Distribution Contraction in Statistical Analysis

• Distribution contraction is a mathematical framework where update rules or flows reduce discrepancies between probability distributions and target objects across various regimes.
• It quantifies convergence in diverse settings—such as Bayesian nonparametrics, MCMC, and PDEs—using metrics like Wasserstein, Hellinger, and relative entropy.
• Applications range from deconvolution and Gibbs sampling to tensor contraction in deep learning, revealing both exponential and polynomial convergence under different structural conditions.

Searching arXiv for recent and foundational uses of “distribution contraction” and closely related contraction terminology to ground the article. {"query":"all:(\"distribution contraction\" OR \"posterior contraction\" OR \"entropy contraction\" OR \"transportation cost contraction\" OR \"contraction rate posterior\")","max_results":10,"sort_by":"relevance"} {"query":"id:(Gao et al., 2015) OR id:(Ascolani et al., 2024) OR id:(Natile et al., 2010) OR id:(Podder et al., 2015) OR id:(Knapik et al., 2014) OR id:(Fan et al., 25 Jan 2026) OR id:(DaeiJavad et al., 2 Jun 2026)","max_results":10,"sort_by":"relevance"} The literature suggests that “distribution contraction” is not a single standardized term but a family of mathematically distinct phenomena in which an update rule, stochastic dynamics, or inferential procedure drives a distribution toward a reference object, a true parameter, or a structurally simpler class. In Bayesian nonparametrics, it denotes posterior concentration around a true mixing distribution or density; in Markov-chain theory, it denotes decay of divergences such as relative entropy or the hockey-stick divergence under a kernel; in transport equations, it denotes nonincrease or exponential decay of transportation costs along Fokker–Planck flows; and in some machine-learning architectures it denotes a regularizing compression of the learned density by multilinear structure (Gao et al., 2015, Ascolani et al., 2024, Natile et al., 2010, Silva et al., 2024).

1. Core definitions and metrics

In Bayesian usage, a sequence $\varepsilon_n$ is a posterior contraction rate if, for a metric $d$ and sufficiently large $M$,

$$\Pi_n\bigl(d(\theta,\theta_0)>M\varepsilon_n\mid X^{(n)}\bigr)\to 0$$

in probability under the true model. In the deconvolution setting of Dirichlet–Laplace mixtures, the relevant metrics are Wasserstein distance $W_p(G,G_0)$ for the latent mixing distribution, Hellinger distance

$h^2(f,f_0)=\int (\sqrt f-\sqrt{f_0})^2\,dx,$

and $L_q$ norms for the mixed density $p_G$ (Gao et al., 2015).

In transport and PDE settings, contraction is formulated through transportation functionals rather than posterior mass. For a continuous nondecreasing cost $h:[0,\infty)\to[0,\infty)$ with $h(0)=0$, the transportation functional is

$d$0

with Kantorovich dual formulation

$d$1

For $d$2, this recovers $d$3 (Natile et al., 2010).

In MCMC, contraction is often expressed through divergences between laws pushed forward by a Markov kernel. For the Gibbs sampler under strong log-concavity, the central quantity is relative entropy

$d$4

while the more recent local/global framework uses the hockey-stick divergence

$d$5

The associated global contraction coefficient of a kernel $d$6 is

$d$7

(Ascolani et al., 2024, DaeiJavad et al., 2 Jun 2026).

A distinct discrete-probabilistic formulation appears for Galton–Watson trees. There the state space is the simplex

$d$8

and contraction means that the recursion operator $d$9 shrinks total variation or Euclidean distance, implying a unique fixed point for the distribution over Ehrenfeucht equivalence classes (Podder et al., 2015).

2. Posterior contraction in Bayesian inverse and deconvolution problems

A canonical example is deconvolution with Dirichlet–Laplace mixtures. Observations satisfy $M$0, with $M$1 on $M$2, latent variables $M$3, noise $M$4 distributed according to the Laplace density $M$5, and observed density $M$6. The base measure $M$7 is supported on $M$8 and has a Lebesgue density bounded away from $M$9 and $$\Pi_n\bigl(d(\theta,\theta_0)>M\varepsilon_n\mid X^{(n)}\bigr)\to 0$$0 on $$\Pi_n\bigl(d(\theta,\theta_0)>M\varepsilon_n\mid X^{(n)}\bigr)\to 0$$1; the true $$\Pi_n\bigl(d(\theta,\theta_0)>M\varepsilon_n\mid X^{(n)}\bigr)\to 0$$2 is supported on the same interval. Under these assumptions, the posterior contracts for the mixing distribution in Wasserstein distance at

$$\Pi_n\bigl(d(\theta,\theta_0)>M\varepsilon_n\mid X^{(n)}\bigr)\to 0$$3

and in particular

$$\Pi_n\bigl(d(\theta,\theta_0)>M\varepsilon_n\mid X^{(n)}\bigr)\to 0$$4

For the mixed density, the posterior contracts at rate $$\Pi_n\bigl(d(\theta,\theta_0)>M\varepsilon_n\mid X^{(n)}\bigr)\to 0$$5 in Hellinger distance and at rate $$\Pi_n\bigl(d(\theta,\theta_0)>M\varepsilon_n\mid X^{(n)}\bigr)\to 0$$6 in $$\Pi_n\bigl(d(\theta,\theta_0)>M\varepsilon_n\mid X^{(n)}\bigr)\to 0$$7, $$\Pi_n\bigl(d(\theta,\theta_0)>M\varepsilon_n\mid X^{(n)}\bigr)\to 0$$8; for $$\Pi_n\bigl(d(\theta,\theta_0)>M\varepsilon_n\mid X^{(n)}\bigr)\to 0$$9, the $W_p(G,G_0)$0 rate matches the Hellinger rate. The paper also states that the mixed-density rate $W_p(G,G_0)$1 matches the minimax rate $W_p(G,G_0)$2 up to a logarithmic factor for densities with Fourier transform decaying at order $W_p(G,G_0)$3, whereas the minimax rate for the latent mixing distribution under $W_p(G,G_0)$4 is unknown (Gao et al., 2015).

The general inverse-problem perspective is given by a modulus-of-continuity theorem. If $W_p(G,G_0)$5 is a continuous injective map and the posterior for the direct parameter $W_p(G,G_0)$6 contracts at rate $W_p(G,G_0)$7 in a metric $W_p(G,G_0)$8, while the posterior concentrates on a sieve $W_p(G,G_0)$9, then the inverse posterior for $h^2(f,f_0)=\int (\sqrt f-\sqrt{f_0})^2\,dx,$0 contracts at the localized modulus

$h^2(f,f_0)=\int (\sqrt f-\sqrt{f_0})^2\,dx,$1

This allows contraction rates for priors not related to the singular value decomposition of the operator, including spline priors and location-mixture priors. The paper recovers, among other examples, the mildly ill-posed sequence-model rate

$h^2(f,f_0)=\int (\sqrt f-\sqrt{f_0})^2\,dx,$2

for Gaussian series priors and the severely ill-posed rate

$h^2(f,f_0)=\int (\sqrt f-\sqrt{f_0})^2\,dx,$3

for truncated Gaussian priors (Knapik et al., 2014).

For nonlinear PDE parameter identification with Gaussian process priors, posterior contraction is established without assuming that the truth lies in the reproducing kernel Hilbert space of the prior. The analysis uses a rescaled Matérn prior, a truncated Karhunen–Loève approximation, and an algebraic balancing of approximation, small-ball, entropy, and testing errors. In the three PDE examples developed in the paper, the resulting concrete rates are $h^2(f,f_0)=\int (\sqrt f-\sqrt{f_0})^2\,dx,$4 for diffusion coefficient identification and $h^2(f,f_0)=\int (\sqrt f-\sqrt{f_0})^2\,dx,$5 for the elliptic and subdiffusion inverse potential problems; the variational posterior contracts at the same rate under the stated truncated Gaussian variational family (Fan et al., 25 Jan 2026).

A recurrent theme in these results is that contraction for the observed or transformed object is faster than contraction for the latent inverse parameter. This suggests that ill-posedness is encoded directly in the gap between direct and inverse contraction rates (Gao et al., 2015, Knapik et al., 2014).

3. Entropy contraction for Gibbs sampling and related MCMC kernels

For the random-scan Gibbs sampler targeting $h^2(f,f_0)=\int (\sqrt f-\sqrt{f_0})^2\,dx,$6 on $h^2(f,f_0)=\int (\sqrt f-\sqrt{f_0})^2\,dx,$7, strong log-concavity and smoothness are expressed by Hessian bounds

$h^2(f,f_0)=\int (\sqrt f-\sqrt{f_0})^2\,dx,$8

or more generally by block smoothness and coordinate-wise strong convexity with coordinate-wise condition number $h^2(f,f_0)=\int (\sqrt f-\sqrt{f_0})^2\,dx,$9. If $L_q$0 is the random-scan Gibbs kernel over $L_q$1 blocks, then

$L_q$2

Iterating yields exponential decay,

$L_q$3

and the paper states that, under the cost model in which one-block conditional evaluations cost $L_q$4 of a full target evaluation, the number of full evaluations needed for mixing grows linearly with the condition number and is independent of the dimension. When strong log-concavity fails, the rate degrades from exponential to polynomial, with

$L_q$5

under the stated convex assumptions (Ascolani et al., 2024).

The local/global contraction framework for MCMC extends this viewpoint from entropy to the full hockey-stick divergence profile. For projected Langevin Monte Carlo on a compact convex domain $L_q$6 of diameter $L_q$7, with Gaussian smoothing scale $L_q$8 and batch-dependent smoothness $L_q$9, the one-step coefficient is

$p_G$0

where

$p_G$1

This gives exponential contraction of the full hockey-stick profile toward the discretized stationary law $p_G$2 for general smooth, possibly non-convex potentials and arbitrary random-batch sampling schemes (DaeiJavad et al., 2 Jun 2026).

The same paper shows that independent Metropolis–Hastings with unbounded importance weight typically has trivial global contraction coefficients. It therefore introduces a local coefficient on the core

$p_G$3

and proves

$p_G$4

For $p_G$5-warm starts, this yields the bound

$p_G$6

When $p_G$7 for some $p_G$8, the rate becomes polynomial in $p_G$9; in a heavy-tailed example with no finite moment of order $h:[0,\infty)\to[0,\infty)$0, the paper obtains the explicit rate $h:[0,\infty)\to[0,\infty)$1 (DaeiJavad et al., 2 Jun 2026).

A common misconception is that one-step global contraction should exist whenever an MCMC kernel is ergodic. The independent Metropolis–Hastings analysis shows that this is generally false for unbounded importance weights, and that a core-and-tail decomposition can be essential (DaeiJavad et al., 2 Jun 2026).

4. Transportation-cost contraction for Fokker–Planck equations

For nonnegative measure-valued solutions of the Fokker–Planck equation

$h:[0,\infty)\to[0,\infty)$2

distribution contraction is expressed through the evolution of transportation costs under a monotone or $h:[0,\infty)\to[0,\infty)$3-monotone drift:

$h:[0,\infty)\to[0,\infty)$4

With the summability condition

$h:[0,\infty)\to[0,\infty)$5

the transportation functional $h:[0,\infty)\to[0,\infty)$6 contracts along two solutions $h:[0,\infty)\to[0,\infty)$7 for every continuous nondecreasing $h:[0,\infty)\to[0,\infty)$8 (Natile et al., 2010).

In the monotone case $h:[0,\infty)\to[0,\infty)$9,

$h(0)=0$0

For $h(0)=0$1-monotone drifts, a space-time rescaling reduces the problem to the monotone case. If $h(0)=0$2 and $h(0)=0$3 satisfies the scaling condition

$h(0)=0$4

then

$h(0)=0$5

In particular, for Wasserstein costs $h(0)=0$6,

$h(0)=0$7

If $h(0)=0$8, any invariant probability measure satisfying the integrability condition is unique, and every solution converges exponentially to it in any cost $h(0)=0$9 with the stated scaling (Natile et al., 2010).

The proof is based on a duality method rather than a gradient-flow or displacement-convexity argument. The key device is a variable-doubling comparison principle for solutions of the backward Kolmogorov equation, transported through the Kantorovich dual constraints. The paper emphasizes that the drift need not be a gradient, no growth condition is imposed, and the method applies directly to distributional solutions (Natile et al., 2010).

The scope is nevertheless specific: the paper does not cover general nonmonotone drifts, and the contraction results are proved for unit diffusion rather than variable-coefficient diffusion matrices (Natile et al., 2010).

5. Recursive and structural contraction on trees and networks

For Galton–Watson trees with $d$00 offspring, a finite set $d$01 of $d$02-move Ehrenfeucht equivalence classes induces a distributional recursion on the simplex

$d$03

If $d$04 denotes the true law of the equivalence class of the rooted tree, then $d$05 is a fixed point of a smooth map $d$06 built from Poisson thinning and a recursion function $d$07. The contraction theorem is explicit in the subcritical regime:

$d$08

For general $d$09, the paper proves that a bounded iterate $d$10 contracts in Euclidean distance,

$d$11

which implies uniqueness of the fixed point and geometric convergence of iterations. The same contraction structure yields smooth dependence of first-order probabilities on the offspring parameter $d$12 via the Implicit Function Theorem (Podder et al., 2015).

A different structural meaning of distribution contraction appears in contracting random networks. For configuration-model networks undergoing random deletion, preferential deletion, or propagating deletion, the evolving degree distribution $d$13 remains uncorrelated and converges toward the Poisson law

$d$14

The relative entropy

$d$15

is nonnegative, vanishes only at the Poisson distribution, and the paper derives a decomposition

$d$16

in which the “trickle-down” contribution satisfies

$d$17

The conclusion is that $d$18 decreases monotonically to zero during contraction, so the degree law converges to Poisson and the graph structure converges to Erdős–Rényi (Tishby et al., 2020).

The related master-equation analysis of deletion-driven network contraction makes the same structural endpoint explicit. Under random deletion,

$d$19

and for an Erdős–Rényi initial network the mean degree decays linearly:

$d$20

Preferential and propagating deletion add redistribution terms that deplete probability above the mean and accelerate the loss of heavy tails and degree-degree correlations, ultimately yielding the same Erdős–Rényi limit (Tishby et al., 2020).

These two examples suggest a common discrete principle: a contraction operator on a simplex or degree space can force uniqueness, convergence, and simplification even when the underlying objects are trees or graphs rather than ordinary probability densities (Podder et al., 2015, Tishby et al., 2020).

6. Representation contraction in machine learning

A distinct, application-specific usage appears in conditional variational autoencoders for tabular data. There, tensor contraction layers map a tokenized feature matrix $d$21 to a new grid by the multilinear operation

$d$22

or equivalently

$d$23

The paper describes this as “distribution contraction”: replacing dense linear mappings on flattened one-hot inputs with multi-linear tensor contractions that enforce low-rank structure, compress parameters, and regularize the learned densities. The proposed mechanism is that TCLs contract the support of $d$24, filter spurious correlations, and produce a parsimonious latent representation for mixed-type tabular data (Silva et al., 2024).

The empirical comparison spans four conditional VAE architectures: a baseline linear VAE, a TCL-focused model, a transformer-focused model, and the hybrid “TensorConFormer.” Averaged across 62 OpenML CC18 classification datasets, TensorConFormer attains the best 1-way marginals score, $d$25, the best pairwise correlations score, $d$26, the best $d$27-Recall, $d$28, the best utility (TSTR), $d$29, and the best fidelity, $d$30; TensorContracted attains the best $d$31-Precision, $d$32 (Silva et al., 2024).

The same paper also reports that the transformer-only VAE fares worse, and that in small-feature and small-sample regimes the hybrid’s decoder transformer can over-diversify, lowering $d$33-recall relative to the baseline VAE. This suggests that, in this usage, “distribution contraction” denotes a regularization strategy rather than a contraction theorem on a metric space (Silva et al., 2024).

7. Synthesis and recurring themes

Across these literatures, distribution contraction is tied to explicit operators and explicit metrics. Posterior contraction uses Hellinger, $d$34, Wasserstein, or $d$35 metrics to quantify concentration toward the truth; Gibbs and Langevin analyses use relative entropy or $d$36 to obtain one-step SDPI bounds; Fokker–Planck theory uses transportation costs generated by continuous nondecreasing $d$37; Galton–Watson recursion uses total variation and Euclidean norms on a simplex; network contraction uses relative entropy to a matching Poisson law; and tensor-contraction models use the term for a low-rank regularization of learned densities (Gao et al., 2015, Ascolani et al., 2024, Natile et al., 2010, Podder et al., 2015, Tishby et al., 2020, Silva et al., 2024).

Several contrasts recur. First, contraction of an observed or mixed distribution is often faster than contraction of a latent or inverse object, as in deconvolution and inverse problems (Gao et al., 2015, Knapik et al., 2014). Second, exponential contraction typically requires strong structure—strong log-concavity, monotonicity, or a compact-domain smoothing effect—whereas weaker structure yields polynomial or merely local rates (Ascolani et al., 2024, DaeiJavad et al., 2 Jun 2026, Natile et al., 2010). Third, when global contraction is unavailable, localization on a sieve, a core set, or a finite-state abstraction can restore quantitative control (Knapik et al., 2014, DaeiJavad et al., 2 Jun 2026, Podder et al., 2015).

A plausible synthesis is that “distribution contraction” names a general mathematical pattern: the existence of a map, flow, or posterior update that decreases an appropriately chosen discrepancy to a target set, often with rates that reveal the geometry, regularity, or ill-posedness of the underlying problem.