DIRT: Deep Inverse Rosenblatt Transport

• Deep Inverse Rosenblatt Transport is a layered framework that composes order-preserving triangular maps with tensor-train approximations to ensure positivity and monotonicity in high-dimensional settings.
• It uses intermediate bridging densities and tempering strategies to decompose geometric complexity, managing concentrated, correlated, and nonlinear distributions efficiently.
• DIRT enables rapid conditional sampling for Bayesian inference and importance sampling in applications like structural reliability and multiscale mechanics, reducing computational overhead.

Deep Inverse Rosenblatt Transport (DIRT) is a multi-layer transport-map framework for characterizing, approximating, and sampling from complicated high-dimensional probability densities through compositions of order-preserving triangular maps. In its canonical form, DIRT combines inverse Rosenblatt transport with tensor-train (TT) approximation of square-root densities, so that positivity and monotonicity are preserved by construction, while a sequence of intermediate or bridging densities distributes geometric complexity across layers. The framework has been developed for high-dimensional Bayesian inference, conditional posterior characterization, and transport-based importance sampling, particularly in settings where the target distribution is concentrated, strongly correlated, nonlinear, or embedded in expensive forward models (Cui et al., 2020, Cui et al., 2021, Tyagi et al., 5 Sep 2025).

1. Historical development and scope

DIRT emerged from work on functional TT approximation of inverse Rosenblatt transports and was formalized as a deep composition framework in 2020. That formulation generalized TT-based inverse Rosenblatt transport from probability densities to a wider class of high-dimensional non-negative functions, including unnormalized probability densities, and introduced the two design features that remain central: squaring a TT approximation of the square root of the density, and composing multiple monotone transport layers in a nested transformation architecture inspired by deep neural networks (Cui et al., 2020).

A 2021 development extended this construction to conditional Bayesian inference by learning the joint law of observables and parameters offline in TT form and then extracting conditional transports online. In that work, the method is described as a scalable conditional version of the same idea and as an explicit extension of the DIRT methodology of Cui et al. through TT function approximation, variable reordering and reparametrization, and layered compositions for high-dimensional joint laws of parameters and observables (Cui et al., 2021).

Subsequent applications placed DIRT in structural reliability and multiscale mechanics. A 2025 study investigated DIRT for reliability analysis in solid mechanics across benchmark problems with dimensionality ranging from $2$ to $250$, while a 2026 multiscale reliability paper coupled DIRT with a physics-augmented Voigt–Reuss Neural Network to estimate posterior failure probabilities in dimensions up to $150$ (Tyagi et al., 5 Sep 2025, Tyagi et al., 20 Apr 2026).

2. Triangular transport and the inverse Rosenblatt construction

The mathematical backbone of DIRT is the Rosenblatt transform, which maps a target law to a reference law by sequential one-dimensional conditional CDFs. For a density

$$f_X(x)=\frac{1}{z}\,\pi(x)\lambda(x), \qquad \lambda(x)=\prod_{i=1}^d \lambda_i(x_i),$$

the Rosenblatt map is lower triangular: $$F(x)=\bigl(F_{X_1}(x_1),\,F_{X_2\mid X_1}(x_2\mid x_1),\,\dots,\,F_{X_d\mid X_{<d}}(x_d\mid x_{<d})\bigr)^\top,$$ and its inverse

$T=F^{-1}$

generates samples from the target distribution by sequential inversion of conditional CDFs (Cui et al., 2020).

In the conditional Bayesian setting, DIRT is formulated on the joint law of observables and parameters,

$$\pi_{Y,\Theta}(y,\theta)=\pi_{Y\mid \Theta}(y\mid \theta)\,\pi_\Theta(\theta),$$

with posterior

$$\pi_{\Theta\mid Y}(\theta\mid y)\propto \pi_{Y,\Theta}(y,\theta).$$

The transport is triangular and order-preserving: $$\left[\begin{array}{c} y\ \theta \end{array}\right] = \mathcal{T}(u_Y,u_\Theta) = \left[\begin{array}{c} \mathcal{T}_Y(u_Y)\ \mathcal{T}_\Theta(u_Y,u_\Theta) \end{array}\right].$$ Once an observation $y$ is available, conditioning is realized by freezing the observation branch and evaluating

$250$0

which produces samples from an approximate posterior (Cui et al., 2021).

This triangularity is not a secondary implementation detail. It is what makes DIRT useful for conditional sampling, sequential inversion, and explicit manipulation of marginals and conditionals. A plausible implication is that DIRT’s practical effectiveness is inseparable from its coordinatewise monotone structure, rather than from generic expressivity alone.

3. Tensor-train parameterization, squaring, and monotonicity preservation

A central obstacle in high-dimensional transport construction is that direct approximation of a nonnegative density by a low-rank tensor expansion can introduce small negative regions after truncation. In the original DIRT formulation, this is addressed by approximating the square root of the target density,

$250$1

with $250$2 in TT form, and then defining a positive surrogate density through squaring: $250$3 with $250$4 a stabilizing constant. Because the approximation is expressed as a constant plus a sum of squares, the induced marginal and conditional densities remain positive, and the resulting inverse Rosenblatt map remains monotone (Cui et al., 2020).

The same principle appears in the scalable conditional formulation, where the square root of the joint density is approximated by a TT decomposition,

$250$5

and the approximate joint density is defined by

$250$6

This converts density approximation into an $250$7-type problem on square roots and yields control of the Hellinger distance between the true joint density and its surrogate (Cui et al., 2021).

The TT representation is used because, when ranks remain moderate, storage and many related operations scale as

$250$8

rather than exponentially in dimension. In the reliability formulation, TT approximations are constructed with TT-cross, using selected fibers and maxvol-based skeleton approximations on tensor unfoldings instead of evaluating the full tensor grid (Tyagi et al., 5 Sep 2025).

From the squared TT factorization, marginals can be obtained analytically by integrating out trailing variables, and conditional PDFs and conditional CDFs follow by ratios of marginal quantities. DIRT then samples by sequential inversion: $250$9 This mechanism is the computational core of the inverse Rosenblatt map in TT form (Tyagi et al., 5 Sep 2025).

4. Deep composition, tempering, and dimension-structuring heuristics

The “deep” in DIRT refers to composition. When the target density is too concentrated or too correlated to be represented accurately by a single TT-based inverse Rosenblatt map, DIRT constructs a sequence of intermediate transports across bridging densities. In the original formulation, this is expressed through a chain of intermediate measures $150$0; in the conditional formulation, one writes

$150$1

with each layer associated with a simpler bridging density, often obtained by tempering,

$150$2

Under the convergence proposition stated in the scalable conditional paper,

$150$3

and, under mild assumptions, uniform temperatures yield $150$4 (Cui et al., 2020, Cui et al., 2021).

This layered strategy is complemented by explicit heuristics for dimension structuring. In the 2021 conditional extension, TT approximation quality is noted to depend strongly on variable ordering, and sensitivity matrices

$150$5

are introduced to quantify dependence. The proposed heuristic sorts variables by decreasing diagonal entries of these matrices and, in Gaussian settings, reparametrizes by leading eigenvectors. The associated reduced surrogate densities satisfy Hellinger-error bounds controlled by the trailing diagonal entries or trailing eigenvalues of the sensitivity matrices. The stated intuition is to place the most strongly coupled coordinates in the middle of the TT chain and push nearly independent directions toward the tails, where TT ranks tend to decay (Cui et al., 2021).

This suggests that DIRT’s scalability is not only a property of TT compression in the abstract; it also depends on choosing coordinates in which the target law exhibits exploitable low-rank structure.

5. Conditional DIRT for Bayesian inference

In Bayesian inverse problems, DIRT is used as an offline–online mechanism for posterior characterization. The offline stage approximates the joint law of observables and parameters; the online stage turns newly observed data into posterior samples by evaluating the conditional branch of the learned triangular map. This eliminates the need to rerun expensive Bayesian inference from scratch for each new observation and yields real-time conditional sampling from a surrogate posterior (Cui et al., 2021).

The 2021 paper emphasizes that, compared with state-of-the-art normalizing flow techniques, the method relies on function approximation and is equipped with a thorough performance analysis. In particular, it proves that small joint Hellinger error implies closeness of posterior expectations with high probability. It also shows that the learned conditional map can be used as a preconditioner for exact MCMC samplers such as pCN and NUTS, reducing integrated autocorrelation times and estimator variances (Cui et al., 2021).

The reported applications illustrate both scale and breadth. In the SIR ODE inverse problem, DIRT was trained offline with TT ranks around $150$6 and several tempering layers; offline construction required about $150$7 joint density evaluations and $150$8 seconds on a CPU, while one conditional sample cost only tens of microseconds online. In a 2D linear elasticity example, the unknown was a $150$9-dimensional log Young’s modulus field with $$f_X(x)=\frac{1}{z}\,\pi(x)\lambda(x), \qquad \lambda(x)=\prod_{i=1}^d \lambda_i(x_i),$$0 observations of displacement, and the method used eigenvector-based reparametrization before reduced-dimensional DIRT construction. In a Besov-prior elliptic PDE example, the parameter was a $$f_X(x)=\frac{1}{z}\,\pi(x)\lambda(x), \qquad \lambda(x)=\prod_{i=1}^d \lambda_i(x_i),$$1-dimensional wavelet coefficient vector with Laplace-like prior, and the proposed variable ordering significantly lowered intermediate TT ranks (Cui et al., 2021).

6. Importance sampling and structural reliability

A second major role of DIRT is transport-based importance sampling for rare-event estimation. In structural reliability, one seeks

$$f_X(x)=\frac{1}{z}\,\pi(x)\lambda(x), \qquad \lambda(x)=\prod_{i=1}^d \lambda_i(x_i),$$2

or its posterior analogue after Bayesian updating. The reliability formulation of DIRT constructs TT approximations of the optimal importance densities and then converts these surrogates into approximate inverse Rosenblatt maps, so that samples from a simple reference law can be transported into regions of posterior or failure mass efficiently (Tyagi et al., 5 Sep 2025).

Because the failure indicator is discontinuous, the reliability papers replace it with a smooth sigmoid surrogate,

$$f_X(x)=\frac{1}{z}\,\pi(x)\lambda(x), \qquad \lambda(x)=\prod_{i=1}^d \lambda_i(x_i),$$3

or, in the multiscale formulation,

$$f_X(x)=\frac{1}{z}\,\pi(x)\lambda(x), \qquad \lambda(x)=\prod_{i=1}^d \lambda_i(x_i),$$4

with $$f_X(x)=\frac{1}{z}\,\pi(x)\lambda(x), \qquad \lambda(x)=\prod_{i=1}^d \lambda_i(x_i),$$5 gradually increased. This smoothing is explicitly used to reduce TT-rank growth and improve numerical stability (Tyagi et al., 5 Sep 2025, Tyagi et al., 20 Apr 2026).

The 2025 reliability study examined three analytical problems and one finite-element example with dimensionality from $$f_X(x)=\frac{1}{z}\,\pi(x)\lambda(x), \qquad \lambda(x)=\prod_{i=1}^d \lambda_i(x_i),$$6 to $$f_X(x)=\frac{1}{z}\,\pi(x)\lambda(x), \qquad \lambda(x)=\prod_{i=1}^d \lambda_i(x_i),$$7. For a linear benchmark, it reported accurate estimation of failure probabilities from $$f_X(x)=\frac{1}{z}\,\pi(x)\lambda(x), \qquad \lambda(x)=\prod_{i=1}^d \lambda_i(x_i),$$8 down to $$f_X(x)=\frac{1}{z}\,\pi(x)\lambda(x), \qquad \lambda(x)=\prod_{i=1}^d \lambda_i(x_i),$$9; for $$F(x)=\bigl(F_{X_1}(x_1),\,F_{X_2\mid X_1}(x_2\mid x_1),\,\dots,\,F_{X_d\mid X_{<d}}(x_d\mid x_{<d})\bigr)^\top,$$0, DIRT produced estimates between $$F(x)=\bigl(F_{X_1}(x_1),\,F_{X_2\mid X_1}(x_2\mid x_1),\,\dots,\,F_{X_d\mid X_{<d}}(x_d\mid x_{<d})\bigr)^\top,$$1 and $$F(x)=\bigl(F_{X_1}(x_1),\,F_{X_2\mid X_1}(x_2\mid x_1),\,\dots,\,F_{X_d\mid X_{<d}}(x_d\mid x_{<d})\bigr)^\top,$$2 with CoV as low as $$F(x)=\bigl(F_{X_1}(x_1),\,F_{X_2\mid X_1}(x_2\mid x_1),\,\dots,\,F_{X_d\mid X_{<d}}(x_d\mid x_{<d})\bigr)^\top,$$3 at $$F(x)=\bigl(F_{X_1}(x_1),\,F_{X_2\mid X_1}(x_2\mid x_1),\,\dots,\,F_{X_d\mid X_{<d}}(x_d\mid x_{<d})\bigr)^\top,$$4, whereas BUS-SuS had CoV values around $$F(x)=\bigl(F_{X_1}(x_1),\,F_{X_2\mid X_1}(x_2\mid x_1),\,\dots,\,F_{X_d\mid X_{<d}}(x_d\mid x_{<d})\bigr)^\top,$$5–$$F(x)=\bigl(F_{X_1}(x_1),\,F_{X_2\mid X_1}(x_2\mid x_1),\,\dots,\,F_{X_d\mid X_{<d}}(x_d\mid x_{<d})\bigr)^\top,$$6. In a 2D plate-with-hole example with KL dimensions $$F(x)=\bigl(F_{X_1}(x_1),\,F_{X_2\mid X_1}(x_2\mid x_1),\,\dots,\,F_{X_d\mid X_{<d}}(x_d\mid x_{<d})\bigr)^\top,$$7, DIRT achieved CoV values around $$F(x)=\bigl(F_{X_1}(x_1),\,F_{X_2\mid X_1}(x_2\mid x_1),\,\dots,\,F_{X_d\mid X_{<d}}(x_d\mid x_{<d})\bigr)^\top,$$8–$$F(x)=\bigl(F_{X_1}(x_1),\,F_{X_2\mid X_1}(x_2\mid x_1),\,\dots,\,F_{X_d\mid X_{<d}}(x_d\mid x_{<d})\bigr)^\top,$$9 for $T=F^{-1}$0, rising to $T=F^{-1}$1 at $T=F^{-1}$2; the study therefore presents both a variance advantage and a clear rank-related limitation (Tyagi et al., 5 Sep 2025).

The 2026 multiscale paper integrated DIRT with a physics-augmented Voigt–Reuss Neural Network that replaces expensive FE$T=F^{-1}$3 homogenization while preserving symmetry, positive definiteness, and Voigt–Reuss bounds. In that workflow, DIRT is used to approximate the optimal importance-sampling densities for the posterior failure numerator and denominator, with a basis size $T=F^{-1}$4, $T=F^{-1}$5 layers, target $T=F^{-1}$6, and adaptive TT ranks reaching $T=F^{-1}$7 at $T=F^{-1}$8. The paper reports posterior mean field reconstruction errors within $T=F^{-1}$9 for all three random fields, a low-dimensional validation in which a reference posterior failure probability $$\pi_{Y,\Theta}(y,\theta)=\pi_{Y\mid \Theta}(y\mid \theta)\,\pi_\Theta(\theta),$$0 was approximated by DIRT as $$\pi_{Y,\Theta}(y,\theta)=\pi_{Y\mid \Theta}(y\mid \theta)\,\pi_\Theta(\theta),$$1 using only $$\pi_{Y,\Theta}(y,\theta)=\pi_{Y\mid \Theta}(y\mid \theta)\,\pi_\Theta(\theta),$$2 forward FEM evaluations, and stable estimates up to $$\pi_{Y,\Theta}(y,\theta)=\pi_{Y\mid \Theta}(y\mid \theta)\,\pi_\Theta(\theta),$$3 with maximum CoV around $$\pi_{Y,\Theta}(y,\theta)=\pi_{Y\mid \Theta}(y\mid \theta)\,\pi_\Theta(\theta),$$4 (Tyagi et al., 20 Apr 2026).

7. Relation to Knothe–Rosenblatt transport, optimal transport, and methodological boundaries

DIRT belongs to the broader family of triangular transports derived from the Knothe–Rosenblatt rearrangement. That connection is explicit in work on Knothe–Rosenblatt Domain Adaptation, where autoregressive density estimators are used to construct a triangular source-to-target map by conditional quantile preservation. The conceptual overlap with DIRT lies in the use of invertible triangular transports grounded in successive one-dimensional conditional transformations, although KRDA is specialized to unsupervised domain adaptation and emphasizes explicit source-to-target alignment rather than TT-based deep composition (Virmaux et al., 2021).

A 2025 optimal transport result provides a separate theoretical bridge. It shows that a weighted quadratic OT problem with a soft target constraint recovers a triangular KR-type map as the weighted-cost parameter $$\pi_{Y,\Theta}(y,\theta)=\pi_{Y\mid \Theta}(y\mid \theta)\,\pi_\Theta(\theta),$$5, and that the true KR map is obtained as the KL-penalty parameter $$\pi_{Y,\Theta}(y,\theta)=\pi_{Y\mid \Theta}(y\mid \theta)\,\pi_\Theta(\theta),$$6. The paper states that this rigorously supports variational and continuation-based methods for learning inverse triangular transports, including DIRT-style algorithms (Baptista et al., 6 Nov 2025).

Several methodological boundaries follow from the published formulations. DIRT is not a generic black-box flow architecture: its defining structure is lower triangular, monotone, and based on explicit marginal and conditional manipulations. Its advertised linear scaling in dimension is also conditional rather than absolute, depending on TT ranks remaining low to moderate; the reliability studies explicitly note that increasing complexity can raise TT ranks, increase forward-model evaluations, and reduce the practical variance advantage (Tyagi et al., 5 Sep 2025). The Rosenblatt component of the name also refers to the Rosenblatt transform rather than to the Rosenblatt process of stochastic-process theory, which is a different object defined as a second-order Wiener–Itô integral (Garzon et al., 2011).