Device-Independent Entanglement Certification

Every ordinary test of entanglement asks you to trust your instruments. An entanglement witness, the PPT criterion, a tomographic reconstruction — each presumes you know which observables the apparatus measures, in what Hilbert space, with what calibration. Device-independent certification throws all of that away. Alice and Bob hold a pair of sealed boxes; each box takes a classical setting (a button press $x$ or $y$ ) and emits a classical outcome ( $a$ or $b$ ), and the only record anyone keeps is the conditional distribution $p(a,b\mid x,y)$ . The claim this article develops is startling: certain such tables of clicks are intrinsically nonclassical — they cannot be produced by any local-hidden-variable model, and within quantum theory they cannot come from any separable state, no matter how the measurements are secretly implemented. The bridge from raw statistics to a certificate is a Bell inequality: the CHSH combination obeys $|S|\le2$ for every local model and every separable state, so $|S|>2$ witnesses entanglement using no assumption beyond no-signalling. This article is the synthesis of the entanglement-certification head — it takes the local bound from aq-qci-chsh-inequality and aq-qci-bell-theorem-local-hidden-variables, the quantum ceiling $2\sqrt2$ from aq-qci-tsirelson-bound, and the free-choice structure from aq-qci-nonlocality-no-signalling, and asks what an observed number actually certifies. The answer is more than a yes/no verdict: the size of the violation quantifies entanglement, converting clicks into a certified lower bound on concurrence and on singlet fidelity — with a guarantee that no construction hidden inside the boxes can fake.

Key takeaways

Device-independent (DI) certification treats the apparatus as black boxes: input classical settings $x,y$ , read classical outcomes $a,b$ , and conclude from the correlation table $p(a,b\mid x,y)$ alone — no trust in dimension, detectors, or state preparation.
If the statistics violate a Bell inequality, $|S|>2$ , then no separable state and no local model can reproduce them; entanglement is certified outright.
The amount of violation quantifies entanglement: an observed $S$ gives a certified concurrence $C\ge\sqrt{S^2/4-1}$ and, via self-testing rigidity, a singlet-fidelity bound, with $S=2\sqrt2$ uniquely fingerprinting a Bell pair plus anticommuting measurements.
Three thresholds bound the story: $2$ (local), $2\sqrt2\approx2.828$ (Tsirelson, quantum), and $4$ (no-signalling, the PR box); DI certification lives in the shell between $2$ and $2\sqrt2$ .
The guarantee is demanding: it needs space-like separation (locality loophole) and detection efficiency above $\eta_{\text{crit}}=2(\sqrt2-1)\approx0.828$ ; loophole-free Bell tests since 2015 made DI certification real.
Key idea: a Bell-inequality violation is a self-contained certificate of entanglement — because every separable state admits a local model, any $p(a,b\mid x,y)$ that violates a Bell inequality cannot come from a separable state, and the size of the violation lower-bounds how much entanglement must be present, with no trust in the devices.

The core insight

A Bell-inequality violation is a self-contained certificate of entanglement: because every separable state admits a local model, any $p(a,b\mid x,y)$ violating a Bell inequality cannot come from a separable state — and the size of the violation lower-bounds how much entanglement must be present, with no trust in the devices.

Where this lives today

The device-independent idea has moved from principle to hardware. In 2015 three groups closed both loopholes at once — the Delft group with nitrogen-vacancy electron spins entangled by photon heralding over $1.3\ \mathrm{km}$ , and Vienna and NIST groups with high-efficiency photon pairs — making $|S|>2$ an unassailable certificate for the first time. On that foundation, device-independent quantum key distribution reached its first demonstrations in 2022 using trapped ions and neutral atoms: the two parties extract a secret key whose security follows from an observed CHSH violation rather than from trusting their sources or detectors, so an adversary who built the boxes still cannot learn the key. The same violation certifies randomness — a value strictly above $2$ guarantees genuine min-entropy, bits that could not have been predetermined, the basis of device-independent random-number generators. Self-testing has become a design tool in its own right, certifying entangling gates, multipartite states, and whole computations from input-output statistics alone. The open frontier is quantitative: pushing DIQKD key rates and randomness-expansion factors up while tolerating realistic loss and finite statistics — the regime where the gap between the ideal $2\sqrt2$ and an achievable $S$ decides whether a protocol closes.

The physical picture

Imagine an inspector who cannot open the boxes and does not even know their dimension. She logs which button she pressed and which light flashed, nothing more. Remarkably, certain logbooks are impossible to produce by any pair of pre-agreed strategies — like two prisoners answering questions in separate cells armed only with a shared cheat sheet written in advance. The CHSH game can be won at most $75\%$ of the time with any cheat sheet (any local hidden variable), but a shared Bell pair wins it with probability $\cos^2(\pi/8)\approx85.4\%$ . Beating $75\%$ is the certificate: no local strategy reaches it, so the boxes must share entanglement. The first figure fixes the scenario — settings in, outcomes out, and space-like separation as the only physical assumption.

Figure

The black-box Bell test: a source shares a pair between Alice and Bob, each box takes a classical setting x or y and returns an outcome a = ±1 or b = ±1; the only record is the correlation table p of a,b given x,y, and space-like separation plus S > 2 certifies entanglement with no trust in the devices

The second twist is the deep one, and it is what makes DI quantitative. Tsirelson's rigidity says that hitting the optimal $85.4\%$ does not merely suggest entanglement — it pins down the state and measurements up to a local change of basis. The correlations at the ceiling leave no freedom: the boxes are forced to contain an embedded copy of a singlet and anticommuting Pauli measurements. So a number computed from clicks can certify a quantum object you never characterized, and how close to the ceiling you sit sets how much entanglement you may claim. The second figure is the object an experimenter actually reasons with: the map from the observed CHSH value to a certified amount of entanglement.

Figure

The certification window: any CHSH value between the local bound S = 2 and Tsirelson's bound S = 2√2 ≈ 2.828 proves entanglement, and the observed S fixes a certified concurrence C ≥ √(S²/4 − 1) reaching C = 1 only at 2√2; the no-signalling maximum S = 4 is the PR box, forbidden by quantum theory

The core equations

The device-independent observable is a single linear combination of correlators built from the four setting pairs. It is estimated purely from click statistics — count how often outcomes agree and disagree for each pair of buttons — and never references a Hilbert space.

S = \langle A_0 B_0\rangle + \langle A_0 B_1\rangle + \langle A_1 B_0\rangle - \langle A_1 B_1\rangle \qquad (1)

where: $A_x,B_y$ — $\pm1$ -valued outcomes for Alice's setting $x\in\{0,1\}$ and Bob's setting $y\in\{0,1\}$ ; $\langle A_xB_y\rangle=\sum_{a,b}ab\,p(a,b\mid x,y)$ — the correlator, estimated from raw statistics only; $S$ — the CHSH parameter, whose value alone certifies and quantifies entanglement. The lone minus sign is what makes the combination sensitive to nonlocality rather than to trivial correlation. The certification threshold is the local bound. Every local-hidden-variable model, and — because a separable state is a convex mixture of product states, each of which is local — every separable state, respects it. Exceeding it is therefore impossible without entanglement.

|S| \le 2 \qquad \text{(local / separable)} \qquad (2)

where: $2$ — the maximum attainable by any local model in any dimension and by any separable quantum state. The inequality is one-sided: $|S|>2$ proves entanglement, but $|S|\le2$ proves nothing (some entangled states admit local models). Derivation A below establishes it for arbitrary separable states. Quantum theory does not let the violation grow without limit. Tsirelson's bound caps the CHSH value at $2\sqrt2$ , and this ceiling — not the algebraic maximum $4$ — is the true edge of the quantum set for this inequality.

S_{\max} = 2\sqrt2 \approx 2.828 \qquad \text{(Tsirelson)} \qquad (3)

where: $2\sqrt2$ — the largest $|S|$ achievable by any quantum state and measurements, saturated only by a maximally entangled two-qubit pair with anticommuting observables. Derivation B proves it by squaring the CHSH operator; the gap between $2$ and $2\sqrt2$ is the entire resource window that certification exploits. The single assumption behind the whole framework is no-signalling: Alice's marginal statistics must not depend on Bob's choice of setting, and vice versa. This is what guarantees the boxes cannot secretly communicate and that the settings are freely chosen inputs.

\sum_b p(a,b\mid x,y) = \sum_b p(a,b\mid x,y') \quad \forall\, y,y' \qquad (4)

where: $p(a,b\mid x,y)$ — the observed black-box behaviour; the equality — Alice's marginal is independent of Bob's setting (and the symmetric statement for Bob). If it fails — a hidden channel between boxes — any $S$ up to $4$ can be faked and the certificate collapses. Now the quantitative heart of the article: the observed violation lower-bounds the concurrence of the underlying two-qubit state. This turns a click table into a certified amount of entanglement, immune to the calibration errors that plague ordinary witnesses.

C(\rho) \ \ge\ \sqrt{\tfrac{S^2}{4} - 1} \qquad (5)

where: $C(\rho)$ — the concurrence (Wootters measure, owned by aq-qci-concurrence-wootters-formula); $S$ — the observed CHSH value. The bound is vacuous at $S=2$ (giving $C\ge0$ ) and forces $C=1$ at $S=2\sqrt2$ . Derivation C obtains it from the Schmidt form of a pure state and extends it to mixtures. Rigidity upgrades quantification into self-testing: a value near the ceiling forces the physical state to be near a Bell pair, up to a local isometry, with a robust fidelity bound.

F = \langle\Phi^+|\,\rho\,|\Phi^+\rangle \ \ge\ \tfrac12\Big(1+\tfrac{S-2}{4(2\sqrt2-2)}\Big) \qquad \text{(robust self-test, schematic)} \qquad (6)

where: $F$ — fidelity of the locally isometric image of the physical state with the Bell pair $|\Phi^+\rangle=\tfrac1{\sqrt2}(|00\rangle+|11\rangle)$ ; $S$ — observed CHSH value. As $S\to2\sqrt2$ the bound drives $F\to1$ : the boxes are certified to contain a singlet, not merely "some" entanglement. Finally, the experimental gatekeeper. A real detector misses events, and an adversary who controls which events are recorded can fake a violation by post-selection unless efficiency clears a threshold.

\eta_{\text{crit}} = 2(\sqrt2-1) \approx 0.828 \qquad (7)

where: $\eta_{\text{crit}}$ — the minimum single-detector efficiency below which the detection loophole reopens for symmetric CHSH on a maximally entangled pair; $\eta$ — the actual efficiency. Below threshold a purely local model reproduces $|S|>2$ ; above it, no local fake exists, which is why heralded and high-efficiency platforms are mandatory for a genuine DI claim.

Derivation AWhy every separable state respects the local bound

The goal is to prove the certificate direction cleanly: that any separable state, in any dimension and with any measurements, satisfies $|S|\le2$ , so observing $|S|>2$ leaves entanglement as the only possible explanation. We write the state as a convex mixture of product states, show each product component reduces to a purely classical CHSH expression in real numbers, bound that expression by elementary algebra, and close with convexity. That no quantum feature survives is exactly the point — for this observable, separable states are indistinguishable from local classical ones.

Step 1. Write the separable state as $\rho_{AB}=\sum_k p_k\,\rho_A^{(k)}\otimes\rho_B^{(k)}$ with $\{p_k\}$ a probability distribution, and note that on a product component the joint correlator factors: $\langle A_xB_y\rangle_k=\mathrm{Tr}(\rho_A^{(k)}A_x)\,\mathrm{Tr}(\rho_B^{(k)}B_y)\equiv a_x^{(k)}b_y^{(k)}$ . Why: on a product state Alice's and Bob's measurements act on independent factors, so the expectation value of the tensor product is the product of expectation values — the entanglement that could couple them is absent.

Step 2. Each $a_x^{(k)},b_y^{(k)}$ is a real number in $[-1,1]$ (the expectation of a $\pm1$ observable), so the CHSH value of one component is $S_k=a_0^{(k)}(b_0^{(k)}+b_1^{(k)})+a_1^{(k)}(b_0^{(k)}-b_1^{(k)})$ , precisely the classical CHSH expression. Why: the operators have vanished and we are left with an algebraic combination of numbers in $[-1,1]$ — a local-deterministic-type quantity with no quantum content.

Step 3. Bound it: since $|a_x^{(k)}|\le1$ , $|S_k|\le|b_0^{(k)}+b_1^{(k)}|+|b_0^{(k)}-b_1^{(k)}|=2\max(|b_0^{(k)}|,|b_1^{(k)}|)\le2$ . Why: for any two reals in $[-1,1]$ the sum of $|b_0+b_1|$ and $|b_0-b_1|$ equals twice the larger magnitude, which cannot exceed $2$ — the same identity that caps a classical strategy.

Step 4. Average over the mixture using convexity: $|S(\rho)|=|\sum_k p_k S_k|\le\sum_k p_k|S_k|\le2$ . Why: the absolute value of an average never exceeds the average of the absolute values, so if every component obeys the bound, so does every mixture — no separable state, in any dimension, exceeds $2$ . *Why the algebra closes: one can check directly that $|b_0+b_1|+|b_0-b_1|\le2$ for all $b_0,b_1\in[-1,1]$ , confirming the bound reduces to the deterministic local ceiling.*

Go deeper · Derivation B · 1 minTsirelson's ceiling from squaring the CHSH operator

Derivation BTsirelson's ceiling from squaring the CHSH operator

The goal is to prove that quantum mechanics caps the violation at $2\sqrt2$ , well short of the algebraic maximum of $4$ , and to expose which operator inequality does the capping — because that same inequality, saturated, is what forces the state and measurements at the ceiling and so drives self-testing. We build the CHSH operator from $\pm1$ observables, square it, watch the cross terms collapse to a single commutator, bound that commutator by the observables' norms, and read off the operator norm, which caps the expectation value in any state.

Step 1. Define the Hermitian CHSH operator $\mathcal B=A_0\otimes(B_0+B_1)+A_1\otimes(B_0-B_1)$ , where $A_x,B_y$ are $\pm1$ observables so $A_x^2=B_y^2=\mathbb 1$ , and Alice's commute with Bob's, $[A_x,B_y]=0$ . Why: $S=\langle\mathcal B\rangle$ , so bounding the operator norm $\|\mathcal B\|$ bounds every possible expectation value at once, independent of the state.

Step 2. Square it and expand. Using $A_x^2=B_y^2=\mathbb 1$ and $[A_x,B_y]=0$ , the diagonal terms sum to $4\,\mathbb 1$ and the cross terms collapse to a single commutator product: $\mathcal B^2=4\,\mathbb 1-[A_0,A_1]\otimes[B_0,B_1]$ . Why: the messy sixteen-term expansion telescopes because most pairs commute and square to the identity, leaving only the genuinely non-commuting piece — the commutators that measure how incompatible the two settings are.

Step 3. Bound each commutator. Since $\|A_x\|=\|B_y\|=1$ , the triangle inequality gives $\|[A_0,A_1]\|\le\|A_0A_1\|+\|A_1A_0\|\le2$ and likewise $\|[B_0,B_1]\|\le2$ . Why: a commutator of two unit-norm operators can be no larger than twice their product norm, so the correction term to $4\,\mathbb 1$ is bounded by $2\cdot2=4$ .

Step 4. Assemble: $\|\mathcal B^2\|\le4+4=8$ , hence $\|\mathcal B\|\le\sqrt8=2\sqrt2$ , and therefore $|S|=|\langle\mathcal B\rangle|\le\|\mathcal B\|\le2\sqrt2$ . Why: the norm of a Hermitian operator caps its expectation value; equality needs $[A_0,A_1]$ and $[B_0,B_1]$ to reach norm $2$ , i.e. anticommuting observables on a Bell pair — the unique saturating configuration. *Why the numbers close: the squaring identity gives $\mathcal B^2=4-[A_0,A_1]\otimes[B_0,B_1]$ , and taking the operator norm yields $\|\mathcal B\|\le\sqrt8=2\sqrt2\approx2.828$ .*

Go deeper · Derivation C · 2 minFrom an observed violation to a certified concurrence

Derivation CFrom an observed violation to a certified concurrence

The goal is to convert the CHSH number into an amount of entanglement — the quantitative claim that distinguishes device-independent certification from a mere yes/no test. We compute the maximal CHSH value of a pure two-qubit state from its Schmidt decomposition, find it is exactly $2\sqrt{1+C^2}$ in the concurrence $C$ , invert to read $C$ off the violation, and argue that a mixed state can only under-report — so the pure-state equality becomes a rigorous lower bound for any state, the Verstraete and Wolf result restated constructively.

Step 1. Write any pure two-qubit state in Schmidt form $|\psi\rangle=\cos\alpha\,|00\rangle+\sin\alpha\,|11\rangle$ , whose concurrence is $C=2\cos\alpha\sin\alpha=\sin2\alpha$ . Form the correlation matrix $T_{ij}=\langle\psi|\sigma_i\otimes\sigma_j|\psi\rangle$ ; a direct computation gives $T=\mathrm{diag}(\sin2\alpha,\,-\sin2\alpha,\,1)$ . Why: the Schmidt form is the canonical two-parameter description of a pure two-qubit state, and $T$ is the object that the maximal Bell violation depends on, so this reduces the problem to three numbers.

Step 2. Invoke the Horodecki criterion: the maximal CHSH value over all measurements is $S_{\max}(\rho)=2\sqrt{t_1^2+t_2^2}$ , where $t_1\ge t_2$ are the two largest singular values of $T$ . Here the singular values are $\{1,\,|\sin2\alpha|,\,|\sin2\alpha|\}$ , so $t_1=1$ and $t_2=|\sin2\alpha|=C$ . Why: the Horodecki formula is the optimal-measurement result for CHSH, and reading its two largest inputs off our diagonal $T$ ties the violation directly to the concurrence.

Step 3. Substitute to get the pure-state identity $S_{\max}=2\sqrt{1+C^2}$ , which inverts to $C=\sqrt{S_{\max}^2/4-1}$ . Sanity checks: $C=0\Rightarrow S_{\max}=2$ (product state, no violation) and $C=1\Rightarrow S_{\max}=2\sqrt2$ (Bell pair, Tsirelson). Why: this exact relation is the backbone — it says the violation and the concurrence carry the same information for pure states, so an observed $S$ names a specific $C$ .

Step 4. Extend to mixed states. For any two-qubit $\rho$ of concurrence $C$ the achievable violation cannot exceed the pure-state value, $S\le S_{\max}(\rho)\le2\sqrt{1+C^2}$ , so rearranging gives the certificate $C(\rho)\ge\sqrt{S^2/4-1}$ . Why: mixing degrades nonlocality faster than it degrades concurrence, so a state that shows violation $S$ must carry at least the concurrence a pure state would need — the inequality is the device-independent guarantee. *Why the endpoints check out: $T=\mathrm{diag}(\sin2\alpha,-\sin2\alpha,1)$ has singular values $\{1,C,C\}$ , so $S_{\max}=2\sqrt{1+C^2}$ inverts to $C=\sqrt{S^2/4-1}$ , giving $C=1$ at $S=2\sqrt2$ and $C=0$ at $S=2$ .*

Limiting cases

The three thresholds of the CHSH parameter mark the boundaries between distinct physical theories and distinct certification verdicts, and one extra row records the crucial one-sidedness of the test.

Case	Result	Meaning
Local / classical	$\lvert S\rvert \le 2$	Reproducible by local hidden variables and by every separable state; no entanglement certified. The bound is the maximum over deterministic local strategies.
Quantum (Tsirelson)	$2 < \lvert S\rvert \le 2\sqrt2$	Entanglement certified device-independently; observed $S$ gives $C\ge\sqrt{S^2/4-1}$ . The ceiling $2\sqrt2\approx2.828$ is reached only by a maximally entangled pair with anticommuting measurements.
No-signalling / super-quantum	$2\sqrt2 < \lvert S\rvert \le 4$	Forbidden by quantum mechanics but allowed by no-signalling alone; the algebraic maximum $S=4$ is the PR box (owned by aq-qci-pr-boxes-superquantum). Inaccessible to any physical state.
Entangled yet local-modelled	$S\le2$ with $\rho$ entangled	Werner states below a visibility threshold are entangled but admit an explicit local model, so non-violation is not proof of separability — the certificate is strictly one-sided.

Common trapA CHSH violation always certifies entanglement, loopholes or not.

It does not. The certificate is only valid once two physical loopholes are closed. The locality loophole: if Alice's and Bob's measurement events are not space-like separated, a hidden influence — a slow signal, a shared clock, a leaked setting — can manufacture any value up to the algebraic maximum $4$ with no entanglement whatsoever, because the boxes are effectively one device. The detection loophole: if detection efficiency falls below $\eta_{\text{crit}}=2(\sqrt2-1)\approx0.828$ , a purely local model that decides which runs to report can post-select a fake violation above $2$ ; only efficiency above threshold forecloses this. And the converse trap is just as important: failing to violate ( $S\le2$ ) does not certify separability, since some entangled states (certain Werner states) admit local models and never exceed $2$ . A DI claim is only as strong as the loophole that is closed and the assumption — no-signalling — that is genuinely enforced.

Worked examples

1 · The singlet saturates Tsirelson's bound (fully worked)

Take the singlet $|\Psi^-\rangle=\tfrac1{\sqrt2}(|01\rangle-|10\rangle)$ and measure spin along directions $\hat a_x,\hat b_y$ with $A_x=\hat a_x\cdot\vec\sigma$ , $B_y=\hat b_y\cdot\vec\sigma$ . Using the angles $a_0=0^\circ,\ a_1=90^\circ,\ b_0=45^\circ,\ b_1=-45^\circ$ , compute $S$ and the CHSH-game success probability.

Step 1. The singlet correlator is direction-antisymmetric: $\langle A_xB_y\rangle=-\hat a_x\cdot\hat b_y=-\cos\theta_{xy}$ , where $\theta_{xy}$ is the angle between the two measurement directions. This is the one physical input; everything else is arithmetic.

Step 2. Read off the four angles. Between $(a_0,b_0)$ , $(a_0,b_1)$ and $(a_1,b_0)$ the separation is $45^\circ$ ; between $(a_1,b_1)$ it is $90^\circ-(-45^\circ)=135^\circ$ . Hence $\langle A_0B_0\rangle=\langle A_0B_1\rangle=\langle A_1B_0\rangle=-\cos45^\circ=-\tfrac1{\sqrt2}$ and $\langle A_1B_1\rangle=-\cos135^\circ=+\tfrac1{\sqrt2}$ .

Step 3. Assemble $S$ using the lone minus sign on the last term: $S=-\tfrac1{\sqrt2}-\tfrac1{\sqrt2}-\tfrac1{\sqrt2}-\big(+\tfrac1{\sqrt2}\big)=-\tfrac{4}{\sqrt2}$ . Therefore $\lvert S\rvert=\tfrac{4}{\sqrt2}=2\sqrt2\approx\mathbf{2.828}$ , saturating Tsirelson's ceiling.

Step 4. Translate to the game. A CHSH value $S$ corresponds to a winning probability $P_{\text{win}}=\tfrac12+\tfrac{S}{8}$ , so $P_{\text{win}}=\tfrac12+\tfrac{2\sqrt2}{8}=\tfrac12+\tfrac{\sqrt2}{4}=\cos^2(\pi/8)\approx\mathbf{0.854}$ . The best any local cheat sheet can do is $S=2$ , i.e. $P_{\text{win}}=0.75$ . The singlet reaches $\lvert S\rvert=2\sqrt2$ and wins $85.4\%$ of the time — clear of the $75\%$ local ceiling, which is the certificate.

2 · Reading a certified concurrence off the data

An experiment reports $S=2.6$ with the locality and detection loopholes closed. How much entanglement is certified, and how much would remain if the value were only $S=2.3$ ?

Step 1. Apply equation (5): $C\ge\sqrt{S^2/4-1}$ . For $S=2.6$ , $S^2=6.76$ , so $S^2/4=1.69$ and $C\ge\sqrt{1.69-1}=\sqrt{0.69}\approx\mathbf{0.83}$ . At least $0.83$ units of concurrence must be present, with no trust in the apparatus.

Step 2. Repeat for $S=2.3$ : $S^2=5.29$ , $S^2/4=1.3225$ , $C\ge\sqrt{0.3225}\approx\mathbf{0.57}$ . The same relation shows how quickly the certified entanglement bleeds away as the violation shrinks toward the local bound — halving the distance to $2$ roughly halves the certified $C$ .

Step 3. Sanity-check the endpoints. At $S=2$ the bound gives $C\ge0$ (nothing certified, consistent with the one-sidedness of the test); at $S=2\sqrt2\approx2.828$ it gives $C\ge\sqrt{2-1}=1$ , forcing a maximally entangled pair. An observed $S=2.6$ certifies $C\ge0.83$ ; the bound is monotone in $S$ and pins $C=1$ exactly at Tsirelson's value. For either run to count, the experiment must also clear the Eberhard efficiency threshold $\eta_{\text{crit}}=2(\sqrt2-1)\approx0.828$ of equation (7) — a single-arm efficiency of $0.90$ certifies, but $0.80$ leaves the detection loophole open and no violation, however large, is trustworthy.

Physics of the portrait

Device-independent certification rests on a single sharp number, $2\sqrt2$ , and that number has an author. The quantum ceiling on Bell violations was the bound that turned nonlocality from a qualitative surprise into a quantitative resource.

Origins · 1 minBoris Tsirelson and how this was discovered

Boris Tsirelson

1950–2020 · quantum bounds on nonlocal correlations

Boris Semyonovich Tsirelson (also transliterated Cirel'son) was a Russian-Israeli mathematician whose 1980 paper Quantum generalizations of Bell's inequality did for the quantum side of nonlocality what Bell had done for the classical side. Bell (1964) and CHSH (1969) had shown that local hidden variables cannot exceed $S=2$ and that quantum mechanics can violate this — but how far? Tsirelson proved the exact answer: no quantum state and no measurements can push the CHSH combination past $2\sqrt2$ , by the operator-norm argument of Derivation B. The result did three things at once. It located quantum theory strictly between the local polytope ( $S\le2$ ) and the no-signalling polytope ( $S\le4$ ), exposing that quantum mechanics is not as nonlocal as causality alone would allow — the puzzle later crystallized by Popescu and Rohrlich's PR box. It fixed the resource window, $2$ to $2\sqrt2$ , in which every device-independent protocol operates. And, most consequentially here, its tightness is a rigidity: because $2\sqrt2$ is reached by an essentially unique configuration up to local isometry, an observed value near the ceiling forces the state to be near a Bell pair — the seed of self-testing, later formalized by Mayers and Yao. Tsirelson worked broadly in probability and functional analysis (Tsirelson's space in Banach-space theory is his too), but in quantum information his name marks the one inequality that every certification, DIQKD, and randomness-expansion protocol quietly depends on.

History

1964

— Bell derives an inequality that local hidden variables cannot violate

1969

— Clauser, Horne, Shimony and Holt cast Bell's bound in testable CHSH form

1980

— Tsirelson proves the quantum ceiling S ≤ 2√2

1998

— Mayers and Yao introduce self-testing of an uncharacterized quantum apparatus

2015

— loophole-free Bell tests close the locality and detection loopholes together

Going further

Texts. Nielsen and Chuang, Quantum Computation and Quantum Information, §2.6 (Bell inequalities) · Preskill, Lecture Notes on Quantum Information, Ch. 4 · Watrous, The Theory of Quantum Information, chapter on nonlocal games · Wilde, Quantum Information Theory, 2nd ed., chapter on entanglement · Kitaev, Shen and Vyalyi, Classical and Quantum Computation · Mermin, Quantum Computer Science. The local and separable bound $|S|\le2$ is established via the deterministic factorization in Derivation A. The quantum ceiling $\|\mathcal B\|\le2\sqrt2$ follows from $\mathcal B^2=4-[A_0,A_1]\otimes[B_0,B_1]$ , as shown in Derivation B. The optimal-angle singlet value $4/\sqrt2=2\sqrt2$ and the game probability $\cos^2(\pi/8)\approx0.8536$ are worked out in Worked Example 1. The concurrence bound $C\ge\sqrt{S^2/4-1}$ is the Verstraete–Wolf/Horodecki relation, derived in Derivation C from the Schmidt form $T=\mathrm{diag}(\sin2\alpha,-\sin2\alpha,1)$ with singular values $\{1,C,C\}$ . The Eberhard threshold $2(\sqrt2-1)\approx0.828$ , the no-signalling framework, the 2015 loophole-free Bell tests, and the 2022 DIQKD demonstrations are standard physics milestones.

Prerequisites. CHSH Inequality and Its Quantum Violation · Bell's Theorem and Local Hidden Variables · Nonlocality and the No-Signalling Principle · Separable versus Entangled States · Tsirelson's Bound. Leads to. PR Boxes and Super-Quantum Correlations · Concurrence and Wootters' Formula · Entanglement Witnesses.

Ask the tutor: Ask me to walk through how an observed CHSH value of your choosing turns into a certified concurrence and singlet fidelity, or how closing each loophole changes what the number is allowed to claim.

What exactly makes a certification "device-independent"?

The conclusion uses nothing about the internals of the apparatus — not the Hilbert-space dimension, not what observables the detectors measure, not how the state was prepared. The only inputs are the classical setting labels you chose and the classical outcome labels the boxes returned, summarized in the table $p(a,b\mid x,y)$ , plus the physical fact of no-signalling. If those statistics violate a Bell inequality, entanglement follows as a theorem, so even boxes built by an adversary who wants to fool you cannot fake the certificate. This is the sharp contrast with an entanglement witness, which certifies entanglement only under the assumption that you measured the operators you think you did.

If violation proves entanglement, why doesn't non-violation prove separability?

Because the CHSH test is one-sided. A violation is a sufficient condition for entanglement, but not a necessary one: there exist entangled states — the classic examples are Werner states below a critical visibility — that admit an explicit local-hidden-variable model and therefore never exceed $S=2$ . Such a state is genuinely entangled yet produces a "local-looking" click table for CHSH. So $S\le2$ is consistent with both separable and entangled states, and you may conclude nothing from it. Certification runs in one direction only: clicks above $2$ certify; clicks below $2$ are silent.

How can a number computed from clicks pin down an uncharacterized quantum state?

Through rigidity. Tsirelson's bound is not merely an inequality but a tight one: the value $2\sqrt2$ is attainable by an essentially unique configuration — a maximally entangled two-qubit state measured with anticommuting observables — up to a local isometry that you can never detect and never need to. So an observed value near $2\sqrt2$ leaves the boxes almost no freedom; they must contain an embedded copy of a Bell pair. Self-testing makes this quantitative and robust: equation (6) turns the gap $2\sqrt2-S$ into a lower bound on the singlet fidelity. The clicks certify the object because the optimal correlations could have come from almost nothing else.

Why is the no-signalling assumption unavoidable, and what breaks without it?

No-signalling — Alice's marginals not depending on Bob's setting, and vice versa — is the minimal statement that the two boxes are genuinely separate parties with freely chosen inputs. It is enforced physically by space-like separation: if the measurement events are causally disconnected, no influence can pass between them within the run. Drop it and the certificate is worthless — a hidden channel lets one box learn the other's setting and coordinate outcomes, and such a local-plus-communication strategy can reach any $S$ up to the algebraic maximum $4$ with no entanglement at all. This is precisely the locality loophole, and closing it is why the 2015 experiments went to such lengths over timing and distance.

What is the difference between certifying entanglement and quantifying it?

Certifying is the yes/no verdict: $|S|>2$ says "entangled." Quantifying is the stronger, and more useful, claim: the magnitude of the violation lower-bounds a specific entanglement measure. Equation (5) turns an observed $S$ into a certified concurrence $C\ge\sqrt{S^2/4-1}$ , and self-testing turns it into a certified singlet fidelity. This is what makes device-independent estimation a real alternative to tomography: you report not just that entanglement is present but at least how much, and that number is immune to calibration errors and measurement misalignment that would corrupt a witness-based estimate.

Why does detection efficiency have a hard threshold near 0.83?

Because an adversary — or simply an unlucky detector — that controls which runs get recorded can bias the reported statistics. If many photons go undetected, a purely local model can arrange to "answer" only on the runs where its predetermined strategy happens to look nonlocal, discarding the rest, and thereby post-select a fake violation above $2$ . Eberhard showed that for symmetric CHSH on a maximally entangled pair this trick works whenever the single-detector efficiency falls below $\eta_{\text{crit}}=2(\sqrt2-1)\approx0.828$ . Above the threshold no such local fake exists, so the fair-sampling assumption can be dropped and the claim becomes genuinely device-independent. Below it, the detection loophole is open and no violation, however large, is trustworthy.