∑Richard Hummel Math

INLA

Integrated Nested Laplace Approximation — a fast, deterministic alternative to MCMC for a broad class of Bayesian models, trading some generality for speed and reliable convergence.

INLA computes marginal posteriors for latent Gaussian models without drawing an MCMC chain
1. Modely depends on hidden xx is Gaussian given θ2. Approximatefind posterior modesfit local Gaussians3. Integratecombine over θ gridreturn marginalstarget: posterior marginals π(xᵢ | y) and π(θⱼ | y)deterministic approximation, specialized to latent Gaussian structure

Latent Gaussian model

yi | xi, θ and x | θ ∼ N(μ(θ), Q(θ)-1)

What INLA avoids

No posterior random walk; no Monte Carlo sampling noise.

Tradeoff

Fast when the model fits the template; less general than MCMC.
Definition

INLA (Integrated Nested Laplace Approximation) is a deterministic approximation method for Bayesian latent Gaussian models. Instead of simulating posterior samples like MCMC, INLA directly approximates the marginal posteriors people usually want:

π(xi∣y)andπ(θj∣y).\pi(x_i \mid y) \quad\text{and}\quad \pi(\theta_j \mid y).

Here yy is the observed data, xx is a high-dimensional latent field, and θ\theta is a small vector of hyperparameters. INLA is fast because it uses the Gaussian structure of xx, sparse precision matrices, and Laplace approximations around posterior modes.

The tradeoff is structural: INLA is not a universal Bayesian engine. It is powerful when the model fits the latent Gaussian template, and much less natural when the hidden layer is not conditionally Gaussian.

When INLA shines

A spatial disease-mapping model might observe counts yiy_i in many regions and include an unobserved spatial risk xix_i for each region. Nearby regions are assumed to have similar risk, so the latent field xx is modeled with a sparse Gaussian precision matrix. MCMC may need a long run and convergence checks; INLA can often compute the region-level posterior marginals in seconds or minutes.

Try it

If MCMC is more general than INLA, why would anyone choose INLA over MCMC?

Solution

Because for latent Gaussian models, INLA targets the desired marginal posteriors directly and deterministically. There is no chain to tune, no burn-in period, and no Monte Carlo error. The price is that the answer is an approximation, and the method depends on the model having the right conditional Gaussian structure.

Related concepts

View in full concept graph →