LSST Informatics and Statistics Science Collaboration

I. Science Justification [limit 4 pages/3200 words]

Note: Attach any figures to the figures section.

A. Describe the science goals of the proposed collaboration and provide the context for overall significance to astronomy/physics.

The unprecedented science opportunities arising from LSST's wide-field cosmic cinematography will be accompanied by equally unprecedented data analysis challenges, due to the huge size and synoptic scope of LSST data products. While the most obvious challenges are those due to the petabyte scale of fundamental LSST databases, in fact LSST will present astronomers with new and difficult analysis problems spanning a broad range of sizes, types, and complexity, and requiring a matching breadth of methodological research to address. The challenges of science analysis---deriving knowledge and understanding from the vast LSST datasets---will require methods, algorithms, and codes that extend or transcend those in common use in astronomy today.

We propose to create a new LSST Informatics and Statistics Science Collaboration (ISSC) to help coordinate interdisciplinary interactions and pursue data analysis methodology research in collaboration with other Science Collaborations (SCs) and the broader LSST user community; we will also offer our services to the LSST Data Management team. The ISSC will be interdisciplinary, with membership split (currently 60%/40%) between astronomers and information scientists who will bring to LSST expertise gained in other communities facing large and complex data analysis challenges. By design, the ISSC includes a number of astronomer members of other SCs, to facilitate coordination of collaborative and cross-cutting activities. We have created a proposal support web site that includes a categorized team list that displays our team breakdown more clearly than is possible with the proposal personnel entry form; this site is at:

http://inference.astro.cornell.edu/lsst/proposal09/

To quickly see the need for the ISSC, envision a "methods cube," where the first dimension is binned by application (e.g., science problems covered by existing SCs), the second dimension is binned by problem scale (measured by the logarithm of the number of objects or database size being analyzed), and the third dimension is binned by task type (e.g., source detection, classification, multivariate density estimation, etc.). Each LSST science project corresponds to an entry in one or more cells in the methods cube. The observation motivating our effort to form the ISSC is that projecting out the application dimension produces a 2-D task/scale grid where relatively few occupied cells will have only one entry. That is, nearly every type of task will be used in multiple application areas, and many task/scale combinations will arise in multiple application areas. Each such coincidence represents an opportunity where research alliances cutting across astronomy application areas can benefit LSST science. Further, other disciplines have their own methods cubes, with numerous applications occupying the same task/scale cells as LSST applications. Each of these coincidences represents an opportunity where interdisciplinary research may benefit LSST science. The ISSC will work to identify promising areas of methodological overlap between applications and disciplines, bring existing methods to bear on these overlap problems, and build alliances pursuing research on new methods where needed. These activities promise to significantly improve the science return from LSST.

We elaborate here on this cross-cutting, interdisciplinary perspective to establish the need for the ISSC and delineate its roles in more detail. We first describe the diversity of problem scales, and then the diversity of analysis tasks that LSST scientists will face, across diverse application areas. This will motivate the main goals of the ISSC:

1. To enable new LSST science by leveraging recent developments in information sciences;
2. To focus consultation from experts in astrostatistics and astroinformatics to serve the full LSST science effort, covering advanced methods and high computational efficiency;
3. To stimulate and guide new, astronomy-focused information science research for application to LSST issues;
4. To improve cross-fertilization and reduce duplication on methodological development across the LSST community.

To establish the context for and significance of ISSC activities, we begin by surveying the range of problem scales spanned by LSST science. We distinguish three scales for the number of objects or samples ("entities") being analyzed (each possibly with multiple attributes). The three scales differ by factors of ~10**3, roughly delineating regimes where computational resource constraints lead to fundamental shifts in how one thinks about data analysis problems.

• "Kiloscale" (10**2 to 10**4 entities) - The fundamental kiloscale LSST problem is analysis of multicolor, multi-epoch photometry for a single object in the Object Catalog, where the entity is a photometric measurement (the actual samples will be obtained from associated entries in the Source Catalog). Much of this will be study of source variability in several bands which falls under the large rubric of multivariate time series analysis. A second class of kiloscale problem arises in population-level analysis of catalog data for modest-sized populations; e.g., trans-Neptunian objects (TNOs; ~20k expected), gamma-ray burst (GRB) hosts (potentially ~100 per year), microlensing events, and relatively rare classes of stars, galaxies, or clusters. Kiloscale problems are generally not challenging in data processing or storage, but will benefit from a variety of statistical techniques such as outlier and change point detection, survival analysis (treatment of upper limits), periodicity searches, and statistical inference involving heteroscedastic (different for each sample) measurement errors.
• "Megascale" (10**5 to 10**7 entities) - Megascale problems include population-level analysis of larger populations (e.g., quasars; variable Galactic stars; low redshift galaxies); population-level re-analysis of previous survey catalogs (e.g., SDSS, FIRST) supplemented by LSST follow-up; and analysis of multicolor, multi-epoch image data for extended objects where the entity is a pixel. Megascale problems require processing datasets occupying storage of several megabytes to gigabytes, corresponding to "low-volume" queries against LSST data products. In this regime, statistical methods must be computationally efficient and advanced data visualization techniques are needed.
• "Gigascale" (10**8 to 10*10 entities) - A class of gigascale problem using Level-One LSST data products is analysis of calibrated image data for a single LSST field or a modest number of fields. A second class, using Level-Two object catalogs, includes population-level analysis of very large samples of stars (about 20 billion total expected) or galaxies (about 10 billion total expected). A third class includes the search for rare or serendipitous objects based on queries over large parts of the Level-One source catalog. Gigascale problems require very efficient (and thus limited) processing of data streams from datasets occupying terabytes of storage, corresponding to "high-volume" queries against LSST data products. Several nightly DMS pipeline tasks also fall into this category.

Roughly speaking, kiloscale problems will raise challenges that are essentially statistical: devising methods for handling novel types of data and models that optimally extract information from LSST data and provide careful uncertainty quantification. Gigascale problems will raise challenges that are more essentially computational, in the realm of informatics: finding algorithms that make it possible to do specific, relatively straightforward tasks across enormous databases; sophisticated statistical modeling will typically not be possible. Megascale problems occupy a middle ground, where significant innovation may be required on both the statistical and computational fronts.

Many important problems have a hierarchy of scales. A key example is nightly pipeline processing: localized image processing is needed to identify sources and associate them with previously detected objects; but this apparently kiloscale problem must be executed for billions of sources per night, making the overall problem gigascale and constraining the level of statistical sophistication.

Having identified representative scales for LSST data analysis problems, we now identify representative tasks, many occurring across a range of scales.

Discovery - These tasks address detecting and identifying astronomical objects, structures, and events:

• Adjusting thresholds to control the number of false detections or associations, taking into account the huge multiplicity of tests that a large survey performs. This is a fundamental DMS task, determining the behavior of nightly detection pipeline processing. Control of the False Discovery Rate (FDR) under multiple testing is a hotbed of statistics research to treat, for example, bioinformatics issues arising in the use of DNA microarray technology. The InCA group (described below) has described initial application of FDR control to astronomical source detection in modest-sized catalogs, pointing to the value of this line of research for LSST astronomy.
• Faint source detection in multicolor/multi-epoch data cubes. The DMS deep detection pipeline will address this task for Data Releases (DRs). It will also arise in re-analyses of Level-One data products that attempt to push the boundary of LSST to the dimmest sources, e.g., for TNO, faint star (e.g., white dwarf) and faint (high redshift) galaxy and AGN studies, particularly using "deep drilling" data. While in many cases, the faintest sources will be found in images merged from many epochs, in cases where observing conditions change or transients are present, the faintest sources may be found in localized portions of the data cubes. Finding these sources is both a statistical and computational challenge.
• Classification of objects within a population, including both supervised classification (assigning objects to previously-identified classes) and unsupervised classification (where the number and characteristics of classes are derived from the data), arising in the study of all sizable populations, from minor planets to distant galaxies. A wide variety of multivariate classification tools are available (see, for instance, the text `Pattern Classification' by Duda, Hart and Stork 2002), often far more capable than the traditional color-index cuts familiar in photometric survey analysis. The options are reduced for gigascale problems due to computational limitations.
• Flexible and adaptive study of variability and transients in time series. The detection of short-lived transients or marginal variability in sources is a problem in statistical inference where the heteroscedastic measurement errors due to changing observing conditions will play an important role. Once variability is established, both time domain and frequency domain time series techniques are relevant. A number of algorithms for establishing autocorrelated and/or periodic variations can be applied to the time series in both a single photometric band or all bands simultaneously. Simple nonparametric measures like the partial autocorrelation function or the minimal string length measure may be appropriate for scanning for variability and periodicity in gigascale LSST catalogs.
• Cross-matching within and between large catalogs, with accurate accounting for astrometric uncertainties. This task arises in all application areas involving correlative analysis, and becomes particularly challenging if the target catalog has significant direction uncertainties. Current astrostatistics research efforts are addressing this task by marrying directional statistics, product partition models, and Monte Carlo methods for searching the space of likely associations and accounting for multiple comparisons. These procedures can take into account knowledge of the local densities of objects in the catalogs.

Modeling & Analysis - These tasks involve analysis of images of known sources, or of catalogs of classified objects or events:

• Image analysis, including multi-frame super-resolution, and multiscale deconvolution with uncertainty quantification. When instrumental or atmospheric distortions cause blurring, likelihood-based deconvolution based on the known point spread function can sharpen the image. Large volumes of image segments can be reduced and characterized using sparse representations such as wavelets, curvelets or compressive sensing. Density estimation techniques like locally weighted regression can reconstruct image regions with reduced noise and with variance images to help adjudicate the statistical significance of features.
• Design and comparison of photometric redshift (photo-z) algorithms, including calibration of redshift uncertainties needed for a wide variety of extragalactic research, ranging from luminosity function estimation to dark energy studies using Type Ia supernovae (SNe Ia) and weak lensing. Relevant information science developments are in multivariate density estimation, Bayesian and neural network clustering, and nonparametric regression and density estimation. These are all areas with many relevant recent developments in machine learning and statistics. Also, statistical study can provide guidance for spectroscopic measurements to improve predictions.
• Flexible modeling of multivariate distributions for populations, accounting for selection effects including truncation in a single survey, censoring in a follow-up survey, and measurement error in all surveys. This class of task arises in nearly all population modeling applications; e.g., orbit/size/color distributions of minor planets; color-magnitude diagrams of stars; distance estimator calibration for galaxies (Tully-Fisher and fundamental plane relations); luminosity functions of galaxies, AGN, and transient populations. Relevant, active research frontiers in information sciences include semiparametric and nonparametric inference with heteroscedastic measurement error, and nonparametric and parametric survival analysis.
• Characterizing multicolor variability. A wide variety of temporal behavior will be discovered and studied by LSST. Examples include: periodic variability of minor planets (to study rotation, geometry, and composition); periodic variations of stars due to rotation and pulsation including Cepheid and RR Lyrae period-luminosity-color correlations for use as low-z distance indicators; characterizing the smooth multicolor light curves of SNe Ia for use as distance indicators in dark energy studies; characterizing the wide range of smooth and chaotic variability across AGN populations. Astronomers have already developed a variety of methods for periodicity searches from unevenly spaced data, although research is needed for incorporation of measurement errors. Relevant frontier information science studies include: nonparametric harmonic analysis; multivariate nonparametric regression with sparse Gaussian processes and PCA-based dimension reduction; and space-time modeling (translated to wavelength-time) with multivariate stochastic processes built over overcomplete bases (for sparse representation).

In addition to the specific information science research frontiers identified above, the general area of algorithm development for large-database proximity searches is having a profound impact on the ability to deploy increasingly sophisticated techniques on ever-larger datasets. Examples include the use of kd-trees and metric ball trees to accelerate nearest-neighbor search, k-means clustering, and Gaussian process regression. These developments play key roles in recent and current analyses of SDSS data.

All areas of LSST research will also need advanced data visualization techniques. For kiloscale problems, "Grand Tour" movies of rotating datacubes with interactive brushing and classification are useful. For megascale and gigascale problems, quantile contour maps and shaded density maps must be rapidly produced from portions of the data. Shaded parallel coordinate maps with brushing may also be powerful interactive visualization tools for multivariate data with more than three dimensions.

The list above is hardly exhaustive, neither in the tasks listed nor in the applications cited. Incomplete though it may be, it already makes clear that LSST will pose an almost dizzying variety of research-level data and science analysis challenges. It also makes clear that there are many opportunities to share expertise and research resources across applications, science collaborations, and disciplines.

In summary, nearly every frontier LSST science effort will face significant astroinformatics or astrostatistics challenges arising from the scale, dimension, and novelty of LSST data. Astronomers have faced similar challenges with other data sets in recent decades. Two new disciplines have emerged in response to the uniquely interdisciplinary nature of these challenges:

• Astrostatistics weds the tools of statistics to the needs of astronomers. It focuses on probabilistic modeling of data and quantification of uncertainty.
• Astroinformatics addresses the wide variety of concerns arising in managing, exploring, and analyzing extremely large data sets whose size thwarts straightforward application of standard methods (including optimal astrostatistical methods). It relies heavily on developments in the emerging fields of informatics and knowledge discovery from databases (KDD).

The emergence of these disciplines marks a growing realization within the astronomy data analysis community that the information science disciplines---primarily statistics and computer science---have significant expertise to offer astronomers. It also marks the realization among information scientists that astronomy poses compelling research-level challenges that push the boundaries of information science.

The ISSC will help focus the attention and link the resources of these new disciplines for the needs of LSST. Building this relationship will be synergistic, improving LSST science, and also fostering the growth of astroinformatics and astrostatistics, as disciplines within astronomy, and as disciplines within the information sciences.

The ISSC will be unique among current LSST SCs in that it will focus its work on methodological issues arising in the pursuit of a wide variety of astronomical science, rather than on a specific astronomical research area. By design, all of its work will overlap with that of other SCs (which will be facilitated by the fact that some of the ISSC members are already members of other LSST Science Collaboration teams); in many cases, ISSC efforts will overlap with the research interests of multiple other SCs. We ask that proposal reviewers and LSST leaders consider the unique role of the ISSC, understanding that it differs in structure and purpose from collaborations devoted to specific astronomical problems.

B. What aspects of LSST data and/or features of operation do you need to realize your scientific goals?

For the catalog-based research activities of our group, access to the LSST science database object catalog, source catalog, and (VO)event catalog is essential. We will make use of the 100+ science attributes per object for analysis and mining. Multi-dimensional and temporal data analyses require this.

For image-based research activities, access to the LSST images and pixel data is required, both in the set of calibrated images and in the annual data release sky template image set.

C. Will additional data or information (not provided by LSST) be needed to realize your scientific goals?

The ISSC itself will not directly need additional data or information not provided by LSST to realize its goals.

D. Provide us with some information on your background, skills and experience that are relevant to the general area described above.

Proposing members of the ISSC include a large fraction of U.S. scholars active in astrostatistics and astroinformatics today. It includes members of major cross-disciplinary collaborations including the California-Harvard Astrostatistics Collaboration (CHASC), the International Computational Astrostatistics group (InCA), the Center for Astrostatistics at Penn State, and ongoing collaborations between Cornell astronomers and Duke U. statisticians. The membership includes senior information scientists with decades of experience collaborating with scientists in many fields. Among them are members of the National Academy of Sciences, MacArthur Fellows, fellows of major societies, and distinguished professors. Others are mid-career astronomers who have developed expertise in statistics or computer science. Yet others are young researchers who have begun their careers on the astronomy-statistics-informatics interfaces. Further details about the team appear in Section IV.

E. Provide a quantitative estimate of the level of effort you are willing to dedicate to this work, and when you will begin.

The level of effort devoted to ISSC activities will vary greatly across the ISSC membership, ranging from leading vigorous LSST-focused methodology research programs, to offering occasional consultation to astronomers seeking to benefit from the expertise of ISSC team members. Section VI, describing our management plan, presents more detailed information about the various levels of effort of team members.

The ISSC hopes to begin its work with a "live" team meeting at the January 2010 AAS meeting, which will be attended by many team members.