1) Scientists have found that maps arising across many fields of science appear to share a common bias toward simplicity in their outputs, even though one might expect nature's maps to reflect a random distribution of complexities.
2) This finding emerges from applying ideas from algorithmic information theory to mathematical mappings. The theory suggests outputs become exponentially less likely as their Kolmogorov complexity increases, resulting in a bias toward simpler outputs.
3) Studies of maps from RNA sequences to structures, ordinary differential equations, and neural networks support this idea, finding the majority of outputs tend to be relatively simple despite many possible complex alternatives.
1) Scientists have found that maps arising across many fields of science appear to share a common bias toward simplicity in their outputs, even though one might expect nature's maps to reflect a random distribution of complexities.
2) This finding emerges from applying ideas from algorithmic information theory to mathematical mappings. The theory suggests outputs become exponentially less likely as their Kolmogorov complexity increases, resulting in a bias toward simpler outputs.
3) Studies of maps from RNA sequences to structures, ordinary differential equations, and neural networks support this idea, finding the majority of outputs tend to be relatively simple despite many possible complex alternatives.
1) Scientists have found that maps arising across many fields of science appear to share a common bias toward simplicity in their outputs, even though one might expect nature's maps to reflect a random distribution of complexities.
2) This finding emerges from applying ideas from algorithmic information theory to mathematical mappings. The theory suggests outputs become exponentially less likely as their Kolmogorov complexity increases, resulting in a bias toward simpler outputs.
3) Studies of maps from RNA sequences to structures, ordinary differential equations, and neural networks support this idea, finding the majority of outputs tend to be relatively simple despite many possible complex alternatives.
A ‘map’ in ordinary usage is a representation The question Dingle and colleagues discrete mapping by discretizing the system of one thing by another, usually simpler address is whether anything similar can be of ordinary differential equations, which thing. Mathematicians define a map more established for maps in general. Any map can has 15 tunable parameters, and solving the formally as a function, f, taking the elements be considered to carry out a computation, resulting numerical model to get outputs as of one set into another. Using today’s and so can be considered in terms of the discrete time trajectories. Again, they found computing terminology, one might also computer code that would be required to that the outputs were strongly biased toward think of inputs and outputs — a map acts on implement it. Following this approach, the simplicity, and the likelihood of outputs of an input to produce an output. Maps of this authors derive an analogue of the classic different complexity could be accurately kind run through all of science, from plasma result discussed above, but applying to any estimated using the output complexity only. physics to linguistics. of a broad class of discrete maps. The result This work may take us part way to Scientists are experts in the maps of asserts that the probability P(x) a randomly understanding why so many aspects of nature their own special areas. Earth scientists chosen input will generate an output x is can be understood on the basis of relatively understand basic models of plate tectonics, bounded by 2–K(x). Again, there’s a natural simple models, and indeed why the world just as biochemists have great familiarity weighting toward simpler outputs. seems understandable at all. It’s a fundamental with models of cell signalling. Yet one might mathematical bias at work. Even so, there are wonder if all these different maps may share many open questions. The authors note that Maps that arise certain general properties, beyond simple a general map could conceivably have any properties such as continuities. Recent work across many distribution of complexities of its outputs. A suggests that the answer is yes — maps that fields of science map might be designed, for example, to have arise across many fields of science appear to appear to share a almost all complex outputs, and very few share a common and rather surprising bias common and simple outputs. In this case, even if simpler toward simplicity in their outputs. That is, rather surprising outputs have a higher relative probability than while we might naively expect nature’s maps bias toward more complex ones, the imbalance toward to reflect a random sample of all possible simplicity in their complexity in the sheer number of outputs maps, they actually tend strongly toward outputs. will win out, leading mostly to complex those with many simple outputs. In this outputs. But in the applications they’ve sense, nature really seems to be simpler examined so far, this doesn’t seem to happen. than it could be. It may seem almost miraculous that Most maps do seem to favour simpler output This idea emerges from an application of one can derive expectations about maps — although exactly why remains mysterious. the ideas of algorithmic information theory in a general setting just by considering The authors also believe the work may help to mathematical mappings (K. Dingle, computational aspects of how maps process to explain some observed features of so-called C. Q. Camargo & A. A. Louis, Nat. Commun. 9, information. But it does seem to work, as deep learning neural networks, particularly 761; 2018). Algorithmic information theory Dingle and colleagues demonstrate in a why they seem to generalize so well and fit aims to quantify the inherent complexity of a number of real world examples. An example data well outside of the examples on which string of digits. In the case of binary digits, for from biophysics is the map connecting RNA they’ve been trained, and despite the over example, a repeating string — 010101. — is nucleotide sequences or genotypes to so-called parametrization of the models — the number inherently simple, whereas a highly erratic RNA secondary structures — the structures of parameters they contain is also much larger string is (in general, though not always) these linear molecules naturally assume than the training data. In a preprint (https:// more complex; it contains more information due to interactions between different bases. arxiv.org/abs/1805.08522), Camargo and and is more difficult to describe. A key idea Dingle and colleagues examined the case Louis, working with Guillermo Valle-Pérez, in the field is the Kolmogorov complexity n =55 — RNAs of length 55 — and first used suggest this might be another example of the K(x) of a binary string x, defined (leaving a popular software package to calculate the bias toward simpler outputs. Deep learning aside some technicalities) as the length of minimum free energy structure for a large models may generalize well because they’re the shortest programme able to generate x number of random sequences. They then biased in the same way that data from real when run on a universal Turing machine. turned to a commonly employed shorthand world problems are — toward simple outputs. This definition centres on a mapping. for describing the secondary structures to turn The work of Dingle, Camargo and Louis It goes from an output string x to another each structure into a binary sequence. Finally, reminds me of the related observation by binary string giving the length of the by estimating the complexity of each resulting Mark Transtrum and colleagues (preprint shortest programme generating x. One output, they found that the results clearly at https://arxiv.org/abs/1501.07668) interesting question is the following: what is showed a bias toward simpler sequences. concerning the surprising effectiveness of the probability P(x) that a randomly chosen A more general example is ordinary 'sloppy models' in so many areas of science. input will generate x? Mathematicians have differential equations, which yield discrete Why is the world understandable? We may developed estimates for this quantity, finding models when coarse-grained by discretizing never know for sure. But the mystery may be that it goes crudely in proportion to 2–K(x). input parameters and outputs, as happens starting to shed its secrets. ❐ Hence, output strings become exponentially in applications throughout science. Here the less likely with increasing Kolmogorov authors tested a representative model drawn Mark Buchanan complexity. This is a general result for from biochemisty and used in modelling one particular mapping of importance in the biochemical pathways controlling Published online: 4 December 2018 algorithmic information theory. circadian rhythms. They first generated a https://doi.org/10.1038/s41567-018-0370-y 1154 Nature Physics | VOL 14 | DECEMBER 2018 | 1154 | www.nature.com/naturephysics