Status of optimal inference methods in physics?

[Fra]

Hello everyone, this is my first post on here.

Many discussions seem to circle around string theory (or not), and the not seems to often be LQG and similar approaches.

But there are also other (more or less similar to the above) approaches. For example those that are based on the principles of optimal inference, various generalised "maximum entropy principles". From the searches I've done it seems the amount of reasearch in this direction is small compared to other directions. Why?

I simply wonder if anyone on here is into this, and knows of any papers or websites that sort of sums up the current state of research in this direction?

This is a hobby for me and I have taken a 10 year break from physics contemplation and are now resuming the problems.

I've seen some of Ariel Caticha's papers and various scattered papers relating to the concepts. Enough to see that there are a few people onto this path, but from what I've found it seems not much progress has been made?? Or has it?

There are plenty of fundamental issues that I expect this track to be able to shed light on.

Deriving the laws of physics as the dynamics of information updates. For example general relativity like Ariel Caticha seemed to be after. For example, did anyone complete that aim so far?

/Fredrik

 

[marcus]

... various generalised "maximum entropy principles".

Welcome Fra!

the connection I make with these words is perhaps only random and inappropriate. it is a kind of "minimum symmetry principle" that one can get from Julian Barbour's ideas.

I will get a link and a quote and see if we can bridge the gap between what I have in mind and what you are talking about

Here:
http://arxiv.org/abs/hep-th/0507235
what I refer to in this article is NOT loop quantum gravity but some philosophical reflections where Smolin is channeling Barbour's ideas, and some joint work he did with Barbour.
I will get a page reference, so you can see the part I mean. Page 33, for instance.

 

[Fra]

Hello Marcus! Your signature coincides with someone else from another forum, you don't happen to be elsewhere too? :) I picked Fra because Fredrik was already taken.

I am grateful to any links you have. I'll check it and see if we are talking about the same things.

The principles I talk about is a kind of optimal inference, in order to make progress on admittedly incomplete data. Sort of like the scientific method abstracted, but the idea is that this can be extened to the laws of physics themselves, so that physical interactions can be interpreted in this way. The concept is relational from start to end, which is a feature. But there are issues with it... and from what I know nonone solved them yet.

/Fredrik

 

[marcus]

Hi Fra,
we had a short thread discussing this relatism business earlier
https://www.physicsforums.com/showthread.php?t=83885
several people here contributed good ideas to it, I think

BTW about names the best Fredrik in my estimation was the holyromanemperor Fred II
born around 1195 if I remember right and king of Sicily and Germany
he had a lot of good ideas and also a giraffe.

 

[Fra]

I'll check that thread later and comment.

Thanks!

/Fredrik

 

[Fra]

I quickly went through Smolin's paper and he asks good questions.

What strikes me during reading is the obsessive narrowing of the discussion to ordinary space and time only. But then during the section 6 he asks what I was waiting for

"I would now like to broaden the discussion by asking: Does the relational view have implications broader than the nature of space and time? I will argue that it does."

I couldn't agree more. I'd even extend it to touch realism, and I am looking for a information theoretic model, in which I think "physical evolution and dynamics" and learning(AI) can be unified. But he doesn't mention much of any specific optimal inference methods although they IMO address many of his posed considerations. This makes me think either it's not very popular or controversial, or has been perceived to harbour too much problems?

I have not found any consistent summary yet myslef, but to give a hint of direction read this

"Towards a Statistical Geometrodynamics"
-- http://arxiv.org/abs/gr-qc/0301061
Ariel Caticha

I am still rethinking this myself and there may be several issues in that paper I take issue with, but that's beyond this early point. I think he is in any case addressing the right questions and heading an interesting way. Also the elaborations is basic, because it seems nonone yet has taken this ideas very far. That's why I was curious if I missed the obvious place to look.

Ariel is also very much constrained to space. But I suggest taking it another step and just call it data. I think there will show to be a way to allow the dimensionality to evolve from the data, just given the learning rules. Thus the answer to "why" is that, that's simply the optimally inferred structure from the data fed to us. There is not why beyond that.

/Fredrik

 

[Fra]

Marcus, I see from your posts that you seem to be into reporting scientific news and stuff. What is your view of the mentioned approach? Have you seen much along these lines? Basically "what's wrong with it" from the point of view of the bulk of the research community so to speak?

/Fredrik

 

[Alamino]

I made my Ph.D. with Ariel's brother Nestor and I know a little about his papers. He also is writing an interesting book on Information Physics. When I first read Smolin's paper it reminded me at once Ariel's paper.

Ariel's approach is classical and I don't know how it should be quantized. It is based on the idea that the distance in spacetime is originated by the distance in a manifold of probability distributions. However, as far as I know, the idea has not been developed beyond Caticha's paper.

 

[Fra]

Thanks for responding. Do you know if Ariel is still working on it? I think that the approach should be taken further. I very much like his mindset as read out from his papers. His papers are admittedly far from a unification, but no matter what the status is he should be encouraged to keep up the work.

I think the ideas can be combined with a unification and learning approach, this is what I currenly have in sight. Which would contained an optimally infered dynamic of the model itself. I think the quantization can be introduced by a generalization of an ensemble approach, which includes also the kinematics. I think this will have the "quantization" built into it from scratch, and it can be viewed as a generalization of statistical mechanics. I think it can be given a more abstract but neverthelss more natural probabilistic interpretation. I have some intuitive ideas howto devise canonical ensembles for generalized eventspaces. It would take the take inference philosophy to another level. But there is a lot of work do be done, so I was very happy to see that more share the visions.

I definitely see potential in Ariels ideas beyond classical regime. But at this point most of it is informal and a lot of intuition.

Any idea when his book will come, and where it will be published?

/Fredrik

 

[Alamino]

I have a draft of the book, which is still incomplete. I can ask his brother if he has an updated version and send to you by email.

I share with you (and Nestor and Ariel as well) the feeling that this information/inference approach can lead to a unification path somehow. Nestor is a big supporter of that idea and I believe that he convinced me about that as well.

But I think that a better understanding of fundamental information concepts are still lacking in physics. All we have is a confusing framework separated in classical and quantum information. If you read careful, you will see that 'information' has a very different meaning in different works. For example, in the 'blackhole information loss' paradox, the concept of loss of information is not exactly the same as in information theory. And there are a lot of other issues that would ask for a better understanding and formulation.

Roberto.

 

[Fra]

I have a draft of the book, which is still incomplete. I can ask his brother if he has an updated version and send to you by email.

That would be awfully nice. I would be very interested to read it in detail! I never tried it but I think you can email be through the webforum, I have accepted emails from forum users. It's a hotmail address I use online to not get spam, if you have problems email me and I can respond to you with my ISP email.

If I have any specific comments that I think could possibly be constructive I'll definitely feed that back you/Ariel when I read it.

But I think that a better understanding of fundamental information concepts are still lacking in physics. All we have is a confusing framework separated in classical and quantum information. If you read careful, you will see that 'information' has a very different meaning in different works. For example, in the 'blackhole information loss' paradox, the concept of loss of information is not exactly the same as in information theory. And there are a lot of other issues that would ask for a better understanding and formulation.

Indeed, I agree that the proper formalism is still lacking. But never stopped us before. We will need to invent the appropriate relational "information formalism" along the way. Clearly information is relative as well, and the probabilistic treatment should be relational, and can't single out a preferred prior like is done in some treatments. I think we're on the same page there.

I'll be happy to receive an email from you.

/Fredrik