The Wolfram Model & Wolfram Physics Project

  • #26
721
507
Sorry, but this is not correct. I am a mathematician working in dynamical systems theory, classified here (2010) and here (2020) by the AMS. Suggesting that "experimental mathematics" and "dynamical systems" are synonyms is wrong.

"Experimental mathematics" can suggest directions of research in dynamical systems theory (and so can many other fields in mathematics and the sciences), and dynamical systems theory can suggest new mathematical experiments (analogous remarks apply), but you cannot equate them.
Pardon me, they aren't synonymous and I am not equating them, but instead relating them in exactly the way that you say: huge swats of experimental mathematics are literally completely part of the basic conventional methodology within dynamical systems theory (analogous to calculus being completely part of the basic conventional methodology in physics). In fact, much of experimental mathematics methodology used in dynamical systems theory tends to be seen as so pedestrian to dynamical systems researchers that they themselves don't even bother to give these methods a name, much less refer to them as 'experimental mathematics'.

It would of course be more correct to regard experimental mathematics as a subfield within mathematics itself, but that completely misses the main point I am trying to make, namely that large parts of experimental mathematics (e.g. doing computational experiments, applied bifurcation theory, stability analysis etc) are in fact conventional methodologies within dynamical systems research in actual practice, whether or not dynamical systems researchers use or recognize the terminology 'experimental mathematics'.

In other words, if a physicist wants to know how to use experimental mathematics in his own research it is productive (or at least, it was in my case) to talk to a dynamical systems theorist or to peruse the dynamical systems theory literature, instead of talking to a (pure) mathematician or perusing the mathematics literature which in my experience is quite counterproductive because many mathematicians don't even seem to recognize dynamical systems theory as proper mathematics.
 
  • #27
721
507
I'm sorry, but Auto-Didact you are putting a lot of words in Wolfram's mouth. I feel what you are saying is unrecognizable to the actual content of what he wrote.

It would be easier to discuss these things if you clearly separated ideas this inspires for you (where you think this will grow as a field, what it may accomplish, and other speculations) from what Wolfram is actually saying. It is already hard enough to discuss because Wolfram is still fleshing out his ideas. Please do not mix in another layer of speculation on top of this.
I can give specific references if you want, where is the exact confusion?
 
  • #28
16
4
Mitchel Porter already brought up an explicit example.

Furthermore, you are expanding the hype beyond what even Wolfram is claiming, with your added layer of abstraction to a black box that can automatically unify any mathematical research we shove in the box
Even stronger, you can even use it to automatically find mathematical frameworks which naturally unify (parts of) different research programs, such as all possible unified mathematical frameworks that underly both string theory and loop quantum gravity; the possibilities seem to be practically endless.
First, it is incredibly difficult to figure out how to extract physical meaning from the hypergraphs. This is not something that just falls out of a black box, or occurs by iterating some machine learning program trying to maximize some fitness function. Extracting meaning from the hypergraphs is something that comes from interpretation, figured out by hard work of a human, and placed on there by a human.

Second, if anything, the "input" is an initial condition and a graph evolution rule. He then is looking at what happens and trying to extract meaning. With luck he can extract QM or something. This is the opposite of how you are trying to present it.

Third, let me remind you that Wolfram is claiming his series of ideas for how to interpret the graph evolution make SR and even GR come out inevitably. This means you cannot have Newtonian gravity with its perfect Kepler orbits (no GR corrections), or Newtonian mechanics with Galilean symmetry (because his ideas _require_ a finite information speed limit), or many other mathematically consistent possibilities. What you are hyping doesn't seem to line up with what he is suggesting at all.

Anyway, my point was to be careful to separate your speculations from Wolfram's.
 
  • #29
721
507
HBrown said:
Mitchel Porter already brought up an explicit example.
More specific, do you mean the connection with algebraic geometry or computational complexity theory (or something else)?
HBrown said:
Furthermore, you are expanding the hype beyond what even Wolfram is claiming, with your added layer of abstraction to a black box that can automatically unify any mathematical research we shove in the box
That isn't exactly my intention; my intention is to make Wolfram's claims as bluntly as possible, instead of dancing around the claims as he does by cryptically hiding the main points in hundreds of pages, walls of blogtexts and hours of video (just look at his blogs, youtube posts, books and so on).
HBrown said:
First, it is incredibly difficult to figure out how to extract physical meaning from the hypergraphs. This is not something that just falls out of a black box, or occurs by iterating some machine learning program trying to maximize some fitness function. Extracting meaning from the hypergraphs is something that comes from interpretation, figured out by hard work of a human, and placed on there by a human.
Agreed, doing this requires serious experience and familiarity with applied discrete mathematics as well as knowledge how it relates to fields such as dynamical systems theory and algebraic geometry; it just so happens that this issue was exactly one of the core issues of the main research topic of a research group that I am leading. If I am speaking in a too as-a-matter-of-fact sort of way that is a consequence of my intimate familiarity with the material in question.
HBrown said:
Second, if anything, the "input" is an initial condition and a graph evolution rule. He then is looking at what happens and trying to extract meaning. With luck he can extract QM or something. This is the opposite of how you are trying to present it.
His input isn't merely an initial condition of some system like in calculus/physics, it is something of a completely different nature; this is de facto the key point that he is dancing around by shrouding it in unnecessary mystery.
HBrown said:
Third, let me remind you that Wolfram is claiming his series of ideas for how to interpret the graph evolution make SR and even GR come out inevitably. This means you cannot have Newtonian gravity with its perfect Kepler orbits (no GR corrections), or Newtonian mechanics with Galilean symmetry (because his ideas _require_ a finite information limit), or many other mathematically consistent possibilities. What you are hyping doesn't seem to line up with what he is suggesting at all.
He is being severely careful in his wording, which leads to the length of the manuscript of 400pp what can be easily stated in about 20pp. The point is though that many different limiting models of deeper theories can be identified by a combination of statistical analysis over the limits by varying the input models and then identifying the unique correct derivation through the method of experimental mathematics e.g. using Mathematica; this strategy itself is a completely conventional method in machine learning.
HBrown said:
Anyway, my point was to be careful to separate your speculations from Wolfram's.
Fair point.
 
  • #30
Demystifier
Science Advisor
Insights Author
Gold Member
10,908
3,600
That of course is true. They do have value. But no one would say that they had a break through in say elliptic differential equations if they couldn't prove a single statement about these equations. But it seems like that in the case of Wolfram.
Did Wolfram said that his approach solves a specific problem in physics? Maybe I missed something, but it seems to me that he only claims to have a general framework for a "theory of everything".
 
  • Like
Likes Auto-Didact
  • #31
S.G. Janssens
Science Advisor
Education Advisor
894
664
Pardon me, they aren't synonymous and I am not equating them
You actually were, in what I quoted in my post #23.
namely that large parts of experimental mathematics (e.g. doing computational experiments, applied bifurcation theory, stability analysis etc) are in fact conventional methodologies within dynamical systems research in actual practice, whether or not dynamical systems researchers use or recognize the terminology 'experimental mathematics'.
Applied bifurcation theory and stability analysis may or may not involve experimental mathematics, but they are not part of it.
In other words, if a physicist wants to know how to use experimental mathematics in his own research it is productive (or at least, it was in my case) to talk to a dynamical systems theorist or to peruse the dynamical systems theory literature, instead of talking to a (pure) mathematician or perusing the mathematics literature which in my experience is quite counterproductive because many mathematicians don't even seem to recognize dynamical systems theory as proper mathematics.
Dynamical systems theory is a proper branch of pure and applied mathematics. I do not know of any reputable mathematician (in dynamical systems or a different field) that would seriously argue otherwise. (The MSC by the AMS is the subject classification used throughout mathematics, e.g. in any journal of a pure or applied nature.)

Good luck pursuing your interests.
 
  • #32
195
16
Sorry, but this is not correct. I am a mathematician working in dynamical systems theory, classified here (2010) and here (2020) by the AMS. Suggesting that "experimental mathematics" and "dynamical systems" are synonyms is wrong.
Is the Wolfram approach anything like Category Theory but applied to physics? My understanding is that Category Theory uses digraphs to create abstractions of underlying mathematical proceses whereas Wolfram is using hypergraphs to create abstractions of underlying physical processes.
 
Last edited:
  • #33
1,184
349
This is precisely Wolfram's biggest problem: he doesn't say that he is using evolutionary algorithms and/or machine learning, while he is.
So can you show somewhere that he is using these methods, even if he doesn't call them by these names?
Wolfram's latest work is directly related to a novel subdiscipline in computational complexity theory, namely (algebro-)geometric complexity theory, see the Wikipedia page and arxiv references. Wolfram's approach in this field seems to be completely novel; moreover, geometric complexity theory is related to the still semi-nascent interdisciplinary field called information geometry (i.e. algebro-geometric information theory), which is of course related to thermodynamics and so theoretical physics; there are even more specific problems that can be experimentally approached, e.g. in theoretical biophysics.
I've studied geometric complexity theory. It is the proposed application of algebraic geometry to the separation of complexity classes (separation means, e.g., proving P is distinct from NP). I call it a proposed application because only a handful of separations have been proven this way so far, although there has been a lot of work aimed at eventually dealing with the more difficult cases.

The GCT method involves constructing algebraic-geometric objects that encode complexity classes, and showing that there are obstructions to embedding one such object into another, which shall in turn imply that the corresponding complexity class of the first object cannot be reduced to the complexity class of the second. This method is not on display anywhere in the Wolfram literature. Indeed the only claim so far is a claim of reduction (equivalence), not of separation, namely the claim that P=BQP, and the methods hinted at, from what I can see, are not even geometric, let alone related to the use of representation theory etc as done in GCT per se.

Wolfram aside, I am wondering whether GCT can possibly count as a branch of information geometry. What I understand of information geometry is that involves geometrizing information spaces, e.g. putting a metric on them. I suppose GCT geometrizes and/or algebrizes complexity classes, but there's presently a huge gap between its specific aims and methods, and what anyone else does.
 
  • #34
195
16
After listening to the videos, Wolfram makes claims that the photon may not be massless, and he can explain why anti-matter does not have anti-mass. It will be interesting to see whether his system can make predictions or explain phenomena.
 
  • #35
martinbn
Science Advisor
1,844
558
Did Wolfram said that his approach solves a specific problem in physics? Maybe I missed something, but it seems to me that he only claims to have a general framework for a "theory of everything".
That's what I was asking. If it doesn't then is it physics or metaphysics?
 
  • #36
721
507
You actually were, in what I quoted in my post #23.
I maintain what I said in #23, if it is understood in the same sense as that calculus is 'part of physics', while in actuality it is part of analysis; this isn't something to get too caught up on because using language to communicate necessarily brings with it some unavoidable vagueness.

In fact, any results found by varying parameters of any iterative map in a computer algebra system is experimental mathematics. Wolfram is a pioneer in this field, because he of course has not merely created Mathematica but fully masters it.

As a side note, the key thing I learned in medicine is that if specific vagueness within some context can be made precise, then there is no problem whatsoever apart from the purely subjective feeling of being uncomfortable around vagueness in a setting without context.
Applied bifurcation theory and stability analysis may or may not involve experimental mathematics, but they are not part of it.
You are again strictly correct but missing the point, namely that using a computer algebra system, such as Mathematica, to actually carry out computational analyses in bifurcation theory and stability analysis in the context of dynamical systems research in practice is de facto doing experimental mathematics.

The original discovery of cellular automata by Wolfram, the original discovery of chaos by Lorenz and the discovery of the Mandelbrot set by Mandelbrot were all mathematical results found by computational experiments i.e. were all instances of experimental mathematics.

Almost all dynamical system theory research is done using computers at some point, instead of using pen and paper for carrying out computations; just recall the key work of Feigenbaum, Lorenz, Mandelbrot, Smale et al. In the universities that I work, the actual subject is taught as a branch of applied mathematics, while the actual experimental research is done using mostly Mathematica.

Of course there are also things done with only pen and paper or chalk and blackboard alone but that is usually more on the theory side (often even done by theoretical and/or mathematical physicists) and even then it is typically based on data gathered from computational experiments.
Dynamical systems theory is a proper branch of pure and applied mathematics. I do not know of any reputable mathematician (in dynamical systems or a different field) that would seriously argue otherwise. (The MSC by the AMS is the subject classification used throughout mathematics, e.g. in any journal of a pure or applied nature.)
The mathematicians I am speaking about are certainly reputable but they only tend to argue pejoratively about the field in private, never in public; of course, those that do look down at dynamical systems tend not to actually be very familiar with the field; they of course accept publications and are more critical of the form than of the content.

They are usually just instinctively criticizing much of the non-rigorous style of research as too foreign from what they themselves do or consider as proper mathematics; in fact, they usually 'insult' the field by saying things like "so you see, the fact that experiments play a key role in the work of those researchers proves that what they are doing is actually not really mathematics, but instead physics".
Good luck pursuing your interests.
Thanks. I never realized that you were a dynamical systems theorist, we seem to be a rare breed on the physicsforums. Judging from your name I'm assuming you are from the Netherlands, do you by any chance happen to have ever met either Ruelle or Takens?
Is the Wolfram approach anything like Category Theory but applied to physics? My understanding is that Category Theory uses digraphs to create abstractions of underlying mathematical proceses whereas Wolfram is using hypergraphs to create abstractions of underlying physical processes.
I get the same feeling, but I'm not fully comfortable answering because I am not a category theorist myself. Maybe ask John Baez? He is active on Twitter.
So can you show somewhere that he is using these methods, even if he doesn't call them by these names?
In the introductory pages of Wolfram's 450 page manuscript he explicitly doesn't mention what the nodes are in his model, but leaves it abstract as 'element': w.r.t. the communication to physicists this seems to be done on purpose to maintain some mystery. His evolution rules of his graphs are described in section 2.6 (page 10) and section 2.7; by actually stating all the axioms of his graphs he essentially has given away what this subject is about.

Moreover, see page 72 of his manuscript in which he shows that the graphs that he constructs purely computationally are statistically indistinguishable from the result of a search done on some actual phenomena in the real world using machine learning methods.

As a side note, I'm getting the feeling that most people commenting here and IRL don't seem to be making these connections. Maybe I should write a review instead of trying to address these things on here? Then again I'm already writing a corona management guideline for primary care at the moment... I don't think my wife will appreciate me taking on more work as it is.
I've studied geometric complexity theory. It is the proposed application of algebraic geometry to the separation of complexity classes (separation means, e.g., proving P is distinct from NP). I call it a proposed application because only a handful of separations have been proven this way so far, although there has been a lot of work aimed at eventually dealing with the more difficult cases.

The GCT method involves constructing algebraic-geometric objects that encode complexity classes, and showing that there are obstructions to embedding one such object into another, which shall in turn imply that the corresponding complexity class of the first object cannot be reduced to the complexity class of the second.
To offer my honest perspective as an outside researcher: the field of GCT is brand new; how many active researchers are there realistically speaking? The fact that there are any meaningful results at all seems to me itself quite amazing.

As a side note, seeing you are familiar with GCT: in how far do the obstructions in GCT map onto the cohomological obstructions in Abramsky's work on non-locality?
This method is not on display anywhere in the Wolfram literature. Indeed the only claim so far is a claim of reduction (equivalence), not of separation, namely the claim that P=BQP, and the methods hinted at, from what I can see, are not even geometric, let alone related to the use of representation theory etc as done in GCT per se.
As I said earlier, Wolfram likes to dance around the point instead of getting straight to the point; perhaps he expects others to flesh out his vaguer points through their own research? I mean, he has after all opened this Project to the universities for all others to participate.
Wolfram aside, I am wondering whether GCT can possibly count as a branch of information geometry. What I understand of information geometry is that involves geometrizing information spaces, e.g. putting a metric on them. I suppose GCT geometrizes and/or algebrizes complexity classes, but there's presently a huge gap between its specific aims and methods, and what anyone else does.
The implication I am making is that they use the same methodology for different purposes i.e. the methods have the same form but do not necessarily refer to the same (type of) content. Nevertheless the interesting question naturally arises whether or not parts of these matured methodologies in one field are directly applicable in another field, e.g. as in the case with the cohomological obstructions in Abramsky's work.

With respect to the geometry of information spaces, the applications are way more obvious because of the direct and central roles that information and entropy play in thermodynamics, mathematical statistics (Fisher metric), machine learning and biophysics.
 
Last edited:
  • #37
721
507
That's what I was asking. If it doesn't then is it physics or metaphysics?
Excuse me, but since when is the mathematical study of (variants of) equations from physics in order to generalize them suddenly not physics anymore?

By that logic, when Dirac took the Schrodinger equation and generalized it purely mathematically into the equation which now bears his name, he wasn't doing physics either.
 
  • #38
martinbn
Science Advisor
1,844
558
Excuse me, but since when is the mathematical study of (variants of) equations from physics in order to generalize them suddenly not physics anymore?

By that logic, when Dirac took the Schrodinger equation and generalized it purely mathematically into the equation which now bears his name, he wasn't doing physics either.
Can you give me an example where the black box did this?

Ps writing down equations is easy, finding useful equations is harder and happens rarely, but deriving the consequences of the equations is the hardest part and the essence.
 
  • Like
Likes Auto-Didact
  • #39
721
507
@mitchell porter & @HBrown if I recall correctly, Wolfram talks about the link to computational complexity, machine learning and so on here and here, among other places.
Can you give me an example where the black box did this?
For example, Wolfram claims on his blog that his black box reproduces the vacuum EFE and full EFE.
Ps writing down equations is easy, finding useful equations is harder and happens rarely, but deriving the consequences of the equations is the hardest part and the essence.
With this I agree to a certain extent, because the step of deriving the consequences can often be translated into a straightforward procedure, sometimes even capturable in a flowchart, i.e. inside some definite often pre-defined framework.

Any method that can be described in this manner isn't very impressive, since it only requires hard work of learning the content inside some given framework, instead of both the hard work of learning the content in some framework and the intuition to go beyond the known framework.

Also, by 'finding equations' I of course mean 'finding useful equations', but I would argue that even finding the wrong equations is useful in the broader theoretical context of gaining scientific understanding; for example, starting with the Schrodinger equation and arriving at the Klein-Gordon equation first when one is actually trying to arrive at the Dirac equation.

In fact, if I question a grad student that is attempting to carry out this derivation and he doesn't know about the route to the Klein-Gordon equation, nor does he even recognize the Klein-Gordon equation, I would take away points and doubt his understanding of what he is actually doing, especially if he wants to become a mathematical physicist.
 
  • #40
721
507
Here is a visual summary of the main results of the Wolfram Physics Project:
visual-summary-4k.jpg
 
  • #41
1,184
349
As a side note, seeing you are familiar with GCT: in how far do the obstructions in GCT map onto the cohomological obstructions in Abramsky's work on non-locality?
At the most concrete level, they don't look to be very similar. Abramsky's obstructions seem to be about preventing a kind of foliation, whereas GCT's obstructions prevent an embedding of one algebraic variety into another.
if I recall correctly, Wolfram talks about the link to computational complexity, machine learning and so on here and here, among other places.
In the first link he's only talking about complexity but not computational complexity; but in the second one he does mention computational complexity. He says first of all that in his systems, amount of fundamental computation is anchored to fundamental physics. (I will note in passing that this has its analogues in conventional physics, e.g. the Bekenstein bound or the Landauer limit.) Then he says

"there’ll be an analog of curvature and Einstein’s equations in rulial space too—and it probably corresponds to a geometrization of computational complexity theory and questions like P?=NP."

Also later he enthuses about how "it almost seems like everyone has been right all along" and his framework has "hints of" every major quantum gravity research program, and aligns naturally with numerous modern mathematical ideas - and here he mentions GCT again. But the previous comment is the most substantive.

Here, specifically with respect to computational complexity, I think he's simply being naive. This part does indeed sound like information geometry. But the whole difficulty of computational complexity theory, is not in quantifying properties of the algorithms, it lies in showing that one problem can not be mapped onto another from a putatively different complexity class. If you think of how difficult problems in geometry and topology can be (e.g. Poincare conjecture); that's also how problems like P?=NP look, when expressed geometrically. It seems like they'll need that full armoury of Fields-Medal-level techniques, and beyond. Simply expressing the problem in a particular context (like Wolfram's directed hypergraphs) will not itself be a silver bullet.
 
  • Informative
Likes Auto-Didact
  • #42
721
507
At the most concrete level, they don't look to be very similar. Abramsky's obstructions seem to be about preventing a kind of foliation, whereas GCT's obstructions prevent an embedding of one algebraic variety into another.
Thanks, that saves me alot of reading up on the obstructions in GCT; I must say that the demonstration of prevention of embedding intuitively makes alot of sense as an approach to finding a proof, which also incidentally makes the traditional (more set theoretic) viewpoint of having one class be a subclass of the other seem almost pedestrian in comparison.
Here, specifically with respect to computational complexity, I think he's simply being naive. This part does indeed sound like information geometry. But the whole difficulty of computational complexity theory, is not in quantifying properties of the algorithms, it lies in showing that one problem can not be mapped onto another from a putatively different complexity class.
I never truly got into computational complexity theory precisely because of its non-geometric flavour, but I do know both graph theory and series acceleration from numerical analysis, and also that both of these fields eventually tie into issues from computational complexity theory (at least in so far as the standard texts mention that they do). Moreover, both of these fields tie into CCT in a different fashion, where the series acceleration connection clearly is about quantifying properties of algorithms, while the graph theoretic connection not so much.

To be blunt, the beauty of graph theory is that it naturally contains certain methods to transform an entire topic into a different discipline, e.g. a particular subdiscipline of graph theory can be transformed into set theory and similarly certain particular kinds of graph theoretic methods are de facto really just algebraic topology in disguise. This, or something very similar, is the connection I think that Wolfram is making after having read his manuscript and blogs.

A possibility to do CCT in a geometric fashion seems to be therefore even more interesting, because it opens up the field to alot of researchers who would otherwise most likely just ignore it, e.g. purely because of its traditional non-geometric form; such biases of only being receptive to certain forms of presentation may sound silly, but they seem to play an enormous role in the practice of mathematics, physics and science more generally.
If you think of how difficult problems in geometry and topology can be (e.g. Poincare conjecture); that's also how problems like P?=NP look, when expressed geometrically. It seems like they'll need that full armoury of Fields-Medal-level techniques, and beyond. Simply expressing the problem in a particular context (like Wolfram's directed hypergraphs) will not itself be a silver bullet.
I don't doubt that it requires Field-Medal-level techniques. What I do doubt is whether or not it has been sufficiently creatively approached from all possible mathematical angles that are already available today, instead of only being creatively approached from a filtered audience of mathematicians who don't mind working on non-geometric forms of mathematics; if the pre-filter group has a higher creativity e.g. because of their geometric intuition, then it's no wonder that the problem hasn't been solved yet.
 
Top