slideshow 1 slideshow 2 slideshow 3 slideshow 4 slideshow 5 slideshow 6

You are here

Applicant -- Early steps down a messy Bayesian path: what we really know about our solar system and its history

Tuesday, December 3, 2019

Planetary science is full of assumptions. There is much that we surmise, that we know only with high uncertainty. In this talk, tying together my work in a wide array of areas, I focus on how we can wade through messy data and great uncertainty to tease out what we know. I explore this first in the context of terrestrial seismology. I look at how we reveal earthquake signals buried deep in the noise. I look at how the tomography work I've done on the Cascadia region depends on the starting model, how to separate out what we can actually constrain about the Earth. I talk briefly about potential applications to Mars, how these terrestrial techniques allow new ways to pull Mars structure from the noise that dominates event-free martian seismograms. I use my work on neural networks to talk about how and when results with high error can be meaningful. Then, I move to crater counting. I explore how we can make sense of the highly uncertain cratering rate in the Saturn system by calculating relative scaling relationships between the moons, under a wide array of assumptions. I show how we can more explicitly model crater counting error, moving beyond the often wrong sqrt(N) approximation. Finally, I confront the full brunt of the simplifying assumptions in crater counting statistics. There will be an uncomfortable conversation about some of the demons we have left lurking beneath our basic assumptions about what crater counting tells us about the chronology of solar system surfaces, about how we can really tie crater counts to actual dates. I summarize some of the recent radical proposals for smashing the old order of crater counting, from adopting frequentist statistics to rejecting even Poisson statistics and just bootstrapping error. Instead of those approaches, I propose a different path forwards: a fully Bayesian model that incorporates, well, everything. One that is honest about what we know, how well we know it, and what it actually means about our hypotheses and theories. Building this model won't be simple, but there are concrete steps we can take to get there.

Page maintained by
jnspitale [at] (J. Spitale)

PSI, a Nonprofit Corporation 501(c)(3), and an Equal Opportunity/M/F/Vet/Disabled/Affirmative Action Employer.
Corporate Headquarters: 1700 East Fort Lowell, Suite 106 * Tucson, AZ 85719-2395 * 520-622-6300 * FAX: 520-622-8060
Copyright © 2020 . All Rights Reserved.