Jorma Rissanen's Festschrift
- (P15) "In my view statistics needs a solid foundation rather than just a collection of isolated techniques however clever. It is not enough to claim to have found a method which works well or better than other methods on some data. It’s even not enough to prove that the technique works on data generated by an imagined ’true’ distribution. We need to understand why a technique works as it does, and why it is better than the competing approaches. This is what’s missing in current statistics, where all sorts of criteria for the model selection problem have been proposed. …"
- (P16) Jorma claims traditional statistics is unable to formalize concepts needed to capture properties in complex data, and that there is more in data than what traditional statistical theory can explain.
- (P16) Minimum Description Length (MDL) theory gives an approach to statistics free of the untenable assumption of a true data generating distribution. In MDL theory, the selection of priors is also restricted so that they are encodable [#DOUBT Doesn't this mean low complexity priors?], meanwhile in the Bayesian philosophy nothing prevents someone from putting all (prior) mass to the data. The irresistable desire to peek into the data has created concepts like empirical priors, which contradict the foundations of Bayesianism.
- (P16) Concepts like universal models, noise, statistical information, complexity which are central to MDL theory have no meaning in Bayesian philosophy [#DOUBT What about the relation between complexity and the marginal likelihood].
- (P17) There is a mention of a "chain link method" that can encode closed curves on a plane and assign a probability to that curve. [#TODO Find an actual reference to this].