Simple Algorithm
Quantum Computing · Rydberg State · 28 min

Tsirelson's Bound

Why quantum mechanics wins CHSH at most cos²(π/8) of the time.

Classical players are stuck at S=2S=2. No-signalling alone would allow S=4S=4. Quantum mechanics stops at exactly 222\sqrt2. Three short proofs show why, and each opens a different door: operator algebra, semidefinite programming, and the search for physical principles that single out quantum theory.

What you’ll learn

  • State Tsirelson's bound precisely, for arbitrary states, dimensions and the commuting-operator model.
  • Prove it via Landau's identity B2=4I−[A0,A1]⊗[B0,B1]\mathcal B^2=4\id-\comm{A_0}{A_1}\otimes\comm{B_0}{B_1} and via a sum-of-squares certificate.
  • Use Tsirelson's vector characterization of correlators and relate XOR games to semidefinite programs and Grothendieck's inequality.
  • Compute the maximal CHSH value of any two-qubit state (Horodecki criterion) and locate the Werner-state thresholds.
  • Describe PR boxes, the geometry of local, quantum and no-signalling correlations, and principles proposed to explain the quantum boundary.

The statement

A quantum strategy for CHSH is a state ρ\rho on HA⊗HB\Hil_A\otimes\Hil_B of any dimension and, for each input, a two-outcome measurement. Encode the measurements as ±1\pm1-valued observables Ax=Π0x−Π1xA_x=\Pi^x_0-\Pi^x_1 and ByB_y: Hermitian, with Ax2=By2=IA_x^2=B_y^2=\id. General two-outcome POVMs reduce to this case by Naimark dilation, which enlarges the local spaces without changing any statistics.

The CHSH value is S=Tr⁡(ρ B)S=\Tr(\rho\,\mathcal B) for the Bell operator

B=A0⊗B0+A0⊗B1+A1⊗B0−A1⊗B1.\mathcal B=A_0\otimes B_0+A_0\otimes B_1+A_1\otimes B_0-A_1\otimes B_1.
ModelCorrelations P(a,b∣x,y)P(a,b\mid x,y)max⁡S\max Smax⁡pwin\max p_{\text{win}}
Local (LHV)∑λp(λ)P(a∣x,λ)P(b∣y,λ)\sum_\lambda p(\lambda)P(a\mid x,\lambda)P(b\mid y,\lambda)223/43/4
QuantumTr⁡[ρ (Πax⊗Πby)]\Tr\big[\rho\,(\Pi^x_a\otimes\Pi^y_b)\big]222\sqrt2cos⁡2(π/8)≈0.854\cos^2(\pi/8)\approx0.854
No-signallingany, as long as each marginal ignores the distant input4411
Three nested sets of correlations, separated by the CHSH value.

Two algebraic proofs

DerivationProof 1: Landau's identity
  1. B=A0⊗(B0+B1)+A1⊗(B0−B1)\mathcal B=A_0\otimes(B_0+B_1)+A_1\otimes(B_0-B_1)

    Regroup, exactly as in the classical proof.

  2. (B0+B1)2+(B0−B1)2=2B02+2B12=4 I(B_0+B_1)^2+(B_0-B_1)^2=2B_0^2+2B_1^2=4\,\id

    The cross terms ±{B0,B1}\pm\acomm{B_0}{B_1} cancel.

  3. (B0+B1)(B0−B1)=−[B0,B1],(B0−B1)(B0+B1)=[B0,B1](B_0+B_1)(B_0-B_1)=-\comm{B_0}{B_1},\qquad(B_0-B_1)(B_0+B_1)=\comm{B_0}{B_1}

    Expand using By2=IB_y^2=\id.

  4. B2=4 I−A0A1⊗[B0,B1]+A1A0⊗[B0,B1]=4 I−[A0,A1]⊗[B0,B1]\mathcal B^2=4\,\id-A_0A_1\otimes\comm{B_0}{B_1}+A_1A_0\otimes\comm{B_0}{B_1}=4\,\id-\comm{A_0}{A_1}\otimes\comm{B_0}{B_1}

    Square, using Ax2=IA_x^2=\id for the diagonal terms.

  5. ∥B∥2=∥B2∥≤4+∥[A0,A1]∥ ∥[B0,B1]∥≤4+2⋅2=8\norm{\mathcal B}^2=\norm{\mathcal B^2}\le4+\norm{\comm{A_0}{A_1}}\,\norm{\comm{B_0}{B_1}}\le4+2\cdot2=8

    B\mathcal B is Hermitian, so ∥B2∥=∥B∥2\norm{\mathcal B^2}=\norm{\mathcal B}^2. Each commutator has norm at most 2∥A0∥∥A1∥=22\norm{A_0}\norm{A_1}=2.

  6. ∣Tr⁡(ρ B)∣≤∥B∥≤22■\abs{\Tr(\rho\,\mathcal B)}\le\norm{\mathcal B}\le2\sqrt2\qquad\blacksquare

    An expectation value never exceeds the operator norm.

DerivationProof 2: a sum of squares
  1. P0=A0⊗I−I⊗B0+B12,P1=A1⊗I−I⊗B0−B12P_0=A_0\otimes\id-\id\otimes\frac{B_0+B_1}{\sqrt2},\qquad P_1=A_1\otimes\id-\id\otimes\frac{B_0-B_1}{\sqrt2}

    Two Hermitian operators built from the strategy.

  2. P02=I−2 A0⊗(B0+B1)+2 I+{B0,B1}2P_0^2=\id-\sqrt2\,A_0\otimes(B_0+B_1)+\frac{2\,\id+\acomm{B_0}{B_1}}{2}

    Expand; Alice's and Bob's operators commute, so the two cross terms add.

  3. P12=I−2 A1⊗(B0−B1)+2 I−{B0,B1}2P_1^2=\id-\sqrt2\,A_1\otimes(B_0-B_1)+\frac{2\,\id-\acomm{B_0}{B_1}}{2}

    The anticommutator now enters with the opposite sign.

  4. 22 I−B=12(P02+P12)⪰02\sqrt2\,\id-\mathcal B=\frac{1}{\sqrt2}\left(P_0^2+P_1^2\right)\succeq0

    Add the two lines and divide by 2\sqrt2. The gap to 222\sqrt2 is a sum of squares of Hermitian operators, hence positive semidefinite in every dimension. Applying the same argument with signs flipped bounds −B-\mathcal B.

  5. ⟨ψ|B|ψ⟩=22  ⟺  P0∣ψ⟩=P1∣ψ⟩=0\mel{\psi}{\mathcal B}{\psi}=2\sqrt2\iff P_0\ket\psi=P_1\ket\psi=0

    For a pure state, equality forces both squares to annihilate ∣ψ⟩\ket\psi: A0∣ψ⟩=B0+B12∣ψ⟩A_0\ket\psi=\tfrac{B_0+B_1}{\sqrt2}\ket\psi and A1∣ψ⟩=B0−B12∣ψ⟩A_1\ket\psi=\tfrac{B_0-B_1}{\sqrt2}\ket\psi. These relations drive the rigidity theorem of Nonlocal Games & Self-Testing.

Check your understanding

Suppose Alice's observables commute, [A0,A1]=0\comm{A_0}{A_1}=0, while Bob's are arbitrary. What is the largest possible ⟨B⟩\expval{\mathcal B}?

  1. 222\sqrt2
  2. 22
  3. 44
  4. It grows with the dimension of Bob's system
Show the answer

B. 22

Landau's identity gives B2=4I\mathcal B^2=4\id, so ∥B∥=2\norm{\mathcal B}=2. Physically, commuting observables can be measured jointly, so Alice effectively has a single measurement, and a party with one setting can always be simulated by a local model.

Vectors, XOR games and two-qubit states

DerivationProof 3: CHSH in two lines of geometry
  1. S=u⃗0⋅(v⃗0+v⃗1)+u⃗1⋅(v⃗0−v⃗1)≤∥v⃗0+v⃗1∥+∥v⃗0−v⃗1∥S=\vec u_0\cdot(\vec v_0+\vec v_1)+\vec u_1\cdot(\vec v_0-\vec v_1)\le\norm{\vec v_0+\vec v_1}+\norm{\vec v_0-\vec v_1}

    Cauchy–Schwarz with unit vectors u⃗x\vec u_x.

  2. ∥v⃗0+v⃗1∥+∥v⃗0−v⃗1∥≤2(∥v⃗0+v⃗1∥2+∥v⃗0−v⃗1∥2)=2⋅4=22\norm{\vec v_0+\vec v_1}+\norm{\vec v_0-\vec v_1}\le\sqrt{2\left(\norm{\vec v_0+\vec v_1}^2+\norm{\vec v_0-\vec v_1}^2\right)}=\sqrt{2\cdot4}=2\sqrt2

    Use a+b≤2(a2+b2)a+b\le\sqrt{2(a^2+b^2)}, then the parallelogram law: ∥v⃗0+v⃗1∥2+∥v⃗0−v⃗1∥2=2∥v⃗0∥2+2∥v⃗1∥2=4\norm{\vec v_0+\vec v_1}^2+\norm{\vec v_0-\vec v_1}^2=2\norm{\vec v_0}^2+2\norm{\vec v_1}^2=4.

  3. Equality requires v⃗0⊥v⃗1\vec v_0\perp\vec v_1, u⃗0∝v⃗0+v⃗1\vec u_0\propto\vec v_0+\vec v_1 and u⃗1∝v⃗0−v⃗1\vec u_1\propto\vec v_0-\vec v_1: two orthonormal pairs rotated by 45°45°, the Bloch-sphere picture of the optimal strategy. Classical strategies correspond to one-dimensional "vectors" ux,vy∈{±1}u_x,v_y\in\{\pm1\}, which give 22.

XOR games. CHSH is an XOR game: the referee checks only a⊕ba\oplus b. For any XOR game with question distribution π\pi and target bits s(x,y)s(x,y), the bias β=2pwin−1\beta=2p_{\text{win}}-1 is ∑x,yπ(x,y)(−1)s(x,y)Exy\sum_{x,y}\pi(x,y)(-1)^{s(x,y)}E_{xy}. Classically one maximizes over Exy=axbyE_{xy}=a_xb_y with ax,by=±1a_x,b_y=\pm1, an NP-hard problem in general. Quantumly, by Tsirelson's theorem, one maximizes over u⃗x⋅v⃗y\vec u_x\cdot\vec v_y: a semidefinite program, solvable in polynomial time (Cleve, Høyer, Toner & Watrous, 2004).

Grothendieck's inequality bounds the gap: for every XOR game β∗/β≤KG\beta^*/\beta\le K_G, the real Grothendieck constant, known only to lie between about 1.6761.676 and 1.7831.783. For CHSH the ratio is 1/21/2=2\tfrac{1/\sqrt2}{1/2}=\sqrt2.

Two-qubit states. Write a state's correlation matrix Tij=Tr⁡[ρ (σi⊗σj)]T_{ij}=\Tr[\rho\,(\sigma_i\otimes\sigma_j)] and take qubit observables Ax=a⃗x⋅σ⃗A_x=\vec a_x\cdot\vec\sigma, By=b⃗y⋅σ⃗B_y=\vec b_y\cdot\vec\sigma. Then ⟨Ax⊗By⟩=a⃗x TT b⃗y\expval{A_x\otimes B_y}=\vec a_x^{\,T}T\,\vec b_y. Writing b⃗0±b⃗1\vec b_0\pm\vec b_1 as 2cos⁡θ c⃗2\cos\theta\,\vec c and 2sin⁡θ c⃗ ′2\sin\theta\,\vec c^{\,\prime} with orthonormal c⃗,c⃗ ′\vec c,\vec c^{\,\prime} and optimizing gives the Horodecki criterion (1995):

Smax⁡(ρ)=2t12+t22,t1≥t2 the two largest singular values of TS_{\max}(\rho)=2\sqrt{t_1^2+t_2^2},\qquad t_1\ge t_2\ \text{the two largest singular values of } T
A two-qubit state violates CHSH for some choice of measurements iff t12+t22>1t_1^2+t_2^2>1.
Predict before you look

For the Werner state ρp=p ∣Ψ−⟩ ⁣⟨Ψ−∣+(1−p) I4\rho_p=p\,\dyad{\Psi^-}+(1-p)\,\tfrac{\id}{4}, above which value of pp does some CHSH measurement violate S≤2S\le2?

Show the answer

0.71

The singlet has T=−IT=-\id, so T(ρp)=−p IT(\rho_p)=-p\,\id and t1=t2=pt_1=t_2=p. Then Smax⁡=22 p>2S_{\max}=2\sqrt2\,p>2 iff p>1/2≈0.707p>1/\sqrt2\approx0.707.

Beyond quantum: PR boxes

With two inputs and two outputs per party, the no-signalling correlations form a polytope with 24 vertices: the 16 local deterministic boxes and 8 PR boxes (one up to relabelings). The local polytope has the 16 deterministic vertices, and its only non-trivial facets are the 8 CHSH inequalities (Fine, 1982). The quantum set lies strictly in between and is not a polytope: its boundary is curved. In correlator space it is described exactly by

∣arcsin⁡E00+arcsin⁡E01+arcsin⁡E10−arcsin⁡E11∣≤π\big|\arcsin E_{00}+\arcsin E_{01}+\arcsin E_{10}-\arcsin E_{11}\big|\le\pi
Together with the three variants where the minus sign moves to another term (Tsirelson, Landau, Masanes). The optimal strategy saturates it: 3⋅π4−(−π4)=π3\cdot\tfrac\pi4-(-\tfrac\pi4)=\pi.

Super-quantum correlations would have strange consequences. With PR boxes, any distributed Boolean function can be computed with a single bit of communication, so communication complexity collapses (van Dam). Brassard and co-authors (2006) showed the same collapse, probabilistically, for any box that wins CHSH with probability above 3+66≈0.908\tfrac{3+\sqrt6}{6}\approx0.908. Information causality (Pawłowski et al., 2009), the principle that mm bits of communication reveal at most mm bits of the sender's data even with pre-shared correlations, rules out every noisy PR box (mixed with white noise) beyond Tsirelson's bound.

Key takeaways

  • Tsirelson: ∣⟨B⟩∣≤22\abs{\expval{\mathcal B}}\le2\sqrt2 for every state and dimension (and in the commuting-operator model); ω∗(CHSH)=cos⁡2(π/8)\omega^*(\mathrm{CHSH})=\cos^2(\pi/8).
  • Landau's identity B2=4I−[A0,A1]⊗[B0,B1]\mathcal B^2=4\id-\comm{A_0}{A_1}\otimes\comm{B_0}{B_1}: violations need noncommuting observables on both sides, and saturation forces anticommutation.
  • The SOS certificate 22I−B=12(P02+P12)2\sqrt2\id-\mathcal B=\tfrac1{\sqrt2}(P_0^2+P_1^2) is a dual SDP solution; the NPA hierarchy generalizes it to any Bell expression.
  • Correlators are inner products of unit vectors (Tsirelson), so XOR-game values are SDPs and exceed classical ones by at most KGK_G; for two qubits Smax⁡=2t12+t22S_{\max}=2\sqrt{t_1^2+t_2^2}.
  • No-signalling allows S=4S=4 (PR box). Information causality and related principles explain part of the gap, but no known principle singles out the quantum set exactly.

Check yourself

In the interactive lesson these are scored with your confidence, and what you miss comes back for review.

Question 1

Landau's identity reads B2=4I−[A0,A1]⊗[B0,B1]\mathcal B^2=4\id-\comm{A_0}{A_1}\otimes\comm{B_0}{B_1}. Which conclusion follows directly from it?

  1. ∣S∣≤2\abs{S}\le2 whenever Bob's two observables commute
  2. ∣S∣≤2\abs{S}\le2 whenever the shared state is pure
  3. SS can approach 44 in high enough dimension
  4. Every entangled state violates CHSH for suitable observables
Show the answer

A. ∣S∣≤2\abs{S}\le2 whenever Bob's two observables commute

If [B0,B1]=0\comm{B_0}{B_1}=0 the second term vanishes, B2=4I\mathcal B^2=4\id, and ∥B∥=2\norm{\mathcal B}=2. Noncommutativity on both sides is necessary for a violation.

Question 2

A Bell-diagonal two-qubit state has weights 0.7,0,0.2,0.10.7,0,0.2,0.1 on Φ+,Φ−,Ψ+,Ψ−\Phi^+,\Phi^-,\Psi^+,\Psi^-, so T=diag⁡(0.8,−0.6,0.4)T=\diag(0.8,-0.6,0.4). Its fidelity with ∣Φ+⟩\ket{\Phi^+} is 0.7>120.7>\tfrac12, so it is entangled. What is its maximal CHSH value?

  1. 22×0.7≈1.982\sqrt2\times0.7\approx1.98
  2. 22: no violation
  3. 20.64+0.36+0.16≈2.152\sqrt{0.64+0.36+0.16}\approx2.15
  4. 2(0.8+0.6)=2.82(0.8+0.6)=2.8
Show the answer

B. 22: no violation

The two largest singular values are 0.80.8 and 0.60.6, so Smax⁡=20.64+0.36=2S_{\max}=2\sqrt{0.64+0.36}=2. This entangled state cannot violate CHSH with any measurements.

Question 3

Which statement about the Popescu–Rohrlich box is correct?

  1. It allows Alice to signal to Bob
  2. It can be realized with a maximally entangled pair of qutrits
  3. It reaches exactly Tsirelson's bound 222\sqrt2
  4. It reaches S=4S=4 while respecting no-signalling
Show the answer

D. It reaches S=4S=4 while respecting no-signalling

Each marginal is uniform whatever the distant input, yet a⊕b=x∧ya\oplus b=x\wedge y always holds, so S=4S=4. No quantum system in any dimension can realize it, since quantum strategies obey S≤22S\le2\sqrt2.

Question 4

What does the identity 22 I−B=12(P02+P12)2\sqrt2\,\id-\mathcal B=\tfrac{1}{\sqrt2}(P_0^2+P_1^2) establish?

  1. Tsirelson's bound for qubit systems only
  2. That 222\sqrt2 is attained by every entangled state
  3. ⟨B⟩≤22\expval{\mathcal B}\le2\sqrt2 for every state, in every dimension, including the commuting-operator model
  4. Tsirelson's bound for pure states only
Show the answer

C. ⟨B⟩≤22\expval{\mathcal B}\le2\sqrt2 for every state, in every dimension, including the commuting-operator model

A sum of squares of Hermitian operators is positive semidefinite, whatever the Hilbert space. Only Ax2=By2=IA_x^2=B_y^2=\id and the commutation of Alice's and Bob's operators were used.

Further reading

  • B. S. Cirel'son (Tsirelson), "Quantum generalizations of Bell's inequality", Lett. Math. Phys. 4, 93 (1980).
  • S. Popescu & D. Rohrlich, "Quantum nonlocality as an axiom", Found. Phys. 24, 379 (1994).
  • M. Navascués, S. Pironio & A. Acín, "Bounding the set of quantum correlations", Phys. Rev. Lett. 98, 010401 (2007).
  • M. Pawłowski, T. Paterek, D. Kaszlikowski, V. Scarani, A. Winter & M. Żukowski, "Information causality as a physical principle", Nature 461, 1101 (2009).
  • N. Brunner, D. Cavalcanti, S. Pironio, V. Scarani & S. Wehner, "Bell nonlocality", Rev. Mod. Phys. 86, 419 (2014).