XXV. Discussion

Abstract

I have been asked to make a written contribution to the discussion, but have nothing new to say. My views will be found in the 1961 version of The Earth and in my lecture to the Royal Astronomical Society last October, which will be published in their Quarterly Journal. My main points are that the only type of imperfection of elasticity considered in convection and drift theories is the elastico-viscous law, which has been found to lead to numerous contradictions when confronted with actual evidence. Different phenomena led to values of the effective viscosity differing by factors of millions. On the other hand, a modified law, chosen to fit two quantitative data and applied far beyond the range of periods related to those data, has steered its way nicely among the other evidence for some sort of imperfection of elasticity, without giving any contradiction. But it does forbid convection and continental drift. I should be disposed to agree that inability to explain an alleged phenomenon is not necessarily a disproof of that phenomenon; but it does require a higher standard of scrutiny of the evidence for that phenomenon. The standard actually applied to evidence for continental drift seems to be considerably lower than is usual for a new phenomenon, and is not associated with any alternative explanations of things that can be explained.

A comment on priors for Bayesian occupancy models

“…However, this prior still displays slight bimodality at 0 and 1 and thus has the potential to affect posterior distributions. An exact prior distribution that is completely invariant to transformation, such as between the logit and probability scale, is the Jeffry’s prior (Jeffreys 1946), which can be derived for occupancy models, but also will likely be informative and thus might not be more appropriate than a Normal distribution with small variance. We note that recent publications in statistical ecology describing how to fit occupancy models in a Bayesian framework provide some suggestions for priors that should reduce the concerns we raise here.…”

Fast methods for training Gaussian processes on large datasets

“…As well as maximising, we can also consider marginalising over σ f [16]. As we are marginalising over a scale parameter we use the (improper) Jeffreys prior P (σ f ) = c/σ f [17]. The result is equal to the maximised form, up to a multiplicative constant,…”