Made small adjustment to mixSQP().

DESCRIPTION

@@ -13,8 +13,8 @@ Authors@R: c(person("Matthew","Stephens",role="aut",
              person("Jason","Willwerscheid",role="aut"),
              person("Nan","Xiao",role="aut"),
              person("Mazon","Zeng",role="ctb"))
-Version: 2.2-35
-Date: 2019-04-17
+Version: 2.2-37
+Date: 2019-04-30
 Title: Methods for Adaptive Shrinkage, using Empirical Bayes
 Description: The R package 'ashr' implements an Empirical Bayes
     approach for large-scale hypothesis testing and false discovery

R/mix_opt.R

@@ -102,11 +102,6 @@ mixSQP <- function (matrix_lik, prior, pi_init = NULL,
   if (is.null(pi_init))
     pi_init <- rep(1,k)
 
-  # Always as a small "tweak" to the initial estimates of the mixture
-  # weights to ensure more reliable convergence.
-  pi_init <- pi_init + 0.01
-  pi_init <- pi_init/sum(pi_init)
-
   # Add in observations corresponding to the prior.
   A <- rbind(diag(k),matrix_lik) 
   w <- c(prior - 1,weights)
@@ -116,11 +111,16 @@ mixSQP <- function (matrix_lik, prior, pi_init = NULL,
   # Fit the mixture weights using the mix-SQP algorithm.
   control0 <- list(verbose = FALSE)
   control  <- modifyList(control0,control,keep.null = TRUE)
-  out <- mixsqp::mixsqp(A,w,pi_init,control = control)
-  
+  out      <- mixsqp::mixsqp(A,w,pi_init,control = control)
+
   # Return the fitted mixture weights, and some other information
   # about the optimization step.
-  return(list(pihat     = out$x,
+  #
+  # Note that the mixture weights may not be normalized (i.e., they
+  # may not sum to 1) if the mix-SQP algorithm terminates prematurely,
+  # so to be extra cautious, the mix-SQP solution is normalized before
+  # returning it.
+  return(list(pihat     = out$x/sum(out$x),
               niter     = nrow(out$data),
               converged = (out$status == mixsqp.status.converged),
               control   = control))