R语言中mle()与fitdist()极大似然估计报错求助：初始值非有限

Alright, let's tackle this initial value in 'vmmin' is not finite error you're hitting with mle() and fitdist()—I’ve seen this pop up a lot when working with custom CDFs for maximum likelihood estimation, so let’s break down the most likely fixes step by step.

1. First, audit your custom distribution functions (PDF/CDF) for non-finite values #

You mentioned using a CDF structured like 1-((1-q)*...)—the first thing to check is whether this calculation returns valid, finite values when paired with your data and initial parameter guesses. Common pitfalls here include:

• Division by zero in hidden parts of your formula
• Logarithms of non-positive numbers (if your PDF/CDF uses logs)
• Exponentiation that blows up to Inf or collapses to -Inf

Test this manually with your initial parameter guesses:

# Replace with your actual CDF/PDF function and parameter values
test_cdf_vals <- your_custom_cdf(data, param1 = 0.5, param2 = 2)
any(!is.finite(test_cdf_vals)) # If this returns TRUE, your function has a problem

Also, remember: fitdist() requires a valid PDF (specified via ddistname), not just a CDF. If you’ve only defined the CDF and skipped the PDF, or the PDF calculation is incorrect, this will break the likelihood calculation immediately.

2. Fix your initial parameter guesses (vstart) #

This error almost always ties back to poor initial parameter values that send the optimization routine (which both mle() and fitdist() use under the hood) into a region where the likelihood is non-finite. Here’s how to fix this:

• Use theoretical bounds: If your parameters have natural limits (e.g., a probability parameter must be between 0 and 1, a scale parameter must be positive), start with values inside those ranges.
• Use moment estimation: Calculate sample moments (mean, variance, median) of your data, then reverse-engineer rough parameter values that align with those moments for your distribution.
• Avoid extreme values: Steer clear of 0, 1e6, or other edge cases that might trigger singularities in your distribution formula.

3. Add parameter bounds to restrict optimization #

If your parameters have hard limits, don’t let optim() wander into invalid territory. Use the lower and upper arguments in fitdist() or mle() to enforce these bounds:

# Example: param1 is a probability (0-1), param2 is a positive scale parameter
fitdist(
  data = your_data,
  distr = "your_custom_dist",
  start = list(param1 = 0.3, param2 = 1.5),
  lower = c(0, 0.01), # Prevent param1=0 or param2=0
  upper = c(1, Inf)
)

This keeps the optimization routine within valid parameter space, eliminating non-finite likelihood calculations at the start.

4. Verify your likelihood function setup (for mle()) #

If you’re using mle() with a custom likelihood function, double-check that you’re calculating the negative log-likelihood correctly. The standard structure should be:

neg_log_lik <- function(params) {
  # Extract parameters
  param1 <- params[1]
  param2 <- params[2]
  # Calculate PDF values for all data points
  pdf_vals <- your_custom_pdf(your_data, param1, param2)
  # Return negative sum of log-PDFs (optim minimizes this)
  -sum(log(pdf_vals))
}

If you accidentally use the CDF instead of the PDF, or flip the sign (maximizing instead of minimizing), you’ll get non-finite values or failed optimization.

5. Debug directly with optim() #

If you’re still stuck, bypass mle()/fitdist() and run the underlying optim() call manually to get clearer error context:

optim_result <- optim(
  par = your_vstart_params,
  fn = neg_log_lik,
  lower = c(0, 0.01),
  upper = c(1, Inf),
  method = "L-BFGS-B" # Use bounded method if you have limits
)

# Check the output for clues
print(optim_result$message)
print(optim_result$value)

This will tell you exactly where the optimization is failing—whether it’s the initial value, a parameter boundary issue, or a problem with your likelihood function.

内容的提问来源于stack exchange，提问作者Elena Glamozdina