Ken Wharton

You mean, interventionist causation in general? That's a big blind spot for many physicists, sadly. We all "know" about it at a gut level, like when we tell students that external forces *cause* accelerations, rather than the other way around. But many physicists probably couldn't articulate why.

If there is a law-like restriction instead of an uncertainty principle, that’s like taking the retrocausal model and reinterpreting the hidden variables as a gauge, where the gauge itself seems to prevent signaling faster than light. But that would defeat the purpose of the paper, wouldn’t it?

The first model (2 posts back) is forward-causal, but with a strange restriction on what I’m allowed to do in the future. The last model is retrocausal, but with an epistemic restriction on the initial state, tuned just right to prevent me from being able to send a known signal faster than light.

If a different model allows those future experiments, it must then it must in turn forbid my complete causal control of the initial state. Maybe, as you say, with some law-like restriction on the initial state itself, or maybe with some uncertainty principle, forbidding paradox-sufficient access.

If I had control of a system that could signal faster than light, and Lorentz transformations are correct when boosting frames, then it’s pretty clear what sort of experiment I could set up to make a paradox. But if a model forbids my causal control of those future experiments, no more paradoxes.

Yes, there are lots of different causal models which correspond to the very same equations. (Most obvious example here is the Ideal Gas Law; depending on which variables you’re allowed to control, you get very different casual pathways.) Different intervention freedoms = different models.

You can't see this problem if all you look at is 'existence and uniqueness'. Instead, you have to ask questions about the input-output structure of the model -- and that often requires more analysis than just writing down the bare equations.

In this case, with a restriction on (S) due to later events, the issue isn’t *forward* causation; it’s *backward*! If I can set something in the future that would constrain allowed past values of (S), that’s retrocausal. And without restricting access to S, such a model could signal back in time.

In many models (1) and (2) are the same thing, because the inputs are assumed to be the initial state. But in some models, like the one in this paper, this isn’t the case. If there are rules telling you that the initial state (S) can’t be freely set, from outside the model, then (S) isn’t an input.

This brings us to question (1), which is where the true causal analysis lies. But you can’t answer this from the bare equations; the model needs to specify the causal structure. What are the model’s “inputs”? What events are we allowed to “set”, from outside the model, independently?

After all, correlation is not causation. Our causal instincts are “interventionist”. We ask ourselves, if I set event A to this value, instead of that (counterfactual) value, is there an effect at B? We’re not just asking about correlations, we’re asking about input-output relationships.

The only nice thing about (2) is that you can analyze it in terms of the “bare equations”, without making any causal assumptions. Given the equations, the answer to (2) is easy to assess -- it’s just a question of which events are correlated. But does it have anything to do with “causation”?

Fun stuff! Still, it’s important not to conflate these two questions, for any given model: (1) Are there inputs to the model at event A which have an effect at event B? vs. (2) Is there an initial event A that is correlated with some other event B? These questions come apart in cases like this.

So my conclusion is almost opposite theirs. Instead of banning signaling-in-principle to rule out "all-at-once" models, it just means that any good all-at-once model must be able to generate its own “uncertainty principle” limit, to forbid any signaling-in-practice. A challenge, not a no-go theorem.

And if “signaling-in-principle” was only possible beyond the model’s own restriction, then there’s nothing amiss. In fact, if one thinks there is a realistic way to explain Bell inequality violations, then of course there *would* be signaling if you could see all the hidden variables. Yet you can't.

I’ll have to think about it some more, but my gut tells me that while such arguments might work in some cases, I don’t see how you could generalize it. All-at-once models would generically lead to restrictions on what sort of states could be prepared, each with their own ‘uncertainty principle’.

Their analysis of Schulman raises a very interesting question: Can you logically combine an all-at-once model (where the probabilities are assigned to entire histories) with conventional preparations (where the probabilities are assigned to initial instantaneous states)? How would that work?

Their real target here isn’t Bohm, but rather the models I favor: Models that break measurement independence by allowing the future setting to constrain the hidden past, solving everything “all at once”. And the framing they use here is the trusty Schulman model! I'm very happy to see this… 😀

Their discussion of Bohmian mechanics doesn’t make this point, but instead notes the possibility of a non-equilibrium initial distribution. Okay, maybe that’s fair, but only because (in that particular model) such a distribution seems (barely) possible. But what about a different model?

But I’m not at all sure about the jump from “signaling in principle” to “testable in principle”. For instance, in Bohmian mechanics, you could signal at a spacelike distance if Alice and Bob could examine their states at a level below what the uncertainty principle allows. But that’s not testable.

Interesting new paper by Guido Bacciagaluppi and colleagues, raising some deep issues with Bell’s Theorem. On one level I think they’re right that there’s a connection between realistic violations of Bell inequalities and “nonlocal signaling in principle”. They seem to go together.

He was there, for sure, but I don't think anyone based at Chapman gave a talk.

YouTube video by Institute for Quantum Studies john templeton foundation at chapman iqs conference day 1 720p

The talks from a recent quantum foundations conference at Chapman U. are now online... Each day is its own 5-hour video, I guess. My talk, "A realistic alternative to the wavefunction", happened to be the first on day 1 (skip to the 10 minute mark). www.youtube.com/watch?v=FPwj...

"Is it wrong to wish on space hardware?"

(Billy Bragg really wants to know...)

What did you think?

Bold text reading 'CANCEL CHATGPT EDU.' with a red banner stating 'INVEST IN HUMANS' over a dark board background.

The CSU/OpenAI contract is set to expire June 30, 2026.

Sign this petition: https://actionnetwork.org/petitions/cancel-chatgpt-edu-invest-in-humans/ for the CSU NOT to renew the contract and to use the savings to protect jobs at CSU campuses facing layoffs.

The Universe is not a Computer When we want to predict the future, we compute it from what we know about the present. Specifically, we take a mathematical representation of observed reality, plug it into some dynamical equations, a...

If anyone wants some follow-up reading, here's a link to my popular-science-level essay on the topic that Daniel mentioned in the podcast: arxiv.org/abs/1211.7081

Right, and the moment you learn anything you didn't already know (even your own choice of measurement basis), it's logically incoherent *not* to Bayesian update your state of knowledge. (a state which includes the wavefunction, imho)

One path forward, motivated by this analysis, would be to “collapse” the state *twice*; once when you decide on the measurement procedure (basis, timing, etc), and then again when the outcome is learned. There’s no formal framework that does this in QM, but there is elsewhere. (Ising models,etc.)

Good points all. Ideal measurement operators “live” on instantaneous hypersurfaces (in some special frame). But most real-life measurements don’t conform to this idealization. For an even more dramatic violation, consider arrival-time measurements, which “live” on timelike surfaces!