# A mathematical proof of Ockham’s razor?

Ockham’s razor is a principle often used to dismiss out of hand alleged phenomena deemed to be too complex. In the philosophy of religion, it is often invoked for arguing that God’s existence is extremely unlikely to begin with owing to his alleged incredible complexity. A geeky brain is desperately required before entering this sinister realm.

In a earlier post I dealt with some of the most popular justifications for the razor and made the following distinction:

Methodological Razor: if theory A and theory B do the same job of describing all known facts C, it is preferable to use the simplest theory for the next investigations.

Epistemological Razor: if theory A and theory B do the same job of describing all known facts C, the simplest theory is ALWAYS more likely.”

Like the last time, I won’t address the validity of the Methodological Razor (MR) which might be an useful tool in many situations.

My attention will be focused on the epistemological glade and its alleged mathematical grounding.

## Example: prior probabilities of models having discrete variables

To illustrate how this is supposed to work, I built up the following example. Let us consider the result $Y$ of a random experiment depending on a measured random variable $X$. We are now searching for a good model (i.e. function  $f(X)$  ) such that the distance $d = Y - f(X)$ is minimized with respect to constant parameters appearing in $f$. Let us consider the following functions: $f1(X,a1)$$f2(X,a1,a2)$$f3(X,a1,a2,a3)$  and  $f4(X,a1,a2,a3,a4)$ . which are the only possible models aiming at representing the relation between $Y$ and $X$. Let n1 = 1, n2 = 2, n3 =3 and n4 = 4 be their number of parameters. In what follows, I will neutrally describe how objective Bayesians justify Ockham’s razor in that situation.

### The objective Bayesian reasoning

Objective Bayesians apply the principle of indifference, according to which in utterly unknown situations every rational agent assigns the same probability to each possibility.

Let be $pi_{total} = p( f i)$ , the probability that the function is the correct description of reality. It follows from that assumption that $p1_{total}=p2_{total} = p3_{total} = p4_{total} = p = \frac{1}{4}$ owing to the the additivity of the probabilities.

Let us consider that one constant coefficient $ai$ can only take on five discrete values  1, 2, 3, 4 and 5. Let us call $p1$  $p2$$p3$  and  $p4$ the probabilities that one of the four models is right with very specific values of the coefficient (a1, a2, a3, a4). By applying once again the principle of indifference, one gets: $p1(1) = p1(2) = p1(3) = p1(4) = p1(5) = \frac{1}{5}p1_{total} = 5^{-n1}p$ In the case of the second function which depends on two variable a, we have 5*5 doublets of values which are possible: (1,1) (1,2),…..(3,4)….(5,5) From indifference, it follows that $p2(1,1)=p2(1,2) = ... = p2(3,4) = ....p2(5,5) = \frac{1}{25} p2_{total} = 5^{-n2}p$ There are 5*5*5 possible values for f3.

Indifference entails that $p3(1,1,1)=p3(1,,12)=... =p3(3,,2,4)=....p3(5,5,5)= \frac{1}{125} p3_{total} = 5^{-n3}p$ f4 is characterized by four parameters, so that a similar procedure leads to $p4(1,1,1,1)=p4(1,1,1,2) =...=p4(3, 2,1,4)=....p4(5,5,5,5)=\frac{1}{625}p4_{total}= 5^{-n4}p$ Let us now consider four wannabe solutions to the parameter identification problem: S1 = a1 S2 = {b1, b2} S3 = {c1, c2, c3} S4 = {d1, d2, d3, d4} each member being an integer between 1 and 5. The prior probabilities of these solutions are equal to  the quantities we have just calculated above. Thus $p(S1)= 5^{-n1}p$ $p(S2)= 5^{-n2}p$ $p(S3)= 5^{-n3}p$ $p(S4)= 5^{-n4}p$ From this, it follows that  $\frac{p(Si)}{p(Sj)}= 5^{nj - ni}$ or $O(i,j)= \frac{p(Si)}{p(Sj)} =5^{nj - ni}$ If one compares the first and the second model, $O(1,2) = 5^{2-1} = 5$ which means that the fit with the first model is (a priori) 5 times as likely as that with the second one .

Likewise, O(1,3) = 25 and O(1,4) = 125 showing that the first model is (a priori) 25 and 125 times more likely than the third and fourth model, respectively. If the four model fits the model with the same quality (in that for example fi(X, ai) is perfectly identical to Y), Bayes theorem will preserve the ratios for the computation of the posterior probabilities.

In other words, all things being equal, the simplest model f1(X,a1) is five times more likely than f2(X,a1,a2), 25 times more likely than f3(X,a1,a2,a3) and 125 times more likely than f4(X,a1,a2,a3,a4) because the others contain a greater number of parameters.

For this reason O(i,j) is usually referred to as an Ockham’s factor, because it penalizes the likelihood of complex models. If you are interested in the case of models with continuous real parameters, you can take a look at this publication. The sticking point of the whole demonstration is its heavy reliance on the principle of indifference.

## The trouble with the principle of indifference

I already argued against the principle of indifference in an older post. Here I will repeat and reformulate my criticism.

### Turning ignorance into knowledge

The principle of indifference is not only unproven but also often leads to absurd consequences. Let us suppose that I want to know the probability of certain coins to land odd. After having carried out 10000 trials, I find that the relative frequency tends to converge towards a given value which was 0.35, 0.43, 0.72 and 0.93 for the four last coins I investigated. Let us now suppose that I find a new coin I’ll never have the opportunity to test more than one time. According to the principle of indifference, before having ever started the trial, I should think something like that:

Since I know absolutely nothing about this coin, I know (or consider here extremely plausible) it is as likely to land odd as even.

I think this is magical thinking in its purest form. I am not alone in that assessment.

The great philosopher of science Wesley Salmon (who was himself a Bayesian) wrote what follows. “Knowledge of probabilities is concrete knowledge about occurrences; otherwise it is uselfess for prediction and action. According to the principle of indifference, this kind of knowledge can result immediately from our ignorance of reasons to regard one occurrence as more probable as another. This is epistemological magic. Of course, there are ways of transforming ignorance into knowledge – by further investigation and the accumulation of more information. It is the same with all “magic”: to get the rabbit out of the hat you first have to put him in. The principle of indifference tries to perform “real magic”. “

Objective Bayesians often use the following syllogism for grounding the principle of indifference.

1)If we have no reason for favoring one outcomes, we should assign the same probability to each of them

2) In an utterly unknown situation, we have no reason for favoring one of the outcomes

3) Thus all of them have the same probability.

The problem is that (in a situation of utter ignorance) we have not only no reason for favoring one of the outcomes, but also no grounds for thinking that they are equally probable.

The necessary condition in proposition 1) is obviously not sufficient.

This absurdity (and other paradoxes) led philosopher of mathematics John Norton to conclude:

“The epistemic state of complete ignorance is not a probability distribution.”

The Dempter Shafer theory of evidence offers us an elegant way to express indifference while avoiding absurdities and self-contradictions. According to it, a conviction is not represented by a probability (real value between 0 and 1) but by an uncertainty interval [ belief(h) ; 1 – belief(non h) ] , belief(h) and belief(non h) being the degree of trust one has in the hypothesis h and its negation.

For an unknown coin, indifference according to this epistemology would entail  belief(odd) = belief(even) = 0, leading to the probability interval [0 ; 1].

### Non-existing prior probabilities

Philosophically speaking, it is controversial to speak of the probability of a theory before any observation has been taken into account. The great philosopher of evolutionary biology Elliot Sober has a nice way to put it: ““Newton’s universal law of gravitation, when suitably supplemented with plausible background assumptions, can be said to confer probabilities on observations. But what does it mean to say that the law has a probability in the light of those observations? More puzzling still is the idea that it has a probability before any observations are taken into account. If God chose the laws of nature by drawing slips of paper from an urn, it would make sense to say that Newton’s law has an objective prior. But no one believes this process model, and nothing similar seems remotely plausible.”

It is hard to see how prior probabilities of theories can be something more than just subjective brain states.

## Conclusion

The alleged mathematical demonstration of Ockham’s razor lies on extremely shaky ground because:

1) it relies on the principle of indifference which is not only unproven but leads to absurd and unreliable results as well

2) it assumes that a model has already a probability before any observation.

Philosophically this is very questionable. Now if you are aware of other justifications for Ockham’s razor, I would be very glad if you were to mention them.

## 3 thoughts on “A mathematical proof of Ockham’s razor?”

1. Andy Schueler says:

Now if you are aware of other justifications for Ockham’s razor, I would be very glad if you were to mention them.

If there are two competing explanations with equal explanatory power, and the assumptions of the simpler explanations are a subset of the more complex explanation, then the simpler can explanation logically can only be more likely or equally likely than the more complex one, but cannot possibly be less likely.
Example: if explanation 1 relies on assumptions A and B, and explanation 2 relies on A, B, and C – then the two explanations are equally likely to be correct if the probability of C being true is 1, and explanation 1 is more likely in every other case.
If the simpler explanation relies on assumptions that are NOT a subset of the assumptions for a more complex explanation, then I doubt that there can be anything like a formal proof of Occam´s Razor being valid.

To get the epistemological razor (see my definitions above) you need to prove that the probability of the more complex explanation is strictly inferior to that of the simpler one. The probability of C might very well be unknown, so that we don’t know the extent of the penalty.

I think there is a simpler way to go: if C does not add anything to the explanation of the phenomenon at hand, C is not a part of the explanation .

Therefore it is a separate problem/claim/phenomenon which should be studied in an entirely different context.

I generally doubt, however, that propositions which are not events possess probabilities.
I believe that events which have not occurred (such as technological singularity) or historical happenings have a physical probability but the same cannot be said about the existence of numbers, the soul, God or the truth of string theory.

In such situations, I resort to likelihoodism and consider the quantity p(E|theory), that is to say the probability of the ensemble of our evidence given the truth of the theory .
In this way it is possible to compare theory1, theory2 and theory3 with respect to their agreement with reality.
The theories leading to the worse predictions (in terms of the probabilities of the predictions of the data in the real world) should no longer be pursued.

Concrete example: p(E|Young Earth Creationism) < p(E| old earth creationism) < p(E | Behe's intelligent design) < p(E| evolution).

Doing this is perfectly sufficient for all my concerns as a scientist (and that of all people I know), there is no need to introduce subjective degrees of belief into the picture.

2. This absurdity (and other paradoxes) led philosopher of mathematics John Norton to conclude: “The epistemic state of complete ignorance is not a probability distribution.”

The Dempter Shafer theory of evidence offers us an elegant way to express indifference while avoiding absurdities and self-contradictions.

This is really neat! Bayesian probabilities convolve two sets which ought to be kept separate: evidence-for, and evidence-against. Of course we often combine these sets when we need to act, but we still keep them separate after the action.

After being instigated by a friend, I’ve thought a lot about how ‘unknown’ is a different type or category than ‘true’ or ‘false’. I think you’re really onto something, Lotharson!

To what extent have you examined the criticism of Dempter Shafer theory? I doubt we will ever find the perfect way to evaluate evidence; instead, I predict we’ll just find better and better models, or at least models that work in more and more situations. 🙂