Variational Inference with Implicit Distributions

19 concepts14 papers1 proposal

In Bayesian inference, we often want to compute a posterior distribution $p(\theta|x)$ over latent parameters $\theta$ given some observed data $x$. Because this posterior is usually analytically intractable, Variational Inference (VI) reframes the integration problem as an optimization task. We select a family of distributions $q_\phi(\theta)$ and minimize its divergence—typically the Kullback-Leibler (KL) divergence—to the true posterior.

However, traditional VI comes with a severe mathematical constraint: evaluating the KL divergence or the Evidence Lower Bound (ELBO) requires us to compute the exact log-density $\log q_\phi(\theta)$ for any sampled parameter. This "KL-divergence tax," as described by Rajesh Ranganath et al. in Operator Variational Inference, forces researchers to choose simple, analytically tractable distributions (like Gaussians) for $q_\phi(\theta)$. Unfortunately, real-world posteriors are rarely neat bell curves; they are often multimodal, skewed, or highly constrained.

Defining the Implicit Family

What if we could define a highly flexible variational distribution without ever needing to write down its probability density function? This is the core premise of implicit variational inference.

Instead of specifying $q_\phi(\theta)$ via an analytical formula, we define it as a generative process. We take a simple base distribution (such as standard Gaussian noise $\epsilon \sim \mathcal{N}(0, I)$) and pass it through a highly flexible, parameterizable function like a neural network $f_\phi$:

$$\theta = f_\phi(\epsilon)$$

We can easily draw samples from this distribution by sampling noise and running it forward through the network. However, because the neural network can warp the space in highly non-linear ways, calculating the exact probability density $q_\phi(\theta)$ of the resulting samples becomes mathematically impossible.

As shown by Sean C Plummer et al. in Statistical Guarantees for Transformation Based Models with applications to Implicit Variational Inference, these transformation-based models can approximate virtually any complex density and possess strong statistical convergence guarantees. This flexibility allows the variational posterior to capture complex, non-Gaussian structures across a wide variety of domains, including deep Gaussian processes (Haibin Yu et al. in Implicit Posterior Variational Inference for Deep Gaussian Processes), continual learning (Randy Ardywibowo et al. in VariGrow: Variational Architecture Growing for Task-Agnostic Continual Learning based on Bayesian Novelty), and personalized federated learning (Shangchao Su et al. in Towards Instance-wise Personalized Federated Learning via Semi-Implicit Bayesian Prompt Tuning).

Strategies for Density-Free Optimization

If we cannot compute $\log q_\phi(\theta)$, standard gradient descent on the ELBO breaks down. To bypass this bottleneck, researchers have developed several distinct mathematical strategies that allow optimization using only samples from $q_\phi(\theta)$.

1. Adversarial Ratio Estimation

Instead of computing the density explicitly, we can train a neural network discriminator to compare samples from our implicit model against samples from the target distribution. In Hierarchical Implicit Models and Likelihood-Free Variational Inference, Dustin Tran et al. introduced a GAN-style ratio estimation trick to scale Bayesian inference to massive datasets without analytical likelihoods.

Similarly, Yibo Yang et al. in Conditional deep surrogate models for stochastic, high-dimensional, and multi-fidelity systems and Simón Rodríguez Santana et al. in Adversarial α-divergence Minimization for Bayesian Approximate Inference use adversarial objectives to minimize broader classes of divergences, allowing the approximate posterior to adapt closely to arbitrary shapes without strong parametric assumptions.

2. Stein Gradients and Score Matching

Rather than evaluating the absolute density value, we can look at the score function—the gradient of the log-density with respect to the parameters. In Learning to Draw Samples: With Application to Amortized MLE for Generative Adversarial Learning, Dilin Wang et al. developed SteinGAN, which leverages Stein Variational Gradient Descent (SVGD) to update a black-box generator using only the score function of the target distribution.

To make this process faster, Yihao Feng et al. in Learning to Draw Samples with Amortized Stein Variational Gradient Descent amortized these dynamics by backpropagating the Stein variational gradients directly into the network's parameters, bypassing the need for a tractable proposal density. More recently, Jack Simons et al. in Neural Score Estimation: Likelihood-Free Inference with Conditional Score Based Diffusion Models demonstrated that conditional score-based diffusion models offer immense architectural freedom for modeling complex posteriors without relying on rigid normalizing flows.

3. Auxiliary Variables and Expectations

Sometimes, an intractable density can be made tractable by introducing a hidden helper variable. In Unbiased Implicit Variational Inference, Michalis K. Titsias et al. uncovered a mathematical identity showing that the intractable density gradients of an implicit model can be expressed exactly as expectations over auxiliary variables. By using a Markov Chain Monte Carlo (MCMC) shortcut to sample these auxiliary variables, the authors directly optimize the ELBO without bias, making highly flexible neural network distributions trainable without explicit density functions.

4. Optimal Transport and Geometric Distances

Instead of measuring vertical differences between probability densities (like the KL divergence), we can measure the horizontal distance required to move probability mass from one distribution to another. In Generalized Variational Inference via Optimal Transport, Jinjin Chi et al. combined optimal transport objectives with the reparameterization trick to optimize sample-based distributions geometrically. Similarly, Mingxuan Yi et al. in Sliced Wasserstein Variational Inference replaced the KL divergence with the sliced Wasserstein distance, enabling efficient amortized inference via black-box samplers.

Shifting from Evaluation to Comparison

To understand why implicit variational inference works so well, consider an analogy.

Traditional variational inference is like trying to match two complex mathematical curves by calculating and comparing their exact heights at every single point along the x-axis. To do this, you must possess the exact algebraic formula for both curves. If one curve is generated by a complex simulation or a deep neural network, you don't have the formula, and the method fails.

Implicit variational inference shifts the paradigm from evaluating formulas to comparing samples. Instead of asking "What is the probability density of this exact point?", these algorithms ask "How should we move our generated samples so that their collective distribution matches the target?"

By using discriminators to sense distributional mismatches, score functions to provide a vector field of directional arrows, or optimal transport to map out geometric rearrangements, the optimization signal is entirely sample-driven. This fundamental shift frees variational inference from the prison of simple parametric families, allowing posterior approximations to be as rich, expressive, and complex as the deep neural networks we use to build them.

Key Foundational Work

Operator Variational Inference (2016) — contributed 2 concept elements to this direction.
Hierarchical Implicit Models and Likelihood-Free Variational Inference (2017) — contributed 2 concept elements to this direction.
Learning to Draw Samples: With Application to Amortized MLE for Generative Adversarial Learning (2017) — contributed 2 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_density_free_variational_inference_480797af,
  title = {density-free variational inference},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/density-free-variational-inference/rdr_480797af2afc442fbf801eb48ef5dc75},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_density_free_variational_inference_480797af_v1,
  title = {density-free variational inference},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/density-free-variational-inference/rdr_480797af2afc442fbf801eb48ef5dc75/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake