixed, sd = self$beta, h = self$h),
                                                     precompute2 = pnpdiscnorm(self$data, mu0 = self$mu0fixed, pi0 = self$pi0fixed, sd = self$beta, lower.tail = FALSE, h = self$h))
                         },
                         modified = function(mu0, pi0, beta){
                           if (!missing(mu0)) self$mu0fixed = mu0
                           if (!missing(pi0)) self$pi0fixed = pi0
                           if (!missing(beta)) self$beta = beta
                           private$precompute$precompute1 = pnpdiscnorm(self$data, mu0 = self$mu0fixed, pi0 = self$pi0fixed, sd = self$beta, h = self$h)
                           private$precompute$precompute2 = pnpdiscnorm(self$data, mu0 = self$mu0fixed, pi0 = self$pi0fixed, sd = self$beta, lower.tail = FALSE, h = self$h)
                         },
                         lossfunction = function(mu0, pi0){
                           -sum(log(pnpdiscnorm(self$data, mu0 = mu0, pi0 = pi0, sd = self$beta, h = self$h) + private$precompute$precompute1) * private$precompute$a1 +
                                  log(pnpdiscnorm(self$data, mu0 = mu0, pi0 = pi0, sd = self$beta, h = self$h, lower.tail = FALSE) + private$precompute$precompute2) * private$precompute$a2)
                         },
                         gradientfunction = function(mu, mu0, pi0, order = c(1, 0, 0)){
                           fullden = pnpdiscnorm(self$data, mu0 = mu0, pi0 = pi0, sd = self$beta, h = self$h) + private$precompute$precompute1
                           fullden2 = pnpdiscnorm(self$data, mu0 = mu0, pi0 = pi0, sd = self$beta, h = self$h, lower.tail = FALSE) + private$precompute$precompute2
                           addconst = -sum(private$precompute$precompute1 * private$precompute$a1 / fullden +
                                             private$precompute$precompute2 * private$precompute$a2 / fullden2) +
                             2 * sum(self$w)
                           murep = self$data - rep(mu, rep(self$len, length(mu)))
                           ans = vector("list", 3)
                           names(ans) = c("d0", "d1", "d2")
                           if (order[1] == 1){
                             temp1 = pdiscnorm(murep, sd = self$beta, h = self$h)
                             temp2 = pdiscnorm(murep, sd = self$beta, lower.tail = FALSE, h = self$h)
                             ans$d0 = .colSums(temp1 * private$precompute$a1 / fullden +
                                                 temp2 * private$precompute$a2 / fullden2,
                                               m = self$len, n = length(mu)) * -sum(pi0) + addconst
                           }
                           if (any(order[2:3] == 1)){
                             temp = dnorm(murep + self$h, sd = self$beta) * (private$precompute$a1 / fullden -
                                                                      private$precompute$a2 / fullden2)
                           }
                           if (order[2] == 1){
                             ans$d1 = .colSums(temp, m = self$len, n = length(mu)) * sum(pi0)
                           }
                           if (order[3] == 1){
                             ans$d2 = .colSums((murep + self$h) * temp, m = self$len, n = length(mu)) * sum(pi0) / self$beta^2
                           }

                           ans
                         },
                         computeweights = function(mu0, pi0, newweights, tol = 1e-6){
                           mu0new = c(mu0, newweights)
                           pi0new = c(pi0, rep(0, length(newweights)))
                           sf = pdiscnorm(self$data, mean = rep(mu0new, rep(self$len, length(mu0new))),
                                          sd = self$beta, h = self$h)
                           dim(sf) = c(self$len, length(mu0new))
                           ss = drop(sf %*% pi0new) + private$precompute$precompute1
                           S = sf / ss

                           uf = pdiscnorm(self$data, mean = rep(mu0new, rep(self$len, length(mu0new))),
                                          sd = self$beta, h = self$h, lower.tail = FALSE)
                           dim(uf) = c(self$len, length(mu0new))
                           us = drop(uf %*% pi0new) + private$precompute$precompute2
                           U = uf / us

                           Sma1 = S * private$precompute$a1
                           Uma2 = U * private$precompute$a2
                           S1 = crossprod(S, Sma1) + crossprod(U, Uma2)
                           S2 = colSums(Sma1) + colSums(Uma2)

                           nw = pnnqp(S1, -2 * S2 + crossprod(S, private$precompute$precompute1 * private$precompute$a1 / ss) +
                                        crossprod(U, private$precompute$precompute2 * private$precompute$a2 / us), sum = 1 - sum(self$pi0fixed))$x
                           r = self$checklossfunction(mu0new, pi0new, nw - pi0new, S2)
                           self$collapsemix(r$pt, r$pr, tol)
                         },
                         estpi0dS = function(){
                           private$S1 = list(S1 = pdiscnorm(self$data, sd = self$beta, h = self$h),
                                             S2 = pdiscnorm(self$data, sd = self$beta, lower.tail = FALSE, h = self$h))
                         },
                         estpi0d = function(mu0, pi0){
                           ans = vector("list", 2)
                           names(ans) = c("d2", "d3")
                           S1 = private$S1$S1 / pnpdiscnorm(self$data, mu0 = mu0, pi0 = pi0, sd = self$beta, h = self$h) - 1
                           S2 = private$S1$S2 / pnpdiscnorm(self$data, mu0 = mu0, pi0 = pi0, sd = self$beta, lower.tail = FALSE, h = self$h) - 1
                           ans$d2 = -sum(S1 * private$precompute$a1 + S2 * private$precompute$a2)
                           ans$d3 = sum(S1^2 * private$precompute$a1 + S2^2 * private$precompute$a2)
                           ans
                         },
                         estpi0 = function(val = qAD(0.5, lower.tail = FALSE), mix = NULL, tol = 1e-6, maxiter = 100, verbose = FALSE){
                           self$modified(pi0 = 1 - tol / 2)
                           self$computemixdist(mix = mix, tol = tol, maxiter = maxiter)
                           r1ll = self$result$ll
                           self$modified(pi0 = 0)
                           self$computemixdist(mix = mix, tol = tol, maxiter = maxiter)

                           if (r1ll - sum(self$w) < val){
                             self$result = list(iter = 0,
                                      family = self$type,
                                      max.gradient = self$gradientfunction(0, 0, 1, order = c(1, 0, 0))$d0,
                                      mix = list(pt = 0, pr = 1),
                                      ll = self$lossfunction(mu0 = 0, pi0 = 1),
                                      beta = self$beta,
                                      convergence = 0)
                           }else{
                             self$estpi0dS()
                             private$solveestpi0(init = dnpdiscnorm(0, mu0 = self$result$mix$pt, pi0 = self$result$mix$pr, sd = self$beta, h = self$h) / ddiscnorm(0, sd = self$beta, h = self$h),
                                                 val = -sum(x$w) - val, tol = tol, maxiter = maxiter, verbose = verbose)
                           }
                         }
                       ),
                       private = list(
                         precompute = NULL,
                         flexden = NULL,
                         mflag = "d2",
                         S1 = NULL
                       ))

#' @rdname makeobject
#' @export
makeobject.npnormadw = function(v, mu0, pi0, beta, order = -3){
  npnormadw$new(v, mu0, pi0, beta, order = order)
}