Using SS SER with alpha=0 and beta>0 using L=1 leads to infinite ELBO

[william-denault]

Using SS SER with alpha=0 and beta>0 using L=1 leads to infinite ELBO @pcarbo

library(susieR)
set.seed(1)
n    <- 1000
p    <- 1000
beta <- rep(0,p)
beta[c(1,2,300,400)] <- 1
X   <- matrix(rnorm(n*p),nrow=n,ncol=p)
y   <- X %*% beta + rnorm(n)
res <- susie(X,y,L=1 ,
             min_abs_corr = 0,
             check_null_threshold = -1000,
             estimate_residual_method =  "Servin_Stephens" ,
             alpha0 = .0 ,
             beta0 =.5)
plot(coef(res)[-1],pch = 20)







library(susieR)
set.seed(1)
n    <- 1000
p    <- 1000
beta <- rep(0,p)
beta[c(1,2,300,400)] <- 1
X   <- matrix(rnorm(n*p),nrow=n,ncol=p)
y   <- X %*% beta + rnorm(n)
res <- susie(X,y,L=2 ,
             min_abs_corr = 0,
             check_null_threshold = -1000,
             estimate_residual_method =  "Servin_Stephens" ,
             alpha0 = .0 ,
             beta0 =.5)
plot(coef(res)[-1],pch = 20)

note that both alpha= 0 and beta=0 leads to no problem

library(susieR)
set.seed(1)
n    <- 1000
p    <- 1000
beta <- rep(0,p)
beta[c(1,2,300,400)] <- 1
X   <- matrix(rnorm(n*p),nrow=n,ncol=p)
y   <- X %*% beta + rnorm(n)
res <- susie(X,y,L=1 ,
             min_abs_corr = 0,
             check_null_threshold = -1000,
             estimate_residual_method =  "Servin_Stephens" ,
             alpha0 = .0 ,
             beta0 =.0)
plot(coef(res)[-1],pch = 20)

[pcarbo]

Thank's right, this is expected; the ELBO is not finite when alpha = 0, beta = 0.

I suppose we should throw a warning or error when alpha and beta are zero @alexmccreight @gaow? (I'd have to check, but my guess is that you get an infinite ELBO when either alpha or beta are zero?)