Collaborative Research: Highly Structured Models and Statistical Computation in High-Energy Astrophysics

主要研究人员：David Van Dyk和xiali mengk；机构：加州大学河滨分校和哈佛大学合作研究：高能天体物理学中的高度结构化模型和统计计算摘要加州-哈佛天体统计学合作旨在设计和实现完全基于模型的统计推断方法，以解决高能天体物理学中突出的数据分析问题。该合作的方法明确地模拟了天文来源和新高科技仪器中固有的数据生成机制的复杂性，并充分利用由此产生的高度结构化模型来了解潜在的天文和物理过程。使用这些模型需要复杂的科学计算、先进的统计推断方法和仔细的模型检查程序。协作项目的pi （van Dyk和孟）在开发协作项目正在扩展、采用和宣传的方法方面都有丰富的研究经验：在涉及多层潜在变量和不完整数据的高度结构化模型下的推理和高效计算方法。这种模型非常适合解释高能天体物理学中构成数据生成机制的许多物理和仪器过滤器。该项目的五位顾问（Chiang, Connors,Kashyap， Karovska和Siemiginowska）都在高能天体物理学的仪器和科学方面具有专业知识，并且都与统计学家合作，努力开发适当的方法来解决科学问题。这个项目有两个主要的影响：发展更可靠的统计方法对天文学科学发现的影响，以及新的统计推断和计算方法在广泛的科学领域的影响。当协作组为特定的推理任务开发方法和分发免费软件时，它也教育天文学界小心使用复杂的统计方法的好处。（该组织每年在美国天文学会的会议上组织一到两次特别会议。）预计这项研究的基本影响将是在天文学家中更普遍地接受和更普遍地使用适当的方法。其次，这次合作是统计推断新模式的一个例子。比起使用现成的模型和方法，开发特定于应用程序的模型变得更加可行，这些模型旨在解释手头问题的特定复杂性。协作开发推理和计算方法来处理这种多层次的模型。随着应用特定的多层次模型变得越来越普遍，这些方法将在整个自然科学、社会科学和工程科学中得到应用。近年来，在观测高能天体物理学方面出现了大量新数据。最近发射的或即将发射的天基望远镜，旨在探测和绘制紫外线、x射线和伽马射线的电磁发射，为研究宇宙打开了一扇全新的窗口。由于高能电磁发射的产生需要数百万度的温度，并且是大量储存能量释放的标志，这些仪器对宇宙的热和湍流区域提供了全新的视角。新的仪器允许非常高的分辨率成像，光谱分析和时间序列分析。例如，钱德拉x射线天文台产生的图像比以往任何x射线望远镜都清晰至少30倍。仪器的复杂性、天文来源的复杂性和科学问题的复杂性导致了一个微妙的推理问题，这需要复杂的统计工具。例如，数据受到不均匀的审查，测量误差和背景污染。天文光源表现出复杂和不规则的空间结构。科学家们希望得出这样的结论：物质环境和物质来源的结构，支配行星、恒星和星系的诞生和死亡的过程和规律，最终是宇宙的结构和演化。然而，很少有人努力利用现代统计方法的力量来解决这些问题。加州-哈佛天体统计合作开发统计方法、计算技术和免费软件，以解决高能天体物理学中突出的推理问题。所开发的方法是统计推断新模式的一个例子：与其使用现成的方法，不如开发特定于应用程序的方法，并设计用于解释手头问题的特定复杂性，这变得越来越可行。协作组织为处理这种多层次模型而设计的推理和计算方法在整个自然科学、社会科学和工程科学中都有应用。

Pricipal Investigators: David Van Dyk and Xiao-Li MengInstitutions: UC Riverside and Harvard UniversityCollaborative Research: Highly Structured Models and Statistical Computation in High-Energy AstrophysicsAbstractThe California-Harvard Astrostatistics Collaboration aims to designand implement fully model-based methods of statistical inference tosolve outstanding data analytic problems in high-energyastrophysics. The Collaboration's methods explicitly model thecomplexities of both astronomical sources and the data generationmechanisms inherent in new high-tech instruments and fully utilize theresulting highly structured models in learning about the underlyingastronomical and physical processes. Using these models requiressophisticated scientific computation, advanced methods for statisticalinference, and careful model checking procedures. The PIs of theCollaboration (van Dyk and Meng) both have substantial researchexperience in developing the methods that the Collaboration isextending, employing, and publicizing: inferential and efficientcomputational methods under highly-structured models that involvemultiple levels of latent variables and incomplete data. Such modelsare ideally suited to account for the many physical and instrumentalfilters that compose the data generation mechanism in high-energyastrophysics. The five consultants on the project (Chiang, Connors,Kashyap, Karovska, and Siemiginowska) all have expertise on theinstrumentation and science of high-energy astrophysics, and, all havecollaborated with statisticians in efforts to develop appropriatemethods to address scientific questions. There are two primary impactsof this project: the impact of the development of more reliablestatistical methods on scientific findings in astronomy and the impactof the new statistical inference and computation methods in a widerange of scientific fields. As the Collaboration develops methods anddistributes free software for specific inferential tasks, it alsoeducates the astronomical community as to the benefit of careful useof sophisticated statistical methods. (The Collaboration organizes oneor two special sessions at meetings of the American AstronomicalSociety each year.) It is expected that a fundamental impact of theproposed research will be a more general acceptance and more prevalentuse of appropriate methods among astronomers. Second, theCollaboration is an example of a new mode of statisticalinference. Rather than using off-the-shelf models and methods, it isbecoming ever more feasible to develop application specific modelsthat are designed to account for the particular complexities of aproblem at hand. The Collaboration develops inferential andcomputational methods for handling such multi-level models. Asapplication specific multi-level models become more prevalent, thesemethods will have application throughout the natural, social, andengineering sciences.In recent years, there has been an explosion of new data inobservational high-energy astrophysics. Recently launched orsoon-to-be launched space-based telescopes that are designed to detectand map ultra-violet, X-ray, and gamma-ray electromagnetic emissionare opening a whole new window to study the cosmos. Because theproduction of high-energy electromagnetic emission requirestemperatures of millions of degrees and is an indication of therelease of vast quantities of stored energy, these instruments give acompletely new perspective on the hot and turbulent regions of theuniverse. The new instrumentation allows for very high resolutionimaging, spectral analysis, and time series analysis. The ChandraX-ray Observatory, for example, produces images at least thirty timessharper than any previous X-ray telescope. The complexity of theinstruments, the complexity of the astronomical sources, and thecomplexity of the scientific questions leads to a subtle inferenceproblem that requires sophisticated statistical tools. For example,data are subject to non-uniform censoring, errors in measurement, andbackground contamination. Astronomical sources exhibit complex andirregular spatial structure. Scientists wish to draw conclusions as tothe physical environment and structure of the source, the processesand laws which govern the birth and death of planets, stars, andgalaxies, and ultimately the structure and evolution of theuniverse. Nonetheless little effort has been made to bring thestrength of modern statistical methods to bare on these problems. TheCalifornia-Harvard Astrostatistics Collaboration develops statisticalmethods, computational techniques, and freely available software toaddress outstanding inferential problems in high-energy astrophysics.The methods developed are an example of a new mode of statisticalinference: Rather than using off-the-shelf methods, it is becomingever more feasible to develop methods that are application specificand are designed to account for the particular complexities of aproblem at hand. The inferential and computational methods designed bythe Collaboration for handling such multi-level models haveapplication throughout the natural, social, and engineering sciences.