October 31, 2009

Books to Read While the Algae Grow in Your Fur, October 2009

Attention conservation notice: I have no taste.

Rosemary Kirstein, The Lost Steersman
Sequel to Steerswoman's Road (below); excellent and perfectly continuous, despite a long gap in the writing. The trick of celebrating intelligence while maintaining the tone and color of a good fantasy novel is not something I have encountered elsewhere, and find deeply addictive. — Sequel
Everything else I have to say is a spoiler: This owes a massive debt to Lovecraft's At the Mountains of Madness. The plot-hinge mystery here has to do with "demons", amphibious barrel-shaped creatures with quadrilateral symmetry, very like (though not exactly the same as) Lovecraft's Antarctic Old Ones. There are scenes of dissecting demons under the impression that they are just animals, and realizing they belong to some radically different division of life than familiar terrestrial organisms; an exploring expedition to an unknown part of the world where the demons are found; explorations of demons' cities and observations of their customs, including subterranean chambers used for their rituals, etc.; and the dawning realization that the creatures are in fact sapient. (HPL: "Radiates, vegetables, monstrosities, star spawn — whatever they had been, they were men!" RK does not put such florid outbursts in her characters' mouths; she just has Rowan come to see that the demons are "people".) Kirstein does a better job, in my view, of making the creatures actually alien, in particular starting from giving them a very inhuman sensorium (continual sonar, without any vision) and means of communication (excreting specially shaped lumps of organic material, reminiscent of the pieces of carved soapstone Lovecraft associated with his Old Ones), and building out logically from there. Needless to say, this complicates the ethics of terraforming the steerswomen's planet considerably. — Janus's speeches (pp. 43--44 and p. 356) about the dangers of learning too much about the world also seems drawn from Lovecraft, though they bring to mind the opening of "The Call of Cthulhu" more than Mountains.
Update, 25 March 2012: Kirstein explains some of the genesis of the story.
Jeffrey D. Hart, Nonparametric Smoothing and Lack-of-Fit Tests
A sound, friendly but reasonably theoretical introduction to nonparametric regression, giving about equal attention to kernel-based methods and to series expansions (Fourier series, orthogonal polynomials, etc.). The first half of the book, through ch. 4, introduces these methods, considers their ability to predict new data (emphasizing, naturally, the bias-variance trade-off), and looks at methods for selecting how much smoothing to do based on the data being smoothed, with a fondness for leave-one-out cross-validation and its variants. (I can't recall if k-fold CV is even mentioned.) The second half is about testing parametric regression specifications. Chapter 5 reviews some classical tests for fully-specified and especially for linear-in-the-parameters parametric models, including Neyman's smooth tests: the latter involve, roughly speaking, fitting an orthogonal series to the deviations from the null model, and checking that all the coefficients are small, and so form a bridge to the smoothing-based tests used in the rest of the book. Basically, one can either smooth the parametric residuals, which should have mean zero and constant variance under the null hypothesis, or compare the parametric estimate to the nonparametric smooth. Hart prefers the former approach, and develops tests for regression functions being constants in chapter 7, which in chapter 8 are turned into tests for departures from arbitrary parametric regression models. The distribution of these test statistics is too complicated for anything except bootstrapping, which needs to be done carefully to preserve power. To simplify the math, up to this point Hart assumes that the input variable takes values at a deterministic set of points on the unit interval ("fixed-design univariate regression"); chapter 9 generalizes to random-design and multivariate regressions, as well as lifting some other restrictions. Chapter 10 contains some case studies of real data.
This book should be accessible to anyone who understands parametric inference at the level of, say, All of Statistics; no prior exposure to smoothing methods is really needed. The series-expansion methods will probably go down more easily with some priori exposure to Fourier analysis. People who are serious about using parametric regression models in the real world (cough econometricians cough) owe it to themselves to test them with these methods.
Rosemary Kirstein, Steerswoman's Road
First two books in an epic fantasy series about the scientific method, reprinted in one volume. There are more books, which I now covet powerfully, but the series is not finished.
Spoilers: "Epic fantasy" here is, I am pretty sure, totally misleading. Initially, and from most of the characters' perspectives, the world looks like a bog-standard medieval fantasyland, only with the addition of an itinerant semi-monastic order of geographers and natural philosophers, the eponymous steerswomen. By the end of this volume, however, I am pretty sure that the setting is actually another planet in this universe, with no magic at all. The steerswomens' world is being terraformed; the Guidestars are satellites in geosynchronous orbit. The native ecology (based on "blackgrass" and "redgrass") is being systematically destroyed (by microwave heating from the orbiting Guidestars ("the spell Routine Bioform Clearance"), and by the Outskirters' goats [which may be genetically modified?]) and replaced by terrestrial flora ("greengrass"), microbes and fauna. Wizards are simply the inhabitants of the planet who retain the old technology, such as electricity and explosives. Why most of the colonists have regressed to medieval technology, and why, having done so, they have an institution like the steerswomen, I couldn't tell you. (I can tell you that sailors and steerswomen are immune to some "curses" because they wear rubber-soled, i.e. electrically insulating, boots.) But I am dying to find out.
J. K. Ghosh and R. V. Ramamoorthi, Bayesian Nonparametrics
I have written extensively about the general subject of Bayesian nonparametrics and especially of its consistency elsewhere (here, here, or, indeed, here), so I'll just plunge in. This 2003 monograph is the best overview of Bayesian nonparametrics from the viewpoint of theoretical statistics which I've found, though there has been a great deal of work since it was written, and I know that a number of new books are coming out soon.
The author begin (ch. 1) by reviewing* results on the consistency of Bayesian learning on finite sample spaces and Dirichlet prior distributions. They then carefully (ch. 2) consider the measure-theoretic issues involved in constructing prior probability distributions over infinite-dimensional spaces, especially priors over all probability measures (or all probability densities) on the real line. Chapter 3 describes in detail the properties of Dirichlet-process priors and of Polya tree priors. Chapter 4 is concerned with consistency for Bayesian updating with IID data, emphasizing the "Kullback-Leibler property" (the prior must put sufficient weight on distributions with small relative entropy from the truth) and the exponentially-consistent-testing conditions which go back to Schwartz. Chapter 5 specializes to inferring probability densities; this is the only place they use Gausian process priors. Chapter 6 considers inferring the location parameter of distributions of unknown shape, and outlines (without full detail) the notorious examples, due to Freedman and Diaconis, of how Bayesian learning can fail to be consistent. Chapter 7 considers linear regression with an unknown noise distribution; this is the only departure from assuming IID data made here. The remaining chapters try to construct uniform distributions on infinite-dimensional spaces, look at some issues in survival analysis, and technical aspects of "neutral to the right" priors, ones whose cumulative hazard functions have independent increments. It is assumed throughout that the true, data-generating distribution lies within the support of the prior.
Ghosh and Ramamoorthi focus on mathematical issues, to the exclusion of computational and statistical considerations. (There are no applications to data, or even to elaborate simulations.) The writing is adequate for a work of "theorem-proof, theorem-proof" math, but no more. Those proofs, however, are really clear and clean, without tricks or complications. I recommend the book for those who want to understand, in depth, the technicalities of constructing priors on infinite-dimensional spaces, and of establishing their consistency when updated with IID data. There are a handful of exercises at the end of the book, but I do not think it would be suitable as a classroom textbook. It could work as the first part of an advanced graduate seminar, or for self-study for motivated and mathematically mature readers.
*: Actually, it's my impression that lots of introductions to Bayesian statistics, even at the graduate level, do not cover these results. This is, I think, something of a scandal for the profession. That goes double if it's due to the attitude which Ghosh and Ramamoorthi (p. 122) paraphrase as "the prior and the posterior given by Bayes theorem [sic] are imperatives arising out of axioms of rational behavior --- and since we are already rational why worry about one more" criterion, namely convergence to the truth. One does indeed find pernicious relativism and epistemic nihilism everywhere these days!
Nadia Gordon, Lethal Vintage
Continuing amateur-sleuthing adventures of a Napa Valley restaurant-owner and her foodie (and wine-y) friends. No prior acquaintance with the series [1, 2, 3] needed.
Larry Gonick, The Cartoon History of the Modern World, Part II: From the Bastille to Baghdad
My parents got the first part of the Cartoon History of the Universe (in its original, shorter edition) for my brother and I in 1981, when I was seven. We loved it so much they ended up having to get us two copies. I have thus been reading the History, as it came out, all my conscious life. (And re-reading it, without any visitations from the Suck Fairy so far.) This latest volume is, as always a delight, but not a pure one, because it's also the last. I can understand wanting to be finished with the work of a lifetime, especially one which in the nature of things could be spun out indefinitely; but I can't help wishing for more.
G. Willow Wilson and M. K. Perker, Air, vol. 2: Flying Machine
James Sethna, Statistical Mechanics: Entropy, Order Parameters, and Complexity
The best introductory statistical mechanics book I have ever seen. (Meaning: advanced undergraduates, not the graduate level of Landau and Lifshitz.) The reader is supposed to have some familiarity with classical and quantum mechanics, a little electromagnetism, and the very barest rudiments of thermodynamics, the latter not going beyond what's in a good first-year physics course. Beyond the basics of differential equations and linear algebra, the only real pieces of math used here are Fourier transforms and elementary probability (such as one sees in undergraduate quantum mechanics). On this basis Sethna erects classical (and, in one chapter, quantum) statistical mechanics, emphasizing the modern applications of the theory and physical intuition.
The exposition begins with random walks, including diffusion and the central limit theorem. The micro-canonical ensemble comes next, along with a very nice chapter on its ergodic basis and failures of ergodicity (such as KAM theory). The other ensembles are derived from imposing the micro-canonical ensemble on the whole system, and looking at marginal distribution of sub-systems. The elaborate axiomatic structure of pure thermodynamics is touched on only briefly; thermodynamic quantities are seen, quite properly, as derivative of statistical-mechanical ones. The question of what macroscopic variables need to be included in the free energy leads naturally to a superb chapter on the meaning and identification of order parameters. This in turn is followed by a really lucid treatment of the connections between spontaneous fluctuations, the decay correlations, response to external forces, and the dissipative approach to equilibrium. The whole is capped off by chapters on abrupt (e.g., ice-water, water-steam) and continuous (e.g., magnetic) phase transitions, including a nice hand-waving discussion of the renormalization group. In addition to the main thread of exposition, each chapter has a large collection of problems, ranging from mathematical proofs through calculations to simulation challenges, which contain a lot of neat applications and special topics, and should at least be read if not attempted.
There are a few places where I would quibble — per Lebowitz, surely the Boltzmann entropy is more useful out of equilibrium than the Gibbs?; couldn't he have been more explicit about the probabilistic foundations of renormalization? — but mostly I just wish this book had been written sixteen years ago when I was taking stat. mech.
Disclaimer: Friends of mine used to work for Sethna, and he's lectured at the SFI summer school (the chapter on order parameters began as a lecture there in 1991), but I've never met him, and have no stake in the success of the book.
Update: Thanks to T. A. Abinandanan for alerting me to the fact that there's a free PDF of the whole book!
Laura E. Reeve, Vigilante
Sequel to Peacekeeper, with an even more awful and totally-misleading cover. (The synposes at the link are accurate, however.) Tasty mind-candy.
C. L. Anderson, Bitter Angels
Space opera about active struggles to prevent war, and other morally-compromising endeavors; military science fiction that lets me respect myself in the morning. The climax, where it becomes clear what is going on, and why and how, and what the peace-keepers will do about it, with what consequences, was very fine indeed. Picked up after reading the author's self-advertisement on Scalzi's blog, which has more.
Madeleine E. Robins, Petty Treason
More alternate-history Regency England private-eye detection (romance-free this time). Very enjoyable; I wish there were more. — And now there is.

Books to Read While the Algae Grow in Your Fur; Enigmas of Chance; Scientifiction and Fantastica; Pleasures of Detection, Portraits of Crime; Writing for Antiquity; The Great Transformation; Physics; Complexity; Cthulhiana; Bayes, anti-Bayes

Posted at October 31, 2009 23:59 | permanent link

Three-Toed Sloth