Evo-Info 4: Non-conservation of algorithmic specified complexity

Introduction to Evolutionary Informatics, by Robert J. Marks II, the “Charles Darwin of Intelligent Design“; William A. Dembski, the “Isaac Newton of Information Theory“; and Winston Ewert, the “Charles Ingram of Active Information.” World Scientific, 332 pages.
Classification: Engineering mathematics. Engineering analysis. (TA347)
Subjects: Evolutionary computation. Information technology–Mathematics.

The greatest story ever told by activists in the intelligent design (ID) socio-political movement was that William Dembski had proved the Law of Conservation of Information, where the information was of a kind called specified complexity. The fact of the matter is that Dembski did not supply a proof, but instead sketched an ostensible proof, in No Free Lunch: Why Specified Complexity Cannot Be Purchased without Intelligence (2002). He did not go on to publish the proof elsewhere, and the reason is obvious in hindsight: he never had a proof. In “Specification: The Pattern that Signifies Intelligence” (2005), Dembski instead radically altered his definition of specified complexity, and said nothing about conservation. In “Life’s Conservation Law: Why Darwinian Evolution Cannot Create Biological Information” (2010; preprint 2008), Dembski and Marks attached the term Law of Conservation of Information to claims about a newly defined quantity, active information, and gave no indication that Dembski had used the term previously. In Introduction to Evolutionary Informatics, Marks, Dembski, and Ewert address specified complexity only in an isolated chapter, “Measuring Meaning: Algorithmic Specified Complexity,” and do not claim that it is conserved. From the vantage of 2018, it is plain to see that Dembski erred in his claims about conservation of specified complexity, and later neglected to explain that he had abandoned them.

Algorithmic specified complexity is a special form of specified complexity as Dembski defined it in 2005, not as he defined it in earlier years. Yet Marks et al. have cited only Dembski’s earlier publications. They perhaps do not care to draw attention to the fact that Dembski developed his “new and improved” specified complexity as a test statistic, for use in rejecting a null hypothesis of natural causation in favor of an alternative hypothesis of intelligent, nonnatural intervention in nature. That’s obviously quite different from their current presentation of algorithmic specified complexity as a measure of meaning. Ironically, their one and only theorem for algorithmic specified complexity, “The probability of obtaining an object exhibiting \alpha bits of ASC is less then [sic] or equal to 2^{-\alpha},” is a special case of a more general result for hypothesis testing, Theorem 1 in A. Milosavljević’s “Discovering Dependencies via Algorithmic Mutual Information: A Case Study in DNA Sequence Comparisons” (1995). It is odd that Marks et al. have not discovered this result in a literature search, given that they have emphasized the formal similarity of algorithmic specified complexity to algorithmic mutual information.

Lately, a third-wave ID proponent by the name of Eric Holloway has spewed mathe­matical­istic nonsense about the relationship of specified complexity to algorithmic mutual information. [Note: Eric has since joined us in The Skeptical Zone, and I sincerely hope to see that he responds to challenges by making better sense than he has previously.] In particular, he has claimed that a conservation law for the latter applies to the former. Given his confidence in himself, and the arcaneness of the subject matter, most people will find it hard to rule out the possibility that he has found a conservation law for specified complexity. My response, recorded in the next section of this post, is to do some mathematical investigation of how algorithmic specified complexity relates to algorithmic mutual information. It turns out that a demonstration of non-conservation serves also as an illustration of the sense­less­ness of regarding algorithmic specified complexity as a measure of meaning.

Let’s begin by relating some of the main ideas to pictures. The most basic notion of nongrowth of algorithmic information is that if you input data x to computer program p, then the amount of algorithmic information in the output data p(x) is no greater than the amounts of algorithmic information in input data x and program p, added together. That is, the increase of algorithmic information in data processing is limited by the amount of algorithmic information in the processor itself. The following images do not illustrate the sort of conservation just described, but instead show massive increase of algorithmic specified complexity in the processing of a digital image x by a program p that is very low in algorithmic information. At left is the input x, and at right is the output p(x) of the program, which cumulatively sums of the RGB values in the input. Loosely speaking, the n-th RGB value of Signature of the Id is the sum of the first n RGB values of Fuji Affects the Weather.

Effects of Fuji on Weather
Fuji Affects the Weather
[13 megabits of “meaning”]
Awesome Glory of the Id
Signature of the Id
[24 megabits of “meaning”]

What is most remarkable about the increase in “meaning,” i.e., algorithmic specified complexity, is that there actually is loss of information in the data processing. The loss is easy to recognize if you understand that RGB values are 8-bit quantities, and that the sum of two of them is generally a 9-bit quantity, e.g.,

    \begin{align*}    \phantom{+}& \phantom{01}11111111 \\             + & \phantom{01}00000001 \\             = & \phantom{0}100000000 \end{align*}

The program p discards the leftmost carry (either 1, as above, or 0) in each addition that it performs, and thus produces a valid RGB value. The discarding of bits is loss of information in the clearest of operational terms: to reconstruct the input image x from the output image p(x), you would have to know the bits that were discarded. Yet the cumulative summation of RGB values produces an 11-megabit increase in algorithmic specified complexity. In short, I have provided a case in which methodical corruption of data produces a huge gain in what Marks et al. regard as meaningful information.

An important detail, which I cannot suppress any longer, is that the algorithmic specified complexity is calculated with respect to binary data called the context. In the calculations above, I have taken the context to be the digital image Fuji. That is, a copy of Fuji has 13 megabits of algorithmic specified complexity in the context of Fuji, and an algorithmically simple corruption of Fuji has 24 megabits of algorithmic specified complexity in the context of Fuji. As in Evo-Info 2, I have used the methods of Ewert, Dembski, and Marks, “Measuring Meaningful Information in Images: Algorithmic Specified Complexity” (2015). My work is recorded in a computational notebook that I will share with you in Evo-Info 5. In the meantime, any programmer who knows how to use an image-processing library can easily replicate my results. (Steps: Convert fuji.png to RGB format; save the cumulative sum of the RGB values, taken in row-major order, as signature.png; then subtract 0.5 Mb from the sizes of fuji.png and signature.png.)

As for the definition of algorithmic specified complexity, it is easiest to understand when expressed as a log-ratio of probabilities,

    \begin{equation*}    A(x; c,P) = \log_2 \frac{ M_c(x) }{ P(x) }, \end{equation*}

where x and c are binary strings (finite sequences of 0s and 1s), and P and M_c are distributions of probability over the set of all binary strings. All of the formal details are given in the next section. Speaking loosely, and in terms of the example above, M_c(x) is the probability that a randomly selected computer program converts the context c into the image x, and P(x) is the probability that image x is the outcome of a default image-generating process. The ratio of probabilities is the relative likelihood of x arising by an algorithmic process that depends on the context, and of x arising by a process that does not depend on the context. If, proceeding as in statistical hypothesis testing, we take as the null hypothesis the proposition that x is the outcome of a process with distribution P, and as an alternative hypothesis the proposition that x is the outcome of a process with distribution M_c, then our level of confidence in rejecting the null hypothesis in favor of the alternative hypothesis depends on the value of A(x; c,P). The one and only theorem that Marks et al. have given for algorithmic specified complexity tells us to reject the null hypothesis in favor of the alternative hypothesis at confidence level 2^{-\alpha} when A(x; c,P) = \alpha.

What we should make of the high algorithmic specified complexity of the images above is that they both are more likely to derive from the context than to arise in the default image-generating process. Again, Fuji is just a copy of the context, and Signature is an algorithmically simple corruption of the context. The probability of randomly selecting a program that cumulatively sums the RGB values of the context is much greater than the probability of generating the image Signature directly, i.e., without use of the context. So we confidently reject the null hypothesis that Signature arose directly in favor of the alternative hypothesis that Signature derives from the context.

This embarrassment of Marks et al. is ultimately their own doing, not mine. It is they who buried the true story of Dembski’s (2005) redevelopment of specified complexity in terms of statistical hypothesis testing, and replaced it with a fanciful account of specified complexity as a measure of meaningful information. It is they who neglected to report that their theorem has nothing to do with meaning, and everything to do with hypothesis testing. It is they who sought out examples to support, rather than to refute, their claims about meaningful information.

Algorithmic specified complexity versus algorithmic mutual information

This section assumes familiarity with the basics of algorithmic information theory.

Objective. Reduce the difference in the expressions of algorithmic specified complexity and algorithmic mutual information, and provide some insight into the relation of the former, which is unfamiliar, to the latter, which is well understood.

Here are the definitions of algorithmic specified complexity (ASC) and algorithmic mutual information (AMI):

    \begin{align*}    A(x; c, P)  &:=  -\!\log P(x) - K(x|c) &\text{(ASC)} \\    I(x: c)    &:=  K(x) - K(x|c^*),  &\text{(AMI)} \\ \end{align*}

where P is a distribution of probability over the set \{0, 1\}^* of binary strings, x and c are binary strings, and binary string c^* is a shortest program, for the universal prefix machine in terms of which the algorithmic complexity K(\cdot) and the conditional algorithmic complexity K(\cdot|\cdot) are defined, outputting c. The base of the logarithm is 2. Marks et al. regard algorithmic specified complexity as a measure of the meaningful information of x in the context of c.

The law of conservation (nongrowth) of algorithmic mutual information is:

    \begin{equation*}    I(f(x) : z) \leq I(x:z) + K(f) + O(1) \end{equation*}

for all binary strings x and z, and for all computable functions f on the binary strings. The analogous relation for algorithmic specified complexity does not hold. For example, let the probability P(\lambda) = 0, where \lambda is the empty string, and let P(x) be positive for all nonempty binary strings x. Also let f(x) = \lambda for all binary strings x. Then for all nonempty binary strings x and for all binary strings c,

    \begin{equation*}    A(f(x); c, P) - A(x; c, P) = \infty \end{equation*}

because

    \begin{align*}    A(f(x); c, P)       &= -\!\log P(f(x)) - K(f(x) | c) \\       &= -\!\log P(\lambda) - K(\lambda | c) \\       &= -\!\log 0 - K(\lambda | c) \\       &= \infty - K(\lambda | c) \\       &= \infty - O(1) \\ \end{align*}

is infinite and A(x; c, P) is finite. [Edit: The one bit of algorithmic information theory you need to know, in order to check the proof, is that K(x|c) is finite for all binary strings x and c. I have added a line at the end of the equation to indicate that K(\lambda|c) is not only finite, but also constant.] Note that this argument can be modified by setting P(\lambda) to an arbitrarily small positive number instead of to zero. Then the growth of ASC due to data processing f(x) can be made arbitrarily large, though not infinite.

There is no upper bound on the increase of algorithmic specified complexity due to data processing.

There’s a simple way to deal with the different subexpressions K(x|c) and K(x|c^*) in the definitions of ASC and AMI, and that is to restrict our attention to the case of A(x; c^*\!, P), in which the context is a shortest program outputting c.

    \begin{align*}    A(x; c^*\!, P)  &\phantom{:}=  -\!\log P(x) - K(x|c^*) &\text{(ASC)} \\    I(x: c)    &:=  K(x) - K(x|c^*)  &\text{(AMI)} \\ \end{align*}

This is not an onerous restriction, because there is for every string c a shortest program c^* that outputs c. In fact, there would be no practical difference for Marks et al. if they were to require that the context be given as c^*. We might have gone a different way, replacing the contemporary definition of algorithmic mutual information with an older one,

    \begin{align*}    I(x:c) &:= K(x) - K(x|c). &\text{(old AMI)} \\ \end{align*}

However, the older definition comes with some disadvantages, and I don’t see a significant advantage in using it. Perhaps you will see something that I have not.

The difference in algorithmic mutual information and algorithmic specified complexity,

    \begin{equation*}    I(x:c) - A(x; c^*\!, P) = K(x) + \log P(x), \end{equation*}

has no lower bound, because K(x) is non-negative for all x, and \log P(x) is negatively infinite for all x such that P(x) is zero. It is helpful, in the following, to keep in mind the possibility that P(x) = 1 for a string x with very high algorithmic complexity, and that P(x) = 0 for all strings w \neq x. Then the absolute difference in I(w:c) and A_P(w|c) is infinite for all w \neq x, and very large, though finite, for w = x. There is no limit on the difference for x. Were Marks et al. to add a requirement that P(w) be positive for all w, we still would be able, with an appropriate definition of P, to make the absolute difference of AMI and ASC arbitrarily large, though finite, for all w.

There is no upper bound on \min_x |I(x: c) - A(x; c^*\!, P)|. Thus algorithmic mutual information and algorithmic specified complexity are not equivalent in any simple sense.

The remaining difference in the expressions of ASC and AMI is in the terms -\!\log P(x) and K(x). The easy way to reduce the difference is to convert the algorithmic complexity K(x) into an expression of log-improbability, -\!\log M(x), where

    \begin{equation*}    M(x) := 2 ^ {-K(x)} \end{equation*}

for all binary strings x. An elementary result of algorithmic information theory is that the probability function M satisfies the definition of a semimeasure, meaning that \sum_x M(x) \leq 1. In fact, M is called the universal semimeasure. A semimeasure with probabilities summing to unity is called a measure. We need to be clear, when writing

    \begin{align*}    A(x; c, P)  &=  -\!\log P(x) - K(x|c^*) &\text{(ASC)} \\    I(x: c)    &=  -\!\log M(x) - K(x|c^*),  &\text{(AMI)} \\ \end{align*}

that P is a measure, and that M is a semimeasure, not a measure. (However, in at least one of the applications of algorithmic specified complexity, Marks et al. have made P a semimeasure, not a measure.) This brings us to a simple characterization of ASC:

The algorithmic specified complexity A(x; c^*\!, P) is the algorithmic mutual information I(x:c) with an arbitrary probability measure P substituted for the universal semimeasure M.

This does nothing but to establish clearly a sense in which algorithmic specified complexity is formally similar to algorithmic mutual information. As explained above, there can be arbitrarily large differences in the two quantities. However, if we consider averages of the quantities over all strings x, holding c constant, then we obtain an interesting result. Let random variable X take values in \{0, 1\}^* with probability distribution P. Then the expected difference of AMI and ASC is

    \begin{align*}    E[I(X : c) - A(X; c^*\!, P)]       &= E[K(X) + \log P(X)]    \\       &= E[K(X)] - E[-\!\log P(X)]    \\       &= E[K(X)] - H(P). \end{align*}

The expected difference in algorithmic mutual information I(X:c) and algorithmic specified complexity A(X; c^*\!, P) is the difference of the expected algorithmic complexity, E[K(X)], and the entropy H(P) of the probability distribution P.

Introducing a requirement that function P be computable, we get lower and upper bounds on the expected difference from Theorem 10 of Grünwald and Vitányi, “Algorithmic Information Theory“:

    \begin{equation*}    0 \leq E_P[K(X)] - H(P) \leq K(P) + O(1). \end{equation*}

Note that the algorithmic complexity of computable probability distribution P is the length of the shortest program that, on input of binary string x and number q (i.e., a binary string interpreted as a non-negative integer), outputs P(x) with q bits of precision (see Example 7 of Grünwald and Vitányi, “Algorithmic Information Theory“).

If the probability distribution P of random variable X is computable, then the expected difference in value of the algorithmic mutual information I(X:c) and the algorithmic specified complexity A(x; c^*\!, P) is at least zero, and is at most K(P) + O(1). Equivalently,

    \begin{align*}    E[A(x; c^*\!, P)] &\leq E[I(X:c)] \\                  &\leq E[A(x; c^*\!, P)] + K(P) + O(1). \end{align*}

How much the expected value of the AMI may exceed the expected value of the ASC depends on the length of the shortest program for computing P(x).

Next we derive a similar result, but for individual strings instead of expectations, applying a fundamental result in algorithmic information theory, the MDL bound. If P is computable, then for all binary strings x and c,

(MDL)   \begin{align*}    K(x) &\leq -\!\log P(x) + K(P) + O(1)  \\    K(x) + \log P(x) &\leq K(P) + O(1)  \\    I(x:c) - A(x; c^*\!, P) &\leq K(P) + O(1).  \\ \end{align*}

Replacing string x in this inequality with a string-valued random variable X, as we did previously, and taking the expected value, we obtain one of the inequalities in the box above. (The cross check increases my confidence, but does not guarantee, that I’ve gotten the derivations right.)

If the probability distribution P is computable, then the difference in the algorithmic mutual information I(x:c) and the algorithmic specified complexity A(x; c^*\!, P) is bounded above by K(P) + O(1).

Finally, we express K(x|c) in terms of the universal conditional semimeasure,

    \begin{equation*}    M_c(x) = 2 ^ {-K(x|c)}. \end{equation*}

Now K(x|c) = -\!\log_2 M_c(x), and we can express algorithmic specified complexity as a log-ratio of probabilities, with the caveat that \sum M_c(x) < 1.

The algorithmic specified complexity of binary string x in the context of binary string c is

    \begin{equation*}    A(x; c, P) = \log_2 \frac{ M_c(x) }{ P(x) }, \end{equation*}

where P is a distribution of probability over the binary strings \{0, 1\}^* and M_c is the universal conditional semimeasure.

An old proof that high ASC is rare

Marks et al. have fostered a great deal of confusion, citing Dembski’s earlier writings containing the term specified complexity — most notably, No Free Lunch: Why Specified Complexity Cannot Be Purchased without Intelligence (2002) — as though algorithmic specified complexity originated in them. The fact of the matter is that Dembski radically changed his approach, but not the term specified complexity, in “Specification: The Pattern that Signifies Intelligence” (2005). Algorithmic specified complexity derives from the 2005 paper, not Dembski’s earlier work. As it happens, the paper has received quite a bit of attention in The Skeptical Zone. See, especially, Elizabeth Liddle’s “The eleP(T|H)ant in the room” (2013).

Marks et al. furthermore have neglected to mention that Dembski developed the 2005 precursor (more general form) of algorithmic specified complexity as a test statistic, for use in rejecting a null hypothesis of natural causation in favor of an alternative hypothesis of intelligent, nonnatural intervention in nature. Amusingly, their one and only theorem for algorithmic specified complexity, “The probability of obtaining an object exhibiting \alpha bits of ASC is less then or equal to 2^{-\alpha},” is a special case of a more general result for hypothesis testing. The result is Theorem 1 in A. Milosavljević’s “Discovering Dependencies via Algorithmic Mutual Information: A Case Study in DNA Sequence Comparisons” (1995), which Marks et al. do not cite, though they have pointed out that algorithmic specified complexity is formally similar to algorithmic mutual information.

According to the abstract of Milosavljević’s article:

We explore applicability of algorithmic mutual information as a tool for discovering dependencies in biology. In order to determine significance of discovered dependencies, we extend the newly proposed algorithmic significance method. The main theorem of the extended method states that d bits of algorithmic mutual information imply dependency at the significance level 2^{-d+O(1)}.

However, most of the argument — particularly, Lemma 1 and Theorem 1 — applies more generally. It in fact applies to algorithmic specified complexity, which, as we have seen, is defined quite similarly to algorithmic mutual information.

Let P_0 and P_A be probability distributions over sequences (or any other kinds of objects from a countable domain) that correspond to the null and alternative hypotheses respectively; by p_0(t) and p_A(t) we denote the probabilities assigned to a sequence t by the respective distributions.

[…]

We now define the alternative hypothesis P_A in terms of encoding length. Let A denote a decoding algorithm [our universal prefix machine] that can reconstruct the target t [our x] based on its encoding relative to the source s [our context c]. By I_A(t|s) [our K(x|c)] we denote the length of the encoding [for us, this is the length of a shortest program that outputs x on input of c]. We make the standard assumption that encodings are prefix-free, i.e., that none of the encodings represented in binary is a prefix of another (for a detailed discussion of the prefix-free property, see Cover & Thomas, 1991; Li & Vitányi, 1993). We expect that the targets that are similar to the source will have short encodings. The following theorem states that a target t is unlikely to have an encoding much shorter than -\!\log p_0(t).

THEOREM 1 For any distribution of probabilities P_0, decoding algorithm A, and source s,

    \begin{equation*}    P_0\{-\!\log p_0(t) - I_A(t|s) \geq d\} \leq 2^{-d}. \end{equation*}

Here is a specialization of Theorem 1 within the framework of this post: Let X be a random variable with distribution P_0 of probability over the set \{0, 1\}^* of binary strings. Let context c be a binary string. Then the probability of algorithmic specified complexity A(X; c, P_0) \geq d is at most 2^{-d}, i.e.,

    \begin{align*}    \Pr\{ A(X; c, P_0) \geq d \}       &= \Pr\{ -\!\log P_0(X) - K(X|c) \geq d \} \\       &\leq 2^{-d}. \end{align*}

For a simpler formulation and derivation, along with a snarky translation of the result into IDdish, see my post “Deobfuscating a Theorem of Ewert, Marks, and Dembski” at Bounded Science.

I emphasize that Milosavljević does not narrow his focus to algorithmic mutual information until Theorem 2. The reason that Theorem 1 applies to algorithmic specified complexity is not that ASC is essentially the same as algorithm mutual information — we established above that it is not — but instead that the theorem is quite general. Indeed, Theorem 1 does not apply directly to the algorithmic mutual information

    \begin{equation*}    I(x:c) = -\!\log_2 M(x) - K(x|c^*), \end{equation*}

because M is a semimeasure, not a measure like P_0. Theorem 2 tacitly normalizes the probabilities of the semimeasure M, producing probabilities that sum to unity, before applying Theorem 1. It is this special handling of algorithmic mutual information that leads to the probability bound of 2^{-d+O(1)} stated in the abstract, which is different from the bound of 2^{-d} for algorithmic specified complexity. Thus we have another clear indication that algorithmic specified complexity is not essentially the same as algorithmic mutual information, though the two are formally similar in definition.

Conclusion

In 2005, William Dembski made a radical change in what he called specified complexity, and developed the precursor of algorithmic specified complexity. He presented the “new and improved” specified complexity in terms of statistical hypothesis testing. That much of what he did made good sense. There are no formally established properties of algorithmic specified complexity that justify regarding it as a measure of meaningful information. The one theorem that Marks et al. have published is actually a special case of a 1995 result applied to hypothesis testing. In my own exploration of the properties of algorithmic specified complexity, I’ve seen nothing at all to suggest that it is reasonably regarded as a measure of meaning. Marks et al. have attached “meaning” rhetoric to some example applications of algorithmic specific complexity, but have said nothing about counter­examples, which are quite easy to find. In Evo-Info 5, I will explain how to use the methods of “Measuring Meaningful Information in Images: Algorithmic Specified Complexity” to measure large quantities of “meaningful information” in nonsense images derived from the context.

[Some minor typographical errors in the original post have been corrected.]

115 Replies to “Evo-Info 4: Non-conservation of algorithmic specified complexity”

  1. Mung Mung
    Ignored
    says:

    EricMH: I took my definition from Wikipedia:

    I took mine from Probability, Random Variables and Stochastic Processes (4th Ed.).

  2. Tom English Tom English
    Ignored
    says:

    Mung,

    You obscurantist! 😉

  3. Mung Mung
    Ignored
    says:

    Tom English: You obscurantist!

    Can’t wait to meet you in person and buy you a drink!

  4. Joe Felsenstein Joe Felsenstein
    Ignored
    says:

    BruceS: It is possible to introduce randomness into usages of K complexity.For example, one can relate K complexity and Shannon entropy by introducing a probability distribution in forming messages whose complexity is to be analysed.

    This done inCover and Thomas chapter 14.4 which starts

    Thanks. I ought to take a look at Cover and Thomas.

    I note that as you state it, it starts with a probabilistic random variable (strings) and only drags in K complexity after that. K complexity does not itself involve any random variables.

  5. Joe Felsenstein Joe Felsenstein
    Ignored
    says:

    Tom English: When Ewert first talked about algorithmic specified complexity, at Jonathan Bartlett’s 2012 Engineering and Metaphysics conference,

    Actually, Dembski talked about it as one of the ways you could “cash out” specified information, in No Free Lunch. He also allowed other ways including viability. Occasionally after that one would find Dembski talking about K complexity as the way one could cash out SI. But not invariably. So it was a little confusing.

  6. Corneel Corneel
    Ignored
    says:

    Tom English: I suspect that a biological application of ASC would look somewhat like the application to the Game of Life, and that the emphasis would be on use of the context, whatever it might be, to describe biologically functional configurations of matter. There’s far too much I might say on this issue, so I need to end things here, however unsatisfyingly, and encourage you to read “Algorithmic Specified Complexity in the Game of Life.”

    Thanks. Conway’s Game of Life is something familiar at least, so will help me wrapping my head around this stuff.

    But without any populations, genomes and heritable characters it’s pretty hard to see what relevance ASC has for criticisms of natural selection as a source of biological complex information. I am really curious how Eric is going to translate this argument to Joe’s toy example.

  7. Joe Felsenstein Joe Felsenstein
    Ignored
    says:

    Tom English: When Ewert first talked about algorithmic specified complexity, at Jonathan Bartlett’s 2012 Engineering and Metaphysics conference,

    Actually, Dembski talked about it as one of the ways you could “cash out” specified information, in No Free Lunch. He also allowed other ways including viability. Occasionally after that one would find Dembski talking about K complexity as the way one could cash out SI. But not invariably. So it was a little confusing.

  8. Joe Felsenstein Joe Felsenstein
    Ignored
    says:

    EricMH: That said, I’ll probably still return here and to Peaceful Science occasionally, because in the midst of everyone’s attempt to obfuscate ID some interesting points are still made.

    It is unfortunate that Eric Holloway intends to absent himself from TSZ — I thought that the discussion of his mutual information argument was gradually moving toward a real explanation of it, so that we can understand and evaluate it. I look forward to Holloway’s return(s).

    I hope in the near future, when other tasks are less pressing, to put up a post asking specifically how one applies Holloway’s mutual information criterion to the simple model of natural selection at multiple loci that I talked about earlier in this thread. Because that is a model that seems to show that selection can result in CSI. And I don’t see how Holloway’s argument applies there. I hope that Holloway will return for that discussion.

  9. Joe Felsenstein Joe Felsenstein
    Ignored
    says:

    Corneel: OK, one question. What is supposed to be the context for the calculations of the ASC in a biological example of evolution? My gut feeling tells me it is supposed to be the allele frequencies of the initial population, but without any biological example I am just guessing. Do you know? Does Joe?

    No, Joe does not. One of the posts I intend to make in the near future will raise exactly that issue in a separate thread. I can’t for the life of me see why defining a target region by algorithmic (non-)complexity is a sensible thing to do. I hope to say that again in a separate post, and maybe that will lead to somebody explaining why it makes sense. Till then, it makes about as much sense to me as defining the target region by blueness instead of some fitness-related adaptation.

  10. Corneel Corneel
    Ignored
    says:

    Joe Felsenstein: No, Joe does not.

    🙂
    If neither Tom nor you can tell me what variables feed into the formulas or how the argument is supposed to go in principle, then I feel a bit less stupid about not understanding it.

  11. phoodoo
    Ignored
    says:

    Corneel:
    If neither Tom nor you can tell me what variables feed into the formulas orhow the argument is supposed to go in principle, then I feel a bit less stupid about not understanding it.

    It seems Tom has neither answered this adequately, nor this:

    EricMH
    Ignored
    says:
    December 8, 2018 at 6:48 pm

    Tom English: 1. Do you understand that my demonstration of non-conservation of ASC involves no random variable?

    This is where we are talking past each other. My understanding of ASC is that it is testing whether x is sampled from a random variable defined by P. That is what I(x) = -log_2 P(x) is measuring. As such, your proof does not work.

    If we don’t agree on what x is, then we probably don’t have much to say to each other as we are talking about different things.

    Seems strange to remove this post from being featured and Tom just allowing these to go unanswered.

  12. Corneel Corneel
    Ignored
    says:

    phoodoo: It seems Tom has neither answered this adequately

    Well, isn’t it up to Eric, not Tom, to make that connection clear?

    phoodoo: , nor this:

    I won’t pretend to understand what Eric is saying here, so I don’t know whether he is voicing a legitimate concern. It annoys me no end that I can’t even make that basic distinction, so I am working my way through the “Game of Life” paper now.

  13. DNA_Jock
    Ignored
    says:

    phoodoo,
    Here’s a summary of the conversation in simpler terms.

    Eric: ASC is conserved
    Tom: Here is an example where it is not.
    Eric: That’s not what I use ASC for. Oh, and you are a boring obfuscating bully. Gotta run now.

    Unclear to me what sort of response you expect from Tom, since Eric has made no effort to contest Tom’s demonstration that ASC is NOT conserved, merely waffled about random variables being labels.

  14. phoodoo
    Ignored
    says:

    Corneel: I won’t pretend to understand what Eric is saying here, so I don’t know whether he is voicing a legitimate concern. It annoys me no end that I can’t even make that basic distinction, so I am working my way through the “Game of Life” paper now.

    Right, but we also can’t pretend to know what Tom is saying, and since this is Tom’s post I guess it is up to him to make it clear what he is saying, because he seems to be the only one who knows.

  15. Corneel Corneel
    Ignored
    says:

    phoodoo: Right, but we also can’t pretend to know what Tom is saying, and since this is Tom’s post I guess it is up to him to make it clear what he is saying, because he seems to be the only one who knows.

    I think Tom did a splendid job visualising a major shortcoming of the ASC metric as a measure of meaning, by demonstrating that ASC can increase enormously when an image gets corrupted. I agree that the mathematical treatment is hard to digest, but to be honest, Eric’s OP hasn’t been light reading either, and wasn’t very illuminating to boot.

    Anyway, if you are unsure about specific details of the OP; Tom has clearly stated that he is anxious to answer our questions, even dumb ones. 🙂

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.