[KongchangAI]
ConceptDobrushin contraction

Dobrushin Contraction

A coefficient measuring the contraction property of a Markov transition kernel, used to prove geometric ergodicity; this work localizes it to apply to algorithms such as HMC.

Timeline (last 90 days)

Nov 17

The paper establishes a high-probability uniform mixing bound for the critically damped Metropolis-adjusted Langevin algorithm (with persistent uniform refresh), requiring O((κ^{1/2} + L_H^{1/2} m^{-3/4}) d^{1/4}) gradient evaluations (ignoring logarithmic factors), where κ = L/m is the condition number and d is the dimension.

Unverified50%
Nov 17

This stability enables the researchers to localize the analysis to the naturally arising fast local mixing region and seamlessly connect it via broad local frameworks such as Dobrushin contraction, Wasserstein contraction, and local Harris theorems.

Unverified50%
Nov 17

The authors specifically develop a formulation of Dobrushin contraction that provides a solid foundation for the coupling argument.

Unverified50%
Nov 17

The Dobrushin contraction coefficient measures the extent to which a Markov transition kernel "pulls together" any two initial distributions: if there exists ρ < 1 such that W₁(δ_x P, δ_y P) ≤ ρ·dist(x, y) for all x, y, then the chain is said to be contracting in the Dobrushin sense.

Unverified50%
Nov 17

The classical Dobrushin contraction version requires globally uniform contraction, which is too stringent for algorithms like HMC that mix quickly only locally and may be very slow globally.

Unverified50%
Nov 17

The key improvement of this paper lies in localizing Dobrushin contraction: verifying contraction only within the naturally arising "fast local mixing region," and then stitching the local conclusions into a global high-probability bound with the help of the weak-(1,1) maximal inequality.

Unverified50%

All Facts (6)

Unverified

The paper establishes a high-probability uniform mixing bound for the critically damped Metropolis-adjusted Langevin algorithm (with persistent uniform refresh), requiring O((κ^{1/2} + L_H^{1/2} m^{-3/4}) d^{1/4}) gradient evaluations (ignoring logarithmic factors), where κ = L/m is the condition number and d is the dimension.

50%
Unverified

This stability enables the researchers to localize the analysis to the naturally arising fast local mixing region and seamlessly connect it via broad local frameworks such as Dobrushin contraction, Wasserstein contraction, and local Harris theorems.

50%
Unverified

The authors specifically develop a formulation of Dobrushin contraction that provides a solid foundation for the coupling argument.

50%
Unverified

The Dobrushin contraction coefficient measures the extent to which a Markov transition kernel "pulls together" any two initial distributions: if there exists ρ < 1 such that W₁(δ_x P, δ_y P) ≤ ρ·dist(x, y) for all x, y, then the chain is said to be contracting in the Dobrushin sense.

50%
Unverified

The classical Dobrushin contraction version requires globally uniform contraction, which is too stringent for algorithms like HMC that mix quickly only locally and may be very slow globally.

50%
Unverified

The key improvement of this paper lies in localizing Dobrushin contraction: verifying contraction only within the naturally arising "fast local mixing region," and then stitching the local conclusions into a global high-probability bound with the help of the weak-(1,1) maximal inequality.

50%

Source Articles