[andrewflnr]

  “But then if you mathematically factor in the fact that
  Verlinde’s prediction doesn’t have any free parameters,
  whereas the dark matter prediction does, then you find
  Verlinde’s model is actually performing slightly better.”

What? How does this work?

[fjarlq]

From the discussion section of Verlinde's paper[1]:

> We have shown that the emergent laws of gravity, when one takes into account the volume law contribution to the entropy, start to deviate from the familiar gravitational laws precisely in those situations where the observations tell us they do. We have only made use of the natural constants of nature, and provided reasonably straightforward arguments and calculations to derive the scales and the behavior of the observed phenomena. [..]

> In our view this undercuts the common assumption that the laws of gravity should stay as they are, and hence it removes the rationale of the dark matter hypothesis. Once there is a conceptual reason for a new phase of the gravitational force, which is governed by different laws, and this is combined with a confirmation of its quantitative behavior, the weight of the evidence tips in the other direction.

[1]: https://arxiv.org/abs/1611.02269

[andrewflnr]

That's just Occam's razor, though. The phrasing made it sound like they had a quantitative argument for why fewer free parameters made it "perform" better.

[krastanov]

Say that you want to estimate what is the chance that observation 1 fits theory A just by accident (the lower this chance is the higher your confidence can be that theory A is "real").

If theory A has parameters you can adjust, it becomes easier for an observation to fit the theory just by accident (because you can adjust the free parameter).

This is a fairly flawed description, but it should be a good enough ELI5.

[imh]

When you have free parameters, you are actually opening the door to a whole family of predictions. You're saying that of this entire family of models, you'll choose the one that best fits the data. Without free parameters, you have to come in with a fully specified model to begin with.

Loosely speaking, with an incorrect family of models, there's a lot of wiggle room to find one that still looks quite a lot like your data, while that freedom is not there when making a single specific prediction. There are semi-rigorous ways to take that into account while comparing models https://methodology.psu.edu/node/504

[gabrielgoh]

principle of parsimony