Transport Map Preconditioning for Markov Chain Monte Carlo

25 concepts17 papers1 proposal

Markov Chain Monte Carlo (MCMC) algorithms, such as Hamiltonian Monte Carlo (HMC) and Langevin dynamics, are the mathematical workhorses of Bayesian inference. They allow us to sample from complex, unnormalized posterior distributions. However, these algorithms suffer from a fundamental geometric limitation: their efficiency drops precipitously when encountering distributions with high curvature, strong correlations, or multiple isolated modes.

When a target distribution features a narrow "funnel" or a highly anisotropic (stretched) geometry, an MCMC sampler gets trapped. To avoid numerical instability or rejected steps, the sampler must reduce its step size globally. As a result, it explores the rest of the distribution at a painful crawl.

Transport map preconditioning solves this problem by changing the coordinate system. Instead of forcing a sampler to navigate a rugged, ill-conditioned landscape, we learn a mathematical transformation—a transport map—that warps the complex target distribution into a simple, isotropic, and near-Gaussian representation. We then run our sampler in this friendly latent space and map the resulting trajectories back to the original domain.


The Intuition of Preconditioning

To understand why this works, consider an analogy from numerical linear algebra. If you want to solve a difficult linear system $Ax = b$ where the matrix $A$ is ill-conditioned, you don't just run standard gradient descent. Instead, you apply a preconditioner $M^{-1}$ to transform the system into $M^{-1}Ax = M^{-1}b$, choosing $M$ such that $M^{-1}A$ is close to the identity matrix. This simple change of variables accelerates convergence by making the optimization landscape perfectly spherical.

As formalised by Matthew D. Parno et al. in Transport Map Accelerated Markov Chain Monte Carlo and B. Zhang et al. in Transport map unadjusted Langevin algorithms: Learning and discretizing perturbed samplers, transport maps act as exact probabilistic preconditioners.

Suppose our target distribution $p_X(x)$ is highly non-linear and poorly scaled. We want to find a smooth, invertible mapping $T$ (the transport map) that links our difficult target variable $x$ to a simple latent variable $z$ drawn from a standard isotropic Gaussian distribution $p_Z(z) = \mathcal{N}(0, I)$:

$$x = T(z) \quad \Longleftrightarrow \quad z = T^{-1}(x)$$

If we possess an approximation of this map, we can reformulate the sampling problem entirely in the $z$-space. Philip Schär et al. demonstrate in Parallel Affine Transformation Tuning of Markov Chain Monte Carlo that MCMC samplers reach their maximum theoretical performance when operating on an isotropic target. By shifting the sampler's environment to the $z$-space, we ensure that the base algorithm operates in its optimal regime, completely insulated from the pathological geometry of the original target.


The Spectrum of Transport Maps

The complexity of the transport map dictates how much of the target's geometric difficulty can be ironed out.

Linear Preconditioning

The simplest transport maps are linear coordinate transformations. As explored by Noah W. Tuchow et al. in The efficiency of geometric samplers for exoplanet transit timing variation models, one can use a pilot MCMC run to estimate the mean and covariance matrix $\Sigma$ of the target distribution. The transport map is then defined via the Cholesky decomposition $\Sigma = L L^T$:

$$x = Lz + \mu$$

This centers and rotates the target space, eliminating linear correlations and global scaling mismatches. While highly effective for near-Gaussian targets, linear maps fail when the distribution exhibits non-linear features like bananas, funnels, or multiple modes.

Non-Linear and Neural Preconditioning

To handle complex, non-linear geometries, researchers leverage Normalizing Flows (NFs)—deep neural networks designed to model bijective, differentiable transformations with tractable Jacobians.

As detailed by Louis Grenioux et al. in On Sampling with Approximate Transport Maps, using a normalizing flow to perform a local reparametrization warps the target distribution into a simpler latent space. This allows gradient-based samplers to seamlessly glide through regions of high curvature. Liam A. Kruse et al. show in Enhanced Importance Sampling Through Latent Space Exploration in Normalizing Flows that this non-linear "gaussianization" aligns the problem's underlying geometry with the basic assumptions of standard sampling algorithms, making even rare-event simulation tractable.


Mechanics: How Latent Space Sampling Works

To sample using a transport map, we do not sample directly from the flow model itself (which might be biased or imperfect). Instead, we use the flow as a bridge to evaluate a pullback density in the latent space, and run an exact MCMC chain on that pullback density.

By the change of variables formula, the probability density of the latent variable $z$ induced by the target $p_X(x)$ is:

$$p_{\text{pullback}}(z) = p_X(T(z)) \cdot \left| \det J_T(z) \right|$$

where $J_T(z)$ is the Jacobian matrix of the transformation.

The preconditioned sampling pipeline proceeds as follows:

 Latent Space (z)                                     Target Space (x)
+------------------+                                 +------------------+
| Isotropic, Smooth|                                 | Rugged, Funnels, |
| Gaussian Geometry|                                 | High Curvature   |
|                  |                                 |                  |
|  (1) Run MCMC    | -- (2) Push forward via T(z) -> |  (3) Collect     |
|      Step        |                                 |      Exact       |
|      in z        | <- (4) Correct via Jacobian --- |      Samples     |
+------------------+                                 +------------------+
  1. Propose a step in Latent Space: The sampler (e.g., HMC or Langevin) proposes a move from $z_t$ to $z_{t+1}$ using the gradient of the log-pullback density $\nabla_z \log p_{\text{pullback}}(z)$.
  2. Push forward: The proposed latent point is mapped to the target space via $x_{t+1} = T(z_{t+1})$.
  3. Accept/Reject: The step is accepted or rejected based on the Metropolis-Hastings rule applied to $p_{\text{pullback}}(z)$. The Jacobian determinant $\left| \det J_T(z) \right|$ acts as a geometric correction factor, ensuring that any imperfections in the learned transport map $T$ are rigorously corrected.

This guarantees that the final samples $x$ are drawn exactly from the true target distribution $p_X(x)$, preserving mathematical exactness even if the neural network is only a rough approximation.

This geometric warping alters the behavior of standard samplers in profound ways. For instance, Richard D. P. Grumitt et al. demonstrate in Deterministic Langevin Monte Carlo with Normalizing Flows for Bayesian Inference that performing gradient updates within this latent space turns a stochastic random walk into a directed, highly efficient optimization-like process. Similarly, Michael Brennan et al. show in Greedy inference with structure-exploiting lazy maps that preconditioning high-dimensional problems with transport maps can rescue acceptance rates from near-zero up to highly practical levels.


Unlocking Advanced Sampler Capabilities

Beyond merely speeding up step sizes, transport preconditioning fundamentally alters the topology of the sampling space, unlocking capabilities that are impossible in the original parameter space.

1. Analytical Trajectories

In Slice Sampling on Hamiltonian Trajectories, Benjamin Bloem-Reddy et al. demonstrate that when a probability integral transformation maps complex priors onto a uniform hypercube, Hamiltonian dynamics simplify to constant-velocity motion. The particles move in straight lines and bounce off boundaries, allowing for the analytical calculation of trajectories without numerical discretization errors.

2. Bridging Disconnected Modes

In standard data spaces, Energy-Based Models (EBMs) and multimodal posteriors suffer from poor mixing because the sampler cannot cross the high-energy barriers separating distinct modes. Erik Nijkamp et al. show in MCMC Should Mix: Learning Energy-Based Model with Neural Transport Latent Space MCMC that a continuous transport map compresses the distance between these modes in the latent space. The sampler can freely traverse a smooth latent energy landscape, enabling the Markov chain to jump between modes that appear completely isolated in the original space.

3. Enabling Variance Reduction

Advanced variance-reduction techniques, such as antithetic sampling, rely on the target distribution being symmetric. If a distribution is highly skewed, these methods fail. Dan Piponi et al. demonstrate in Hamiltonian Monte Carlo Swindles that mapping a target into an approximately symmetric latent space allows these "swindles" to function perfectly, yielding super-efficient estimators where the effective sample size can actually exceed the number of physical MCMC iterations.


The Symbiotic Feedback Loop

A common critique of transport-based sampling is the chicken-and-egg dilemma: to learn a perfect transport map to a target distribution, you must already be able to sample from that target distribution.

Modern research resolves this by framing transport preconditioning as an adaptive, co-evolutionary process. As outlined by Eric Vanden‐Eijnden in Learning to sample better, the map and the sampler can be optimized simultaneously in a variational feedback loop:

+---------------------------------------------------------+
|                                                         |
|     +-----------------------+     Better Samples        |
|     |  MCMC Latent Sampler  | --------------------+     |
|     +-----------------------+                     |     |
|                 ^                                 v     |
|                 |                       +-------------------+
|                 +---------------------- |   Transport Map   |
|                     Smoother Space      | (Normalizing Flow)|
|                                         +-------------------+
|                                                   |
+---------------------------------------------------+

This synergy is explicitly leveraged by Liyi Zhang et al. in Transport Score Climbing: Variational Inference Using Forward KL and Adaptive Neural Transport. Instead of requiring expensive pre-training, an initial poor transport map provides a slightly improved space for HMC. The improved HMC sampler explores the space better, generating higher-quality samples that are used to refine the transport map via variational objectives (such as the forward KL divergence).

Similarly, Youssef M. Marzouk et al. demonstrate in A transport-based multifidelity preconditioner for Markov chain Monte Carlo that one can construct transport maps using cheap, low-fidelity models. This low-fidelity map warps the space into a near-Gaussian geometry, allowing the high-fidelity sampler to run with massive speedups from the very first iteration.

By treating transport maps not as generative models, but as geometric lenses that clarify complex probability spaces, transport-based preconditioning bridges the gap between variational approximation and exact Monte Carlo sampling.

Key Foundational Work

Slice Sampling on Hamiltonian Trajectories (2016) — contributed 1 concept elements to this direction.
A Multiscale Strategy for Bayesian Inference Using Transport Maps (2015) — contributed 1 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_transport_based_preconditioning_5e9efc11,
  title = {transport-based preconditioning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/transport-based-preconditioning/rdr_5e9efc1159d849e9af6f87ffc07edc3b},
  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_transport_based_preconditioning_5e9efc11_v1,
  title = {transport-based preconditioning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/transport-based-preconditioning/rdr_5e9efc1159d849e9af6f87ffc07edc3b/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