Uncommon Descent Serving The Intelligent Design Community

Does information theory support design in nature?

Share
Facebook
Twitter
LinkedIn
Flipboard
Print
Email

Eric Holloway argues at Mind Matters that design theorist William Dembski makes a convincing case, using accepted information theory principles relevant to computer science:

When I first began to look into intelligent design (ID) theory while I was considering becoming an atheist, I was struck by Bill Dembski’s claim that ID could be demonstrated mathematically through information theory. A number of authors who were experts in computer science and information theory disagreed with Dembski’s argument. They offered two criticisms: that he did not provide enough details to make the argument coherent and that he was making claims that were at odds with established information theory.

In online discussions, I pressed a number of them, including Jeffrey Shallit, Tom English, Joe Felsenstein, and Joshua Swamidass. I also read a number of their articles. But I have not been able to discover a precise reason why they think Dembski is wrong. Ironically, they actually tend to agree with Dembski when the topic lies within their respective realms of expertise. For example, in his rebuttal Shallit considered an idea which is very similar to the ID concept of “algorithmic specified complexity”. The critics tended to pounce when addressing Dembski’s claims outside their realms of expertise.

To better understand intelligent design’s relationship to information theory and thus get to the root of the controversy, I spent two and a half years studying information theory and associated topics during PhD studies with one of Dembski’s co-authors, Robert Marks. I expected to get some clarity on the theorems that would contradict Dembski’s argument. Instead, I found the opposite.

Intelligent design theory is sometimes said to lack any practical application. One straightforward application is that, because intelligence can create information and computation cannot, human interaction will improve computational performance.
More.

Also: at Mind Matters:

Would Google be happier if America were run more like China? This might be a good time to ask. A leaked internal discussion document, the “Cultural Context Report” (March 2018), admits a “shift toward censorship.” It characterizes free speech as a “utopian narrative,” pointing out that “As the tech companies have grown more dominant on the global stage, their intrinsically American values have come into conflict with some of the values and norms of other countries.”

Facebook’s old motto was “Move fast and break things.” With the current advertising scandal, it might be breaking itself A tech consultant sums up the problem, “Sadly Facebook didn’t realize is that moving fast can break things…”

AI computer chips made simple Jonathan Bartlett: The artificial intelligence chips that run your computer are not especially difficult to understand. Increasingly, companies are integrating“AI chips” into their hardware products. What are these things, what do they do that is so special, and how are they being used?

The $60 billion-dollar medical data market is coming under scrutiny As a patient, you do not own the data and are not as anonymous as you think. Data management companies can come to know a great deal about you; they just don’t know your name—unless, of course, there is a breach of some kind. Time Magazine reported in 2017 that “Researchers have already re-identified people from anonymized profiles from hospital exit records, lists of Netflix customers, AOL online searchers, even GPS data of New York City taxi rides.” One would expect detailed medical data to be even more revelatory.

George Gilder explains what’s wrong with “Google Marxism”
In discussion with Mark Levin, host of Life, Liberty & Levin, on Fox TV: Marx’s great error, his real mistake, was to imagine that the industrial revolution of the 19th century, all those railways and “dark, satanic mills” and factories and turbine and the beginning of electricity represented the final human achievement in productivity so in the future what would matter is not the creation of wealth but the redistribution of wealth.

Do we just imagine design in nature? Or is seeing design fundamental to discovering and using nature’s secrets? Michael Egnor reflects on the way in which the 2018 Nobel Prize in Chemistry has so often gone to those who intuit or impose desire or seek the purpose of things

Comments
Hi Eric
So, from my brief perusal of Felsenstein’s response, I’ll have to say that Felsenstein has repeated the pattern I continue to detect, that the skeptics are saying straightforward information theory is highly controversial merely because it shows up in Dembski’s argumentation. Otherwise, they just don’t read Dembski’s work very carefully.
Although I am getting comfortable with you proof I need to spend more time with it. I really like your collaboration with gpuccio as I have tremendous respect for him. I think Demski is right and believe the objections by Tom and Joe are just the best they can do trying to refute a solid claim. Question: Would you consider a AA sequence for the same protein in humans as in chimps where there are slight variation as mutual information?bill cole
November 9, 2018
November
11
Nov
9
09
2018
06:29 PM
6
06
29
PM
PDT
@Bill Cole, Joe Felsenstein, and Tom English, With a bit more time on my hands, I took a look at Dr. Felsenstein's claim that Dembski's law of conservation of information (LCCSI) is false: https://pandasthumb.org/archives/2018/11/Eric-Holloway-needs-our-help.html This is Dembski's most important claim, because it proves natural processes cannot generate CSI. I reread Shallit & co.'s 2011 paper: http://web.cecs.pdx.edu/~mperkows/CLASS_479/SearchComplexityPage/CRITICS-OF-DEMBSKI.pdf and to pick out what seems their clearest objection against LCCSI, on pg. 256 they say: To cover this possibility Dembski introduces the universal composition function U, defined by U(i, f ) = f (i) = j. He then argues that the amount of CSI in the pair (i, f ) is at least as much as j. Of course, this claim also suffers from the two problems mentioned above, but now there is yet another error: Dembski does not discuss how to determine the CSI contained in f. First of all, Dembski is correct. Since K(x, U(x)) = K(x), we can show I(x:y) > I(U(x):y), which is the same as showing I(f,i:j) > I(U(f.i):j). This is straightforward algorithmic information theory, so it is unclear why a professor of computer science is taking issue with Dembski's claim here. Second, they call a lack of discussion an 'error.' That is a novel use of the term 'error,' and by that definition there are many relevant pieces of information Shallit does not discuss in his paper, so he makes a multitude of errors as well. Seems like an unhelpful definition of error, and certainly not a usage that shows Dembski's theory is false. It raises the question in my mind, if Shallit says Dembski's theory is in error, does he mean it is false or does he mean Dembski hasn't covered something that Shallit is interested in? Back to Felsenstein's original post, he says the problem with Dembski's argument is the target must be independent, and this is covered when Dembski says the specification must be detachable. So, from my brief perusal of Felsenstein's response, I'll have to say that Felsenstein has repeated the pattern I continue to detect, that the skeptics are saying straightforward information theory is highly controversial merely because it shows up in Dembski's argumentation. Otherwise, they just don't read Dembski's work very carefully.EricMH
November 9, 2018
November
11
Nov
9
09
2018
04:37 PM
4
04
37
PM
PDT
EricMH:: Finally, you say: "Being able to detect new FI seems very useful. For example, say we have a code generation tool. Obviously, it comes with background FI. How can we determine if a code base is only generated by the tool vs a code base written from scratch? This would be detected by a significant amount of new FI compared to the code generation tool baseline." Being able to detect new FI is certainly very useful, and not difficult. For example, in the case of Prickle 1 in my first linked OP, the second part of the protein, the blue part in my OP, suddenly acquires about 600 bits of sequence information that did not exist in pre-vertebrates and that are then fully conserved up to humans, for more than 400 million years. Those 600 bits are a novelty, and a novelty that appears in about 30 million years (the reasonable time span of the transition to vertebrates). Moreover, they are 600 bits of functional information: otherwise, how can we explain that they have been conserved for 400+ million years of exposure to neutral variation? So, we can detect new FI in proteins. And it is indeed very useful. Let's see if I understand well your interesting scenario of the code generation tool. You say: You say: "Obviously, it comes with background FI." That's absolutely true: it's the FI linked to the function of the tool: generating codes. It can reasonably be computed or estimated. You say: "How can we determine if a code base is only generated by the tool vs a code base written from scratch? This would be detected by a significant amount of new FI compared to the code generation tool baseline." Again, perfectly true. A couple of interesting thoughts. a) The tool can probably generate codes that are apparently more complex than the tool otself, in terms of length and apparent FI. That can be done by using the computational abilities programmed in the tool itself. We are here in the same situation as described about programs that compute the digits of pi. At some point, the apparent functional complexity of the output will be greater that the functional complexity of the generating program. But that is really not true: the generating program is indeed a valid compression of the output, so the Kolmogorov complexity of the output, if its apparent functional complexity is greater than the functional complexity of the tool, is by definition the functional complexity of the tool itself. IOWs, the functional complexity of the tool is a higher threshold for the functional complexity of the output. IOWs, a merely computing tool cannot generate FI significantly higher than the FI it already has. (I have added "significantly" because, as we know, low levels of FI can be added incorporating RV in the algorithm). What about programs that use RV + Artificial selection? We know they can achieve good results, in terms of FI. Not so good as a system that includes new design interventions, but good results just the same. We have good demonstration of that in the procedures of protein engineering, that are often rather successful in optimizing proteins, sometimes even in generating some gross new ones. What is the limit here? The limits are two: 1) These are the limited but important powers of Artificial selection. They must not be confused with the extremely limited powers of Natural selection. I have discussed the differences here: Natural Selection vs Artificial Selection https://uncommondescent.com/intelligent-design/natural-selection-vs-artificial-selection/ 2) Artificial selection as implemented by a non conscious tool, however good, can only optimize a function that has already been defined and programmed in the tool. Directly or indirectly. Nothing else. Again, the FI in the tool is the real limit of what the tool can do. Only conscious, intelligent and purposeful agents can think, define and recognize new and original functions. That power comes from their conscious experiences and representations, as I have briefly discussed at #465 here. So, this is not a mathematical law of conservation, but just a very simple and empirical truth: Machines can only do what they have been programmed to do. They can use RV if they have been programmed to do that, but only in the way they have been programmed to do. They can use RV + AS if they have been programmed to do that, but only in the way they have been programmed to do, and for the function they have been programmed to select for. They can incorporate new information form the outside world, or from their inner simulations, but only in the way they have been programmed to do. Very simply, they can compute about already defined functions, because their FI has been designed to do exactly that. But they can never create, recognize or implement new, original functions. Why? Because they are not conscious. They do not understand any meaning. They have no desires. For those who, like me, believe in free will, they have no free will. So, a new functional specification, which cannot be derived from the existing FI or computed from it, and that is complex enough, is an extremely reliable indicator of conscious design.gpuccio
November 9, 2018
November
11
Nov
9
09
2018
03:18 PM
3
03
18
PM
PDT
EricMH: Again, very good questions! This converstation is really interesting. :) First of all, the link to ifnormation jump OPs. Indeed, there are a number of them. This one: Homologies, differences and information jumps is probably mi first explicit application of the procedure, applied tp a very interesting regulatory protein, Prickle 1. https://uncommondescent.com/intelligent-design/homologies-differences-and-information-jumps/ In the second one: The highly engineered transition to vertebrates: an example of functional information analysis https://uncommondescent.com/intelligent-design/the-highly-engineered-transition-to-vertebrates-an-example-of-functional-information-analysis/ I apply the ame procedure to a small group of proteins. In the third one: The amazing level of engineering in the transition to the vertebrate proteome: a global analysis https://uncommondescent.com/intelligent-design/the-amazing-level-of-engineering-in-the-transition-to-the-vertebrate-proteome-a-global-analysis/ I have done the "big work": I have applied the procedure to the whole human genome, 20000 proteins, using it as a probe against the main gorups of organisms in Metazoa. The results are very, very interesting. You are perfectly right: computer code is prescriptive information, and it sollows the same rules as proteins and protein genes: they are all examples of digital prescriptive information. Yes, your procedure is perfectly correct. That's the idea. As you say, with sequences of some length it's impossible to test the function directly, "in a bruteforce manner", as you say. We have to use indirect methods. The same is true for proteins. In the case of some simple algorithm, the procedure would be: a) We observe a program that is functional in a certain context, for example a sorting algorithm. b) We define explicitly the function (taking lists of words as input and sorting them in alphabetical ascending order as output). We can also define a measure for the existence of the function, for example being able to sort a reference list of 1000 words. We can also set a threshold of efficiency (in less than 1 msec). The important thing is that everything is explicit and not ambiguous, so that anyone can apply our definition and procedure to an object (in this case, a sequence of bits in the computer) and verify if that sequence works as a program that can implement the function. c) We compute the search space. You are correct, this is the easy part, and your procedure is perfectly correct. Just a few comments about the problem of length. The simplest way to act is what you have described: we take the length of the reference object (the program we are analyzing for FI) and use that length to define the search space. With computer programs, in order to minimize the search space, it would be fine to use as reference the shortest program available that can do the job. Even if we can never be sure that one program is the shortest, empirically it's perfectly fine to work with what is known and available. That will set a reasonable minimum size of the search space. The existence of possible longer programs that are functional should not be a problem, because in general FI increases with the length of the functional sequence, and we want to find the minimun FI linked to our defined function. d) As you say, the size of the target space is the real difficult step. We have to use all possible strategies to get reasonable approximations. In the case of computer programs, I believe that the target space is usually rather small. IOWs, computer programs should not be very tolerant to bit substitutions. I am not an expert, but it seems reasonable. So, the FI will always be very high if the search space is big enough (that is usually the case even with simple computer programs). Indeed, I cannot think of any complete and independent computer program, however simple, that is so functionally simple that it can be easily found by a random search. Correct me if I am wrong. Of course, it is probably possible to get, by a random search, some simple set of instructions that can be functional in the appropriate context. IOWs, a functional sequence with low FI. For proteins things are rather different, because we know that protein sequences are certainly more tolerant to AA substituions than computer programs are to bit substitutions. How much they are tolerant is key to measure their target space, and indeed that is still a very open problem, one that is often used by neo-darwinists to unleash their imagination, assuming extremely huge functional spaces with wonderful functional connections. All things never observed, of course, because very simply they do not exist. The simple truth is that the functional space of proteins obeys similar rules as the functional space of computer programs: simple functions can abound, but complex functions are rare and isolated in the sequence space. We know that: we know that all functional proteins that we have observed have been classified into 2000 superfamilies of sequences, all of them unrelated at the level of sequence, structure and function. Unlike darwinists, we have no need to imagine the functional islands: they are there, in front of us. However, as said, it is true that those functional islands, even if rare and isolated, are bigger for proteins than they are for computer programs. Just an example that I often use. The alpha and beta chains of ATP synthase are a very good example of highly conserved sequences. In this case, I am not evaluating an information jump in vertebrates, but just the total conserved sequence information between E. coli, a bacterium, and humans. About 2 billion years of separation, reasonably. Well, as said at #148 here, the homology between the two sequences (E. coli andhuman), evaluated by BLAST, is as follows: alpha chain = 561 bits beta chain = 663 bits total = 1224 bits As I have said, there are good reasons to believe that BLAST gives a severe underestimation of the real FI, but let's take it as it is. 1224 bits of FI for these two chains, that are functionally one thing because together they make the main part of the F1 subunit of the protein. OK, but what is the search space? Using our procedure, we can see that the search space for both protein is as follows: alpha chain: length = 553 AAs beta chain: length = 529 AAs total length = 1082 AAs So, as a gross evaluation, the search space is about 20^1082, that is about 4676 bits. Our estimated FI is about 1224 bits. What does that means? It means that we are estimating a target space of: 4676 - 1224 = 3452 bits! IOWs, a target space of 2^3452 sequences. That is certainly an overestimation, but again let's take it as true. So (for those who are not familiar with that kind of computation, not you of course), we have: 2^3452 (target space) / 2^4676 (search space) = 2^-1224 FI = -log2(2^-1224) = 1224 bits I don't believe that the true target space is so big, but I accept the results of the BLAST algorithm, even if they certainly underestimate severely FI, because it is a reliable and easy to use tool to evaluate homologies, universally used. And it's fime with me to underestimate FI and give neo-darwinists some "help". I can be satisfied with 1224 bits of FI for one biological object! :) So, we can see that even huge target spaces are rare and isolated islands of function in the ocean of the sequence space. More in next post.gpuccio
November 9, 2018
November
11
Nov
9
09
2018
02:41 PM
2
02
41
PM
PDT
@gpuccio, how about computer code then, since that is prescriptive? We can use the repl.it I posted, since it is not trivial, but also not very long. https://repl.it/@EricHolloway/KLD-for-FI How would we measure the baseline, and how would we measure function? My naive approach is to measure baseline by len(code)*log(len(alphabet),2). Function is a bit more difficult. I think it is a proportion of how many character strings of that length result in compilable code. It would be intractable to approach in a bruteforce manner, but clearly most such character strings are broken code, so the functional measure would be very small. I think it could be tractably upper bounded by all the lexeme variations that can fit in the string length. Thus, I'd lower bound FI for computer code with: len(code)*log(len(alphabet),2) - log(number of lexeme variations, 2) = very big number. Is my understanding of how to apply FI correct here, gpuccio? Also, do you have a link to your work: > That’s exactly what I do when I analyze information jumps in the transition to vertebrates. Being able to detect new FI seems very useful. For example, say we have a code generation tool. Obviously, it comes with background FI. How can we determine if a code base is only generated by the tool vs a code base written from scratch? This would be detected by a significant amount of new FI compared to the code generation tool baseline.EricMH
November 9, 2018
November
11
Nov
9
09
2018
10:30 AM
10
10
30
AM
PDT
EricMH: Let's go to text. That is easier, because text is digital and symbolic. That helps. However, both images and text are usually examples of descriptive information, not prescriptive information. They convey meanings, rather than implement a function (which is the purpose of prescriptive information). While it is always possible to convert descriptive information into a functional form to measure it, that is often tricky. For example, we can define the function of an image or a text as being able to convey to an observer some specific and well defined meaning, but of course it is not completely obvious how to do that. The comprehension of a text can be difficult to measure objectively, for example. And the same is true for images. This is an objection that is often exploited by neo-darwinists (or in general critics of ID): where is the threshold to say that the cloud is really like a weasel, and not a camel or a whale? Or to conclude, correctly, that it is only a cloud with some simple and contingent resemblance to other things? For FI proper (prescriptive information) it's much easier to define a function and a way to evaluate it. However, some indirect method for FI in a text can be more easily applied. For example, in my OP: An attempt at computing dFSCI for English language https://uncommondescent.com/intelligent-design/an-attempt-at-computing-dfsci-for-english-language/ I have used a simple approach and a minimal functional definition to compute a lower threshold for FI in a Shakepseare sonnet. With good results, IMO. So, it is possible. You ask: "Also, is there a way to identify new FI?" Of course. That's exactly what I do when I analyze information jumps in the transition to vertebrates. "Examples: identifying an out of place object in a photograph" Of course, it is easy to identify an out of place object in a photograph, if you have the reference photograph. The problem is: is that variation complex and functiona? I can imagine many scenario where that is not true, and other where it could be true. Again, it is important to consider the system, and to define a function to measure the FI. "plagiarism in text," That should not be difficult, if we have references. And of course, a plagiarim that is well adapted to the context would be functional and, if complex enought, an indicator of design, while a simple occurrence of some parapgraph out of context could be a random event. "an intrusion in a network, a virus on the computer, steganography in spam, or spies in a population." Again detection is not necessarily difficult. I think that machine learning can deal with some of those scenarios quite efficiently. The problem always remains to find some way of defining a function to measure FI. Again, it's not impossible, but it can be tricky. In the end, the purpose is always to detect a conscious intervention that has assembled specific bits of information to generate an extremely unlikely configuration defined by a complex function.gpuccio
November 9, 2018
November
11
Nov
9
09
2018
09:24 AM
9
09
24
AM
PDT
Antonin: "It should be obvious at a glance what I’m referring to." I don't know if it's obvious. I have not looked much at other threads, while I was busy here. Can you specify better?gpuccio
November 9, 2018
November
11
Nov
9
09
2018
08:26 AM
8
08
26
AM
PDT
ES, a minimally functioning single cell organism is a molecular nanotech based, encapsulated, smart gated automaton that has metabolic systems coupled to code controlled fabricators and a von Neumann kinematic self replicator. "Simple" is exactly what such is not! KFkairosfocus
November 9, 2018
November
11
Nov
9
09
2018
08:20 AM
8
08
20
AM
PDT
EricMH: Very interesting example, thank you. And very good questions. Let's see. a) Photography. This is interesting. Of course, a camera is certainly a designed object with a lot of FI. But let's say that the camera is already part of our system. So, we have a photo, and we want to know if the photo itself is a designed object. Let's say that the photo represents, with good fidelity, a room (furniture, etc). The question is: can we infer design for the object "photo" in our system? My answer is no. There can be no doubt that the photo seems to exhibit a lot of FI: the function can easily be defined in terms of representing well enough what it represents (see later the discussion about text for more details). And certainly, the target space/search space ratio will be extremely small, and the potential FI extremely high. There should be no problems in iferring design... But... there is, indeed, a proble. The problem is necessity. Because our system includes a tool, the camera, that can easily shoot that photo, even without a design intervention from a conscious being (like pushing the shooting button, which however would be a very simple design intervention). For example, however weird it can seem, some material object falling on the button for random reasons could well shoot the photo. This is very interesting, and demonstrates how important it is to include the system in our analysis. If the system already includes some mechanism of necessity that could in principle generate the complex object, maybe as the consequence of some random, but not completely unlikely, event, then design is not necessarily the explanation for the object. To be more clear: a) In our system, the camera is certainly designed: I cannot imagine any explanation for tha camera that does not include a lot of design intervention to produce the camera, with the specification of many, many bits of functional information to get the working configuration. b) Given the camera, the photo of the room does not add any big amount of FI. OK, the act of pushing the button requires usually a simple design intervention, but the FI in it is very low. The point is that the wonderful ability of the photo to represent the room derives directly from the optical properties of the light coming from the room itself, even if thorugh the compex action of the camera. IOWs, given the camera, it is the light itself, thorugh physical laws of necessity, that generates the photo and its informational content. So, given the camera, necessity is here the good explanation, and not design, for the complex configuration of the photo. That is often true for information that directly derives, analogically, from a form already existing in the system. Now, let's say instead that we have not a photo, but a good drawing of the room, made with coloured pencils by a good designer. Here, the reasoning is different, because the nature of the object in our system can easily exclude any necessity origin. IOWs, there is no natural law that converts thorugh light the form of the room into a drawing (if we can exclude that our system includes some complex science fiction machinery designed to draw by coloured pencil). The meaning of all that is that we must be very careful to exclude possible necessity mechanisms that can explain the configuration that we are observing. However, in the absence of a designed camera, I don't think that it is easy to find a system that could generating something similar to a photo just from the natural pattern of light in the environment. But I cannot exclude it, because in a way that information is potentially already there, in the reflected light. More in next post.gpuccio
November 9, 2018
November
11
Nov
9
09
2018
08:20 AM
8
08
20
AM
PDT
...complex functions are extremely rare, and isolated...
And so much else to pick up on! :)Antonin
November 9, 2018
November
11
Nov
9
09
2018
07:54 AM
7
07
54
AM
PDT
@ gpuccio, KF and Upright Biped I may have time over the weekend to respond to outstanding comments but I'm in a bit of a quandary whether to continue here. I don't mind abuse when it's directed at me but I'm more than a little disappointed with the discussion on another thread. It should be obvious at a glance what I'm referring to. If not... Well, that says something in itself, no?Antonin
November 9, 2018
November
11
Nov
9
09
2018
07:52 AM
7
07
52
AM
PDT
EugeneS: "This one is very important. It came up a few times in my own discussions with biologists, who claimed guess what?… that once we have a minimal functioning organism it can achieve all the rest" Why am I not surprised? :)gpuccio
November 9, 2018
November
11
Nov
9
09
2018
07:40 AM
7
07
40
AM
PDT
@gpuccio, thanks a lot. Very clear. I think the uniform hypothesis was my big question. You are right, that the Kullback-Liebler divergence between the uniform distribution and your specification is going to always be smaller than between some arbitrary distribution and your specification, so the uniform minimizes the FI score. Also, the KLD with the uniform simplifies to H(U) - H(P), so I see where Durston gets his formula. Here's a Python simulation to demonstrate the point: https://repl.it/@EricHolloway/KLD-for-FI gives the result: kld(u,p): 0.43 kld(u,q): 0.44 kld(p,q): 0.73 kld(u,p)<kld(p,q): True kld(u,q)<kld(p,q): True Here's a follow on question. How can we apply this method to another domain, say photography or text? Assuming the uniform distribution seems like it'll always give me a very high FI score. Maybe that is because the FI is actually very high. Also, is there a way to identify new FI? Examples: identifying an out of place object in a photograph, plagiarism in text, an intrusion in a network, a virus on the computer, steganography in spam, or spies in a population.EricMH
November 9, 2018
November
11
Nov
9
09
2018
06:40 AM
6
06
40
AM
PDT
GP "But complex functions are extremely rare, and isolated." This one is very important. It came up a few times in my own discussions with biologists, who claimed guess what?... that once we have a minimal functioning organism it can achieve all the rest :)EugeneS
November 9, 2018
November
11
Nov
9
09
2018
05:05 AM
5
05
05
AM
PDT
EricMH: One more point. FI is "rare" only when it is related to a complex function. FI related to simple functions is not rare at all. That's why simple functions (involving a transition of 1 or 2 AAs, like in antibiotic resistance) can easily emerge in a biological system from RV. But complex functions are extremely rare, and isolated.gpuccio
November 9, 2018
November
11
Nov
9
09
2018
03:17 AM
3
03
17
AM
PDT
EricMH: About the probability distribution. Of course, we must always consider the physical system we are analyzing. So, let's talk proteins. In the biological system, where the main source of unguided variation is random point mutations, the true nature of the random system is a random wlalk rather than a random search (while a random search model is still valid for certain types of variation, like frameshift mutations). However, that is no real problem, provided we clarify what we are doing. Of course, in a random walk in the sequence space, sequences that are nearer to the starting point have higher probabilities to be reached. Indeed, with one single point mutation event, only sequences that differ of 1 AA from the starting point can be reached. But we can easily overcome that problem by defining the terms of our analysis. What we are interested in, from an ID point of view, is the emergence of a new complex function from an unrelated sequence. IOWs, we are interested to a transition that involves more than 500 bits of FI. For example, the 2000 protein superfamilies are completely unrelated both at sequence level and at function level (indeed, even at structure level). So, when a new protein superfamily emerges in natural history, it is a real novelty both at sequence level and at function level. And, of course, we must allow for a high number of events to explain the transition. We can allow up to the probabilistic resource of the biological system, that I have computed to be (very generously!), at most, 138.6 bits (2^138.6 events). See here, Table 1: What are the limits of Random Variation? A simple evaluation of the probabilistic resources of our biological world https://uncommondescent.com/intelligent-design/what-are-the-limits-of-random-variation-a-simple-evaluation-of-the-probabilistic-resources-of-our-biological-world/ So, with a high number of events, any sequence in the search space can certainly be reached. Of course, we are not considering here possible effects of NS. Let's say that our model starts with a non coding non functional sequence, so that the effectf of NS are avoided, at least until a selectable function emerges. Now, we can reason this way. There is a certain number of states in the search space that are obviously more likely: indeed, all the states that are significantly related at sequence level with the starting sequence. So, we can ideally divide the search space into two subsets: a) States which are more or less related to the starting sequence, at sequence level. b) States that are completely unrelated to the starting sequence at sequence level (let's say that they show no significant homology with it). Now, subset a) is certanly hugely smaller than subset b), for sequences that are not trivially long. We also know that our target space is somewher in sibset b), because we are considering the emergence of a new sequence, unrelated to what already existed. A new protein superfamily. So, for the moment, let's ignore subset a). My simple point is that all the states in subset b) can be considered approximately equiprobable in a random walk starting form the unrelated sequence. Even if we consider the effect of subset a), it can only lower the average probability of sequences in subset b), because of course sequences in subset a) are more likely to be reached by a random walk. So, considering the probability of each state in the search space as 1/n is a good approximation, indeed a slight overestimation for the probability of unrelated states (and therefore of the target space). Of course, the probabilities of individual sequences can be slighlty different, for example according to their AA composition. The genetic code is not symmetric, some AAs are more likely than others in point mutations, and some mutations are more likely than others. And there are certainly other biochemical factors that can generate non uniformity in the probability of the individual states. But that's not really important, because of course none of those "anomalies" can explain or favour specific functional sequences. And even if they could slightly favour one of them, then they would not favour the others, because, as we know, functional sequences at the superfamily level are sequence unrelated. They have different sequences, different AA composition, and so on. So, for all practical purposes, assuming an uniform distribution in the search space is the best approximation. One final point: my analysis is not limited to protein superfamilies. It is valid for all complex transitions. When I apply my procedure based on homologies and conservation through long evolutionary times, I am always interested in what I call an "information jump". IOWs, I don't consider the absolute conserved information in a protein, but the difference (jump) in human conserved information between two evolutionary nodes: usually, between pre-vertebrates and vertebrates. That is a "jump": it means that I am cosnidering the emergence of a new sequence, even if it is part of an older protein, that already exhibited some human conserved homology. I hope that's clear. Of course I am available to clarify any point, if necessary.gpuccio
November 9, 2018
November
11
Nov
9
09
2018
03:11 AM
3
03
11
AM
PDT
edit workedkairosfocus
November 8, 2018
November
11
Nov
8
08
2018
11:27 PM
11
11
27
PM
PDT
sorry, missing information.kairosfocus
November 8, 2018
November
11
Nov
8
08
2018
11:26 PM
11
11
26
PM
PDT
Nonlin, symbols implies a communication context where distinct symbols in context convey meaningful or functional messages. ASCII-based alphanumeric text strings in English, as in this thread, is an example of this in action. Similarly, D/RNA has meaningful coded symbols used to assemble proteins step by step, again not controversial. If you will go to my Sec A in my always linked you will see an outline of how we get to Shannon Entropy, which is indeed the weighted average info per symbol. This is fairly standard stuff dating to Shannon's work in 1948. The similarity to the Gibbs expression for entropy created much interest and as I noted Jaynes et al have developed an informational perspective on thermodynamics that makes sense of it. Thermodynamic entropy is effectively an index of missing info to specify microstate for a system in a given macrostate. For, there are many ways for particles, momentum/energy etc to be arranged consistent with a macrostate. As for the clip from my online note, I an simply speaking to the longstanding trichotomy of law of mechanical necessity vs blind chance contingency on closely similar initial conditions, vs functionally specific complex organisation and associated information. While we can have an onward discussion of how laws and circumstances of the world are designed [per fine tuning] or that randomness can be used as part of a design, the issue here is to discern on empirically grounded reliable sign, whether the more or less IMMEDIATE cause of an effect is intelligently directed configuration or is plausibly explained on blind chance and/or mechanical necessity. KFkairosfocus
November 8, 2018
November
11
Nov
8
08
2018
11:25 PM
11
11
25
PM
PDT
EricMH: "Separate topic, can anyone reduce FI to simple terms and math to save me having to read gpuccio’s links? I know the basic idea that it is the difference of two SMIs, but what are the underlying distributions the SMIs are measuring, why are they subtracted, and why is FI rare? If no one wants to, that is fine. I will get around to gpuccio’s links sooner or later." well, I can try, even if I am not sure that I understand correctly your questions. So, I will try, and you can clarify if necessary. FI is the probability of the emergence of one object in the target set in a well defined material system, without any design intervention, by the probabilistic resources of the system itself. I think this is in perfect accord with Dembski's original ideas (IOWs, the explanatory system). The target set is defined in relation to one explicitly defined function. Any function definition can be used, provided that it is explicit and that it specifies an explicit way to test any object in the system to asses if it can implement the function or not. IOWs, we must have an objective procedure to assess, at least in principle, for any possible object in the system if the function is present or not. Of course, any measure of FI is related to its explicitly defined function. In most systems, and certainly in most biological systems, we can assume an uniform probability distribution in the system, where the probability fo each object to emerge is approximately the same. Of course, that is not literally true, but as I will try to explain it is the best approximation for our purposes. With that assumption, the probability of the emergence of one object in the target set (IOWs, the FI for that target set defined by the function) is simply the negative of the ratio between the target space and the search space (IOWs between the two numerosities). If we express the two numerosities in base 2, we just take -log2 as the value of FI in bits. Let's take the case of one protein, or class of proteins, with a well defined function. A forst assumption, to make the computations possible, is to assume a reference length for tha protein, which can be derived form the existing functional protein or proteins. In Durston's terms, we compute the search space simply as the value in base 2 of 20^n (where n is the reference length). That is rather simple. The problem is in computing the numerosisty of the target space. There is no problem in principle: for each possible protein sequence there is a way to verify if the function is present or not, as defined in the function definition. But of course, for any length that is not trivial, that is, and always will be, empirically impossible. So, we need indirect methods. As our uinderstanding of the sequence-function relationship in proteins is limited, we cannot just derive the result from a coomputation. Not yet. So, we can use protein conservation. Durston uses one approach, I use a slightly different one. Durston aligns a number of existing sequences in that protein family, from different species, and then computes the Shannon entropy in that set, defined by the common function. I use human protein sequences as a "probe" to evaluate human conserved information in distant groups of organisms, using the BLAST tool, that eveluates the level of homology between sequences. There are pros and cons in both methods, but in the end they measure the same thing, conserved information, and I believe that the results are reasonably similar. I will explain here in more detail, but very briefly, my approach. The basic assumption is that sequence conservation through long evolutionary times is a good measure of functional constraint. That is in perfect accord with the neo-darwinist model, and there are many empirical confirmations for this idea. I will not go into biological details here, but we can discuss the reasons for that if you are interested. In my analyses, I have usually considered the transition to vertebrtaes as a good object of study. The reson for thatis that: a) It is very interesting. b) It took place more than 400 million years ago, which is enough to make my meausres of human conserved information a reliable indicator of functional constraint. OK, I wil say something more about the assumtpion of an uniform probability distribution in next post.gpuccio
November 8, 2018
November
11
Nov
8
08
2018
11:22 PM
11
11
22
PM
PDT
Eric
and why is FI rare?
Depends on the function how rare it is. There appears to be a range. Gpuccio measures "rareness" by amino acid preservation over time. This demonstrates that high sequence specificity was required. gpuccio can describe his method in more detail. There are many highly preserved proteins gpuccio has identified.bill cole
November 8, 2018
November
11
Nov
8
08
2018
09:02 PM
9
09
02
PM
PDT
@ET regarding Felsenstein's example, this sounds the same as Dawkin's weasel program. Why does he not find the standard response by Marks and Ewert convincing, that the CSI is just being propagated by the fitness function, so no creating of CSI is taking place? I keep getting the impression the skeptics are not trying to understand the argument.EricMH
November 8, 2018
November
11
Nov
8
08
2018
07:09 PM
7
07
09
PM
PDT
@Mung, too many terms! In this case I have a precise meaning. ASC has a 'self information' quantity which is the -log of the probability for X according to some chance hypothesis. Maybe I'll call this quantity "point SMI" since it deals with a particular instance instead of the whole distribution, while per your usage SMI refers to the whole distribution. Separate topic, can anyone reduce FI to simple terms and math to save me having to read gpuccio's links? I know the basic idea that it is the difference of two SMIs, but what are the underlying distributions the SMIs are measuring, why are they subtracted, and why is FI rare? If no one wants to, that is fine. I will get around to gpuccio's links sooner or later. And thanks everyone, this has been an illuminating thread!EricMH
November 8, 2018
November
11
Nov
8
08
2018
07:06 PM
7
07
06
PM
PDT
Self-Information is not EntropyMung
November 8, 2018
November
11
Nov
8
08
2018
06:18 PM
6
06
18
PM
PDT
EricMH:
The specific hypothesis would be addressed by the probability distribution used to measure self information.
I think this language is imprecise. There is no probability distribution that measures self-information. Self-information is mutual information. There is no probability distribution that measures mutual information. I'm still working out how to put this all in plain language so forgive me for stopping here. :) self-informationMung
November 8, 2018
November
11
Nov
8
08
2018
06:14 PM
6
06
14
PM
PDT
gpuccio:
What misconception?
That you can reason with someone who cannot be reasoned with.Mung
November 8, 2018
November
11
Nov
8
08
2018
05:57 PM
5
05
57
PM
PDT
Upright BiPed:
Ah, forgot to mention … thanks to all for the kind words upthread.
If I spoke any kind words I can assure you it was not sincere.Mung
November 8, 2018
November
11
Nov
8
08
2018
05:48 PM
5
05
48
PM
PDT
kairosfocus @442, 443 Not sure what your point is. You talk about "weighted average information per symbol", but know that information is not tied to specific symbols. Example: you can represent a 'circle' in English or any other language, as math equation, descriptive, etc. Each one of those employs different symbols. Somehow I have the feeling we're talking past each other again. From your site: "One of these, is that when we see regularities of nature, we are seeing low contingency, reliably observable, spontaneous patterns and therefore scientifically explain such by law-like mechanical necessity: e.g. an unsupported heavy object, reliably, falls by "force of gravity." But, where we see instead high contingency -- e.g., which side of a die will be uppermost when it falls -- this is chance ["accident"] or intent ["design"]. Then, if we further notice that the observed highly contingent pattern is otherwise very highly improbable [i.e. "complex"] and is independently functionally specified, it is most credible that it is so by design, not accident. " Do you understand where we differ on this passage?Nonlin.org
November 8, 2018
November
11
Nov
8
08
2018
05:39 PM
5
05
39
PM
PDT
gpuccio @440, Meanwhile, YOU were the one forced to agree with me. Does that tell you anything? It should. Yes, logic becomes unbearable when you find yourself in a corner.Nonlin.org
November 8, 2018
November
11
Nov
8
08
2018
05:20 PM
5
05
20
PM
PDT
Ah, forgot to mention ... thanks to all for the kind words upthread.Upright BiPed
November 8, 2018
November
11
Nov
8
08
2018
02:23 PM
2
02
23
PM
PDT
1 2 3 4 5 20

Leave a Reply