How does the Occam's razor penalty change if you use an overly broad prior on the extra parameters?

This is a great question about Bayesian model comparison, but it requires some setup to answer properly.

In Bayesian inference, when you add extra parameters to a model, you also specify a prior probability distribution over those new parameters. The "automatic Occam's razor" emerges because the posterior must integrate (average) over all possible values of those extra parameters weighted by both the likelihood and the prior [AG-2026.04-1269]. Here's the key insight: if your prior is very broad—assigning substantial probability density across a wide range of parameter space—the model effectively has more "room to wiggle," which actually *weakens* the Occam penalty rather than strengthening it.

More concretely: a narrower prior on extra parameters concentrates probability density in a smaller region, so the model makes more focused predictions and the posterior mass concentrates more sharply. This leaves less probability "budget" to fit diverse datasets, penalizing the model more heavily when data don't align with those predictions. An overly broad prior spreads that density thin across a large space, so the model can accommodate a wider range of data without the same concentration cost—the automatic razor becomes duller [AG-2026.04-1269].

This isn't just theoretical. In cosmological parameter estimation using effective field theory, careless choice of priors on nuisance parameters (extra degrees of freedom) can introduce systematic bias—not just statistical uncertainty—because the prior volume affects how well the model actually constrains the parameters of real interest [AG-2024.10-1362]. Similarly, in gravitational-wave astronomy, the choice of spin priors can substantially shift inferred population properties when the data themselves are weak [AG-2026.02-276].

The provided preprints don't directly address how to optimally trade off prior width against model complexity, which would be the natural follow-up question.