Knothe–Rosenblatt Rearrangement

• Knothe–Rosenblatt rearrangement is a canonical method that constructs a lower‐triangular map by recursively coupling one-dimensional marginals of probability measures.
• It ensures strict monotonicity and invertibility via conditional quantile transforms, making it highly reliable for tasks in optimal transport and generative modeling.
• Recent computational approaches, including flow-based networks and sparse polynomial approximations, efficiently scale the KR map for high-dimensional and structured distributions.

The Knothe–Rosenblatt (KR) rearrangement is a canonical construction of a lower-triangular transport map between probability measures on $\mathbb{R}^d$. It is uniquely defined by recursively coupling one-dimensional marginals and conditional distributions, resulting in a Monge transport map with strict monotonicity in each variable and robust invertibility properties. The KR rearrangement serves as a critical object in classical probability, optimal transport, flow-based generative modeling, causal transport, and stochastic analysis.

1. Definition and Construction

The KR rearrangement constructs a map $$T^{\mathrm{KR}}: \mathbb{R}^d \rightarrow \mathbb{R}^d$$ pushing a reference measure $\mu$ onto a target measure $\nu$ (assumed absolutely continuous) in a recursively monotone, lower-triangular fashion. For measures with smooth positive densities $f(x_1,\dots,x_d)$ and $g(y_1,\dots,y_d)$, the construction proceeds coordinate by coordinate:

• The first coordinate is mapped as the quantile transform:

$T_1(x_1) = G_1^{-1}(F_1(x_1)),$

where $F_1$ and $G_1$ are cumulative distribution functions for the first marginals of $f$ and $$T^{\mathrm{KR}}: \mathbb{R}^d \rightarrow \mathbb{R}^d$$0.

• For $$T^{\mathrm{KR}}: \mathbb{R}^d \rightarrow \mathbb{R}^d$$1, recursively fix the previous output coordinates $$T^{\mathrm{KR}}: \mathbb{R}^d \rightarrow \mathbb{R}^d$$2, and define

$$T^{\mathrm{KR}}: \mathbb{R}^d \rightarrow \mathbb{R}^d$$3

where $$T^{\mathrm{KR}}: \mathbb{R}^d \rightarrow \mathbb{R}^d$$4 and $$T^{\mathrm{KR}}: \mathbb{R}^d \rightarrow \mathbb{R}^d$$5 denote conditional CDFs. The resulting map is lower-triangular:

$$T^{\mathrm{KR}}: \mathbb{R}^d \rightarrow \mathbb{R}^d$$6

This mapping pushes $$T^{\mathrm{KR}}: \mathbb{R}^d \rightarrow \mathbb{R}^d$$7 forward to $$T^{\mathrm{KR}}: \mathbb{R}^d \rightarrow \mathbb{R}^d$$8, i.e., $$T^{\mathrm{KR}}: \mathbb{R}^d \rightarrow \mathbb{R}^d$$9 (Baptista et al., 6 Nov 2025, Tang et al., 2021, Bonnotte, 2012).

2. Analytical Structure and Properties

The KR rearrangement is characterized by key structural properties:

• Triangularity and Monotonicity: Each $\mu$0 depends only on $\mu$1 and is strictly increasing in $\mu$2 for fixed previous variables, provided conditional densities are positive and continuous (Baptista et al., 6 Nov 2025, Tang et al., 2021).
• Invertibility: The Jacobian determinant is positive everywhere:

$\mu$3

This guarantees global diffeomorphism (Tang et al., 2021).

• Explicit Inverse: The map admits an explicit inversion using conditional quantiles:

$\mu$4

where $\mu$5 (Tang et al., 2021).

• Change of Variables: The induced density transformation satisfies

$\mu$6

useful for density estimation and generative models (Cui et al., 2023, Tang et al., 2021).

3. Optimal Transport, Degenerate Limits, and PDE Connections

The KR map emerges as a singular limit in optimal transport:

• Weighted-Cost Limit: Consider minimizing

$\mu$7

over maps $\mu$8 pushing $\mu$9 to $\nu$0. As $\nu$1, the unique minimizer $\nu$2 converges in $\nu$3 to $\nu$4 (Baptista et al., 6 Nov 2025, Bonnotte, 2012).

• Dynamic Formulation: In Benamou–Brenier dynamic OT, letting the quadratic cost anisotropy $\nu$5 yields a velocity field whose time-1 map recovers KR (Baptista et al., 6 Nov 2025).
• Soft Constraints and Variational Estimation: KR can be approximated via parametric triangular ansatz $\nu$6 by minimizing

$\nu$7

for divergence $\nu$8 (e.g., Kullback–Leibler) and large $\nu$9 (Baptista et al., 6 Nov 2025). The KR map is recovered as $f(x_1,\dots,x_d)$0, $f(x_1,\dots,x_d)$1.

• Degeneration of Brenier Maps: On smooth manifolds (e.g., torus $f(x_1,\dots,x_d)$2), the solution to the degenerate quadratic OT problem with cost diagonalizing to zero on all but one axis approaches the KR rearrangement. The limiting transport potential satisfies a singular PDE, with smooth continuation ensured by Nash–Moser inverse function methods (Bonnotte, 2012).

4. Computational and Approximation Techniques

Numerous practical schemes have been developed for approximating KR maps:

• Flow-Based Models (KRnet, ADDA-KR): Block-triangular parameterizations with affine, scaling, rotation layers and explicit monotonicity constraints ensure global diffeomorphism and tractable log-densities. These models push a base measure (often uniform or Gaussian) forward to target complex distributions and are well-suited for solving high-dimensional PDEs and density estimation (Tang et al., 2021).
• Sparse Polynomial Approximation: The KR map is constructed using least squares on downward-closed tensor-product polynomial bases, which allows analytic marginalization and analytic error bounds (e.g., Hellinger distance). Self-reinforced (compositional) strategies introduce a sequence of bridging densities and layerwise KR transformations to address highly concentrated posteriors, improving error control and scalability (Cui et al., 2023).
• Conditional Quantile Schemes: By matching marginal and conditional quantiles iteratively, explicit coupling is achieved. This approach is especially computationally efficient in 1D and is extended to high-d by triangularization (Tang et al., 2021, Bonnotte, 2012).

Key features and algorithmic underpinnings are summarized below:

Approach	Core Principle	Scalability/Features
Flow-based (KRnet)	Block-triangular, neural coupling layers	Global $f(x_1,\dots,x_d)$3-diffeo, invertible, tractable likelihood
Sparse polynomials	Weighted LS in tensor bases, composition	Error bounds, efficient for $f(x_1,\dots,x_d)$4
Quantile recursion	Recursive CDF transforms	1D/lex order, fast in moderate $f(x_1,\dots,x_d)$5, monotonicity by construction

5. Causal, Bicausal, and Stochastic Extensions

The KR rearrangement functions as a canonical coupling in several stochastic and causal contexts:

• Causal/Bicausal OT: For product-form or Markovian source measures and separable convex costs, the KR map is the optimal causal/bicausal coupling in discrete time, as it respects adaptedness and filtration structure (Veraguas et al., 2016, Backhoff-Veraguas et al., 2022).
• Dynamic Programming Principle: The lower-triangularity of KR aligns with adapted DPP in multistage stochastic programs (Veraguas et al., 2016).
• SDEs and Time Discretization: For scalar time-homogeneous SDEs, time-discrete approximations converge to the synchronous (common driving noise) coupling, which in discrete-time is exactly the KR rearrangement (Backhoff-Veraguas et al., 2022).
• Stochastic Schrödinger Problems: The iterative entropic-regularized coupling, under small-noise limits, converges to the deterministic KR map, unifying entropy-regularized transport and classical monotone rearrangements (Mikami et al., 1 Jan 2025).

6. Generalizations, Variants, and Applications

Significant recent work elucidates the KR rearrangement's roles and extensions:

• Gromov–Wasserstein Triangularization: For certain separable Hadamard product costs, optimal plans converge (as coordinate-weighting degenerates) to a triangular, KR-type rearrangement (Bonet et al., 2021).
• Unsupervised Domain Adaptation: In moderate-dimensional tabular settings, explicit autoregressive (KR-type) mappings align marginal quantiles across datasets, enabling state-of-the-art domain alignment (Virmaux et al., 2021).
• Efficient Sampling and Inverse Problems: Composite KR transformations facilitate efficient sampling from concentrated posteriors in Bayesian inverse problems, especially those involving ODE or PDE constraints (Cui et al., 2023).
• Transport-Information Inequalities: The nested $f(x_1,\dots,x_d)$6 or $f(x_1,\dots,x_d)$7-cost bicausal OT problem admits sharp transport-entropy inequalities, with constants controlled via the recursive KR structure (Veraguas et al., 2016).
• Algorithmic Design: Recent advances enable fast variational estimation of parameterized triangular flows, adaptive learning of variable permutations, and efficient numerical strategies for high-dimensional and structured distributions (Baptista et al., 6 Nov 2025, Cui et al., 2023).

7. Limitations and Ongoing Research

While the KR rearrangement is robustly defined for absolutely continuous measures and ensures monotone, invertible mappings, key limitations and research directions include:

• Variable Ordering: Optimality and complexity depend on the chosen variable ordering, which may not coincide with intrinsic problem symmetries. Adaptive schemes are an active area (Baptista et al., 6 Nov 2025).
• High-Dimensional Concentration: In highly concentrated or ill-conditioned settings, self-reinforced composition and dimensionality reduction become critical (Cui et al., 2023).
• Extensions beyond $f(x_1,\dots,x_d)$8: Extensions to non-Euclidean settings, manifolds, and PDE-constrained inverse problems are under investigation (Baptista et al., 6 Nov 2025, Mikami et al., 1 Jan 2025).
• Comparison with Brenier Map: The KR rearrangement does not in general minimize $f(x_1,\dots,x_d)$9 cost, except in the anisotropic degenerate limit. The relationship and interpolation between KR and Brenier maps remains central in both analysis and computation (Bonnotte, 2012, Baptista et al., 6 Nov 2025).
• Stochastic Dynamics and Multi-Dimensional SDEs: In higher-dimensional stochastic processes, the optimality of KR-type couplings typically fails unless special cost and structure conditions are met (Backhoff-Veraguas et al., 2022).

The KR rearrangement remains a foundational object in the study of transport maps, with ongoing research extending its computational, theoretical, and applied frontiers in probability, statistics, machine learning, and stochastic analysis.