Physicist Amok

I’ll get around to quantum mechanics eventually. Bear with me.

Biology is autonomous from physics. Any change to quantum mechanics will have at most cosmetic implications for biology. Quantum mechanics contributes nothing more than the existence of atoms and molecules, which are necessary for the lossless transmission of information. However, anything that replaces present day quantum mechanics must also predict atoms and molecules. The interactions between the two subjects a minimized because their interface is pinned by experiment.

Within physics, thermodynamics, fluid mechanics, elasticity, and the other macroscopic theories bear the same relation to quantum mechanics. The interfaces between the subjects are experimentally pinned.

A similar pattern appears within biology. Population genetics is built on molecular biology, but their interface is pinned by the existence of genes. We can replace our understanding of an allele’s molecular character, but that allele’s propagation in a population isn’t going to dramatically change.

This pinning is a form of damage control. All fields of science depend on other fields to be able to bootstrap themselves. The only way to keep the structure from falling to shreds is to pin the interfaces. A particle physicist uses macroscopic equipment, which experimentally obeys classical mechanics and electromagnetism. If the relation between his equipment and what he studies weren’t experimentally fixed in the intervening scales, his experiments would be impossible.

It is worth keeping your field as self contained as possible. Black boxes that reach into the heart of entirely different fields are a recipe for disaster.

This is why I dislike the Copenhagen interpretation of quantum mechanics. It posits an observer, whose observation collapses the wave function to an eigenstate of the observation in question. In so doing, it reaches directly to the heart of neuroscience, and builds quantum mechanics on the hardest, most central questions of an even larger area of science.

If the neuroscientists come up with the answer that consciousness isn’t anything special, just a pattern of spikes in neurons, then we have an insoluble problem. Further, there isn’t likely to be a physically robust structure of these firings which is consciousness, so what’s to stop other random patterns of particles from observing and collapsing wave functions? These difficulties are so great, that any interpretation — that is, a connection of the undisputed mathematical structure to reality — which is properly pinned at its boundaries are immediately preferable.

Yet most physicists aren’t willing to accept a new interpretation which requires conceptual gimmicks. Visualizations are tools with local use, not integral parts of theories. Since Heisenberg, we want our theories only to relate observable quantities, not enforce a particular picture. This was the great downfall of the Bohm-de Broglie pilot wave mechanics.

Thankfully, there is an interpretation which is both properly pinned and satisfies the Heisenberg aesthetic. There’s a beautifully written, absolutely simple book on it (Consistent Quantum Mechanics by Robert Griffiths). The book is available for free online.

There’s been a hullabaloo about a New York Times op-ed by one Paul Davies which claims that rational ordering of the universe is an article of faith. Blog posts followed.

Some gems: “The most refined expression of the rational intelligibility of the cosmos is found in the laws of physics.” (Paul Davies) No, the most refined expression of rational intelligibility is a properly randomized experiment. Theoretical physics is not king of the sciences, and most of science won’t change one whit if the entire forefront of theoretical physics reaches a dead end.

“Over the years I have often asked my physicist colleagues why the laws of physics are what they are.” (Paul Davies again) He gives various naive answers. The proper answer is, “I don’t know, and I can’t think of a good way to answer such a question, so I’m going to keep it in the back of my mind in case I do, and get on with questions I can answer.”

“The only problem is that inductive reasoning is not sound.” (first comment on the first blog post I linked to above) Here is someone with little exposure to logic, who doesn’t realize that deductive reasoning isn’t sound either. You can choose any of an infinite number of rules for manipulating sets of well defined strings of symbols. The applicability of any of them is an empirical question. Classical deduction occupies no privileged place. It’s just old.

Then Davies rambles about how the emergence of life is sensitive to the details of the universe. We don’t know this, and no one has thought of any way of finding this out, hallucinations of a few experimentally-challenged string theorists aside, and they have no idea how to construct a science from the ground up. I might even use the word parasite.

Aside from all this, everyone seems to agree that they’re arguing over the proposition “The universe behaves in an orderly and rational manner.” I have no idea what that actually means, and I think there’s a better statement of the required proposition: “There exists some finite level of detail which guarantees the outcome of a protocol/algorithm/recipe.” (I don’t think we have a word for what I have offered three to convey.)

That statement is much weaker, and can be considered in an even weaker form: “For a GIVEN protocol/algorithm/recipe, there is a finite level of detail which guarantees the outcome.”

If we assume that tomorrow will be much like yesterday, then this statement’s converse is falsifiable, though it may not be finitely so. This isn’t perfect, but it’s a long way from a leap of faith.

If we don’t assume that tomorrow will be much like today, we can’t get anywhere. Christians don’t assume this (they expect a Judgement Day, when tomorrow will decidedly not be like yesterday), but fail to realize that you could just as strongly assert that tomorrow there just wouldn’t be a god anymore. So, although I don’t know how to demonstrate the axiom, I don’t have demonstrate it in an argument between science and faith. I would need to demonstrate it in an argument between science and skepticism.

Everyone talks about how the NIH funding system is broken. I rarely see numbers besides falling success rates. No one ever talks about real numbers such as the cost of the current system. How much do grants actually cost?

I have tried for a conservative estimate. The real cost may be anywhere from a third again as large to twice as large.

I have left out airfares for professors to attend study sections, paychecks of the bureaucrats, panels to assign grants to institutes, heads of study sections, time spent in study sections, and a wealth of other costs. Some of these are absorbed in the total award cost NIH reports for grants, since half of most grants disappears into overhead, both for the NIH and the institution which receives it.

How long does writing and reviewing take? How much does it cost? I assume $150/h as the going rate. This is what my mother, a professional science writer, charges. MDs and PhDs on the same projects regularly bill $200-$300 an hour. She estimates producing a document takes 2.5h/page and costs $375/page. Reviewing or editing a document takes 0.2h/page and costs $30/page.

An NIH R01 grant — the backbone of the funding system — is 25 pages long. In 2006, the average success rate (counting all repeat submissions as a single grant) was 0.128 (weighting new and continued grants by the number of each). The average award size with the same weighting was $363,731.08.

Each grant gets four reviewers. 25 pages takes 5h and costs $750. Writing it takes 62.5h and costs $9,375. The total cost for a single grant is 4*$750 + $9375 = $12,375. The time for a single grant is 82.5h. We can assign the reviewing cost without worrying about where in detail it should be counted: the reviewers are grant-seeking scientists as well, and someone must review their grants, so this is a shared cost in the community.

A success rate of 0.128 means we need to submit 7.81 grants in order to get one funded. Each R01 gets three resubmissions (which are still counted as one submission in the NIH statistics). Anecdotally, nothing is getting funded right now on first submission, so we have to multiply the expected number of grants by three, or 23.44 grant equivalents.

The mean cost of a successful grant is $290,039.06. The mean time for a successful grant is 1,933.8h. A full time job of 40h a week, 52 weeks a year is 2080h. Getting a grant is 0.93 of a full time job. It is 0.8 of the first year of the average grant size mentioned above, and remember that most of that money doesn’t make it to the professor at all. A full professor can reasonably ask for $220,000 a year. The first year and a third of the grant is eaten by the costs of getting the grant.

The final summary:
Cost of a grant: $290,039.06
Annual award of a grant: $220,000
Time to get a grant: 1,933.8h
Full time job: 2080h/year

Before we wade in to fix this, let’s set a target for what constitutes “fixed.” I say reduce cost and time to a tenth of their current value. Then we should develop a set of possible systems, and run controlled, randomized experiments to compare them.

BioCurious points to a Nature Physics editorial that increases my contempt for the vanity journals. All but one paragraph is hypocrisy — injunctions not to label your work as ‘ultrashort quantum nanobiology,’ when Nature accepts no work not so labeled — but that one is revealing:

‘Story’ is the concept that should underlie the structure of the entire paper. The clearer and simpler, the more engrossing it is. On that basis, think about relegating technical details — essential to the science but not the narrative — to a Methods section or to Supplementary Information (the latter published online). Similarly, figures should be designed to enhance the telling of the story and each accompanied by a caption that is as short as possible; to an expert reader, the information conveyed in a figure should be clear without needing to consult the main text.

Do exactly the opposite and you have begun well.

Shun captions. When layout was hard and expensive, pictures and graphs were excised from their surrounding text and typeset on their own pages. This time is behind us. Restore your figures to their natural environment. Dismiss the captions you set to watch over them.

Leave your scaffolding in place. When you have jailed your technical information in ‘Materials and Methods’ or exiled it entirely, what is left but verbiage, citations, and graphs taken on faith? The details of your analysis are shunted elsewhere. Even the measurement technique is banished from its point of use.

Supplemental information has one use: source code, analogous to plasmids and strains in genetics. We would attach our plasmids, too, if only we knew how.

Most damning of all, science isn’t stories. We reason about the world from hypotheses we have justified by stringent test. A mixing angle in quantum field theory is not a story, nor the Gibbs distribution, nor template directed synthesis of DNA. Their value is independent of any story around them. They have value as they provide traction for testing other hypotheses. Wrapping them in a story merely imposes on your reader to unwrap them.

I spent a little time this morning looking into source control management systems. Today I watched Linus Torvalds’s talk at Google. I’m going to use Git, for the same reason that C is still a standby in programming.

C is a terrible language. Compared to its contemporaries (ALGOL 68, Lisp 1.5, Smalltalk) it’s primitive. Compared to today’s top languages (Haskell, ML, Scheme) it seems ludicrous. C has something these languages lack. It is the codification of how a couple of smart and experienced programmers liked to write assembly. It provides support where its designers wanted support, and nowhere else. Such a language can never be planned or designed, nor should it, but it provides something to fall back if you can’t find a clean language appropriate for your project.

Git represents Linus implementing a system to support how he wants to do source code control, based on years of a large and particularly chaotic development process. Developing software within a laboratory is chaotic and unstructured. I have trusted Linus for years on the kernel of my operating system. I’ll trust him on version control.

The only other thing of real interest out there is darcs, which is version control based on a theory of patches as noncommuting operations in analogy to Hermitian operators in quantum mechanics. Intellectually, it’s very exciting, but I have become jaded and bitter and I need to finish my PhD, not mess with beautiful computer science.

I don’t want to write raw Java again, but I need to write ImageJ plugins. I’m on the hunt for a new language.

Almost a year ago, I wrote an interface between Kawa and ImageJ, but ImageJ’s design precludes giving Kawa the type annotations it needs to make the interface reasonably fast.

I am no longer appalled by languages like Java. They are muck heaps. The last good language in this family was ALGOL 68. I am appalled by Scala. It tries to be a modern language while still poking its tendrils into the muck heap, but managing the muck has contaminated the language.

Certain things tell me I have been in Haskell too long. I am mildly disturbed by assigning types as values. The case classes, which try to be algebraic data types, miss the point, since you can’t check if functions over them are total.

The primary horrors:

Covariant/contravariant types: Java isn’t strongly typed. You can cast a String to an Object, assign it an integer, and the compiler cannot tell. To escape this, Scala introduces covariant/contravariant types.

Consider the function cons :: a -> [a] -> [a]. The first argument can be any subclass of a. We call it "covariant." The second argument, any superclass of a works in the list. It is "contravariant."

The best solution is to remove inheritance entirely, but the muck monster won’t talk to anyone without inheritance.

Implicit parameters: A library call returns a list. You want to use it as a set, which you have set up as a subclass of list. Write a function which converts lists to sets, annotate it with the implicit keyword, and then you can happily treat the library’s return value as a set.

There are two uses for this. One lets you create typeclasses by defining "traits" (Java interfaces with arbitrary restrictions removed) and then implicitly coercing everything you want into this trait. The other lets you add fields and methods to existing classes and objects.

In the first use, the only claim the creators advance for this approach instead of typeclasses is that implicit parameters are scoped. I learned my lesson about implicit things when I had to maintain a large Perl code base, but this makes the implicit conversions in Perl look like child’s play.

The second use is ludicrous. In Smalltalk, if I want to extend a predefined class C with a method f, I just assign the function body to the field f of C. I can only use the public interface of C, so it isn’t dangerous.