Self-Testing of Quantum States

Every laboratory certification of a quantum state normally rests on a leap of faith: you trust that your detectors measure the observables written on their front panels, and tomography reconstructs a state only after you grant that. Self-testing tears up that contract. The devices become sealed black boxes with two knobs and two lights — an experimenter dials in a classical setting, reads out a classical outcome, and from the bare correlation table $p(a,b\mid x,y)$ alone must deduce the quantum state and measurements hidden inside. Nothing about dimension, purity, or detector alignment is assumed; only quantum theory and the spatial separation of the two labs. The astonishing fact is that this is sometimes enough. Certain correlations — the ones saturating a Bell inequality at its quantum maximum — admit an essentially unique quantum realization. Observing the CHSH value $2\sqrt{2}$ , Tsirelson's bound, forces the shared state to be a maximally entangled qubit pair and the measurements to be anticommuting Pauli-like observables. This rigidity of nonlocal correlations, first formalized by Mayers and Yao, is the strongest certification known — device- and dimension-independent at once. One caveat can never be removed — correlations are blind to appending an unmeasured ancilla and to applying a local unitary in each lab — so the conclusion always reads "up to a local isometry, plus an undetectable junk register." Real data never saturate the bound exactly, so the usable theory is robust self-testing: a near-maximal violation $2\sqrt{2}-\varepsilon$ still pins the real state within $O(\sqrt{\varepsilon})$ of the ideal. That single fact turns a Bell test into a hardware guarantee — the load-bearing wall beneath device-independent cryptography, certified randomness, and the verification of quantum computers.

Key takeaways

Self-testing is the strongest device-independent certification: from the raw input–output statistics $p(a,b\mid x,y)$ alone, a black-box quantum device is shown to contain a specific entangled state and specific measurements.
The flagship example is the CHSH game — observing the quantum maximum $2\sqrt{2}$ certifies a maximally entangled qubit pair $\lvert\Phi^+\rangle$ measured with anticommuting Pauli observables.
The certification is intrinsically up to a local isometry: local unitaries, embeddings into larger Hilbert spaces, and unentangled ancillas are freedoms no correlation can ever resolve.
It is robust — a violation $2\sqrt{2}-\varepsilon$ forces the real objects to be $O(\sqrt{\varepsilon})$ -close to ideal, so noisy real-world data still certify.
It assumes nothing about the devices, only quantum theory and no-signalling between labs, which is why it underpins device-independent QKD, randomness expansion, and delegated verification of quantum computation.
Key idea: maximal violation of a Bell inequality is a fingerprint — the CHSH value $2\sqrt{2}$ can only arise from a maximally entangled qubit pair with anticommuting measurements, up to a local isometry and an unmeasured junk system, so the statistics alone certify the state and the measurements.

The core insight

Maximal violation of a Bell inequality is a fingerprint: the CHSH value $2\sqrt{2}$ can only be produced by a maximally entangled qubit pair with anticommuting measurements, up to a local isometry and an unmeasured junk system — so the raw statistics alone certify both the quantum state and the measurements.

Where this lives today

Self-testing has crossed from theorem to instrument. Loophole-free Bell tests in electron spins, trapped ions, and entangled photons now report CHSH values close enough to $2\sqrt2$ that robust bounds certify device-independent state fidelities without trusting any detector. On that foundation sit three flagship applications. In device-independent QKD, a certified near-maximally-entangled pair means any eavesdropper's purification is decoupled from the key. In certified randomness expansion, the rigidity of near-optimal statistics lower-bounds the min-entropy of the outputs against any side information, turning a short trusted seed into a long guaranteed-random string. And in the verification of quantum computation, a classical verifier self-tests qubits, Bell pairs, and gates to confirm an untrusted prover really ran a quantum computation — the route behind interactive proofs for BQP and the $\mathrm{MIP}^*=\mathrm{RE}$ result. Parallel self-testing, certifying many Bell pairs at once with graceful overhead, is the frontier that makes verifying a whole quantum computer conceivable.

The physical picture

Picture two sealed boxes wired to nothing but a random-number generator and a printer. Alice's box has a knob $x\in\{0,1\}$ and prints $a=\pm1$ ; Bob's, far away, has a knob $y\in\{0,1\}$ and prints $b=\pm1$ . A hidden source once sent each a share of some quantum system, and now they are on their own. The self-testing claim is that if the accumulated correlation table gives the CHSH combination its quantum-maximal value, then — without ever opening a box — you already know what is inside: a maximally entangled pair of qubits, measured along directions $45^\circ$ apart.

Figure

The device-independent scenario: a hidden source sends shares to two sealed boxes. Alice dials setting x and reads outcome a = ±1; Bob dials y and reads b = ±1. Only the correlations of outcomes given settings are observed, from which the CHSH combination ⟨𝓑⟩ = ⟨A₀B₀⟩ + ⟨A₀B₁⟩ + ⟨A₁B₀⟩ − ⟨A₁B₁⟩ is computed

Why is the maximal score a fingerprint and not just a good score? Because Tsirelson's bound is saturated only when the local observables obey strict algebra — each squares to the identity and the two on each side anticommute on the shared state, exactly the defining relations of the Pauli operators $Z$ and $X$ . The statistics do not merely suggest a qubit; they force one to exist on the support of whatever high-dimensional state the boxes hold, and the SWAP gadget of Derivation 3 pulls the certified Bell pair out into a clean register. The second figure is where the whole subject lives: the landscape of CHSH values. Self-testing is not a statement about a region but about a single point — only at the boundary of the quantum set, exactly at $2\sqrt2$ , does the realization become rigid.

Figure

The landscape of CHSH values on one axis: the local-hidden-variable bound 2, the quantum maximum 2√2 ≈ 2.828 (Tsirelson), and the algebraic ceiling 4 (PR box). Self-testing is exact only at the single point 2√2 where the quantum realization becomes rigid; a violation of 2.76 lies fidelity 1 − O(√ε) from ideal with ε = 2√2 − 2.76 ≈ 0.068

The core equations

The CHSH game is built from four two-outcome measurements — two on each side — assembled into a single Hermitian operator whose expectation is the score. Every self-testing statement is written against it.

\mathcal{B} = A_0 B_0 + A_0 B_1 + A_1 B_0 - A_1 B_1 \qquad (1)

where: $A_x$ — Alice's $\pm1$ -valued observable for setting $x\in\{0,1\}$ , Hermitian with $A_x^2=\mathbb{1}$ ; $B_y$ — Bob's $\pm1$ -valued observable, on Bob's tensor factor, commuting with every $A_x$ ; $\mathcal{B}$ — the CHSH operator, whose expectation is the Bell value. Reading it: three terms enter with a plus and one with a minus, the asymmetry that lifts the quantum bound above the classical one. Whatever the box dimension, the quantum value never reaches the algebraic ceiling $4$ ; it stops at Tsirelson's bound, and how it stops is the whole story.

\langle\mathcal{B}\rangle_{\psi} = \langle\psi\rvert\mathcal{B}\lvert\psi\rangle \le 2\sqrt{2} \qquad (2)

where: $\lvert\psi\rangle$ — the shared bipartite state on $\mathcal{H}_A\otimes\mathcal{H}_B$ of unknown dimension; $2\sqrt{2}\approx2.828$ — the quantum maximum, above the local bound $2$ and below the algebraic bound $4$ . Reading it: exceeding $2$ already witnesses entanglement, but only $2\sqrt{2}$ is rigid. Squaring the operator makes the bound transparent: with $A_x^2=B_y^2=\mathbb{1}$ and $[A_x,B_y]=0$ , everything collapses to a single commutator product — the Landau identity.

\mathcal{B}^2 = 4\,\mathbb{1} - [A_0,A_1]\otimes[B_0,B_1] \qquad (3)

where: $[A_0,A_1]=A_0A_1-A_1A_0$ — Alice's commutator, vanishing exactly when her observables commute; $\mathbb{1}$ — identity on the joint space. Reading it: each commutator has norm at most $2$ , so the correction term has norm at most $4$ , giving $\lVert\mathcal{B}^2\rVert\le8$ and $\lVert\mathcal{B}\rVert\le2\sqrt{2}$ . The sign of this term follows the minus in (1) and does not affect the bound. Saturation is where certification is born. Reaching $2\sqrt{2}$ forces the correction term to act with maximal magnitude on $\lvert\psi\rangle$ , which makes the observables on each side anticommute — not everywhere, but on the state.

\langle\mathcal{B}\rangle=2\sqrt{2}\ \Longrightarrow\ \{A_0,A_1\}\lvert\psi\rangle=0,\quad \{B_0,B_1\}\lvert\psi\rangle=0 \qquad (4)

where: $\{A_0,A_1\}=A_0A_1+A_1A_0$ — the anticommutator; $\lvert\psi\rangle$ — the optimal state on which the relation is enforced. Reading it: an observable squaring to $\mathbb{1}$ and anticommuting with a partner is a Pauli operator on its subspace — the statistics have conjured a qubit. The certification is the existence of a local isometry that extracts the ideal pair and quarantines the unobservable into a junk register.

\Phi_A\otimes\Phi_B\,\lvert\psi\rangle = \lvert\Phi^+\rangle_{A'B'}\otimes\lvert\xi\rangle_{AB} \qquad (5)

where: $\Phi_A,\Phi_B$ — local isometries (the SWAP gadgets) acting only within each lab; $\lvert\Phi^+\rangle=\tfrac{1}{\sqrt2}(\lvert00\rangle+\lvert11\rangle)$ — the ideal maximally entangled qubit pair on fresh certified registers $A'B'$ ; $\lvert\xi\rangle$ — the junk state, carrying every degree of freedom the statistics cannot probe. Reading it: "certified" never means "equal to" — it means "isometric to, times junk." Because real experiments only approach $2\sqrt{2}$ , the usable statement is a continuity bound: a small shortfall $\varepsilon$ costs only $O(\sqrt{\varepsilon})$ in fidelity.

F\big(\Phi(\psi),\,\lvert\Phi^+\rangle\otimes\xi\big)\ \ge\ 1 - O(\sqrt{\varepsilon}),\qquad \varepsilon = 2\sqrt{2}-\langle\mathcal{B}\rangle \qquad (6)

where: $F$ — fidelity between the isometry image and the ideal target; $\varepsilon$ — the observed value's deviation from the quantum maximum; $O(\sqrt{\varepsilon})$ — the square-root robustness near the optimum, sharpened to a linear $O(\varepsilon)$ by optimized analyses. Reading it: this single inequality is what makes self-testing an experimental tool, not an idealization. Finally, the abstract qubit is made concrete by naming the effective Pauli operators built from the raw black-box observables.

Z_A = A_0,\quad X_A = A_1,\qquad Z_B=\tfrac{B_0+B_1}{\sqrt2},\quad X_B=\tfrac{B_0-B_1}{\sqrt2} \qquad (7)

where: $Z_A,X_A$ — Alice's effective Paulis, anticommuting on $\lvert\psi\rangle$ ; $Z_B,X_B$ — Bob's, symmetrized so $\{Z_B,X_B\}\lvert\psi\rangle=0$ and each squares to $\mathbb{1}$ on the state. Reading it: on the optimal strategy these reduce to $Z,X$ , and $\lvert\psi\rangle$ becomes the common $+1$ eigenstate of $Z_AZ_B$ and $X_AX_B$ — the stabilizer definition of $\lvert\Phi^+\rangle$ .

Derivation ATsirelson's bound from the squared CHSH operator

The goal is to derive the quantum maximum $2\sqrt{2}$ from nothing but operator algebra — no explicit qubits, no chosen dimension — and to expose the condition under which it is saturated, because that condition is the engine of self-testing. We use only that each observable squares to the identity, that Alice's and Bob's operators commute because the labs are separated, and that $\mathcal{B}$ is Hermitian. The payoff is structural: equality holds only when a commutator reaches its maximal norm, the fingerprint that later pins down a qubit.

Step 1. Square the operator. Writing $\mathcal{B}=A_0(B_0+B_1)+A_1(B_0-B_1)$ and expanding, the four diagonal terms use $A_x^2=B_y^2=\mathbb{1}$ to give $(B_0+B_1)^2+(B_0-B_1)^2=4\,\mathbb{1}$ , while the cross terms assemble into $-[A_0,A_1]\otimes[B_0,B_1]$ , yielding $\mathcal{B}^2=4\,\mathbb{1}-[A_0,A_1]\otimes[B_0,B_1]$ . Why: the square is where the four measurements interact, and the commutator bookkeeping is what quantum theory adds beyond the classical count.

Step 2. Bound each commutator. For Hermitian unitaries, $\lVert[A_0,A_1]\rVert\le\lVert A_0A_1\rVert+\lVert A_1A_0\rVert=2$ since each product is unitary with norm $1$ ; the same holds for Bob. Why: this is the one place the $\pm1$ -valued nature of the measurements enters.

Step 3. Combine the norms. The correction term is a tensor product of two operators each of norm at most $2$ , so $\lVert[A_0,A_1]\otimes[B_0,B_1]\rVert\le4$ and $\lVert\mathcal{B}^2\rVert\le4+4=8$ . Why: submultiplicativity turns two local bounds into one global bound on the squared score.

Step 4. Take the square root. Because $\mathcal{B}$ is Hermitian, $\lVert\mathcal{B}\rVert^2=\lVert\mathcal{B}^2\rVert\le8$ , so $\lVert\mathcal{B}\rVert\le2\sqrt2$ and $\langle\mathcal{B}\rangle\le2\sqrt2$ in any state. Why: a Hermitian operator's norm is the largest value its expectation can reach, so the operator bound is the physical bound.

Step 5. Read off saturation. Equality requires $\langle\psi\rvert[A_0,A_1]\otimes[B_0,B_1]\lvert\psi\rangle=-4$ , its extreme value, possible only if each commutator acts as a norm- $2$ operator on the state — equivalently $\{A_0,A_1\}\lvert\psi\rangle=0$ and $\{B_0,B_1\}\lvert\psi\rangle=0$ . Why: this converts a scalar equality about the score into the qubit algebra self-testing exploits. A direct check confirms attainability: the optimal qubit strategy of Worked Example 1 gives $\langle\mathcal{B}\rangle=4/\sqrt2=2\sqrt2$ .

Go deeper · Derivation B · 2 minAnticommutation forces a virtual qubit

Derivation BAnticommutation forces a virtual qubit

The goal is to show that the relations forced by saturation — each observable squaring to the identity and anticommuting with its partner on the state — are exactly the defining relations of the Pauli algebra, and that a state carrying that algebra on both sides is uniquely the maximally entangled pair. This is the structural heart of self-testing, what upgrades "the score is maximal" into "the state is $\lvert\Phi^+\rangle$ ," and we never choose a basis or dimension for the boxes.

Step 1. Restate saturation as anticommutation. From Derivation 1, $\{A_0,A_1\}\lvert\psi\rangle=0$ and $\{B_0,B_1\}\lvert\psi\rangle=0$ ; combined with $A_x^2=\mathbb{1}$ these are the relations of a Pauli pair on the support of $\lvert\psi\rangle$ . Why: two anticommuting Hermitian involutions generate the single-qubit Pauli algebra, so a qubit is guaranteed inside whatever the boxes contain.

Step 2. Name Alice's effective Paulis. Set $Z_A:=A_0$ , $X_A:=A_1$ ; they satisfy $Z_A^2=X_A^2=\mathbb{1}$ and $\{Z_A,X_A\}\lvert\psi\rangle=0$ , the algebra of $Z$ and $X$ on the relevant subspace. Why: no rotation is needed on Alice's side — her raw observables already play $Z$ and $X$ .

Step 3. Regularize Bob's observables. Bob's optimal settings are $45^\circ$ rotations, so define $Z_B:=(B_0+B_1)/\sqrt2$ , $X_B:=(B_0-B_1)/\sqrt2$ ; on the optimal state these square to $\mathbb{1}$ and anticommute, defining Bob's virtual qubit. Why: the raw $B_0,B_1$ are not aligned with $Z,X$ , but their sum and difference are.

Step 4. Impose the correlation conditions. Maximal violation forces $Z_A\otimes Z_B\lvert\psi\rangle=\lvert\psi\rangle$ and $X_A\otimes X_B\lvert\psi\rangle=\lvert\psi\rangle$ , so $\lvert\psi\rangle$ is a $+1$ eigenstate of both $ZZ$ and $XX$ . Why: a stabilizer pair is a far stronger constraint than a single expectation value — it fixes a unique state.

Step 5. Identify the state. In $\mathbb{C}^2\otimes\mathbb{C}^2$ the only common $+1$ eigenstate of $ZZ$ and $XX$ is $\lvert\Phi^+\rangle=\tfrac{1}{\sqrt2}(\lvert00\rangle+\lvert11\rangle)$ , so on the certified factor $\lvert\psi\rangle$ is exactly the maximally entangled pair. Why: the Schmidt-flat state's uniqueness as the joint stabilizer eigenstate makes the certification sharp, not merely suggestive. A quick check: $ZZ$ and $XX$ commute, and their shared $+1$ eigenspace in four dimensions is one-dimensional — precisely $\lvert\Phi^+\rangle$ , the flat Schmidt spectrum of a maximally entangled pair.

Go deeper · Derivation C · 1 minThe SWAP isometry that extracts the certified pair

Derivation CThe SWAP isometry that extracts the certified pair

The goal is to turn the abstract "there exists a local isometry" of equation (5) into an explicit, runnable circuit: a gadget built only from the certified black-box observables and one fresh ancilla per lab that swaps the hidden qubit into a trusted register and leaves the carrier as junk. Constructing it turns self-testing from an existence theorem into a procedure a verifier can apply — the same primitive that later certifies gates and whole computations. Each party appends $\lvert0\rangle$ and applies a Hadamard-controlled sequence keyed to their effective $Z$ and $X$ .

Step 1. Prepare the ancillas. Map $\lvert\psi\rangle\to\lvert\psi\rangle\otimes\lvert0\rangle_{A'}\lvert0\rangle_{B'}$ by appending one fresh qubit to each lab. Why: the certified pair needs somewhere clean to land, and appending an ancilla is one of the operations correlations can never detect, so it costs nothing.

Step 2. Apply the SWAP gadget. On Alice's side run $H_{A'}$ , controlled- $X_A$ , $H_{A'}$ , controlled- $Z_A$ — a circuit whose gates are the black-box observables themselves — and mirror it on Bob's side. Why: using the certified operators as gates keeps the isometry local and device-independent; it never opens the box.

Step 3. Check on the ideal strategy. On the qubit strategy $\Phi_A\otimes\Phi_B$ sends $\lvert\Phi^+\rangle\otimes\lvert00\rangle_{A'B'}$ to $\lvert\mathrm{junk}\rangle_{AB}\otimes\lvert\Phi^+\rangle_{A'B'}$ : the entanglement moves into the fresh registers, the carriers factor off. Why: verifying on the known case proves the circuit works before trusting it on unknown boxes.

Step 4. Extend to the unknown state. Because the algebra of Derivation 2 holds regardless of the hidden dimension, the same circuit gives $\Phi(\lvert\psi\rangle)=\lvert\xi\rangle_{AB}\otimes\lvert\Phi^+\rangle_{A'B'}$ — the certification statement (5). Why: the gadget depends on the boxes only through certified relations, so it succeeds for every maximal-scoring realization, whatever its size.

Step 5. Carry the measurements through. The same isometry intertwines the raw observables with the ideal Paulis, $\Phi(A_x\otimes B_y\lvert\psi\rangle)=(\sigma_x\otimes\sigma_y)\lvert\Phi^+\rangle\otimes\lvert\xi\rangle$ , certifying state and measurements together. Why: a state certificate without a measurement certificate is useless for cryptography — you must know the boxes measure Paulis. The construction is the standard teleportation-style extraction, confirming the clean junk factorization of equation (5).

Limiting cases

Self-testing lives at the quantum boundary, and its character changes sharply as the Bell value moves.

Case	Result	Meaning
Classical / local bound	$\langle\mathcal{B}\rangle\le 2$	Reproducible by local hidden variables; entanglement need not be present and many strategies give identical statistics, so nothing is self-tested.
Quantum maximum	$\langle\mathcal{B}\rangle = 2\sqrt{2}$	Tsirelson's bound. Saturation forces, up to local isometry, the maximally entangled qubit pair with anticommuting observables — the exact self-test.
Near-optimal (robust)	$\langle\mathcal{B}\rangle = 2\sqrt{2}-\varepsilon$	Fidelity to the ideal is at least $1-O(\sqrt{\varepsilon})$ ; the square-root scaling makes certification survive realistic noise.
No-signalling / PR box	$\langle\mathcal{B}\rangle = 4$	The algebraic ceiling, reached only by a super-quantum Popescu–Rohrlich box; not realizable in quantum theory, so it has no quantum self-testing statement.

Worked examples

1 · The optimal strategy attains Tsirelson's bound (fully worked)

Alice measures $A_0=Z$ and $A_1=X$ ; Bob measures $B_0=(Z+X)/\sqrt2$ and $B_1=(Z-X)/\sqrt2$ ; the shared state is $\lvert\Phi^+\rangle$ . Compute the CHSH value and confirm the boxes saturate the bound.

Step 1. Record the two-qubit correlators of $\lvert\Phi^+\rangle$ : it is a $+1$ eigenstate of $Z\otimes Z$ and $X\otimes X$ and gives zero for mixed products, so $\langle Z\otimes Z\rangle=\langle X\otimes X\rangle=1$ while $\langle Z\otimes X\rangle=\langle X\otimes Z\rangle=0$ .

Step 2. Evaluate the four correlators by linearity in $B_0,B_1$ : $\langle A_0B_0\rangle=\tfrac{1}{\sqrt2}(\langle ZZ\rangle+\langle ZX\rangle)=\tfrac{1}{\sqrt2}$ , and likewise $\langle A_0B_1\rangle=\tfrac{1}{\sqrt2}$ , $\langle A_1B_0\rangle=\tfrac{1}{\sqrt2}$ , while $\langle A_1B_1\rangle=\tfrac{1}{\sqrt2}(\langle XZ\rangle-\langle XX\rangle)=-\tfrac{1}{\sqrt2}$ .

Step 3. Assemble the score. With the minus on the last term, $\langle\mathcal{B}\rangle=\tfrac{1}{\sqrt2}+\tfrac{1}{\sqrt2}+\tfrac{1}{\sqrt2}-\left(-\tfrac{1}{\sqrt2}\right)=\dfrac{4}{\sqrt2}=\mathbf{2\sqrt2\approx2.828}$ , the ceiling of equation (2).

Step 4. Confirm rigidity. Since $A_0=Z$ and $A_1=X$ anticommute and $B_0,B_1$ symmetrize to $Z_B=Z,\ X_B=X$ , the state is stabilized by $Z_AZ_B$ and $X_AX_B$ — the unique such state is $\lvert\Phi^+\rangle$ . So the maximal value $2\sqrt2$ certifies both the Bell pair and the anticommuting measurements, with no assumption on dimension.

2 · From a real violation to a certified fidelity

A loophole-free Bell experiment reports a CHSH value $S=2.76$ . Translate this into a device-independent statement about the produced state.

Step 1. Compute the shortfall: $\varepsilon=2\sqrt2-S=2.82843-2.76=\mathbf{0.0684}$ . This is the only experimental input the robustness bound needs.

Step 2. Take the square root that governs the scaling: $\sqrt{\varepsilon}=\sqrt{0.0684}=\mathbf{0.262}$ . Equation (6) reads $F\ge1-O(0.262)$ , so the extracted state is close to $\lvert\Phi^+\rangle$ up to the isometry.

Step 3. Sharpen with Kaniewski's linear self-test, which replaces the generic $\sqrt{\varepsilon}$ by an $O(\varepsilon)$ slope and gives fidelities well above $0.9$ for $S=2.76$ . The single number $S=2.76$ therefore certifies a near-maximally-entangled qubit pair inside untrusted hardware, exactly what device-independent QKD consumes.

Physics of the portrait

Self-testing is the constructive shadow of a bound Boris Tsirelson wrote down decades before anyone asked it to certify hardware — the ceiling $2\sqrt2$ whose saturation is rigid enough to name the state.

Origins · 1 minBoris Tsirelson and how this was discovered

Boris Tsirelson

1950–2020 · quantum nonlocality and the bound that carries his name

Boris Semyonovich Tsirelson, born in Leningrad in 1950 and later a professor at Tel Aviv University, was a mathematical physicist whose 1980 paper "Quantum generalizations of Bell's inequality" quietly founded a subject. Where John Bell had shown that quantum correlations can exceed the local-hidden-variable bound $2$ , Tsirelson asked how far quantum theory can go. His answer, $\lVert\mathcal{B}\rVert\le2\sqrt2$ , is now called Tsirelson's bound, and his proof — squaring the operator and bounding the commutator — is Derivation 1 of this article. Crucially he characterized the entire set of quantum correlations as a convex body whose CHSH boundary is reached only by essentially unique states and measurements. That uniqueness is precisely the rigidity later authors, from Summers and Werner to Mayers and Yao, turned into self-testing. Tsirelson's problem — whether the commuting-operator and tensor-product definitions of quantum correlations agree — was resolved only in 2020, in the negative via the $\mathrm{MIP}^*=\mathrm{RE}$ theorem, the same year he died. His bound sits where quantum theory is most quantum, and self-testing is the discovery that there the theory hands you a blueprint of the hardware.

History

1964

— Bell shows local hidden variables cap correlations at the value 2

1980

— Tsirelson proves the quantum bound 2√2 and characterizes the quantum correlation set

1987

— Summers and Werner establish maximal violation is achieved by essentially unique states

1992

— Popescu and Rohrlich define the super-quantum box saturating the algebraic value 4

2004

— Mayers and Yao formalize self-testing of a quantum apparatus

2020

— Šupić and Bowles survey the field as a mature device-independent toolkit

Going further

Texts. Nielsen & Chuang, Quantum Computation and Quantum Information (Ch. 2, Bell inequalities) · Preskill, Lecture Notes on Quantum Computation (Ch. 4, nonlocality) · Watrous, The Theory of Quantum Information (nonlocal games) · Wilde, Quantum Information Theory · Kitaev, Shen & Vyalyi, Classical and Quantum Computation · Mermin, Quantum Computer Science · Mayers & Yao, "Self-testing quantum apparatus" (2004) · Šupić & Bowles, "Self-testing of quantum systems: a review" (2020) · Kaniewski, "Analytic and nearly optimal self-testing bounds for the CHSH inequality" (2016). A few sign and value conventions are worth keeping in mind. The Landau identity $\mathcal{B}^2=4\,\mathbb{1}-[A_0,A_1]\otimes[B_0,B_1]$ holds for $\mathcal{B}=A_0B_0+A_0B_1+A_1B_0-A_1B_1$ — the correction term carries a minus sign with this operator convention, though the norm bound $\lVert\mathcal{B}\rVert\le2\sqrt2$ is sign-independent. The optimal qubit strategy $A_0=Z,A_1=X,B_{0,1}=(Z\pm X)/\sqrt2$ on $\lvert\Phi^+\rangle$ attains $2\sqrt2\approx2.828$ via the four correlators $\pm1/\sqrt2$ , and $\lvert\Phi^+\rangle$ is the unique common $+1$ eigenstate of $ZZ$ and $XX$ . The three landmark values are: local $2$ , quantum $2\sqrt2$ , algebraic $4$ . The $O(\sqrt{\varepsilon})$ robustness bound is from Mayers–Yao and Šupić–Bowles; the linear $O(\varepsilon)$ improvement is due to Kaniewski (2016).

Prerequisites. CHSH Inequality and Its Quantum Violation · Tsirelson's Bound · Bell States and EPR Pairs · Bell's Theorem and Local Hidden Variables · Nonlocality and the No-Signalling Principle · Schmidt Decomposition. Leads to. Device-Independent QKD · Certified Randomness Expansion · Verification of Quantum Computation · Entanglement Witnesses.

Ask the tutor: Hand me any CHSH value or measurement strategy and I will walk you through what it certifies — the anticommutation it forces, the qubit it extracts, and how close the real state must be to a Bell pair.

What is the difference between self-testing and ordinary state tomography?

Tomography reconstructs an unknown state from data, but it trusts the measurements: you must know your detectors implement the observables you think they do. Self-testing removes that trust — the devices are black boxes and only the classical correlations are used. The price of assuming so little is that you can only certify near-maximally-violating statistics, up to a local isometry. The reward is that the conclusion holds even if an adversary built the hardware — which is why self-testing is the certification layer under device-independent cryptography.

Why does the value 2√2, and not any value above 2, give a unique realization?

Exceeding the local bound $2$ already witnesses entanglement, but many different states and measurements produce the same intermediate value, so nothing specific is certified. Uniqueness appears only at the extreme point $2\sqrt2$ . There the operator inequality of Derivation 1 must be saturated, which forces the observables to anticommute on the state, which forces the Pauli algebra, which forces $\lvert\Phi^+\rangle$ . Rigidity is a boundary phenomenon.

What exactly is the "junk" register, and can I get rid of it?

The junk $\lvert\xi\rangle$ holds every detail the statistics cannot see: the extra dimensions of the boxes, phases fixed by local unitaries, any unmeasured ancilla the source included. You cannot remove it: doing so would mean distinguishing states with identical correlations, impossible by construction. The isometry deposits the clean certified Bell pair in a fresh register $A'B'$ and leaves the junk in the carriers — it is the part of reality the device-independent setting declares unknowable.

Is self-testing robust enough to use on real, noisy experiments?

Yes — that is the point of robust self-testing. Equation (6) says a shortfall $\varepsilon=2\sqrt2-\langle\mathcal{B}\rangle$ costs only $O(\sqrt{\varepsilon})$ in fidelity, and optimized analyses (Kaniewski) sharpen this to a linear $O(\varepsilon)$ slope. In practice a loophole-free violation of $S\approx2.7$ – $2.8$ certifies fidelities well above the thresholds needed for key generation and randomness — without robustness, the demand for exact saturation would make the theory experimentally empty.

Does self-testing require knowing the dimension of the quantum systems?

No, and this is one of its most striking features. The whole argument — Landau identity, anticommutation, stabilizer identification, SWAP isometry — never chooses a dimension for $\mathcal{H}_A,\mathcal{H}_B$ . Whether the boxes hold qubits, qutrits, or infinite-dimensional systems, saturating CHSH forces a qubit onto the support of the state, and the isometry extracts exactly that qubit. Dimension-independence is why self-testing assumes strictly less than any dimension-bounded method.

How does self-testing connect to the MIP*=RE theorem?

Self-testing is the physics engine inside that result. Interactive proofs with entangled provers rely on the verifier being able to force the provers into specific quantum strategies using only classical questions — which is exactly rigidity. Parallel self-testing certifies many Bell pairs and gates at once, letting a classical verifier pin down an enormous computation. Pushed to the limit, this makes $\mathrm{MIP}^*$ equal the recursively enumerable languages, resolving Tsirelson's problem in the negative — a curiosity about Bell inequalities turned out to bound the limits of verifiable computation.