Some sets of quantum states carry no entanglement at all: each state is just a list of separate single-system states, one per party. Even so, the parties cannot tell which state they hold by measuring locally and talking on the phone. In the strongest form, called strong quantum nonlocality, no group of parties can make any informative first measurement without spoiling the orthogonality that a perfect identification would need. This is currently a finite phenomenon, studied in systems of three or four parties. We ask what happens when the number of parties is an uncountable infinite cardinal $\kappa$. Our main conjecture is as follows. The statement "whenever a set of product states on $\kappa$ parties is strongly nonlocal, some sub-system on fewer than $\kappa$ parties already is" holds exactly when $\kappa$ is a weakly compact cardinal. Weakly compact cardinals are the first large cardinals whose properties always reflect down to smaller pieces.
If the conjecture holds, the payoff runs in both directions. A quantum-information property would become a new characterization of a large cardinal. The existing theory of how large cardinals survive forcing (the method set theorists use to build new models of set theory), and of how they fail inside definable inner models, would then tell us when this quantum reflection principle is robust and when it breaks. If the conjecture fails, and reflection turns out to be provable in ordinary set theory (ZFC), that is also a clean answer. It would mean strong nonlocality is "finitary" even at transfinite scale, which is a structural fact nobody has established.
Core Claim
Nonlocality without entanglement has one feature that makes large cardinals relevant rather than decorative. Local irreducibility is a statement that quantifies over all measurement operators on a subsystem, and over all bipartitions of the parties. The first point needs a definition. A set is locally irreducible for a party when the only measurement elements that keep the states orthogonal are multiples of the identity, so the measurement yields no information. When the parties are indexed by $\kappa$ and the states are coded as a subset $A \subseteq V_\kappa$, strong nonlocality becomes a universally quantified second-order statement about the structure $(V_\kappa, \in, A)$. Here $V_\kappa$ is the collection of all sets built in fewer than $\kappa$ stages. Such statements are exactly the ones a weakly compact cardinal reflects: that is its classical characterization as $\Pi^1_1$-indescribable. The upper bound, that weak compactness implies reflection of strong nonlocality, should follow once the definitions are set up correctly.
The new and hard part is the lower bound. We want to show that if strong nonlocality reflects at $\kappa$, then $\kappa$ is weakly compact. The natural route is to take a $\kappa$-tree with no branch through it, the combinatorial object whose existence refutes weak compactness, and code it into a product-state set. The goal is a set that is globally strongly nonlocal while every proper sub-system on fewer than $\kappa$ parties can be sorted locally.
Definitions We Will Fix
Transfinite tensor products are analytically awkward. An infinite product of inner products with modulus below one tends to zero, so states can become "accidentally orthogonal". We therefore make three modeling choices and ablate each one.
The first choice concerns the states. Each party holds $\mathbb{C}^d$, and a product state is a function assigning each party a vector from a fixed finite set of local vectors. Only fewer than $\kappa$ coordinates may differ from a reference vector $e_0$.
The second choice is strict orthogonality. Two product states count as orthogonal when they are orthogonal on at least one coordinate. This is the combinatorial notion that all finite constructions actually use.
The third choice concerns measurements. A local measurement on a side $S$ of a bipartition is an operator acting nontrivially on fewer than $\kappa$ coordinates of $S$.
A slice of a set onto a party subset $T$ fixes the local vectors on the parties outside $T$ and keeps the states that agree with those fixed vectors. Reflection means that some slice onto fewer than $\kappa$ parties is itself strongly nonlocal.
Research Program
The theory has three stages.
The first stage proves the upper bound through indescribability. The key check is that local irreducibility across all bipartitions really is $\Pi^1_1$ over $V_\kappa$ under the third choice above.
The second stage attacks the lower bound by tree coding. We need finite gadgets: product-state sets on $n$ parties that are strongly nonlocal while every proper slice is not. These gadgets must also glue coherently when a party is added, so that they can be assembled along the levels of a branchless tree. A dead end in the tree must correspond to a slice that loses nonlocality.
The third stage transfers preservation results, and here the large-cardinal direction supplies the method. Once the equivalence is proved, several results become statements about the quantum reflection principle. Large cardinals need not be large in HOD forces a measurable or supercompact cardinal to stay large in $V$ but fail even weak compactness in HOD, the inner model of sets definable from ordinals. It does this by adding a weakly homogeneous $\kappa$-Suslin tree and coding it into HOD. Under the equivalence, this becomes a model where strong nonlocality reflects at $\kappa$ in the full universe but fails to reflect for the definable state sets inside HOD. The coding mechanism, a Suslin tree, is exactly the kind of object our lower-bound gadget turns into a non-reflecting state set.
In the other direction, Indestructibility properties of remarkable cardinals shows that small forcing preserves remarkability and that suitable preparation makes it survive $\mathrm{Add}(\kappa,\theta)$ followed by closed, distributive forcing. It uses embedding-lifting and surgery arguments that we would adapt to the reflection principle directly, in case only the upper bound holds.
The HOD Hypothesis sets a constraint on this program. The HOD Hypothesis and a supercompact cardinal proves that, under that hypothesis, a supercompact cardinal forces measurable cardinals in HOD cofinally below it. So a V-versus-HOD divergence for the quantum principle below a supercompact is impossible under the HOD Hypothesis. Any separation model we build must therefore violate it.
The quantum side supplies the objects and the verification method. Strong quantum nonlocality without entanglement in multipartite quantum systems gives the definition and the explicit constructions. It builds strongly nonlocal sets of $6(d-1)^2$ states in $\mathbb{C}^d\otimes\mathbb{C}^d\otimes\mathbb{C}^d$ and the first four-party complete bases, including an 81-state basis in $(\mathbb{C}^3)^{\otimes 4}$. These are the seeds for our gadgets. Those authors leave open constructions for $n\ge 5$ parties. Our gadget search is, in effect, a structured attack on that question with an added coherence requirement. Indistinguishability of bipartite states by positive-partial-transpose operations in the many-copy scenario is the closest existing result about a property surviving an unbounded family of enlargements. It shows that a subspace containing a PPT-definite operator stays indistinguishable for every number of copies. PPT measurements are a tractable relaxation of local measurement, defined by positivity of the partial transpose. That result is a uniform statement over $k$ copies, whereas ours is uniform over infinitely many parties. It motivates the PPT variant of our principle, discussed below.
Baselines
The naive reference point on the theory side is the ZFC compactness argument. If strong nonlocality at $\kappa$ were always witnessed by finitely many coordinates, reflection would follow from a König-lemma-style argument with no large cardinals at all. We will check this first by trying to prove reflection in ZFC at $\kappa=\omega_1$. Beating this baseline means exhibiting, in ZFC, a non-reflecting set at some non-weakly-compact $\kappa$.
The strongest existing competitor is the indescribability argument itself. It already gives the upper bound for free once the definitions are fixed. The proposal beats it only by proving the matching lower bound, meaning that the tree-coded set is strongly nonlocal while no slice is. Anything less is a restatement.
On the computational side, the naive reference is random orthogonal product sets and the standard product basis, which parties can sort locally. The strongest existing constructions are those of Yuan, Tian and Sun: 24 states versus a full basis of 27 in $3\times3\times3$, 54 versus 64 in $4\times4\times4$, and the 81-state four-party basis. We will run every candidate through the same exact-arithmetic irreducibility checker under the same enumeration budget, meaning the same local vector alphabet and the same maximum number of parties. Beating these constructions means producing a gadget family that none of them provides: a strongly nonlocal $n$-party set with no strongly nonlocal proper slice that extends coherently to $n+1$ parties, for $n$ up to at least 6.
Two ablations isolate the key ingredient. The first replaces support below $\kappa$ with finite support. The prediction is that reflection then becomes ZFC-provable, which would show the large-cardinal strength comes from infinitary support. The second replaces strict orthogonality with the analytic infinite inner product. The prediction is that accidental orthogonality then trivializes the question.
Why not just…?
Why not just use finite semidefinite programs, as the quantum literature does? Semidefinite programs are optimization problems over positive matrices, and they decide finite instances exactly. Entanglement cost of discriminating quantum states under locality constraints shows its hierarchy converges and checks tightness on 35 random $4\times4$ state pairs. But no finite program says anything about whether a property of $\kappa$-party systems reflects to smaller ones. We use such checks as the verification layer for gadgets, not as a substitute for the question.
Why not just work with countably many parties and standard operator algebras? At $\omega$ parties, infinite tensor product algebras are the right tool, and we expect reflection there to be provable. This is a prediction we will test, and it plays the role of the trivial case, because $\omega$ is "compact" in the relevant sense. The large-cardinal question appears only at uncountable $\kappa$ with support below $\kappa$. The finite-support ablation is exactly the experiment that separates the two cases.
Isn't this just weak compactness restated in quantum vocabulary? This is a real threat. If the tree coding turns out to be routine, the upper bound is folklore and the lower bound is a translation. The contribution survives only if the gadgets require something genuinely quantum. Specifically, irreducibility must fail for reasons that the purely combinatorial orthogonality graph does not see. We will test this by checking whether the gadgets remain non-reflecting when local irreducibility is replaced by the purely graph-theoretic condition. If they do, we report the result as a translation. A second route to real content is the PPT variant, which may not be $\Pi^1_1$ at all. For that variant, the V-versus-HOD question would be open even if the local version is settled.
Proposed Experiments
The decisive computational experiment is a gadget search. Seeds come from the 24-state $3\times3\times3$ set and the 81-state four-party basis. We enumerate all slices and all one-party extensions over a fixed local alphabet in $\mathbb{C}^3$ and $\mathbb{C}^4$. Local irreducibility is computed exactly, as the null space of the linear conditions $\langle\psi_i|E\otimes I|\psi_j\rangle=0$ for $i\neq j$ over rational or cyclotomic entries, across every bipartition. The metrics are three: the fraction of slices that remain strongly nonlocal, the branching number of coherent extensions, and the length of the longest extension chain along which no proper slice is strongly nonlocal. As a solver sanity check, the same code base will reproduce the double-trine error probabilities of Revisiting the optimal detection of quantum information: 0.0286 for global measurement, about 0.0647 for two-way local protocols, and about 0.0670 for one-way local protocols. No reported result was found on Lacuna for any system with more than four parties, or for any infinite-party version of strong nonlocality.
The outcome decides the direction of the theory. Suppose non-reflecting gadgets exist and extend coherently through six parties with branching at least two. Then the tree coding is viable, and we proceed to the lower-bound proof and the HOD separation built on the Suslin-tree coding of Cheng, Friedman and Hamkins. Suppose instead that every strongly nonlocal set on five or six parties has a strongly nonlocal slice. That is evidence that reflection is provable in ZFC, and we redirect the theory toward proving it at $\omega_1$. The finite-support and analytic-orthogonality ablations are run in both cases to confirm which ingredient carries the strength.
References
- Large cardinals need not be large in HOD
- Indestructibility properties of remarkable cardinals
- The HOD Hypothesis and a supercompact cardinal
- Strong quantum nonlocality without entanglement in multipartite quantum systems
- Indistinguishability of bipartite states by positive-partial-transpose operations in the many-copy scenario
- Entanglement cost of discriminating quantum states under locality constraints
- Revisiting the optimal detection of quantum information