Dembski's argument in Chicago -- New? Persuasive?

On August 14, William Dembski spoke at the Computations in Science Seminar at the University of Chicago. Was this a sign that Dembski’s arguments for intelligent design were being taken seriously by computational scientists? Did he present new evidence? There was no new evidence, and the invitation seems to have come from Dembski’s Ph.D. advisor Leo Kadanoff. I wasn’t present, and you probably weren’t either, but fortunately we can all view the seminar, as a video of it has been posted here on Youtube.

It turns out that Dembski’s current argument is based on two of his previous papers with Robert Marks (available here and here) so the arguments are not new. They involve considering a simple model of evolution in which we have all possible genotypes, each of which has a fitness. It’s a simple model of evolution moving uphill on a fitness surface. Dembski and Marks argue that substantial evolutionary progress can only be made if the fitness surface is smooth enough, and that setting up a smooth enough fitness surface requires a Designer.

Briefly, here’s why I find their argument unconvincing:

  1. They conside all possible ways that the set of fitnesses can be assigned to the set of genotypes. Almost all of these look like random assigments of fitnesses to genotypes.
  2. Given that there is a random association of genotypes and fitnesses, Dembski is right to assert that it is very hard to make much progress in evolution. The fitness surface is a “white noise” surface that has a vast number of very sharp peaks. Evolution will make progress only until it climbs the nearest peak, and then it will stall. But …
  3. That is a very bad model for real biology, because in that case one mutation is as bad for you as changing all sites in your genome at the same time!
  4. Also, in such a model all parts of the genome interact extremely strongly, much more than they do in real organisms.
  5. Dembski and Marks acknowledge that if the fitness surface is smoother than that, progress can be made.
  6. They then argue that choosing a smooth enough fitness surface out of all possible ways of associating the fitnesses with the genotypes requires a Designer.
  7. But I argue that the ordinary laws of physics actually imply a surface a lot smoother than a random map of sequences to fitnesses. In particular if gene expression is separated in time and space, the genes are much less likely to interact strongly, and the fitness surface will be much smoother than the “white noise” surface.
  8. Dembski and Marks implicitly acknowledge, though perhaps just for the sake of argument, that natural selection can create adaptation. Their argument does not require design to occur once the fitness surface is chosen. It is thus a Theistic Evolution argument rather than one that argues for Design Intervention.

That’s a lot of argument to bite off in one chew. Let’s go into more detail below the fold …

Dembski and Marks’s argument involves defining a new form of information, showing that it is conserved. Evolution can succeed only if this information is already present, so therefore evolution does not bring about new information. In Dembski’s case he goes on from that to make a theological argument (in his recent book), which I gather is basically “In the Beginning is the Information”.

People like to argue about how one ought to define information, but I’m going to ignore most of those arguments, because I think that there is a simpler problem that undercuts the Dembski-Marks argument. My argument here is not new (it has been given before at Panda’s Thumb (here and also here and here). But with a new wave of publicity for Dembski and Marks’s argument, it’s worth pointing out in more detail the flaw in their argument.

A typical fitness surface?

Dembski and Marks have a simple model with genotypes and fitnesses. Of course it is overly simple, but all models are. It is worth examining, because if evolution is in trouble in such a model, we need to know why. Their computation of information is a measure of how smoothly the fitnesses change as one moves from one genotype to another, where neighboring genotypes are those that can be reached from each other by evolutionary processes such as mutation. If the fitness is smooth enough, one can find neighboring genotypes that are better, and the natural selection will tend to move the population to those.

To figure out how common smooth fitness surfaces are, Dembski and Marks invoke Bernoulli’s “Principal of Insufficient Reason”. That basically says that if we can’t think of a reason to consider probabilities of different outcomes unequal, we should consider the probabilities all to be equal. The use of Bernoulli’s Principle underlies all of Dembski and Marks’s calculations.

In the case of fitness surfaces, the outcomes are all the different ways that fitnesses can be assigned to genotypes. So if we have DNA sequences, and the genome is 1000 bases long, there are 4 x 4 x 4 x … x 4 different genotypes, with 1000 factors of 4 in the product. That is about 10 raised to the 602nd power. If each of these possible genotypes has a different fitness, there are also that many fitness values. The Principle of Insufficient Reason says that, lacking any reason to think otherwise, we should give each of the possible ways that the 10-to-the-602 fitnesses could be assigned to the genotypes an equal chance of being true.

A typical one of this vast number of possibilities has fitnesses randomly assigned to genotypes. If that is the case, then when we change a genotype by making a single mutation in it, we arrive at a new genotype that has a fitness that is, in effect, chosen from all the possible fitnesses, at random.

What a mutation does

What if, instead of changing one base, we took the drastic step of mutating all of the bases in the genotype at the same time? If the Bernoulli Principle applied, we would get to a genotype whose fitness was also chosen at random. So in that case, on average, that would be no better and no worse than changing just one base. In other words, when fitnesses are randomly assigned to genotypes making a single typographical error is exactly as bad as changing every letter in the text.

Real biology doesn’t work anything like that. Making one mutation in one of my genes will on average make it worse, though sometimes not. If it produces a protein, a single amino acid change often leaves the protein still functioning. But making changes in every site of its DNA is the same as replacing every protein by a random string of amino acids. Which will be a complete disaster.

Similarly, in statements in English, one typographical error might change “to be or not to be that is the question” into “to be or not to de that is the question”. Changing all letters would give something like “bdglvwujzib lxmoxg rjdg a ohlowugrbl owj”. It should be obvious that the latter is far less functional. The comprehensibility of English sentences is more like the actual fitness of organisms, and not like the fitness of the organisms Dembski and Marks imagine.

Unbelievably strong interactions

In Dembski and Marks’s “white noise” fitness surfaces, another bizarre property is that every part of the genome interacts incredibly strongly with every other part. If they did not interact strongly, we would get cases like this: we might find that changing position number 834 in the DNA from C to T would make the genome somewhat worse. And changing position 95161 from A to C might also make the genome worse. If those two positions in the DNA underwent both of these changes at the same time, we could reasonably expect that this accumulation of two bad changes would be worse yet.

But if the fitnesses are assigned to genotypes at random, that prediction could not be made. The double mutant would have a randomly-chosen fitness and that would have only about a 1/4 chance of being worse than either of the single mutants. In fact, about 1/4 of the time it would actually be better than the original genotype! We can immediately see that this could only happen if the two parts of the genome were tightly interacting in some way. But in the Dembski-Marks white noise model all parts of the genome interact tightly with each other.

No real organism works that way. And there is a simple reason why.

What physics does

The reason is “because physics”. In the physical laws of our universe, interaction at a distance gets weaker and weaker as the distance increases. This is an everyday fact that we rely on all the time. As I type these words my fingers and the keys move. There is an (extremely) slight gravitational and electrostatic effect of those movements on (say) the food in your refrigerator. That effect declines with distance. As a result, you don’t have to worry that my typing is busy rearranging the food in your refrigerator. The eggs will be right where you left them, and this will not depend on whether I type the letter A or the letter B.

Similarly, in the genome, a gene that functions in the growth of your toenails typically shows no strong interaction with another gene that controls nerve connections in your ear. They are physically far from each other and probably function at different times as well.

In Dembski and Marks’s model universe things don’t work that way. If one particular gene has a mutant, we can’t know anything about what its effect is, until we check all other genes. A change in any one of those others will make a major difference in what the effect of the first mutant is. And this is not just something that happens occasionally. It is always true, for all parts of the genome. Every gene, and every base in every gene, interacts incredibly tightly with all other bases in all other genes.

Why the white-noise model prevents evolution

The fitness surfaces implicit in Dembski and Marks’s argument are known as “white noise” fitness surfaces. White noise has a signal whose values are uncorrelated from one time to another. The white noise fitness function is the same – fitnesses of closely similar genotypes are totally dissimilar. Knowing the fitness of your genotype simply provides no prediction as to what the fitness will be if the base at one point in your DNA is changed.

Natural selection with mutation and recombination can work its way uphill on the fitness surface by putting together individually favorable changes. If the fitness surface does not allow any prediction that such combinations will often be better than either change alone, then this is a big problem for evolution. The evolutionary process will frequently get stuck.

Fortunately, “because physics” white noise fitness functions basically don’t exist.

What Dembski and Marks’s argument doesn’t do

It is notable that Dembski and Marks’s argument is not actually an Intelligent Design argument. It argues that a Designer is needed to explain the shape of the fitness surface, but once that surface is smooth enough, natural selection and other evolutionary forces do the rest. So there is no Design Intervention needed.

Is evolution a search? Is it important whether it is?

The audience at Dembski’s talk in Chicago seemed to think that the crucial issue is whether evolution is or is not a “search”. Strictly speaking, in a model of evolution like the one he is using, I think that the answer is no. But it actually is not important whether it is or isn’t. Given the issue of whether a white noise fitness function is the default, Dembski’s argument is invalid even if one allows him the point that evolution is a search.

Has this criticism of Dembski’s arguments been made before?

Dembski also used a white noise fitness function in his No Free Lunch argument, and in the Search For a Search papers he and Marks acknowledge that connection. In the No Free Lunch argument the performance of the search that moves uphill on the fitness surface is extremely poor if averaged over all possible fitness functions. This is the same as its behavior on a typical randomly-chosen fitness function. At least seven major criticisms of Dembski’s No Free Lunch argument have objected that white noise fitness functions are not realistic (links to their articles and posts are given in my 2007 article and in a summary I wrote here at Panda’s Thumb). The criticism goes back to 2002 and has been voiced by all these authors.

Dieb’s argument with Dembski and Marks’s theorem

Mathematical blogger “Dieb” (Dietmar Eben) has raised the issue (here, here and here) of whether Marks and Dembski have actually proven their Horizontal No Free Lunch Theorem. His arguments are interesting and strike me as cogent. But whether or not that theorem is proven, the point remains that evolution will do badly almost all the time on a white noise fitness function. So a smoother fitness function is required. But, as we have seen, the laws of physics make a white noise fitness function unlikely. This is true whether or not the HNFL theorem can be proven rigorously.

New types of information? Important to arguing for Design?

The point about physics and the unlikelihood of white noise fitness functions is also true however we define information, and it is true whether natural selection “creates” information or whether it takes existing information that is implicit in the smoothness of the fitness surface and repackages it in the genome. I suspect that Dembski and Marks’s “active information” will end up not being a helpful concept, but for the purposes of my present critique that issue is not central.

What Richard Dawkins’s “Weasel” model was not intended to do

One should note in passing Dembski’s use of Richard Dawkins’s “Methinks It Is a Weasel” model. In his Chicago talk, Dembski portrays Dawkins as arguing that the Weasel model shows that natural selection can originate information, and portrays Dawkins as claiming that it is a realistic model of evolution. Dawkins was not arguing that it was a realistic model of evolution, or that this evolution originated new information. Dawkins’s model was a teaching example to show why creationist debaters who argue that natural selection is doing a “random” search are disingenuous. The Weasel search succeeds in about 1000 steps, while a truly random search would take astronomical numbers of steps. Dawkins’s model is an effective teaching device. It is routinely misrepresented in the creationist and ID literature as intended to be a realistic model of evolution, and intended to prove assertions about where the information in life originates. Unfortunately Dembski has followed this sad tradition.

Is Dembski’s theology of information central to his argument about evolution?

No, because he’s got to end up arguing that, for the laws of phyics to be the way they are, requires some active Design. But once the laws of physics are admitted, how they got that way is just not part of any argument about evolution. Biologists will certainly decide not to waste time on the issue and to leave it to cosmologists.