A Bayesian lower bound that limits the mean squared error (Bayes risk) of any estimator of a parameter under a specified prior. It expresses that, under standard regularity and integrability conditions on the likelihood and prior, the expected estimation error covariance is bounded below by the inverse of the sum of the Fisher information of the likelihood (averaged over the prior) and the Fisher information contributed by the prior.

Principle

The minimum achievable mean-squared estimation error (averaged over data and prior) cannot be smaller than a bound determined by the combined information from the data (expected Fisher information) and the prior (prior Fisher information); increasing either source of information tightens (lowers) the bound.

Demonstration

Illustrative scenario → A scalar parameter θ has prior p(θ) with finite Fisher information and observations X with likelihood p(X|θ) satisfying differentiability and integrability. Recognition → The estimator designer computes the expected Fisher information E_{p(θ)}[I(θ)] and the prior information J_prior. Action → Apply Van Trees inequality to obtain a lower bound on E[(θ̂(X)−θ)^2]. Consequence → No estimator, regardless of bias, can attain a smaller Bayes mean-squared error than that bound under the stated regularity.

Misapplication

Treating the bound as an attainable equality in all cases or applying it when regularity conditions fail (e.g., parameter-dependent support, non-differentiable likelihood, improper prior) — the error is assuming the inequality yields an achievable equality or is valid without verifying differentiability, integrability and boundary conditions.

Consequence

Provides a principled target for estimator design and experiment selection: improving model or prior information (within validity conditions) reduces the lower bound on Bayes MSE; it also distinguishes what is impossible regardless of algorithmic effort.

Reversal

If regularity conditions fail (for example, the likelihood support depends on the parameter, the prior is improper, or derivatives are undefined), the Van Trees inequality may not hold; in some nonregular Bayesian problems alternative bounds or direct Bayes risk computation are required.

Boundary

Within: parametric estimation problems with proper priors and likelihoods satisfying differentiability, integrability and boundary conditions. Boundary case: priors with heavy tails where prior Fisher information is finite/near-infinite, requiring careful verification. Outside: nonparametric problems without a finite-dimensional parameterization, improper priors, or models with parameter-dependent support violating required integrals.

Semantic Tension

Tension between frequentist bounds (classical Cramér–Rao for unbiased estimators) and Bayesian bounds: Van Trees incorporates prior information but depends on prior choice, while frequentist bounds avoid priors but have different applicability.

Synthesis

Van Trees formalizes how data and prior information jointly limit achievable average estimation accuracy: it is a statement about unavoidable Bayes risk under model regularity rather than a constructive estimator recipe.