Iteration Distribution Transfer (IDT)

• IDT is a collection of methods that iteratively refine the transfer of probability distributions, addressing high-dimensional challenges in diverse application domains.
• It employs strategies like 1D projection for optimal transport and dynamic reweighting for domain adaptation, ensuring efficient convergence and alignment.
• The approach spans stochastic process modeling, image processing, and transformer-based multi-view decomposition, highlighting its practical significance in modern research.

Iteration Distribution Transfer (IDT) refers to a cluster of modern methodologies across probability theory, computer vision, and machine learning that exploit iterative or compositional transfer between probability distributions, typically for tasks where direct transport or adaptation is computationally or statistically inefficient. The acronym IDT and the underlying approaches have three major usages in the research literature: (1) infinitely divisible with respect to time, a concept from the theory of stochastic processes; (2) iterative projection-based distribution transport methods, prominent in high-dimensional optimal transport and color transfer; and (3) iterative reweighting strategies for dynamic distribution alignment in transfer learning and domain adaptation. Each context shares the core principle of transferring or matching distributions through composition or iterative refinement, but differs substantially in formalism, mathematical guarantees, and application domain.

1. Infinitely Divisible with Respect to Time (IDT) Processes

In stochastic process theory, IDT denotes "infinitely divisible with respect to time." An IDT process $X = (X_t; t \geq 0)$ is defined by the scaling property

$$(X_{nt}; t \geq 0) \stackrel{\mathrm{law}}{=} \left( X_t^{(1)} + \ldots + X_t^{(n)}; t \geq 0 \right),$$

where $X^{(1)}, \ldots, X^{(n)}$ are i.i.d. copies of $X$ (Hakassou et al., 2012). This property generalizes the time-additivity of Lévy processes to processes lacking independent increments. Any stochastically continuous IDT process has a unique (in law) associated Lévy process $L$ such that $X_t \stackrel{\mathrm{law}}{=} L_t$ for each $t$, thereby guaranteeing infinitely divisible marginals even if the process dependence structure differs radically from Lévy (Hakassou et al., 2012).

A significant extension is the $\alpha$-IDT process, where the time scaling is replaced by a power law: $$(X_{n^{1/\alpha} t}; t \geq 0) \stackrel{\mathrm{law}}{=} \left(X_t^{(1)} + \ldots + X_t^{(n)}; t \geq 0 \right).$$ This generalization encompasses fractional Brownian motion and self-similar processes with non-trivial scaling exponents (Hakassou et al., 2012).

IDT processes admit explicit construction via sheet methods (Lévy, Sato, Gaussian), an integrated weak Itô formula for expectations, and deep connections to the theory of selfdecomposability, temporal selfdecomposability, and stability. Multiparameter extensions allow for processes parameterized by $\mathbb{R}_+^N$, leading to rich families such as Lévy's multiparameter Brownian motion (Hakassou et al., 2012).

2. Iterative Distribution Transport in High-Dimensional Optimal Transport

In computational statistics and vision, Iterative Distribution Transport refers to efficient algorithms for high-dimensional optimal transport (OT) problems. Standard approaches to OT become computationally intractable in high dimensions due to the curse of dimensionality. The IDT method—introduced for color transfer by Pitié et al. and extended by subsequent work—circumvents this by iteratively projecting the source and target high-dimensional distributions onto 1D subspaces, solving the 1D OT problem (often via monotone rearrangement), and lifting the solution back to the original space. This sequence is repeated using different projection directions until convergence is reached (Alghamdi et al., 2020).

Mathematically, at each step the transformation is

$$(X_{nt}; t \geq 0) \stackrel{\mathrm{law}}{=} \left( X_t^{(1)} + \ldots + X_t^{(n)}; t \geq 0 \right),$$0

where $$(X_{nt}; t \geq 0) \stackrel{\mathrm{law}}{=} \left( X_t^{(1)} + \ldots + X_t^{(n)}; t \geq 0 \right),$$1 and $$(X_{nt}; t \geq 0) \stackrel{\mathrm{law}}{=} \left( X_t^{(1)} + \ldots + X_t^{(n)}; t \geq 0 \right),$$2 are the source and target cumulative distribution functions along the chosen projection. The method is especially effective for patch-based color or feature transfer where each data point is a high-dimensional patch vector, and direct OT is impractical in $$(X_{nt}; t \geq 0) \stackrel{\mathrm{law}}{=} \left( X_t^{(1)} + \ldots + X_t^{(n)}; t \geq 0 \right),$$3 dimensions. Smoothed variants of IDT, applying Nadaraya–Watson kernel regression to the 1D transfer maps, are used to address visual artifacts and improve robustness (Alghamdi et al., 2020).

IDT in this context is closely related to Sliced Wasserstein Distance (SWD), differing mainly in the precise mechanism of 1D matching (histogram-matching vs quantile sorting), and both are foundational in practical distribution alignment in image processing.

3. Dynamic Iterative Distribution Alignment in Domain Adaptation

In domain adaptation and transfer learning, iterative distribution transfer mechanisms dynamically align source and target domains under changing distributional importance during training. Dynamic Distribution Adaptation (DDA), as exemplified in (Wang et al., 2019), introduces an adaptive balance coefficient $$(X_{nt}; t \geq 0) \stackrel{\mathrm{law}}{=} \left( X_t^{(1)} + \ldots + X_t^{(n)}; t \geq 0 \right),$$4 that reweights marginal and conditional distribution alignment losses throughout training: $$(X_{nt}; t \geq 0) \stackrel{\mathrm{law}}{=} \left( X_t^{(1)} + \ldots + X_t^{(n)}; t \geq 0 \right),$$5 Here, $$(X_{nt}; t \geq 0) \stackrel{\mathrm{law}}{=} \left( X_t^{(1)} + \ldots + X_t^{(n)}; t \geq 0 \right),$$6 and $$(X_{nt}; t \geq 0) \stackrel{\mathrm{law}}{=} \left( X_t^{(1)} + \ldots + X_t^{(n)}; t \geq 0 \right),$$7 are marginal distributions, $$(X_{nt}; t \geq 0) \stackrel{\mathrm{law}}{=} \left( X_t^{(1)} + \ldots + X_t^{(n)}; t \geq 0 \right),$$8 and $$(X_{nt}; t \geq 0) \stackrel{\mathrm{law}}{=} \left( X_t^{(1)} + \ldots + X_t^{(n)}; t \geq 0 \right),$$9 are class-conditional distributions, and $X^{(1)}, \ldots, X^{(n)}$0 is a discrepancy metric such as the Maximum Mean Discrepancy. The balance $X^{(1)}, \ldots, X^{(n)}$1 is updated iteratively based on current empirical distances, producing an adaptation process that reacts to the evolving alignment difficulty in marginal and conditional components.

MDDA (shallow manifold-based) and DDAN (deep network-based) realize this strategy in closed-form or stochastic-gradient optimization schemes, respectively. This family of techniques brings iteration into the distribution transfer process both at the statistical matching level (through adaptive reweighting) and at the representation learning level via repeated pseudo-label estimation and feature alignment (Wang et al., 2019).

4. Transformer-Based Iterative Distribution Transfer for Multi-View Decomposition

IDT also refers to the Intrinsic Decomposition Transformer, a novel transformer architecture that performs joint, feed-forward multi-view intrinsic image decomposition, enforcing both physical plausibility and cross-view consistency (Du et al., 29 Dec 2025). The network receives a set of multi-view images $X^{(1)}, \ldots, X^{(n)}$2 and outputs a physically-structured decomposition into shared diffuse reflectance, per-view diffuse shading, and per-view specular shading: $X^{(1)}, \ldots, X^{(n)}$3 The architecture leverages a multi-view transformer encoder for joint latent aggregation and introduces factor-specific adapters to disentangle material and lighting factors, crucially outperforming per-view or iterative generative approaches in view-consistency and decomposition interpretablity. The system is designed as entirely feed-forward, exploiting transformer attention to realize global, iteration-like information transfer across all views in a single inference pass (Du et al., 29 Dec 2025).

5. Mathematical Structures and Formal Guarantees

IDT processes in the probability-theoretic sense admit precise structural theorems. For stochastically continuous IDT processes, fixed-time marginals coincide with those of a unique Lévy process; Gaussian IDT processes are characterized via covariance homogeneity $X^{(1)}, \ldots, X^{(n)}$4; and $X^{(1)}, \ldots, X^{(n)}$5-IDT processes capture a broader class, allowing for fractional Brownian motion and selfsimilar stable processes (Hakassou et al., 2012, Hakassou et al., 2012). Lévy-sheet and Sato-sheet constructions provide methodologies for generating IDT processes with prescribed marginal laws.

In computational transport, iterative projection-based IDT methods are justified by Radon-type logic: aligning all 1D projections suffices to align high-dimensional distributions under mild regularity. Convergence is empirical, though full theoretical guarantees are complex due to non-convex composition in projection updates (Alghamdi et al., 2020).

For iterative domain adaptation, dynamic weighting strategies have empirical validation of improved transfer on hard benchmarks and are supported by ablation studies quantifying the advantage of dynamic versus static distribution alignment (Wang et al., 2019).

6. Application Domains

• Probability theory and stochastic processes: classification and construction of processes with prescribed marginal laws, bridging Lévy processes, fractional Brownian motion, and stable laws.
• Computer vision and graphics: color propagation, domain-invariant feature transfer, decomposing image formation models for editing and relighting.
• Domain adaptation and transfer learning: iterative fine-tuning of classifiers where the domain shift involves marginal and conditional distribution changes, robust transfer in the presence of label ambiguity and class imbalance.

IDT methodologies are foundational to a range of modern empirical and theoretical approaches where direct one-shot mapping between complex distributions is infeasible, and iterative, compositional, or multi-scale transfer provides a tractable solution.

7. Further Directions and Open Problems

Emergent variants such as $X^{(1)}, \ldots, X^{(n)}$6-IDT, multi-parameter and multi-view IDT architectures, and adaptive dynamic weighting raise further questions about the relationships between self-similarity, stability, and decomposability in stochastic processes, and how to optimize convergence, robustness, and controllability in high-dimensional transport. In transformer-based multi-view systems, the challenge persists of extending feed-forward joint reasoning beyond static indoor scenes to highly unconstrained visual environments while preserving interpretability and physical consistency (Du et al., 29 Dec 2025).

The multi-domain and cross-disciplinary relevance of IDT makes advances in one area likely to propagate to others, suggesting continued synthesis between stochastic process theory, computational optimal transport, and modern representation learning frameworks.