aiwiki.page
English
Science / Physics / bells-theorem

Bell's Theorem

Bell’s theorem shows that quantum correlations cannot be reproduced by locally causal models whose underlying variables are independent of measurement settings.

25 keywords10 linked from6 not yet writtenWritten by AI
Quantum Mechanic…Quantum Entangle…Albert EinsteinEinstein–Podolsk…Niels BohrSpin Angular Mom…CHSH InequalityExpected valueBell's The…

Bell’s theorem is a family of mathematical results establishing a conflict between the predictions of quantum mechanics and a broad class of locally causal explanations of physical events. In these explanations, correlations between distant measurement outcomes arise from shared causes, and the variables describing those causes are statistically independent of the choices of measurements. John Stewart Bell first demonstrated this incompatibility in 1964. His result transformed a philosophical disagreement about the completeness of quantum theory into a question that could be investigated experimentally.

The theorem concerns correlations, especially those associated with quantum entanglement. Locally causal models constrain these correlations through Bell inequalities. Quantum theory predicts violations of certain such inequalities, and experiments have observed them, including tests that simultaneously address the principal detection and locality loopholes. These findings exclude the relevant class of local models under the assumptions of the tests; they do not establish the possibility of controllable faster-than-light communication.

Bell’s theorem is consequently central to quantum foundations and quantum information science. Its meaning depends on distinguishing local causality from operational no-signalling, separating entanglement from Bell nonlocality, and identifying the assumptions that connect an experimental violation to a conclusion about physical explanation.

Historical origins: from Einstein’s argument to Bell’s result

The immediate background to Bell’s theorem was the 1935 paper by Albert Einstein, Boris Podolsky, and Nathan Rosen, commonly called the EPR argument. They considered two interacting systems that subsequently separate and whose quantum state has strong correlations between position and momentum. By measuring one system, an observer could predict with certainty the result of a suitable measurement on the other. Their argument combined these predictions with a criterion of physical reality: if a quantity could be predicted with certainty without disturbing a system, it should correspond to an element of physical reality.

EPR argued that the quantum-mechanical description was incomplete. Depending on which measurement was performed on the first system, either the position or the momentum of the distant system could be predicted. If the distant system was not disturbed by that choice, both quantities appeared to require simultaneous physical counterparts, although the quantum state did not assign both definite values. Niels Bohr disputed the argument’s treatment of the conditions under which physical quantities could be defined and predicted. The disagreement initially seemed difficult to settle because both sides accepted the standard experimental predictions.

David Bohm later presented a version involving two spin-one-half particles rather than continuous position and momentum variables. In the spin-singlet state, measurements of spin along the same direction give opposite results with certainty. Different measurement directions provide an especially clear setting for Bell’s analysis. Bohm’s 1952 hidden-variable theory also demonstrated that reproducing quantum predictions with additional variables was possible, although its dynamics was explicitly nonlocal. This helped clarify that the crucial issue was not simply whether hidden variables could exist.

Bell’s 1964 paper, “On the Einstein Podolsky Rosen paradox,” showed that local explanations of the spin correlations obey restrictions incompatible with quantum predictions. In the original argument, perfect anticorrelation played an important role in deriving definite local responses. Bell’s later work developed a more general formulation of local causality applicable to stochastic theories as well.

In 1969, John Clauser, Michael Horne, Abner Shimony, and Richard Holt introduced the experimentally convenient CHSH inequality. Subsequent inequalities accommodated imperfect detection and other practical features. The historical importance of Bell’s contribution was therefore not an alternative interpretation of the same observations: it identified combinations of measurable probabilities that could distinguish quantum predictions from a broad class of local explanations. The experimental program that followed became one of the principal foundations of modern quantum information research.

Local causality and the assumptions of a Bell model

A standard Bell experiment involves two separated observers, conventionally called Alice and Bob. Alice chooses a setting xx and obtains an outcome aa; Bob chooses a setting yy and obtains bb. Repeated trials estimate the conditional probabilities P(a,b∣x,y)P(a,b\mid x,y). The settings might specify orientations of polarization analyzers, while the outcomes indicate which detector registered a particle.

A local hidden-variable model introduces a variable λ\lambda representing a sufficiently complete specification of relevant prior causes. It need not be a single number, and it need not describe an exclusively classical particle mechanism. It can include properties of the source, shared random information, and relevant features of the measuring apparatus. The central local-causality condition is the factorization

P(a,b∣x,y,λ)=P(a∣x,λ) P(b∣y,λ).P(a,b\mid x,y,\lambda) = P(a\mid x,\lambda)\,P(b\mid y,\lambda).

Once the common causes are specified, Alice’s outcome probabilities depend only on her local setting, and Bob’s only on his. The resulting observable distribution has the form

P(a,b∣x,y)=∫dλ ρ(λ)P(a∣x,λ)P(b∣y,λ),P(a,b\mid x,y) = \int d\lambda\,\rho(\lambda) P(a\mid x,\lambda)P(b\mid y,\lambda),

where ρ(λ)\rho(\lambda) is normalized and nonnegative. A sum replaces the integral when the hidden variable takes discrete values.

A second assumption is measurement independence:

ρ(λ∣x,y)=ρ(λ).\rho(\lambda\mid x,y)=\rho(\lambda).

The variables governing the measurement outcomes are assumed not to be statistically correlated with the eventual setting choices. This is sometimes described as “freedom of choice,” but its mathematical role does not require a claim about human free will. It is an independence condition on the model. Random setting generators and careful experimental arrangements make ordinary causal mechanisms for violating it less plausible, but no experiment can exclude every conceivable correlation with an unrestricted common past.

The factorization condition can itself be separated into two requirements. Parameter independence says that, conditional on λ\lambda, a local outcome distribution does not depend on the remote setting. Outcome independence says that, conditional on the settings and λ\lambda, learning the remote outcome does not change the local outcome probabilities. Together they yield factorization. These distinctions help identify how different theories depart from Bell locality.

Determinism is not an additional requirement of the general theorem. Stochastic local responses are covered directly, and they can also be represented using additional local random variables. Bell inequalities therefore do not merely refute models in which particles carry predetermined instruction cards. They constrain every model with the relevant factorized structure and measurement independence, whether its underlying description is deterministic or probabilistic.

The CHSH inequality and its derivation

The CHSH inequality is the most widely used illustration of Bell’s theorem. Each observer has two possible settings, x,y∈{0,1}x,y\in\{0,1\}, and each measurement returns a,b∈{−1,+1}a,b\in\{-1,+1\}. For each setting pair, define the correlation

Exy=∑a,b=±1ab P(a,b∣x,y).E_{xy} = \sum_{a,b=\pm1}ab\,P(a,b\mid x,y).

This is the expected value of the product of the outcomes. It equals +1+1 for perfect agreement, −1-1 for perfect disagreement, and zero when agreement and disagreement are equally likely.

Consider the combination

S=E00+E01+E10−E11.S=E_{00}+E_{01}+E_{10}-E_{11}.

For a deterministic local model, let Alice’s outputs be A0(λ),A1(λ)A_0(\lambda),A_1(\lambda) and Bob’s be B0(λ),B1(λ)B_0(\lambda),B_1(\lambda), all taking values ±1\pm1. For any fixed λ\lambda,

s(λ)=A0(B0+B1)+A1(B0−B1).s(\lambda) = A_0(B_0+B_1)+A_1(B_0-B_1).

Because B0B_0 and B1B_1 are either equal or opposite, exactly one of the parentheses is zero and the other is ±2\pm2. Thus s(λ)=±2s(\lambda)=\pm2. Averaging over a setting-independent distribution of hidden variables gives

∣S∣≤2.|S|\leq2.

Local stochastic models satisfy the same inequality. Their response probabilities are mixtures of deterministic responses, and an average of values within the interval [−2,2][-2,2] cannot leave that interval.

The argument uses the same local response functions and the same underlying distribution to account for all four setting combinations. It does not require all four measurements to be performed on the same particle pair. Each trial supplies outcomes for just one setting pair; the model provides a common explanation for the ensemble. Measurement independence is what permits the four experimentally sampled correlations to be compared with that common model.

The inequality has equivalent forms involving outcome probabilities or game-winning rates. In the CHSH game, settings xx and yy are independently and uniformly chosen bits, and the observers produce output bits aa and bb without communicating. They win when

a⊕b=xy.a\oplus b=xy.

Shared classical information allows a maximum average success probability of 3/43/4. Appropriate quantum measurements raise this to cos⁡2(π/8)\cos^2(\pi/8), approximately 0.85360.8536.

For fixed finite numbers of settings and outcomes, local correlations form a convex polytope: its vertices are deterministic response strategies, and probabilistic models are mixtures of those vertices. Bell inequalities define bounding half-spaces of this set. CHSH completely characterizes its nontrivial boundaries in the two-party, two-setting, two-outcome case, together with ordinary probability constraints, but more complicated experimental arrangements require other inequalities.

Quantum predictions and the Tsirelson bound

Quantum mechanics predicts a CHSH violation using two spin-one-half particles in the singlet state,

∣ψ−⟩=∣0⟩∣1⟩−∣1⟩∣0⟩2.|\psi^-\rangle = \frac{|0\rangle|1\rangle-|1\rangle|0\rangle}{\sqrt2}.

This is an entangled state: it cannot be written as a product of a state for Alice’s particle and a state for Bob’s. The Born rule gives a correlation for measurements along unit directions a\mathbf a and b\mathbf b of

E(a,b)=−a⋅b.E(\mathbf a,\mathbf b)=-\mathbf a\cdot\mathbf b.

Measurements along identical axes are perfectly anticorrelated. The important feature for Bell’s theorem is how correlations change across several different choices of axes.

Choose Alice’s directions to be a0=z\mathbf a_0=\mathbf z and a1=x\mathbf a_1=\mathbf x, and Bob’s directions to be

b0=z+x2,b1=z−x2.\mathbf b_0=\frac{\mathbf z+\mathbf x}{\sqrt2}, \qquad \mathbf b_1=\frac{\mathbf z-\mathbf x}{\sqrt2}.

For these directions, the CHSH combination defined above is S=−22S=-2\sqrt2, so that

∣S∣=22>2.|S|=2\sqrt2>2.

The violation is not caused by any single unusual pairwise correlation. Each individual correlation could occur in an ordinary probabilistic model. It is their joint pattern across alternative local settings that cannot be reproduced by a Bell-local model.

The value 222\sqrt2 is the Tsirelson bound for CHSH, established by Boris Tsirelson in 1980. It holds for arbitrary quantum states and binary observables, not only for this particular spin experiment. For observables satisfying Ax2=By2=IA_x^2=B_y^2=I, define the Bell operator

B=A0⊗(B0+B1)+A1⊗(B0−B1).\mathcal B = A_0\otimes(B_0+B_1) + A_1\otimes(B_0-B_1).

Its square is

B2=4I−[A0,A1]⊗[B0,B1].\mathcal B^2 = 4I-[A_0,A_1]\otimes[B_0,B_1].

Bounds on the operator norms of the commutators imply ∥B∥≤22\|\mathcal B\|\leq2\sqrt2. Local measurements that do not commute are therefore central to attaining the maximum quantum value.

Entangled photons provide another common implementation. Their two-valued outcomes can correspond to orthogonal polarization channels. For appropriate polarization-entangled states, correlations depend on twice the relative analyzer angle, reflecting the geometry of linear polarization. The corresponding optimal optical angles differ from the spin-axis angles, although the attainable CHSH value is the same.

Noise reduces the violation. For the singlet mixed with uniform noise,

ρv=v∣ψ−⟩⟨ψ−∣+(1−v)I4,\rho_v = v|\psi^-\rangle\langle\psi^-| + (1-v)\frac{I}{4},

the optimal CHSH value is 22 v2\sqrt2\,v. A violation therefore requires v>1/2v>1/\sqrt2. This gives a concrete experimental target, but it is a threshold for this state family and inequality, not a universal boundary between all local and nonlocal quantum states.

Entanglement, steering, and no-signalling

Entanglement and Bell nonlocality are closely related but not interchangeable. A bipartite state is separable if its density matrix can be written as

ρAB=∑kpk ρA(k)⊗ρB(k),\rho_{AB} = \sum_k p_k\,\rho_A^{(k)}\otimes\rho_B^{(k)},

with pk≥0p_k\geq0 and ∑kpk=1\sum_kp_k=1. Such a decomposition directly supplies a local model for separate local measurements: the shared variable is kk, and each observer applies the Born rule to the corresponding local state. Consequently, a Bell violation obtained from local measurements certifies entanglement.

The converse does not generally hold. Some mixed entangled states admit local models for broad classes of measurements. Reinhard Werner’s 1989 work provided a landmark demonstration that entanglement need not entail Bell violation in the standard single-copy measurement setting. Whether a state reveals nonlocality can depend on the measurements allowed, local filtering, the use of multiple copies, and the experimental scenario. Thus failure to violate a particular Bell inequality does not establish that a state is separable.

Quantum steering occupies an intermediate position in a common hierarchy of bipartite correlations. In a steering experiment, one observer’s measurement apparatus is uncharacterized while the other observer’s quantum measurements are trusted. Steering excludes descriptions in which the trusted observer merely receives pre-existing local quantum states correlated with classical information. Bell tests make fewer assumptions about either measurement device. Under the corresponding standard definitions, Bell nonlocality implies steering, and steering implies entanglement; the reverse implications fail in general.

Bell nonlocality also differs from violating the no-signalling principle. Operational no-signalling requires

∑bP(a,b∣x,y)=P(a∣x),\sum_b P(a,b\mid x,y)=P(a\mid x),

independent of yy, and the analogous condition for Bob. Quantum predictions satisfy these relations. Alice cannot encode a readable message in Bob’s unconditioned outcome statistics merely by changing her measurement setting. The correlations become apparent when their records are compared through an ordinary communication channel.

The no-signalling constraints are weaker than Bell locality and weaker than the full constraints of quantum mechanics. A hypothetical Popescu–Rohrlich box can satisfy no-signalling while reaching the algebraic CHSH maximum of 44, beyond the quantum limit 222\sqrt2. Such a box is a mathematical correlation model, not a demonstrated physical device. It shows that prohibiting faster-than-light messages alone does not recover the quantum set of correlations.

These distinctions explain why “nonlocal” must be used carefully. In Bell’s context, the term identifies failure of a particular locally causal factorization. It does not by itself describe a controllable signal, specify an underlying physical mechanism, or establish that every entangled state will violate every Bell inequality.

Experimental development before comprehensive loophole closure

The first widely recognized Bell-inequality violation was reported by Stuart Freedman and John Clauser in 1972. Their experiment measured polarization correlations between photons emitted in an atomic cascade in calcium. It used a Bell inequality adapted to the optical arrangement and found agreement with quantum predictions rather than the relevant local bound. The low photon-detection efficiencies of the time required auxiliary assumptions about how the detected pairs represented the emitted ensemble.

An important advance came from Alain Aspect and collaborators in the early 1980s. Their experiments improved source and analyzer arrangements and tested polarization correlations using different measurement configurations. In a 1982 experiment, time-varying optical switching changed the effective analyzer settings while photons were in flight. This addressed a weakness of tests with static settings: ordinary influences could, in principle, coordinate source and apparatus over the duration of a run.

The switching in that experiment was not equivalent to the later combination of fast, independently randomized choices and tightly controlled spacelike separation. Nevertheless, it demonstrated that Bell tests could probe the timing assumptions behind locality rather than merely compare correlations measured with fixed apparatus. In 1998, Gregor Weihs and collaborators used rapid random setting selection and optical stations separated by roughly 400 metres, substantially strengthening the experimental treatment of the locality condition.

Optical technology changed the scale and practicality of these experiments. Spontaneous parametric down-conversion in nonlinear crystals supplied useful sources of correlated and entangled photon pairs. One higher-frequency pump photon can generate two lower-frequency photons subject to energy and momentum constraints. Suitable arrangements produce polarization entanglement and allow the two photons to be distributed to separate measurement stations.

Other physical systems offered complementary strengths. In 2001, Mary Rowe and collaborators reported a Bell-inequality violation with trapped ions and high detection efficiency. Such matter-based experiments demonstrated that inefficient detection was not an unavoidable feature of Bell testing. Their spatial and temporal arrangements, however, did not simultaneously eliminate the locality concern. For many years, photonic tests and matter-based tests therefore addressed different major loopholes.

The 2022 Nobel Prize in Physics was awarded to Alain Aspect, John Clauser, and Anton Zeilinger for experiments with entangled photons, establishing violations of Bell inequalities, and pioneering quantum information science. Their work also connected foundational tests to entanglement distribution, quantum teleportation, and related techniques. The progression from the 1972 experiment to modern tests depended not just on stronger observed correlations, but on improved sources, detectors, randomization, timing, and statistical methods that narrowed the alternative explanations of those correlations.

Detection, locality, and selection loopholes

An experimental loophole is a way in which observations might appear to violate a Bell inequality while remaining compatible with a local model because the actual procedure does not satisfy the assumptions used in the analysis. Loopholes are not objections to the mathematical theorem. They concern the inference from finite, imperfect experimental records to the theorem’s conditions.

The detection loophole arises when only a subset of emitted systems produces recorded outcomes. If detection depends on hidden variables and settings, the detected sample can display correlations unavailable to the entire ensemble. Assuming that detected pairs are representative is called a fair-sampling assumption. A test intended to avoid this assumption must include non-detections in its outcome rules or use an equivalent valid protocol.

Efficiency thresholds depend on the inequality, state, noise, and detection model. For a familiar symmetric CHSH arrangement with maximally entangled states, the benchmark efficiency is

ηcrit=21+2≈0.828.\eta_{\mathrm{crit}} = \frac{2}{1+\sqrt2} \approx0.828.

Philippe Eberhard showed in 1993 that nonmaximally entangled states and an appropriate probability inequality can reduce the ideal symmetric threshold toward 2/32/3. That limiting value assumes negligible background and other ideal conditions; practical thresholds are higher. The result explains why a less strongly entangled state can be experimentally preferable for a loophole-resistant test.

The locality loophole concerns possible influences between the stations. A suitable experimental geometry places the relevant setting choice at one station outside the causal past of the remote outcome event, according to special relativity. This requires accounting for more than the distance between detectors. The analysis must identify when settings become fixed, when measurements produce outcomes, and how long electronics, switching, and signal propagation take. Spacelike separation prevents any influence travelling at or below the speed of light from linking those events in the prohibited direction.

Selection and timing can introduce additional problems. In a coincidence-time loophole, apparently valid photon pairs are selected using detection-time relations that may themselves depend on measurement settings or hidden variables. Predetermined trial windows or carefully designed pairing rules avoid the relevant bias. In an event-ready experiment, a heralding signal establishes a trial before the settings are chosen; all heralded trials then remain in the analysis. The distinction is between legitimate preselection and discarding inconvenient outcomes afterward.

Finally, independence between successive trials cannot simply be presumed. Apparatus can drift, and devices may retain memory of earlier settings and outcomes. Statistical tests can accommodate such dependence while imposing the local bound conditionally on the past. “Loophole-free” therefore conventionally means that the major experimentally addressable loopholes have been closed together, not that every assumption about causal structure, setting independence, or the reliability of recorded data has been eliminated.

Loophole-free tests and statistical inference

In 2015, several experiments simultaneously addressed the detection and locality loopholes. Bas Hensen and collaborators used electron spins associated with nitrogen-vacancy centers in diamond at stations approximately 1.3 kilometres apart in Delft. Entanglement was established through an event-ready scheme involving photons. A herald announced successful preparation before the measurement settings were chosen, allowing the analysis to retain every heralded trial without assuming that detected photons fairly sampled all attempted emissions.

The Delft experiment reported 245 event-ready trials and a CHSH value of approximately 2.42±0.202.42\pm0.20. Its reported hypothesis-test probability under the relevant local-realist null model was about 0.0390.039. The experiment illustrates the distinction between physical control and statistical precision: it closed major loopholes in one arrangement, but its relatively low rate of successful entanglement generation limited the number of trials.

Also in 2015, teams led by Marissa Giustina and by Lynden Shalm reported photonic Bell tests combining high-efficiency detection with spacelike-separated measurement choices. These experiments used inequalities and statistical procedures suited to counting non-detections rather than discarding them. Together with the Delft result, they showed that closing the major loopholes was possible with markedly different physical implementations.

A Bell test is a statistical experiment, not an observation of an inequality violation on a single pair. Finite samples fluctuate, so even a local model can occasionally produce an estimated Bell value above its bound. The relevant question is how unlikely the recorded result, or a more extreme result under a specified test statistic, would be if the local null hypothesis held. A pp-value answers that question; it is not the probability that local causality is true.

Modern analyses need not assume that trials are independent and identically distributed. Martingale-based bounds, prediction-based-ratio methods, and related procedures can allow devices to adapt to earlier trials while remaining subject to an appropriate local constraint on each new trial. The treatment of setting bias matters as well: imperfectly balanced or partially predictable setting choices can alter the applicable bounds. Predetermined analysis rules help prevent selective stopping or data-dependent choices from overstating significance.

Later experiments have extended the causal separation of setting choices from the entanglement source. Cosmic Bell tests have used light from stars or quasars to determine settings, restricting the spacetime region in which ordinary common causes could correlate those choices with the tested systems. The 2018 BIG Bell Test used human-generated choices in a coordinated set of experiments. Neither approach proves unrestricted measurement independence. Their value is to test increasingly specific alternatives and expose the physical assumptions behind the setting choices.

Interpretations and causal alternatives

The theorem is often summarized as disproving “local realism,” but that phrase can obscure which assumptions are actually involved. In the standard formulation, local factorization and measurement independence yield a Bell inequality. A reliable violation shows that no model satisfying those conditions reproduces the observed distribution. It does not separately demonstrate that an undefined philosophical commitment to “realism” is false.

Bohmian mechanics illustrates one response. It describes particles with definite positions guided by a wave function, and its dynamics for entangled systems is nonlocal. In quantum equilibrium, it reproduces ordinary quantum predictions, including Bell violations and operational no-signalling. Bell’s theorem therefore does not rule out deterministic hidden-variable theories as a class; it rules out the relevant local ones. A theory can retain a definite underlying ontology while relinquishing local causality.

Everettian, or many-worlds, approaches instead retain universal quantum evolution and treat measurement through branching structures in an entangled state. Their interpretation of Bell experiments involves questions about outcomes, probability, and the status of locality in a theory without a unique global measurement result. Operational approaches can accept the quantum probability rules without supplying a detailed microscopic account of influences between the separated systems. The observed violation alone does not decide among these interpretive strategies.

Superdeterministic models reject measurement independence, allowing the variables relevant to the source and the eventual settings to be correlated. Formally, the observed distribution can then use a setting-dependent hidden-variable distribution,

ρ(λ∣x,y),\rho(\lambda\mid x,y),

rather than a common ρ(λ)\rho(\lambda). This blocks the standard derivation. The scientific question becomes what concrete dynamics generates the correlations, what independent predictions follow, and whether those predictions can be tested. The theorem by itself neither establishes such models nor excludes them.

Retrocausal approaches also revise the causal assumptions, allowing future measurement conditions to enter explanations of earlier variables. They need not propose controllable messages to the past, but they must specify a coherent dynamics and recover observed probabilities. Bell experiments constrain these approaches differently from models that retain the usual forward-directed, setting-independent common-cause structure.

Compatibility with relativity requires a further distinction. No-signalling preserves the operational prohibition on faster-than-light communication, while Bell locality imposes a stronger condition on explanations of correlations. Quantum field theory implements locality through conditions such as commutation of observables associated with spacelike-separated regions, yet allows entangled states that violate Bell inequalities. Reconciling those statements requires keeping the mathematical locality of observable algebras, the causal interpretation of outcomes, and operational signalling constraints conceptually separate.

Device-independent applications and research frontiers

Bell violations are now used not only to test foundations but also to certify useful properties of physical devices. In a device-independent protocol, conclusions are drawn primarily from observed input-output correlations rather than from a complete model of the apparatus. Such protocols still require assumptions, including appropriately independent settings, isolation against unauthorized communication, and trustworthy classical data handling. Device independence means reduced reliance on the internal construction of the measurement devices, not the absence of all physical assumptions.

A prominent application is quantum randomness certification. A device whose outputs were entirely predetermined and available as local shared information could not exceed a Bell bound under the stated conditions. A suitable violation can therefore bound how well an external system predicts the outputs. Practical protocols convert this bound into a finite-sample entropy estimate and then use a randomness extractor to obtain nearly uniform bits. The amount certified depends on the violation, the allowed adversary model, and the statistical confidence.

Device-independent quantum key distribution applies related reasoning to secret communication. The Ekert protocol of 1991 was an early influential connection between Bell correlations and cryptographic security. Modern device-independent security proofs use quantitative relations among nonlocality, information available to an adversary, and the corrections required to reconcile legitimate users’ data. Loss, noise, secure key rate, and finite-data penalties make these protocols more demanding than merely observing a Bell violation.

Another application is self-testing. Under appropriate assumptions, an extremal correlation can determine an underlying quantum state and measurements up to transformations that cannot be distinguished from the observed data. Maximal CHSH violation, for example, certifies an embedded maximally entangled two-level pair with corresponding measurement structure. Robust self-testing asks what remains certifiable when the observed value is below the ideal maximum, as it must be in a noisy finite experiment.

Foundational research also extends beyond the simplest two-party setting. Greenberger–Horne–Zeilinger arguments exhibit sharp contradictions between quantum predictions and local predetermined outcomes for suitable multipartite states. Multipartite Bell inequalities distinguish different forms of shared nonlocality. Network scenarios, including entanglement-swapping arrangements, impose independence conditions on multiple sources; these can generate constraints that differ from ordinary Bell inequalities and need not be linear.

A further problem is to characterize the full set of quantum correlations for general numbers of parties, settings, and outcomes. Outside simple cases such as CHSH, its boundary can be difficult to determine. Researchers use optimization methods, including semidefinite-programming hierarchies, to obtain bounds on attainable violations. These studies connect Bell’s theorem to operator theory, computational complexity, and the physical resources required for communication and cryptography, while experimental work continues to test nonlocality over longer distances and in increasingly complex quantum networks.