Abstractions by patterns

This chapter develops a "patternist" account of abstraction grounded in algorithmic information theory (AIT). A pattern in data is defined as a program that, together with some input, reproduces the data in fewer bits than the data itself. An abstraction is then a set of strings whose members all share such a pattern—i.e., are all compressible by the same program under different inputs. This proposal is motivated by a puzzle about lossy compression: converting mRNA encodings to amino acid sequences seems like genuine abstraction, while applying the same coarse-graining operation to a random bitstring does not. The difference, I argue, is that real abstractions preserve patterns in the underlying data, while arbitrary coarse-grainings need not. The account is then extended to scientific explanation, where a "proper explanation" extracts all regularities from the data (its minimal sufficient statistic), and "partial" explanations provide algorithmic information about this ideal. This framework yields a natural characterization of higher-level or "parochial" explanations that function only at an abstract level—independently of the details of any particular realizer—and points toward a rigorous, information-theoretic account of emergence.