Can putting information systems through a "functional bottleneck" to approximate their Algorithmic Information reduce their attack and argument surfaces substantially?

Software surface may be defined in terms of its complexity, both as attack surface and as argument surface. While cyber attack surface is well known in the industry, and is a key security metric, argument surface is term coined by Nick Szabo in the context of information system governance, e.g. smart contracts, blockchain software, etc., including those aspects that involve humans as well as software. Such arguments involve the software process itself (e.g. discussions regarding Bitcoin algorithm trustworthiness), such as consensus about requirements. Operationalizing these requirements eventuates in Q/A acceptance test suites as part of quality assurance.

A reasonable argument may be made that a minimum complexity software system meeting all operationalized requirements will have not only a minimum attack surface, but will by definition, result in a minimum complexity description of the processes required to fulfill the operationally defined requirements.

An analogy may be drawn here with the formalization of Occam's Razor as an approximation of the Algorithmic Information of a given dataset. Under Solomonoff's theory of inductive inference, the minimum length algorithm that outputs exactly the given dataset is most likely to generalize to new data generated by the same process. By analogy, the minimum length program capable, independent computational complexity (ie: ignoring execution time*), of meeting all functional requirements may be most likely to generalize to new functional requirements. Here "functional" is in the sense of functional programming or mathematical functions, as opposed to the real world response time requirements as "function".

The connection here to the philosophy of science, as in Solomonoff's proof or as in computer science, seems clear enough.

What is not so clear is the degree to which putting whole software systems through a bottleneck stage (that objectively measures its complexity) may help discipline computer science in accord with Solomonoff's proof of inductive inference based on Algorithmic Information: The approximation thereof in lossless compression using Turing complete codes (in contrast to mere "universal" codes of MDL which are not Turing complete hence cannot even in theory be used to describe dynamical systems, such as the real world).

Although choice of UTM is left unspecified by Algorithmic Information Theory, it is no less constrained than is any scientific model constrained to ground itself in the axioms of arithmetic so as to compute testable predictions. For example, one may not alter the axioms of arithmetic to include empirical observations in order to get around Occam's Razor -- nor (equivalently) examine the very observations one is to predict in the process of formulating a theory to predict those observations. And, yes, this blatant sophistry is routinely used as an "argument" against the "argument surface" of Algorithmic Information Theory's applicability. We've all heard it countless times and even from Turing Awardees in artificial intelligence no less!

* One may imagine various techniques to address other pragmatic issues -- not the least of which are program "pragmas" that might be included in the original source, but ignored in creating the Algorithmic Information approximation. Such "pragmas" generalize to any directives given to the compile/build and deployment processs to expand the source to meet time requirements.



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.