Next Article in Journal
Recognizing the Repeatable Configurations of Time-Reversible Generalized Langton’s Ant Is PSPACE-Hard
Previous Article in Journal
Univariate Cubic L1 Interpolating Splines: Spline Functional, Window Size and Analysis-based Algorithm
Previous Article in Special Issue
Ray Solomonoff, Founding Father of Algorithmic Information Theory
Article Menu

Export Article

Algorithms 2010, 3(4), 329-350; doi:10.3390/a3040329

A Complete Theory of Everything (Will Be Subjective)
Marcus Hutter
SoCS, RSISE, IAS, CECS, Australian National University, Canberra, ACT, 0200, Australia
Received: 30 August 2010 / Accepted: 22 September 2010 / Published: 29 September 2010


Increasingly encompassing models have been suggested for our world. Theories range from generally accepted to increasingly speculative to apparently bogus. The progression of theories from ego- to geo- to helio-centric models to universe and multiverse theories and beyond was accompanied by a dramatic increase in the sizes of the postulated worlds, with humans being expelled from their center to ever more remote and random locations. Rather than leading to a true theory of everything, this trend faces a turning point after which the predictive power of such theories decreases (actually to zero). Incorporating the location and other capacities of the observer into such theories avoids this problem and allows to distinguish meaningful from predictively meaningless theories. This also leads to a truly complete theory of everything consisting of a (conventional objective) theory of everything plus a (novel subjective) observer process. The observer localization is neither based on the controversial anthropic principle, nor has it anything to do with the quantum-mechanical observation process. The suggested principle is extended to more practical (partial, approximate, probabilistic, parametric) world models (rather than theories of everything). Finally, I provide a justification of Ockham’s razor, and criticize the anthropic principle, the doomsday argument, the no free lunch theorem, and the falsifiability dogma.
world models; observer localization; predictive power; Ockham’s razor; universal theories; inductive reasoning; simplicity and complexity; universal self-sampling; no-free-lunch; computability
“... in spite of it’s incomputability, Algorithmic Probability can serve as a kind of ‘Gold Standard’ for induction systems”
— Ray Solomonoff (1997)
“There is a theory which states that if ever anyone discovers exactly what the Universe is for and why it is here, it will instantly disappear and be replaced by something even more bizarre and inexplicable. There is another theory which states that this has already happened.”
— Douglas Adams, Hitchhikers guide to the Galaxy (1979)

1. Introduction

This paper uses an information-theoretic and computational approach for addressing the philosophical problem of judging theories (of everything) in physics. In order to keep the paper generally accessible, I’ve tried to minimize field-specific jargon and mathematics, and focus on the core problem and its solution.
By theory I mean any model which can explain≈describe≈predict≈compress [1] our observations, whatever the form of the model. Scientists often say that their model explains some phenomenon. What is usually meant is that the model describes (the relevant aspects of) the observations more compactly than the raw data. The model is then regarded as capturing a law (of nature), which is believed to hold true also for unseen/future data.
This process of inferring general conclusions from example instances is call inductive reasoning. For instance, observing 1000 black ravens but no white one supports but cannot prove the hypothesis that all ravens are black. In general, induction is used to find properties or rules or models of past observations. The ultimate purpose of the induced models is to use them for making predictions, e.g., that the next observed raven will also be black. Arguably inductive reasoning is even more important than deductive reasoning in science and everyday life: for scientific discovery, in machine learning, for forecasting in economics, as a philosophical discipline, in common-sense decision making, and last but not least to find theories of everything. Historically, some famous, but apparently misguided philosophers [2,3], including Popper and Miller, even disputed the existence, necessity or validity of inductive reasoning. Meanwhile it is well-known how minimum encoding length principles [4,5], rooted in (algorithmic) information theory [6], quantify Ockham’s razor principle, and lead to a solid pragmatic foundation of inductive reasoning [7]. Essentially, one can show that the more one can compress, the better one can predict, and vice versa.
A deterministic theory/model allows from initial conditions to determine an observation sequence, which could be coded as a bit string. For instance, Newton mechanics maps initial planet positions+velocities into a time-series of planet positions. So a deterministic model with initial conditions is “just” a compact representation of an infinite observation string. A stochastic model is “just” a probability distribution over observation strings.
Classical models in physics are essentially differential equations describing the time-evolution of some aspects of the world. A Theory of Everything (ToE) models the whole universe or multiverse, which should include initial conditions. As I will argue, it can be crucial to also localize the observer, i.e., to augment the ToE with a model of the properties of the observer, even for non-quantum-mechanical phenomena. I call a ToE with observer localization, a Complete ToE (CToE).
That the observer itself is important in describing our world is well-known. Most prominently in quantum mechanics, the observer plays an active role in ‘collapsing the wave function’. This is a specific and relatively well-defined role of the observer for a particular theory, which is not my concern. I will show that (even the localization of) the observer is indispensable for finding or developing any (useful) ToE. Often, the anthropic principle is invoked for this purpose (our universe is as it is because otherwise we would not exist). Unfortunately its current use is rather vague and limited, if not outright unscientific [8]. In Section 6 I extend Schmidhuber’s formal work [9] on computable ToEs to formally include observers. Schmidhuber [9] already discusses observers and mentions sampling universes consistent with our own existence, but this part stays informal. I give a precise and formal account of observers by explicitly separating the observer’s subjective experience from the objectively existing universe or multiverse, which besides other things shows that we also need to localize the observer within our universe (not only which universe the observer is in).
In order to make the main point of this paper clear, Section 2 first traverses a number of models that have been suggested for our world, from generally accepted to increasingly speculative and questionable theories. Section 3 discusses the relative merits of the models, in particular their predictive power (precision and coverage). We will see that localizing the observer, which is usually not regarded as an issue, can be very important. Section 4 gives an informal introduction to the necessary ingredients for CToEs, and how to evaluate and compare them using a quantified instantiation of Ockham’s razor. Section 5 gives a formal definition of what accounts for a CToE, introduces more realistic observers with limited perception ability, and formalizes the CToE selection principle. The Universal ToE is a sanity critical point in the development of ToEs, and will be investigated in more detail in Section 6. Extensions to more practical (partial, approximate, probabilistic, parametric) theories (rather than ToEs) are briefly discussed in Section 7. In Section 8 I show that Ockham’s razor is well-suited for finding ToEs and briefly criticize the anthropic principle, the doomsday argument, the no free lunch theorem, and the falsifiability dogma. Section 9 concludes.

2. Theories of Something, Everything & Nothing

A number of models have been suggested for our world. They range from generally accepted to increasingly speculative to outright unacceptable. For the purpose of this work it doesn’t matter where you personally draw the line. Many now generally accepted theories have once been regarded as insane, so using the scientific community or general public as a judge is problematic and can lead to endless discussions: for instance, the historic geo↔heliocentric battle; and the ongoing discussion of whether string theory is a theory of everything or more a theory of nothing. In a sense this paper is about a formal rational criterion to determine whether a model makes sense or not. In order to make the main point of this paper clear, below I will briefly traverse a number of models. Space constraints prevent to explain these models properly, but most of them are commonly known; see e.g. [10,11] for surveys. The presented bogus models help to make clear the necessity of observer localization and hence the importance of this work.
(G) Geocentric model. In the well-known geocentric model, the Earth is at the center of the universe and the Sun, the Moon, and all planets and stars move around Earth. The ancient model assumed concentric spheres, but increasing precision in observations and measurements revealed a quite complex geocentric picture with planets moving with variable speed on epicycles. This Ptolemaic system predicted the celestial motions quite well for its time, but was relatively complex in the common sense and in the sense of involving many parameters that had to be fitted experimentally.
(H) Heliocentric model. In the modern (later) heliocentric model, the Sun is at the center of the solar system (or universe), with all planets (and stars) moving in ellipses around the Sun. Copernicus developed a complete model, much simpler than the Ptolemaic system, which interestingly did not offer better predictions initially, but Kepler’s refinements ultimately outperformed all geocentric models. The price for this improvement was to expel the observers (humans) from the center of the universe to one out of 8 moving planets. While today this price seems small, historically it was quite high. Indeed we will compute the exact price later.
(E) Effective theories. After the celestial mechanics of planets have been understood, ever more complex phenomena could be captured with increasing coverage. Newton’s mechanics unifies celestial and terrestrial gravitational phenomena. When unified with special relativity theory one arrives at Einstein’s general relativity, predicting large scale phenomena like black holes and the big bang. On the small scale, electrical and magnetic phenomena are unified by Maxwell’s equations for electromagnetism. Quantum mechanics and electromagnetism have further been unified to quantum electrodynamics (QED). QED is the most powerful theory ever invented, in terms of precision and coverage of phenomena. It is a theory of all physical and chemical processes, except for radio-activity and gravity.
(P) Standard model of particle physics. Salam, Glashow and Weinberg extended QED to include weak interactions, responsible for radioactive decay. Together with quantum chromo dynamic [12], which describes the nucleus, this constitutes the current standard model (SM) of particle physics. It describes all known non-gravitational phenomena in our universe. There is no experiment indicating any limitation (precision, coverage). It has about 20 unexplained parameters (mostly masses and coupling constants) that have to be (and are) experimentally determined (although some regularities can be explained [13]). The effective theories of the previous paragraph can be regarded as approximations of SM, hence SM, although founded on a subatomic level, also predicts medium scale phenomena.
(S) String theory. Pure gravitational and pure quantum phenomena are perfectly predictable by general relativity and the standard model, respectively. Phenomena involving both, like the big bang, require a proper final unification. String theory is the candidate for a final unification of the standard model with the gravitational force. As such it describes the universe at its largest and smallest scale, and all scales in-between. String theory is essentially parameter-free, but is immensely difficult to evaluate and it seems to allow for many solutions (spatial compactifications). For these and other reasons, there is currently no uniquely accepted cosmological model.
(C) Cosmological models. Our concept of what the universe is, seems to ever expand. In ancient times there was Earth, Sun, Moon, and a few planets, surrounded by a sphere of shiny points (fixed stars). The current textbook universe started in a big bang and consists of billions of galaxy clusters each containing billions of stars, probably many with a planetary system. But this is just the visible universe. According to inflation models, which are needed to explain the homogeneity of our universe, the “total” universe is vastly larger than the visible part.
(M) Multiverse theories. Many theories (can be argued to) imply a multitude of essentially disconnected universes (in the conventional sense), often each with their own (quite different) characteristics [14]. In Wheeler’s oscillating universe a new big bang follows the assumed big crunch, and this repeats indefinitely. Lee Smolin proposed that every black hole recursively produces new universes on the “other side” with quite different properties. Everett’s many-worlds interpretation of quantum mechanics postulates that the wave function doesn’t collapse but the universe splits (decoheres) into different branches, one for each possible outcome of a measurement. Some string theorists have suggested that possibly all compactifications in their theory are realized, each resulting in a different universe.
(U) Universal ToE. The last two multiverse suggestions contain the seed of a general idea. If theory X contains some unexplained elements Y (quantum or compactification or other indeterminism), one postulates that every realization of Y results in its own universe, and we just happen to live in one of them. Often the anthropic principle is used in some hand-waving way to argue why we are in this and not that universe [8]. Taking this to the extreme, Schmidhuber [9,15] postulates a multiverse (which I call universal universe) that consists of every computable universe (note there are “just” countably many computer programs). Clearly, if our universe is computable, then it is contained in the universal universe, so we have a ToE already in our hands. Similar in spirit but neither constructive nor formally well-defined is Tegmark’s mathematical multiverse [16].
(R) Random universe. Actually there is a much simpler way of obtaining a ToE. Consider an infinite sequence of random bits (fair coin tosses). It is easy to see that any finite pattern, i.e., any finite binary sequence, occurs (actually infinitely often) in this string. Now consider our observable universe quantized at e.g. Planck level, and code the whole space-time universe into a huge bit string. If the universe ends in a big crunch, this string is finite. (Think of a digital high resolution 3D movie of the universe from the big bang to the big crunch). This big string also appears somewhere in our random string, hence our random string is a perfect ToE. This is reminiscent of the Boltzmann brain idea that in a sufficiently large random universe, there exist low entropy regions that resemble our own universe and/or brain (observer) [17, Sec.3.8].
(A) All-a-Carte models. The existence of true randomness is controversial and complicates many considerations. So ToE (R) may be rejected on this ground, but there is a simple deterministic computable variant. Glue the natural numbers written in binary format, 1,10,11,100,101,110,111,1000,1001,... to one long string.
1101110010111011110001001 . . .
The decimal version is known as Champernowne’s number. Obviously it contains every finite substring by construction. Indeed, it is a Normal Number in the sense that it contains every substring of length n with the same relative frequency ( 2 n ). Many irrational numbers like 2 , π, and e are conjectured to be normal. So Champernowne’s number and probably even 2 are perfect ToEs.
Remarks. I presume that every reader of this section at some point regarded the remainder as bogus. In a sense this paper is about a rational criterion to decide whether a model is sane or insane. The problem is that the line of sanity differs for different people and different historical times.
Moving the earth out of the center of the universe was (and for some even still is) insane. The standard model is accepted by nearly all physicists as the closest approximation to a ToE so far. Only outside physics, often by opponents of reductionism, this view has been criticized. Some respectable researchers including Nobel Laureates go further and take string theory and even some Multiverse theories serious. Universal ToE also has a few serious proponents. Whether Boltzmann’s random noise or my All-a-Carte ToE find adherers needs to be seen. For me, Universal ToE (U) is the sanity critical point. Indeed UToE will be investigated in greater detail in later sections.
References to the dogmatic Bible, Popper’s misguided falsifiability principle [2,3], and wrong applications of Ockham’s razor are the most popular pseudo justifications of what theories are (in)sane. More serious arguments involving the usefulness of a theory will be discussed in the next section.

3. Predictive Power & Observer Localization

In the last section I have enumerated some models of (parts or including) the universe, roughly sorted in increasing size of the universe. Here I discuss their relative merits, in particular their predictive power (precision and coverage). Analytical or computational tractability also influences the usefulness of a theory, but can be ignored when evaluating its status as a ToE. For example, QED is computationally a nightmare, but this does not at all affect its status as the theory of all electrical, magnetic, and chemical processes. On the other hand, we will see that localizing the observer, which is usually not regarded as an issue, can be very important. The latter has nothing to do with the quantum-mechanical measuring process, although there may be some deeper yet to be explored connection.
Particle physics. The standard model has more power and hence is closer to a ToE than all effective theories (E) together. String theory plus the right choice of compactification reduces to the standard model, so has the same or superior power. The key point here is the inclusion of the “right choice of compactification”. Without it, string theory is in some respect less powerful than SM.
Baby universes. Let us now turn to the cosmological models, in particular Smolin’s baby universe theory, in which infinitely many universes with different properties exist. The theory “explains” why a universe with our properties exist (since it includes universes with all kinds of properties), but it has little predictive power. The baby universe theory plus a specification in which universe we happen to live would determine the value of the inter-universe variables for our universe, and hence have much more predictive power. So localizing ourselves increases the predictive power of the theory.
Universal ToE. Let us consider the even larger universal multiverse. Assuming our universe is computable, the multiverse generated by UToE contains and hence perfectly describes our universe. But this is of little use, since we can’t use UToE for prediction. If we knew our “position” in this multiverse, we would know in which (sub)universe we are. This is equivalent to knowing the program that generates our universe. This program may be close to any of the conventional cosmological models, which indeed have a lot of predictive power. Since locating ourselves in UToE is equivalent and hence as hard as finding a conventional ToE of our universe, we have not gained much.
All-a-Carte models also contain and hence perfectly describe our universe. If and only if we can localize ourselves, we can actually use it for predictions. (For instance, if we knew we were in the center of universe 001011011 we could predict that we will ‘see’ 0010 when ‘looking’ to the left and 1011 when looking to the right.) Let u be a snapshot of our space-time universe; a truly gargantuan string. Locating ourselves means to (at least) locate u in the multiverse. We know that u is the u’s number in Champernowne’s sequence (interpreting u as a binary number), hence locating u is equivalent to specifying u. So a ToE based on normal numbers is only useful if accompanied by the gargantuan snapshot u of our universe. In light of this, an “All-a-Carte” ToE (without knowing u) is rather a theory of nothing than a theory of everything.
Localization within our universe. The loss of predictive power when enlarging a universe to a multiverse model has nothing to do with multiverses per se. Indeed, the distinction between a universe and a multiverse is not absolute. For instance, Champernowne’s number could also be interpreted as a single universe, rather than a multiverse. It could be regarded as an extreme form of the infinite fantasia land from the NeverEnding Story, where everything happens somewhere. Champernowne’s number constitutes a perfect map of the All-a-Carte universe, but the map is useless unless you know where you are. Similarly but less extreme, the inflation model produces a universe that is vastly larger than its visible part, and different regions may have different properties.
Egocentric to Geocentric model. Consider now the “small” scale of our daily life. A young child believes it is the center of the world. Localization is trivial. It is always at “coordinate” (0,0,0). Later it learns that it is just one among a few billion other people and as little or much special as any other person thinks of themself. In a sense we replace our egocentric coordinate system by one with origin (0,0,0) in the center of Earth. The move away from an egocentric world view has many social advantages, but dis-answers one question: Why am I this particular person and not any other? (It also comes at the cost of constantly having to balance egoistic with altruistic behavior.)
Geocentric to Heliocentric model. While being expelled from the center of the world as an individual, in the geocentric model, at least the human race as a whole remains in the center of the world, with the remaining (dead?) universe revolving around us. The heliocentric model puts Sun at (0,0,0) and degrades Earth to planet number 3 out of 8. The astronomic advantages are clear, but dis-answers one question: Why this planet and not one of the others? Typically we are muzzled by questionable anthropic arguments [8,18]. (Another scientific cost is the necessity now to switch between coordinate systems, since the ego- and geocentric views are still useful.)
Heliocentric to modern cosmological model. The next coup of astronomers was to degrade our Sun to one star among billions of stars in our milky way, and our milky way to one galaxy out of billions of others. It is generally accepted that the question why we are in this particular galaxy in this particular solar system is essentially unanswerable.
Summary. The exemplary discussion above has hopefully convinced the reader that we indeed lose something (some predictive power) when progressing to too large universe and multiverse models. Historically, the higher predictive power of the large-universe models (in which we are seemingly randomly placed) overshadowed the few extra questions they raised compared to the smaller ego/geo/helio-centric models. (we’re not concerned here with the psychological disadvantages/damage, which may be large). But the discussion of the (physical, universal, random, and all-a-carte) multiverse theories has shown that pushing this progression too far will at some point harm predictive power. We saw that this has to do with the increasing difficulty to localize the observer.

4. Complete ToEs (CToEs)

A ToE by definition is a perfect model of the universe. It should allow to predict all phenomena. Most ToEs require a specification of some initial conditions, e.g. the state at the big bang, and how the state evolves in time (the equations of motion). In general, a ToE is a program that in principle can “simulate” the whole universe. An All-a-Carte universe perfectly satisfies this condition but apparently is rather a theory of nothing than a theory of everything. So meeting the simulation condition is not sufficient for qualifying as a Complete ToE. We have seen that (objective) ToEs can be completed by specifying the location of the observer. This allows us to make useful predictions from our (subjective) viewpoint. We call a ToE plus observer localization a subjective or complete ToE. If we allow for stochastic (quantum) universes we also need to include the noise. If we consider (human) observers with limited perception ability we need to take that into account too. So
A complete ToE needs specification of
initial conditions
state evolution
localization of observer
random noise
perception ability of observer
We will ignore noise and perception ability in the following and resume to these issues in Section 7 and Section 5, respectively. Next we need a way to compare ToEs.
Epistemology. I assume that the observers’ experience of the world consists of a single temporal binary sequence which gets longer with time. This is definitely true if the observer is a robot equipped with sensors like a video camera whose signal is converted to a digital data stream, fed into a digital computer and stored in a binary file of increasing length. In humans, the signal transmitted by the optic and other sensory nerves could play the role of the digital data stream. Of course (most) human observers do not possess photographic memory. We can deal with this limitation in various ways: digitally record and make accessible upon request the nerve signals from birth till now, or allow for uncertain or partially remembered observations. Classical philosophical theories of knowledge [19] (e.g., as justified true belief) operate on a much higher conceptual level and therefore require stronger (and hence more disputable) philosophical presuppositions. In my minimalist “spartan” information-theoretic epistemology, a bit-string is the only observation, and all higher ontologies are constructed from it and are pure “imagination”.
Predictive power and elegance. Whatever the intermediary guiding principles for designing theories/models (elegance, symmetries, tractability, consistency), the ultimate judge is predictive success. Unfortunately we can never be sure whether a given ToE makes correct predictions in the future. After all we cannot rule out that the world suddenly changes tomorrow in a totally unexpected way (cf. the quote at beginning of this article). We have to compare theories based on their predictive success in the past. It is also clear that the latter is not enough: For every model we can construct an alternative model that behaves identically in the past but makes different predictions from, say, year 2020 on. Popper’s falsifiability dogma is little helpful. Beyond postdictive success, the guiding principle in designing and selecting theories, especially in physics, is elegance and mathematical consistency. The predictive power of the first heliocentric model was not superior to the geocentric one, but it was much simpler. In more profane terms, it has significantly less parameters that need to be specified.
Ockham’s razor suitably interpreted tells us to choose the simpler among two or more otherwise equally good theories. For justifications of Ockham’s razor, see [6] and Section 8. Some even argue that by definition, science is about applying Ockham’s razor, see [20]. For a discussion in the context of theories in physics, see [21]. It is beyond the scope of this paper to repeat these considerations. In Section 4 and Section 8 I will show that simpler theories more likely lead to correct predictions, and therefore Ockham’s razor is suitable for finding ToEs.
Complexity of a ToE. In order to apply Ockham’s razor in a non-heuristic way, we need to quantify simplicity or complexity. Roughly, the complexity of a theory can be defined as the number of symbols one needs to write the theory down. More precisely, write down a program for the state evolution together with the initial conditions, and define the complexity of the theory as the size in bits of the file that contains the program. This quantification is known as algorithmic information or Kolmogorov complexity [6] and is consistent with our intuition, since an elegant theory will have a shorter program than an inelegant one, and extra parameters need extra space to code, resulting in longer programs [4,5].
From now on I identify theories with programs and write Length ( q ) for the length=complexity of program=theory q.
Standard model versus string theory. To keep the discussion simple, let us pretend that standard model (SM) + gravity (G) and string theory (S) both qualify as ToEs. SM+Gravity is a mixture of a few relatively elegant theories, but contains about 20 parameters that need to be specified. String theory is truly elegant, but ensuring that it reduces to the standard model needs sophisticated extra assumptions (e.g., the right compactification).
SM+G can be written down in one line, plus we have to give 20+ constants, so lets say one page. The meaning (the axioms) of all symbols and operators require another page. Then we need the basics, natural, real, complex numbers, sets (ZFC), etc., which is another page. That makes 3 pages for a complete description in first-order logic. There are a lot of subtleties though: (a) The axioms are likely mathematically inconsistent, (b) it’s not immediately clear how the axioms lead to a program simulating our universe, (c) the theory does not predict the outcome of random events, and (d) some other problems. So to transform the description into a C program simulating our universe, needs a couple of pages more, but I would estimate around 10 pages overall suffices, which is about 20,000 symbols=bytes. Of course this program will be (i) a very inefficient simulation and (ii) a very naive coding of SM+G. I conjecture that the shortest program for SM+G on a universal Turing machine is much shorter, maybe even only one tenth of this. The numbers are only a quick rule-of-thumb guess. If we start from string theory (S), we need about the same length. S is much more elegant, but we need to code the compactification to describe our universe, which effectively amounts to the same. Note that everything else in the world (all other physics, chemistry, etc.) is emergent.
It would require a major effort to quantify which theory is the simpler one in the sense defined above, but I think it would be worth the effort. It is a quantitative objective way to decide between theories that are (so far) predictively indistinguishable.
CToE selection principle. It is trivial to write down a program for an All-a-Carte multiverse (A). It is also not too hard to write a program for the universal multiverse (U), see Section 6. Lengthwise (A) easily wins over (U), and (U) easily wins over (P) and (S), but as discussed (A) and (U) have serious defects. On the other hand, these theories can only be used for predictions after extra specifications: Roughly, for (A) this amounts to tabling the whole universe, (U) requires defining a ToE in the conventional sense, (P) needs 20 or so parameters and (S) a compactification scheme. Hence localization-wise (P) and (S) easily win over (U), and (U) easily wins over (A). Given this trade-off, it now nearly suggests itself that we should include the description length of the observer location in our ToE evaluation measure. That is,
among two CToEs, select the one that has shorter overall length
Length ( i ) + Length ( e ) + Length ( l )
For an All-a-Carte multiverse, the last term contains the gargantuan string u, catapulting it from the shortest ToE to the longest CToE, hence (A) will not minimize (1).
ToE versus UToE. Consider any (C)ToE and its program q, e.g., (P) or (S). Since (U) runs all programs including q, specifying q means localizing (C)ToE q in (U). So (U)+q is a CToE whose length is just some constant bits (the simulation part of (U)) more than that of (C)ToE q. So whatever (C)ToE physicists come up with, (U) is nearly as good as this theory. This essentially clarifies the paradoxical status of (U). Naked, (U) is a theory of nothing, but in combination with another ToE it excels to a good CToE, albeit slightly longer=worse than the latter.
Localization within our universe. So far we have only localized our universe in the multiverse, but not ourselves in the universe. To localize our Sun, we could, e.g., sort (and index) stars by their creation date, which the model (i)+(e) provides. Most stars last for 1-10 billion years (say an average of 5 billion years). The universe is 14 billion years old, so most stars may be 3rd generation (Sun definitely is), so the total number of stars that have ever existed should very roughly be 3 times the current number of stars of about 10 11 × 10 11 . Probably “3” is very crude, but this doesn’t really matter for sake of the argument. In order to localize our Sun we only need its index, which can be coded in about log 2 ( 3 × 10 11 × 10 11 ) 75 bits. Similarly we can sort and index planets and observers. To localize earth among the 8 planets needs 3 bits. To localize yourself among 7 billion humans needs 33 bits. Alternatively one could simply specify the ( x , y , z , t ) coordinate of the observer, which requires more but still only very few bits. These localization penalties are tiny compared to the difference in predictive power (to be quantified later) of the various theories (ego/geo/helio/cosmo). This explains and justifies theories of large universes in which we occupy a random location.

5. Complete ToE - Formalization

This section formalizes the CToE selection principle and what accounts for a CToE. Universal Turing machines are used to formalize the notion of programs as models for generating our universe and our observations. I also introduce more realistic observers with limited perception ability.
Objective ToE. Since we essentially identify a ToE with a program generating a universe, we need to fix some general purpose programming language on a general purpose computer. In theoretical computer science, the standard model is a so-called Universal Turing Machine ( UTM ) [6]. It takes a program coded as a finite binary string q { 0 , 1 } * , executes it and outputs a finite or infinite binary string u { 0 , 1 } * { 0 , 1 } . The details do not matter to us, since drawn conclusions are typically independent of them. In this section we only consider q with infinite output
UTM ( q ) = u 1 q u 2 q u 3 q . . . = : u 1 : q
In our case, u 1 : q will be the space-time universe (or multiverse) generated by ToE candidate q. So q incorporates items (i) and (e) of Section 4. Surely our universe doesn’t look like a bit string, but can be coded as one as explained in Section 2 and Section 7. We have some simple coding in mind, e.g. u 1 : N q being the (fictitious) binary data file of a high-resolution 3D movie of the whole universe from big bang to big crunch, augmented by u N + 1 : q 0 if the universe is finite. Again, the details do not matter.
Observational process and subjective complete ToE. As we have demonstrated it is also important to localize the observer. In order to avoid potential qualms with modeling human observers, consider as a surrogate a (conventional not extra cosmic) video camera filming=observing parts of the world. The camera may be fixed on Earth or installed on an autonomous robot. It records part of the universe u denoted by o = o 1 : . (If the lifetime of the observer is finite, we append zeros to the finite observation o 1 : N ).
I only consider direct observations like with a camera. Electrons or atomic decays or quasars are not directly observed, but with some (classical) instrument. It is the indicator or camera image of the instrument that is observed (which physicists then usually interpret). This setup avoids having to deal with any form of informal correspondence between theory and real world, or with subtleties of the quantum-mechanical measurement process. The only philosophical presupposition I make is that it is possible to determine uncontroversially whether two finite binary strings (on paper or file) are the same or differ in some bits.
In a computable universe, the observational process within it, is obviously also computable, i.e., there exists a program s { 0 , 1 } * that extracts observations o from universe u. Formally
UTM ( s , u 1 : q ) = o 1 : s q
where we have extended the definition of UTM to allow access to an extra infinite input stream u 1 : q . So o 1 : s q is the sequence observed by subject s in universe u 1 : q generated by q. Program s contains all information about the location and orientation and perception abilities of the observer/camera, hence specifies not only item (l) but also item (o) of Section 4.
A Complete ToE (CToE) consists of a specification of a (ToE,subject) pair ( q , s ) . Since it includes s it is a Subjective ToE.
CToE selection principle. So far, s and q were fictitious subjects and universe programs. Let o 1 : t t r u e be the past observations of some concrete observer in our universe, e.g., your own personal experience of the world from birth till today. The future observations o t + 1 : t r u e are of course unknown. By definition, o 1 : t contains all available experience of the observer, including e.g., outcomes of scientific experiments, school education, read books, etc.
The observation sequence o 1 : s q generated by a correct CToE must be consistent with the true observations o 1 : t t r u e . If o 1 : t s q would differ from o 1 : t t r u e (in a single bit) the subject would have ‘experimental’ evidence that ( q , s ) is not a perfect CToE. We can now formalize the CToE selection principle as follows
Among a given set of perfect  ( o 1 : t s q = o 1 : t t r u e )  CToEs  { ( q , s ) }
select the one of smallest length  Length ( q ) + Length ( s )
Minimizing length is motivated by Ockham’s razor. Inclusion of s is necessary to avoid degenerate ToEs like (U) and (A). The selected CToE ( q * , s * ) can and should then be used for forecasting future observations via . . . o t + 1 : f o r e c a s t = UTM ( s * , u 1 : q * ) .

6. Universal ToE - Formalization

The Universal ToE is a sanity critical point in the development of ToEs, and will formally be defined and investigated in this section.
Definition of Universal ToE. The Universal ToE generates all computable universes. The generated multiverse can be depicted as an infinite matrix in which each row corresponds to one universe.
q UTM ( q ) ϵ u 1 ϵ   u 2 ϵ   u 3 ϵ   u 4 ϵ   u 5 ϵ . . . 0 u 1 0   u 2 0   u 3 0   u 4 0 . . . . . . 1 u 1 1   u 2 1   u 3 1 . . . . . . 00 u 1 00 u 2 00 . . . . . .  ⋮  ⋮ 
To fit this into our framework we need to define a single program q ˘ that generates a single string corresponding to this matrix. The standard way to linearize an infinite matrix is to dovetail in diagonal serpentines though the matrix:
u ˘ 1 : : = u 1 ϵ u 1 0 u 2 ϵ u 3 ϵ u 2 0 u 1 1 u 1 00 u 2 1 u 3 0 u 4 ϵ u 5 ϵ u 4 0 u 3 1 u 2 00 . . .
Formally, define a bijection i = q , k between a (program, location) pair ( q , k ) and the natural numbers I N i , and define u ˘ i : = u k q . It is not hard to construct an explicit program q ˘ for UTM that computes u ˘ 1 : = u 1 : q ˘ = UTM ( q ˘ ) .
One might think that it would have been simpler or more natural to generalize Turing machines to have matrix “tapes”. But this is deceiving. If we allow for Turing machines with matrix output, we also should allow for and enumerate all programs q that have a matrix output. This leads to a 3d tensor that needs to be converted to a 2d matrix, which is no simpler than the linearization above.
Partial ToEs. Cutting the universes into bits and interweaving them into one string might appear messy, but is unproblematic for two reasons: First, the bijection i = q , k is very simple, so any particular universe string u q can easily be recovered from u ˘ . Second, such an extraction will be included in the localization / observational process s, i.e., s will contain a specification of the relevant universe q and which bits k are to be observed.
More problematic is that many q will not produce an infinite universe. This can be fixed as follows: First, we need to be more precise about what it means for UTM ( q ) to write u q . We introduce an extra symbol ‘#’ for ‘undefined’ and set each bit u i q initially to ‘#’. The UTM running q can output bits in any order but can overwrite each location # only once, either with a 0 or with a 1. We implicitly assumed this model above, and similarly for s. Now we (have to) also allow for q that leave some or all bits unspecified.
The interleaving computation UTM ( s , UTM ( q ) ) = o of s and q works as follows: Whenever s wants to read a bit from u 1 : q that q has not (yet) written, control is transferred to q until this bit is written. If it is never written, then o will be only partially defined, but such s are usually not considered. (If the undefined location is before t, CToE ( q , s ) is not perfect, since o 1 : t t r u e is completely defined.)
Alternatively one may define a more complex dynamic bijection · , · that orders the bits in the order they are created, or one resorts to generalized Turing machines [9,22] which can overwrite locations and also greatly increase the set of describable universes. These variants (allow to) make u ˘ and all u q complete (no ‘#’ symbol, tape { 0 , 1 } ).
ToE versus UToE. We can formalize the argument in the last section of simulating a ToE by UToE as follows: If ( q , s ) is a CToE, then ( q ˘ , s ˜ ) based on UToE q ˘ and observer s ˜ : = r q s , where program r extracts u q from u ˘ and then o s q from u q , is an equivalent but slightly larger CToE, since UTM ( s ˜ , u ˘ ) = o q s = UTM ( s , u q ) by definition of s ˜ and Length ( q ˘ ) + Length ( s ˜ ) = Length ( q ) + Length ( s ) + O ( 1 ) .
The best CToE. Finally, one may define the best CToE (of an observer with experience o 1 : t t r u e ) as
UCToE : = arg min q , s { Length ( q ) + Length ( s ) : o 1 : t s q = o 1 : t t r u e }
where o 1 : s q = UTM ( s , UTM ( q ) ) . This may be regarded as a formalization of the holy grail in physics; of finding such a TOE.

7. Extensions

Our CToE selection principle is applicable to perfect, deterministic, discrete, and complete models q of our universe. None of the existing sane world models is of this kind. In this section I extend the CToE selection principle to more realistic, partial, approximate, probabilistic, and/or parametric models for finite, infinite and even continuous universes.
Partial theories. Not all interesting theories are ToEs. Indeed, most theories are only partial models of aspects of our world.
We can reduce the problem of selecting good partial theories to CToE selection as follows: Let o 1 : t t r u e be the complete observation, and ( q , s ) be some theory explaining only some observations but not all. For instance, q might be the heliocentric model and s be such that all bits in o 1 : t t r u e that correspond to planetary positions are predicted correctly. The other bits in o 1 : t q s are undefined, e.g. the position of cars. We can augment q with a (huge) table b of all bits for which o i q s o i t r u e . Together, ( q , b , s ) allows to reconstruct o 1 : t t r u e exactly. Hence for two different theories, the one with smaller length
Length ( q ) + Length ( b ) + Length ( s )
should be selected. We can actually spare ourselves from tabling all those bits that are unpredicted by all q under consideration, since they contribute the same overall constant. So when comparing two theories it is sufficient to consider only those observations that are correctly predicted by one (or both) theories.
If two partial theories ( q , s ) and ( q , s ) predict the same phenomena equally well (i.e., o 1 : t q s = o 1 : t q s o 1 : t t r u e ), then b = b and minimizing (4) reduces to minimizing (3).
Approximate theories. Most theories are not perfect but only approximate reality, even in their limited domain. The geocentric model is less accurate than the heliocentric model, Newton’s mechanics approximates general relativity, etc. Approximate theories can be viewed as a version of partial theories. For example, consider predicting locations of planets with locations being coded by (truncated) real numbers in binary representation, then Einstein gets more bits right than Newton. The remaining erroneous bits could be tabled as above. Errors are often more subtle than simple bit errors, in which case correction programs rather than just tables are needed.
Celestial example. The ancient celestial models just capture the movement of some celestial bodies, and even those only imperfectly. Nevertheless it is interesting to compare them. Let us take as our corpus of observations o 1 : t t r u e , say, all astronomical tables available in the year 1600, and ignore all other experience.
The geocentric model q G more or less directly describes the observations, hence s G is relatively simple. In the heliocentric model q H it is necessary to include in s H a non-trivial coordinate transformation to explain the geocentric astronomical data. Assuming both models were perfect, then, if and only if q H is simpler than q G by a margin that is larger than the extra complications due to the coordinate transformation ( Length ( q G ) Length ( q H ) > Length ( s H ) Length ( s G ) ), we should regard the heliocentric model as better.
If/since the heliocentric model is more accurate, we have to additionally penalize the geocentric model by the number of bits it doesn’t predict correctly. This clearly makes the heliocentric model superior.
Probabilistic theories. Contrary to a deterministic theory that predicts the future from the past for sure, a probabilistic theory assigns to each future a certain chance that it will occur. Equivalently, a deterministic universe is described by some string u, while a probabilistic universe is described by some probability distribution Q ( u ) , the a priori probability of u. (In the special case of Q ( u ) = 1 for u = u and 0 else, Q describes the deterministic universe u.) Similarly, the observational process may be probabilistic. Let S ( o | u ) be the probability of observing o in universe u. Together, ( Q , S ) is a probabilistic CToE that predicts observation o with probability P ( o ) = u S ( o | u ) Q ( u ) . A computable probabilistic CToE is one for which there exist programs (of lengths Length ( Q ) and Length ( S ) ) that compute the functions Q ( · ) and S ( · | · ) .
Consider now the true observation o 1 : t t r u e . The larger P ( o 1 : t t r u e ) the “better” is ( Q , S ) . In the degenerate deterministic case, P ( o 1 : t t r u e ) = 1 is maximal for a correct CToE, and 0 for a wrong one. In every other case, ( Q , S ) is only a partial theory that needs completion, since it does not compute o 1 : t t r u e . Given P, it is possible to code o 1 : t t r u e in | log 2 P ( o 1 : t t r u e ) | bits (arithmetic or Shannon-Fano code). Assuming that o 1 : t t r u e is indeed sampled from P, one can show that with high probability this is the shortest possible code. So there exists an effective description of o 1 : t t r u e of length
Length ( Q ) + Length ( S ) + | log 2 P ( o 1 : t t r u e ) |
This expression should be used (minimized) when comparing probabilistic CToEs. The principle is reminiscent of classical two-part Minimum Encoding Length principles like MML and MDL [4,5]. Note that the noise corresponds to the errors and the log term to the error table of the previous paragraphs.
Probabilistic examples. Assume S ( o | o ) = 1 o and consider the observation sequence o 1 : t t r u e = u 1 : t t r u e = 11001001000011111101101010100 . If we assume this is a sequence of fair coin flips, then Q ( o 1 : t ) = P ( o 1 : t ) = 2 t are very simple functions, but | log 2 P ( o 1 : t ) | = t is large. If we assume that o 1 : t t r u e is the binary expansion of π (which it is), then the corresponding deterministic Q is somewhat more complex, but | log 2 P ( o 1 : t t r u e ) | = 0 . So for sufficiently large t, the deterministic model of π is selected, since it leads to a shorter code (5) than the fair-coin-flip model.
Quantum theory is (argued by physicists to be) truly random. Hence all modern ToE candidates (P+G, S, C, M) are probabilistic. This yields huge additive constants | log 2 P ( o 1 : t t r u e ) | to the otherwise quite elegant theories Q. Schmidhuber [9,15] argues that all apparent physical randomness is actually only pseudo random, i.e., generated by a small program. If this is true and we could find the random number generator, we could instantly predict all apparent quantum-mechanical random events. This would be a true improvement of existing theories, and indeed the corresponding CToE would be significantly shorter. In [20, Sec.8.6.2] I give an argument why believing in true random noise may be an unscientific position.
Theories with parameters. Many theories in physics depend on real-valued parameters. Since observations have finite accuracy, it is sufficient to specify these parameters to some finite accuracy. Hence the theories including their finite-precision parameters can be coded in finite length. There are general results and techniques [4,5] that allow a comfortable handling of all this. For instance, for smooth parametric models, a parameter accuracy of O ( 1 / n ) is needed, which requires 1 2 log 2 n + O ( 1 ) bits per parameter. The explicable O ( 1 ) term depends on the smoothness of the model and prevents ‘cheating’ (e.g. zipping two parameters into one).
Infinite and continuous universes. So far we have assumed that each time-slice through our universe can be described in finitely many bits and time is discrete. Assume our universe were the infinite continuous 3+1 dimensional Minkowski space I R 4 occupied by (tiny) balls (“particles”). Consider all points ( x , y , z , t ) I R 4 with rational coordinates, and let i = x , y , z , t be a bijection to the natural numbers similarly to the dovetailing in Section 6. Let u i = 1 if ( x , y , z , t ) is occupied by a particle and 0 otherwise. String u 1 : is an exact description of this universe. The above idea generalizes to any so-called separable mathematical space. Since all spaces occurring in established physical theories are separable, there is currently no ToE candidate that requires uncountable universes. Maybe continuous theories are just convenient approximations of deeper discrete theories. An even more fundamental argument put forward in this context by [9] is that the Loewenheim-Skolem theorem (an apparent paradox) implies that Zermelo-Fraenkel set theory (ZFC) has a countably infinite model. Since all physical theories so far are formalizable in ZFC, it follows they all have a countable model. For some strange reason (possibly an historical artifact), the adopted uncountable interpretation seems just more convenient.
Multiple theories. Some proponents of pluralism and some opponents of reductionism argue that we need multiple theories on multiple scales for different (overlapping) application domains. They argue that a ToE is not desirable and/or not possible. Here I give a reason why we need one single fundamental theory (with all other theories having to be regarded as approximations): Consider two Theories ( T 1 and T 2 ) with (proclaimed) application domains A 1 and A 2 , respectively.
If predictions of T 1 and T 2 coincide on their intersection A 1 A 2 (or if A 1 and A 2 are disjoint), we can trivially “unify” T 1 and T 2 to one theory T by taking their union. Of course, if this does not result in any simplification, i.e., if Length ( T ) = Length ( T 1 ) + Length ( T 2 ) , we gain nothing. But since nearly all modern theories have some common basis, e.g. use natural or real numbers, a formal unification of the generating programs nearly always leads to Length ( q ) < Length ( q 1 ) + Length ( q 2 ) .
The interesting case is when T 1 and T 2 lead to different forecasts on A 1 A 2 . For instance, particle versus wave theory with the atomic world at their intersection, unified by quantum theory. Then we need a reconciliation of T 1 and T 2 , that is, a single theory T for A 1 A 2 . Ockham’s razor tells us to choose a simple (elegant) unification. This rules out naive/ugly/complex solutions like developing a third theory for A 1 A 2 or attributing parts of A 1 A 2 to T 1 or T 2 as one sees fit, or averaging the predictions of T 1 and T 2 . Of course T must be consistent with the observations.
Pluralism on a meta level, i.e., allowing besides Ockham’s razor other principles for selecting theories, has the same problem on a meta-level: which principle should one use in a concrete situation? To argue that this (or any other) problem cannot be formalized/quantized/mechanized would be (close to) an anti-scientific attitude.

8. Justification of Ockham’s Razor

We now prove Ockham’s razor under the assumptions stated below and compare it to the No Free Lunch myth. The result itself is not novel [9]. The intention and contribution is to provide an elementary but still sufficiently formal argument, which in particular is free of more sophisticated concepts like Solomonoff’s a-priori distribution.
Ockham’s razor principle demands to “take the simplest theory consistent with the observations”.
Ockham’s razor could be regarded as correct if among all considered theories, the one selected by Ockham’s razor is the one that most likely leads to correct predictions.
Assumptions. Assume we live in the universal multiverse u ˘ that consists of all computable universes, i.e., UToE is a correct/true/perfect ToE. Since every computable universe is contained in UToE, it is at least under the computability assumption impossible to disprove this assumptions. The second assumption we make is that our location in the multiverse is random. We can divide this into two steps: First, the universe u q in which we happen to be is chosen randomly. Second, our “location” s within u q is chosen at random. We call these the universal self-sampling assumption. The crucial difference to the informal anthropic self-sampling assumption used in doomsday arguments is discussed below.
Recall the observer program s ˜ : = r q s introduced in Section 5. We will make the simplifying assumption that s is the identity, i.e., restrict ourselves to “objective” observers that observe their universe completely: UTM ( s ˜ , u ˘ ) = o q s = UTM ( s , u q ) = u q = UTM ( q ) . Formally, the universal self-sampling assumption can be stated as follows:
A priori it is equally likely to be in any of the universes u q generated by some program q { 0 , 1 } * .
To be precise, we consider all programs with length bounded by some constant L, and take the limit L .
Counting consistent universes. Let o 1 : t t r u e = u 1 : t t r u e be the universe observed so far and
Q L : = { q : Length ( q ) L  and UTM ( q ) = u 1 : t t r u e * }
be the set of all consistent universes (which is non-empty for large L), where * is any continuation of u 1 : t t r u e . Given u 1 : t t r u e , we know we are in one of the universes in Q L , which implies by the universal self-sampling assumption a uniform sampling in Q L . Let
q m i n : = arg min q { Length ( q ) : q Q L }  and  l : = Length ( q m i n )
be the shortest consistent q and its length, respectively. Adding (unread) “garbage” g after the end of a program q does not change its behavior, i.e., if q Q L , then also q g Q L provided that Length ( q g ) L . Hence for every g with Length ( g ) L l , we have q m i n g Q L . Since there are about 2 L l such g, we have | Q L | 2 L l . It is a deep theorem in algorithmic information theory [6] that there are also not significantly more than 2 L l programs q equivalent to q m i n . The proof idea is as follows: One can show that if there are many long equivalent programs, then there must also be a short one. In our case the shortest one is q m i n , which upper bounds the number of long programs. Together this shows that
| Q L | 2 L l
Probabilistic prediction. Given observations u 1 : t t r u e we now determine the probability of being in a universe that continues with u t + 1 : n , where n > t . Similarly to the previous paragraph we can approximately count the number of such universes:
Q L n : = { q : Length ( q ) L and UTM ( q ) = u 1 : t t r u e u t + 1 : n * } Q L q m i n n : = arg min q { Length ( q ) : q Q L n } and l n : = Length ( q m i n n ) | Q L n | 2 L l n
The probability of being in a universe with future u t + 1 : n given u 1 : t t r u e is determined by their relative number
P ( u t + 1 : n | u 1 : t t r u e ) = | Q L n | | Q L | 2 ( l n l )
which is (asymptotically) independent of L.
Ockham’s razor. Relation (6) implies that the most likely continuation u ^ t + 1 : n : = argmax u t + 1 : n P ( u t + 1 : n | u 1 : t t r u e ) is (approximately) the one that minimizes l n . By definition, q m i n is the shortest program in Q L = u t + 1 : n Q L n . Therefore
P ( u ^ t + 1 : n | u 1 : t t r u e ) P ( u t + 1 : n q m i n | u 1 : t t r u e )
The accuracy of ≈ is clarified later. In words
We are most likely in a universe that is (equivalent to) the simplest universe consistent with our past observations.
This shows that Ockham’s razor selects the theory that most likely leads to correct predictions, and hence proves (under the stated assumptions) that Ockham’s razor is correct.
Ockham’s razor is correct under the universal self-sampling assumption.
Discussion. It is important to note that the universal self-sampling assumption has not by itself any bias towards simple models q. Indeed, most q in Q L have length close to L, and since we sample uniformly from Q L this actually represents a huge bias towards large models for L .
The result is also largely independent of the uniform sampling assumption. For instance, sampling a length l I N w.r.t. any reasonable (i.e., slower than exponentially decreasing) distribution and then q of length l uniformly leads to the same conclusion.
How reasonable is the UToE? We have already discussed that it is nearly but not quite as good as any other correct ToE. The philosophical, albeit not practical advantage of UToE is that it is a safer bet, since we can never be sure about the future correctness of a more specific ToE. An a priori argument in favor of UToE is as follows: What is the best candidate for a ToE before, i.e., in absence of any observations? If somebody (but how and who?) would tell us that the universe is computable but nothing else, universal self-sampling seems like a reasonable a priori UToE.
Comparison to anthropic self-sampling. Our universal self-sampling assumption is related to anthropic self-sampling [18] but crucially different. The anthropic self-sampling assumption states that a priori you are equally likely any of the (human) observers in our universe. First, we sample from any universe and any location (living or dead) in the multiverse and not only among human (or reasonably intelligent) observers. Second, we have no problem of what counts as a reasonable (human) observer. Third, our principle is completely formal.
Nevertheless the principles are related since (see inclusion of s) given o 1 : t t r u e we also sample from the set of reasonable observers, since o 1 : t t r u e includes snapshots of other (human) observers.
No Free Lunch (NFL) myth. Wolpert [23] considers algorithms for finding the minimum of a function, and compares their average performance. The simplest performance measure is the number of function evaluations needed to find the global minimum. The average is taken uniformly over the set of all functions from and to some fixed finite domain. Since sampling uniformly leads with (very) high probability to a totally random function (white noise), it is clear that on average no optimization algorithm can perform better than exhaustive search, and no reasonable algorithm (that is one that probes every function argument at most once) performs worse. That is, all reasonable optimization algorithms are equally bad on average. This is the essence of Wolpert’s NFL theorem and all variations thereof I am aware of, including the ones for less uniform distributions.
While NFL theorems are nice observations, they are obviously irrelevant, since we usually do not care about the maximum of white noise functions, but functions that appear in real-world problems. Despite NFL being the holy grail in some research communities, the NFL myth has little to no practical implication [24].
An analogue of NFL for prediction would be as follows: Let u 1 : n { 0 , 1 } n be uniformly sampled, i.e., the probability of u 1 : n is λ ( u 1 : n ) = 2 n . Given u 1 : t t r u e we want to predict u t + 1 : n . Let u t + 1 : n p be any deterministic prediction. It is clear that all deterministic predictors p are on average equally bad (w.r.t. symmetric performance measures) in predicting uniform noise ( λ ( u t + 1 : n p | u 1 : t t r u e ) = 2 ( n t ) ) .
How does this compare to the positive result under universal self-sampling? There we also used a uniform distribution, but over effective models=theories=programs. A priori we assumed all programs to be equally likely, but the resulting universe distribution is far from uniform. Phrased differently, we piped uniform noise (via M, see below) through a universal Turing machine. We assume a universal distribution M, rather than a uniform distribution λ.
Just assuming that the world has any effective structure breaks NFL down, and makes Ockham’s razor work [25]. The assumption that the world has some structure is as safe as (or I think even weaker than) the assumption that e.g. classical logic is good for reasoning about the world (and the latter one has to assume to make science meaningful).
Some technical details * . Readers not familiar with Algorithmic Information Theory might want to skip this paragraph. P ( u ) in (6) tends for L to Solomonoff’s a priori distribution M ( u ) . In the definition of M [26] only programs of length = L , rather than L are considered, but since lim L 1 L l = 1 L a l = lim L a L if the latter exists, they are equivalent. Modern definitions involve a 2 l ( q ) -weighted sum of prefix programs, which is also equivalent [6]. Finally, M ( u ) is also equal to the probability that a universal monotone Turing machine with uniform random noise on the input tape outputs a string starting with u [20]. Further, l Length ( q m i n ) = K m ( u ) is the monotone complexity of u : = u 1 : t t r u e . It is a deep result in Algorithmic Information Theory that K m ( u ) log 2 M ( u ) . For most u equality holds within an additive constant, but for some u only within logarithmic accuracy [6]. Taking the ratio of M ( u ) 2 K m ( u ) for u = u 1 : t t r u e u t + 1 : n and u = u 1 : t t r u e yields (6).
The argument/result is not only technical but also subtle: Not only are there 2 L l programs equivalent to q m i n but there are also “nearly” 2 L l programs that lead to totally different predictions. Luckily they don’t harm probabilistic predictions based on P, and seldomly affect deterministic predictions based on q m i n in practice but can do so in theory [27]. One can avoid this problem by augmenting Ockham’s razor with Epicurus principle of multiple explanations, taking all theories consistent with the observations but weigh them according to their length. See [6,20] for details.

9. Discussion

Summary. I have demonstrated that a theory that perfectly describes our universe or multiverse, rather than being a Theory of Everything (ToE), might also be a theory of nothing. I have shown that a predictively meaningful theory can be obtained if the theory is augmented by the localization of the observer. This resulted in a truly Complete Theory of Everything (CToE), which consists of a conventional (objective) ToE plus a (subjective) observer process. Ockham’s razor quantified in terms of code-length minimization has been invoked to select the “best” theory (UCToE).
Assumptions. The construction of the subjective complete theory of everything rested on the following assumptions: ( i ) The observers’ experience of the world consists of a single temporal binary sequence o 1 : t t r u e . All other physical and epistemological concepts are derived. ( i i ) There exists an objective world independent of any particular observer in it. ( i i i ) The world is computable, i.e., there exists an algorithm (a finite binary string) which when executed outputs the whole space-time universe. This assumption implicitly assumes (i.e., implies) that temporally stable binary strings exist. ( i v ) The observer is a computable process within the objective world. ( v ) The algorithms for universe and observer are chosen at random, which I called universal self-sampling assumption.
Implications. As demonstrated, under these assumptions, the scientific quest for a theory of everything can be formalized. As a side result, this allows to separate objective knowledge q from subjective knowledge s. One might even try to argue that if q for the best ( q , s ) pair is non-trivial, this is evidence for the existence of an objective reality. Another side result is that there is no hard distinction between a universe and a multiverse; the difference is qualitative and semantic. Last but not least, another implication is the validity of Ockham’s razor.
Conclusion. Respectable researchers, including Nobel Laureates, have dismissed and embraced each single model of the world mentioned in Section 2, at different times in history and concurrently. (Excluding All-a-Carte ToEs which I haven’t seen discussed before.) As I have shown, Universal ToE is the sanity critical point.
The most popular (pseudo) justifications of which theories are (in)sane have been references to the dogmatic Bible, Popper’s limited falsifiability principle, and wrong applications of Ockham’s razor. This paper contained a more serious treatment of world model selection. I introduced and discussed the usefulness of a theory in terms of predictive power based on model and observer localization complexity.

List of Notation

G,H,E,P,S,C,M,U,R,A,...   specific models/theories defined in Section 2
T { G,H,E,P,S,C,M,U,R,A,...} theory/model
ToETheoy of Everything (in any sense)
ToE candidatea theory that might be a partial or perfect or wrong ToE
UToEUniversal ToE
CToEComplete ToE (i+e+l+n+o)
UCToEUniversal Complete ToE
theorymodel which can explain≈describe≈predict≈compress observations
universetypically refers to visible/observed universe
multiverseun- or only weakly connected collection of universes
predictive powerprecision and coverage
precisionthe accuracy of a theory
coveragehow many phenomena a theory can explain/predict
predictionrefers to unseen, usually future observations
computability assumption: that our universe is computable
q T { 0 , 1 } * the program that generates the universe modeled by theory T
u q { 0 , 1 } the universe generated by program q: u q = UTM ( q )
UTM Universal Turing Machine
s { 0 , 1 } * observation model/program. Extracts o from u.
o q s { 0 , 1 } Subject’s s observations in universe u q : o q s = UTM ( s , u q )
o 1 : t t r u e True past observations
q ˘ , u ˘ Program and universe of UToE
S ( o | u ) Probability of observing o in universe u
Q ( u ) Probability of universe u (according to some prob. theory T)
P ( o ) Probability of observing o


  1. Hutter, M. Human Knowledge Compression Prize, 2006. Availible on line: (accessed on 20 September 2010).
  2. Stove, D.C. Popper and After: Four Modern Irrationalists; Pergamon Press: Oxford, UK, 1982. [Google Scholar]
  3. Gardner, M. A Skeptical Look at Karl Popper. Skeptical Inquirer 2001, 25, 13–14,72. [Google Scholar]
  4. Wallace, C.S. Statistical and Inductive Inference by Minimum Message Length; Springer: Berlin, German, 2005. [Google Scholar]
  5. Grünwald, P.D. The Minimum Description Length Principle; The MIT Press: Cambridge, MA, USA, 2007. [Google Scholar]
  6. Li, M.; Vitányi, P.M.B. An Introduction to Kolmogorov Complexity and its Applications, 3rd ed.; Springer: Berlin, German, 2008. [Google Scholar]
  7. Hutter, M. On Universal Prediction and Bayesian Confirmation. Theor. Comput. Sci. 2007, 384, 33–48. [Google Scholar] [CrossRef]
  8. Smolin, L. Scientific Alternatives to the Anthropic Principle. 2004; arXiv:hep-th/0407213v3. [Google Scholar]
  9. Schmidhuber, J. Algorithmic Theories of Everything. Report IDSIA-20-00, IDSIA, Manno (Lugano), Switzerland. 2000; arXiv:quant-ph/0011122. [Google Scholar]
  10. Harrison, E. Cosmology: The Science of the Universe, 2nd ed.; Cambridge University Press: Cambridge, UK, 2000. [Google Scholar]
  11. Barrow, J.D.; Davies, P.C.W.; Harper, C.L. Science and Ultimate Reality; Cambridge University Press: Cambridge, UK, 2004. [Google Scholar]
  12. Hutter, M. Instantons in QCD: Theory and Application of the Instanton Liquid Model. PhD thesis, Faculty for Theoretical Physics, LMU Munich, Munich, German, 1996. Translation from the German original. [Google Scholar]
  13. Blumhofer, A.; Hutter, M. Family Structure from Periodic Solutions of an Improved Gap Equation. Nucl. Phys. 1997, B484, 80–96, Missing figures in B494 (1997) 485. [Google Scholar] [CrossRef]
  14. Tegmark, M. Parallel Universes. In Science and Ultimate Reality; Cambridge University Press: Cambridge, UK, 2004; pp. 459–491. [Google Scholar]
  15. Schmidhuber, J. A Computer Scientist’s View of Life, the Universe, and Everything. In Foundations of Computer Science: Potential - Theory - Cognition; Springer: Berlin, German, 1997; Vol. 1337, LNCS; pp. 201–208. [Google Scholar]
  16. Tegmark, M. The Mathematical Universe. Found. Phys. 2008, 38, 101–150. [Google Scholar] [CrossRef]
  17. Barrow, J.; Tipler, F. The Anthropic Cosmological Principle; Oxford Univ. Press: Oxford, UK, 1986. [Google Scholar]
  18. Bostrom, N. Anthropic Bias; Routledge: Oxford, UK, 2002. [Google Scholar]
  19. Alchin, N. Theory of Knowledge, 2nd ed.; John Murray Press: London, UK, 2006. [Google Scholar]
  20. Hutter, M. Universal Artificial Intelligence: Sequential Decisions based on Algorithmic Probability; Springer: Berlin, German, 2005. [Google Scholar]
  21. Gell-Mann, M. The Quark and the Jaguar: Adventures in the Simple and the Complex; W.H. Freeman & Company: New York, NY, USA, 1994. [Google Scholar]
  22. Schmidhuber, J. Hierarchies of Generalized Kolmogorov Complexities and Nonenumerable Universal Measures Computable in the Limit. Int. J. Found. Comput. Sci. 2002, 13, 587–612. [Google Scholar] [CrossRef]
  23. Wolpert, D.H.; Macready, W.G. No Free Lunch Theorems for Optimization. IEEE Trans. Evolut. Comput. 1997, 1, 67–82. [Google Scholar] [CrossRef]
  24. Stork, D. Foundations of Occam’s Razor and Parsimony in Learning. NIPS 2001 Workshop. 2001. Availible online:∼ stork/OccamWorkshop.html (accessed on 20 September 2010).
  25. Schmidhuber, J.; Hutter, M. Universal Learning Algorithms and Optimal Search. NIPS 2001 Workshop. 2002. Availible online: (accessed on 20 September 2010).
  26. Solomonoff, R.J. A Formal Theory of Inductive Inference: Parts 1 and 2. Inf. Contr. 1964, 7, 1–22 and 224–254. [Google Scholar] [CrossRef]
  27. Hutter, M. Sequential Predictions based on Algorithmic Complexity. J. Comput. Syst. Sci. 2006, 72, 95–117. [Google Scholar] [CrossRef]
Algorithms EISSN 1999-4893 Published by MDPI AG, Basel, Switzerland RSS E-Mail Table of Contents Alert
Back to Top