Can't sample from posterior. Error: Log probability evaluates to log(0), i.e. negative infinity

Modeling
#1

I am trying to denoise N points (CTCP_1,…,CTCP_N) given the noisy inputs (ECTCP_1,…,ECTCP_N). Each point CTCP_i is distributed normally with mean zero and one of C standard deviations gamma_1, …, gamma_C. Conditional on CTCP_i, ECTCP_i is normal with mean CTCP_i and standard deviation beta_i, where the input data
ETSE_i^2 = beta_i^2 + gamma_i^2
I am having difficulty sample from the posterior given this structure. Any help would be greatly appreciated.
Thanks.

data {
  int<lower=1> N; // Number of data points
  int<lower=1> K; // Number of dimensions
  int<lower=1> C; // Number of 
  vector[K] ECTCP[N]; // Input data
  vector<lower=0>[K] ETSE[N]; // Estimation error
  int<lower=1, upper=C> mem[N]; // Membership for each datapoint, choices from 1 to C
}
parameters {
  vector[K] CTCP[N];
  vector<lower=0>[K] gamma[N];
  cholesky_factor_corr[K] L_Omega; //Cholesky factorization of Area Correlation
}
model {
  matrix[K, K] L_ASD[C]; //Area Covariance Cholesky Factor
  vector[K] beta[N];
  
  L_Omega ~ lkj_corr_cholesky(1);
  
  for (c in 1:C) {
    for (k in 1:K) {
      gamma[c][k] ~ std_normal();
    }
    L_ASD[c] = diag_pre_multiply(gamma[c], L_Omega);
  }
  
  for (n in 1:N) {
    for (k in 1:K) {
      beta[n][k] = sqrt(square(ETSE[n][k])-square(gamma[mem[n]][k]));
    }
  }

  for (n in 1:N) {
    CTCP[n] ~ multi_normal_cholesky([0,0]', L_ASD[mem[n]]);
    for (k in 1:K) {
      ECTCP[n][k] ~ normal(CTCP[n][k], beta[n][k]);
    }
  }
}
generated quantities { //Like transformed parameters, but included in output
 corr_matrix[K] Omega;
 Omega = multiply_lower_tri_self_transpose(L_Omega);
}

Edited by @jsocolar for syntax highlighting.

#2

Are you able to either add some print statements or use ReddingStan to see what parameter values lead to the -Inf target?

1 Like