서브메뉴
검색
Randomized Integrators and Adaptive Step-Size Selection for Hamiltonian Monte Carlo
Randomized Integrators and Adaptive Step-Size Selection for Hamiltonian Monte Carlo
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202104744
- ISBN
- 9798290650869
- DDC
- 574
- 서명/저자
- Randomized Integrators and Adaptive Step-Size Selection for Hamiltonian Monte Carlo
- 발행사항
- [Sl] : Stanford University, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 167 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-03, Section: B.
- 주기사항
- Advisor: Diaconis, Persi;Ying, Lexing.
- 학위논문주기
- Thesis (Ph.D.)--Stanford University, 2025.
- 초록/해제
- 요약This thesis concerns the algorithmic task of sampling from a distribution of known (unnormalized) density µ(dx) ∝ e −U(x)dx on Rd through Markov Chain Monte Carlo (MCMC), a cornerstone method in Bayesian statistics [45], machine learning [3], and computational statistical mechanics [60, 62]. In particular, we consider Hamiltonian Monte Carlo (HMC) [72] and its related variants. Recall that HMC requires the integration of Hamilton's Equations of Motion.This thesis concerns the interplay between the numerical techniques used to integrate (1) - (2), the probabilistic properties of the resulting transition kernel, and the computational cost required to produce an accurate approximation of the target distribution µ. Much of this work is contained in [19, 15, 13].The first two chapters of this thesis provide essential background material on Hamiltonian dynamics and numerical integration. We also present the Gibbs Self-Tuning(GIST) framework of [15], which allows us to consider variants of HMC that select additional "tuning parameters" in each step while maintaining µ-reversibility.Chapters 3 and 4 present new theoretical upper bounds on the computational complexity of sampling via HMC when randomized discretizations of (1) - (2) are used and the function U is assumed both L-gradient Lipschitz and K-strongly convex. We carefully consider the cost of sampling as a function of the dimension d, condition number κ = L K, and accuracy ε.In Chapter 3, we study unadjusted HMC using a randomized integrator we call "stratified Monte Carlo" (sMC). This material is adapted from joint work with Nawaf Bou-Rabee [18]. Under the above Lipschitzness and convexity assumptions we show by coupling arguments that a suitably tuned version of unadjusted HMC with sMC integration produces an ε-accurate approximation of µ in Wasserstein distance using O((d/K) 1/3κ 5/3ε −2/3) gradient evaluations. This result improves upon the corresponding result for unadjusted HMC with Verlet integration, which requires additional smoothness assumptions and gives a complexity that scales with dimension as Ω(d1/2).Chapter 4 considers HMC with a different randomized integrator, which we call the Randomized Splitting Integrator. Unlike the sMC integrator in Chapter 3, the Randomized Splitting Integrator of Chapter 4 allows one to efficiently incorporate a Metropolis-Hastings adjustment to ensure the transition kernel is µ-reversible. We call the resulting algorithm Randomized Splitting Integrator HMC (RSI-HMC). Using the Conductance Profile technique from [28], we show under the Lipschitness and convexity assumptions above that RSI-HMC provided with a β-warm starting distribution, produces an ε-accurate approximation of µ in Total Variation distance using O ( max κ 5/3d 11/12 log ( log(β)/ε ) , κ2/3d5/12log(β) )) gradient evaluations. This matches the upper bound for HMC with Verlet integration obtained in [28] under the additional assumptions that U is LH-Hessian Lipschitz and κ = O(d2/3).Lastly, Chapter 5 presents a novel variant of the No-U-Turn Sampler which incorporates local adaptation of the step-size h in each iteration. This material also appears in [13]. Using the GIST framework, we prove the proposed transition kernel is µ-reversible. We also provide numerical evidence for the efficacy of the resulting sampler on challenging distributions.
- 일반주제명
- Adaptation
- 일반주제명
- Families & family life
- 일반주제명
- Markov analysis
- 일반주제명
- Statistics
- 기타저자
- Stanford University.
- 기본자료저록
- Dissertations Abstracts International. 87-03B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017358736
■00520260202104744
■006m o d
■007cr#unu||||||||
■020 ▼a9798290650869
■035 ▼a(MiAaPQ)AAI32149739
■035 ▼a(MiAaPQ)Stanfordwm808gq7886
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a574
■1001 ▼aMarsden, Milo Steven, Jr.
■24510▼aRandomized Integrators and Adaptive Step-Size Selection for Hamiltonian Monte Carlo
■260 ▼a[Sl]▼bStanford University▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a167 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-03, Section: B.
■500 ▼aAdvisor: Diaconis, Persi;Ying, Lexing.
■5021 ▼aThesis (Ph.D.)--Stanford University, 2025.
■520 ▼aThis thesis concerns the algorithmic task of sampling from a distribution of known (unnormalized) density µ(dx) ∝ e −U(x)dx on Rd through Markov Chain Monte Carlo (MCMC), a cornerstone method in Bayesian statistics [45], machine learning [3], and computational statistical mechanics [60, 62]. In particular, we consider Hamiltonian Monte Carlo (HMC) [72] and its related variants. Recall that HMC requires the integration of Hamilton's Equations of Motion.This thesis concerns the interplay between the numerical techniques used to integrate (1) - (2), the probabilistic properties of the resulting transition kernel, and the computational cost required to produce an accurate approximation of the target distribution µ. Much of this work is contained in [19, 15, 13].The first two chapters of this thesis provide essential background material on Hamiltonian dynamics and numerical integration. We also present the Gibbs Self-Tuning(GIST) framework of [15], which allows us to consider variants of HMC that select additional "tuning parameters" in each step while maintaining µ-reversibility.Chapters 3 and 4 present new theoretical upper bounds on the computational complexity of sampling via HMC when randomized discretizations of (1) - (2) are used and the function U is assumed both L-gradient Lipschitz and K-strongly convex. We carefully consider the cost of sampling as a function of the dimension d, condition number κ = L K, and accuracy ε.In Chapter 3, we study unadjusted HMC using a randomized integrator we call "stratified Monte Carlo" (sMC). This material is adapted from joint work with Nawaf Bou-Rabee [18]. Under the above Lipschitzness and convexity assumptions we show by coupling arguments that a suitably tuned version of unadjusted HMC with sMC integration produces an ε-accurate approximation of µ in Wasserstein distance using O((d/K) 1/3κ 5/3ε −2/3) gradient evaluations. This result improves upon the corresponding result for unadjusted HMC with Verlet integration, which requires additional smoothness assumptions and gives a complexity that scales with dimension as Ω(d1/2).Chapter 4 considers HMC with a different randomized integrator, which we call the Randomized Splitting Integrator. Unlike the sMC integrator in Chapter 3, the Randomized Splitting Integrator of Chapter 4 allows one to efficiently incorporate a Metropolis-Hastings adjustment to ensure the transition kernel is µ-reversible. We call the resulting algorithm Randomized Splitting Integrator HMC (RSI-HMC). Using the Conductance Profile technique from [28], we show under the Lipschitness and convexity assumptions above that RSI-HMC provided with a β-warm starting distribution, produces an ε-accurate approximation of µ in Total Variation distance using O ( max κ 5/3d 11/12 log ( log(β)/ε ) , κ2/3d5/12log(β) )) gradient evaluations. This matches the upper bound for HMC with Verlet integration obtained in [28] under the additional assumptions that U is LH-Hessian Lipschitz and κ = O(d2/3).Lastly, Chapter 5 presents a novel variant of the No-U-Turn Sampler which incorporates local adaptation of the step-size h in each iteration. This material also appears in [13]. Using the GIST framework, we prove the proposed transition kernel is µ-reversible. We also provide numerical evidence for the efficacy of the resulting sampler on challenging distributions.
■590 ▼aSchool code: 0212.
■650 4▼aAdaptation
■650 4▼aFamilies & family life
■650 4▼aMarkov analysis
■650 4▼aStatistics
■653 ▼aBayesian statistics
■690 ▼a0463
■690 ▼a0800
■71020▼aStanford University.
■7730 ▼tDissertations Abstracts International▼g87-03B.
■790 ▼a0212
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17358736▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


