1 Motivation and Definitions↩︎

The Greenberger–Horne–Zeilinger (GHZ) theorem [1], [2] shows that quantum mechanics conflicts with local realism in a deterministic way for three or more parties, requiring no inequalities and no averages over runs. A single algebraic identity relating four products of dichotomic outcomes is enough to produce a contradiction [3]–[5]. Although the qualitative content of this identity is celebrated, the quantitative question how strongly does a given state \(\rho\) exhibit GHZ-type correlations? has remained somewhat scattered across the literature, with answers expressed through Mermin polynomials, fidelities with \(\ket{\mathrm{GHZ}}\), the three-tangle [6], or convex-roof extensions of various entanglement monotones [7]. The purpose of this paper is to formulate and analyse a single functional \(I(\vec{n}_1,\vec{n}_2)\) that

1. exactly reproduces the GHZ algebraic paradox when \(\vec{n}_1=\hat{x}\), \(\vec{n}_2=\hat{y}\);

2. satisfies a tight upper bound, \(|I|\le 2\), that is universal across all three-qubit states;

3. saturates this bound only on the GHZ orbit under local unitaries, thereby providing a witness of GHZ-type genuine tripartite entanglement;

4. admits closed-form evaluation on canonical state families (Acín form, W class, biseparable states); and

5. extends naturally from qubits to qudits.

Position with respect to prior work. Our quantity sits at the intersection of three lines of research. First, in contrast to the standard Mermin polynomial [3]–[5], [8] and its genuine-tripartite refinement by Svetlichny and Seevinck [9], [10], the functional \(I\) is multiplicative in the correlators rather than linear; the algebraic identity 2 replaces the inequality. Multiplicative Bell-type quantities have been considered abstractly by Te’eni et al. [11] and (in quadratic form) by Uffink [12], but never specialised to the GHZ stabiliser. Second, in contrast to the inequality-free GHZ-vs-W paradoxes of Cabello [13], our construction yields a single LU-invariant scalar that orders all three-qubit states on a \([0,1]\) scale. Third, in contrast to the three-tangle \(\tau_3\) [6], GHZ fidelity witnesses [14], [15], the genuine multipartite negativity [16], and the GME concurrence of Ma et al. [17], our \(\mathcal{E}_{\mathrm{GHZ}}\) is non-zero on both the W class and biseparable states, with sharp ordering \(\mathcal{E}_{\mathrm{GHZ}}(\mathrm{GHZ})=1>\mathcal{E}_{\mathrm{GHZ}}(\mathrm{W})=35/54> 1/2 \ge \mathcal{E}_{\mathrm{GHZ}}(\textrm{bisep})\).

Notation. Three qubits live in a Hilbert space \(\mathcal{H}=\mathbb{C}^2_A\otimes\mathbb{C}^2_B \otimes\mathbb{C}^2_C\), with state \(\rho^{ABC} \in \mathcal{S}(\mathcal{H})\). Marginals are \(\rho^{AB}=\mathrm{Tr}_C\rho^{ABC}\), etc. For a unit vector \(\vec{n}\in\mathbb{R}^3\), the spin observable on a single qubit is \(\sigma_{\vec{n}}=\vec{n}\cdot\boldsymbol{\sigma}=n_x\sigma_x+n_y\sigma_y+n_z\sigma_z\), with \(\sigma_x,\sigma_y,\sigma_z\) the standard Pauli matrices. For three direction labels \(\vec{n}_a,\vec{n}_b,\vec{n}_c\) we denote the tripartite local observable \[\sigma(\vec{n}_a,\vec{n}_b,\vec{n}_c)\;=\;\sigma_{\vec{n}_a}\otimes\sigma_{\vec{n}_b}\otimes\sigma_{\vec{n}_c},\] and its expectation \(\langle\sigma(\vec{n}_a,\vec{n}_b,\vec{n}_c)\rangle =\mathrm{Tr}[\sigma(\vec{n}_a,\vec{n}_b,\vec{n}_c)\rho^{ABC}]\).

The functional. For two unit vectors \(\vec{n}_1,\vec{n}_2\in\mathbb{R}^3\) we define, using the short-hand \(E_{abc}\equiv\langle\sigma(\vec{n}_{a},\vec{n}_{b},\vec{n}_{c})\rangle\), \[\label{eq:Idef} I(\vec{n}_1,\vec{n}_2) = E_{111} - E_{122} E_{212} E_{221} .\tag{1}\] The quantity \(I\) is a real-valued function of the state \(\rho\) and the two directions; only product correlations of the spin observables on each subsystem appear, so \(I\) is, in principle, accessible to device-independent estimation from local Pauli-string measurements.

Local hidden-variable identity. For any deterministic local hidden-variable (LHV) model assigning values \(A(\vec{n}),B(\vec{n}),C(\vec{n})\in\{\pm1\}\) to the three qubits, every realisation of the underlying parameter \(\lambda\) obeys the algebraic identity \[\begin{gather} \label{eq:LHV-identity} \bigl[A(\vec{n}_1)B(\vec{n}_2)C(\vec{n}_2)\bigr]\bigl[A(\vec{n}_2)B(\vec{n}_1)C(\vec{n}_2)\bigr]\\ \times\bigl[A(\vec{n}_2)B(\vec{n}_2)C(\vec{n}_1)\bigr]=A(\vec{n}_1)B(\vec{n}_1)C(\vec{n}_1), \end{gather}\tag{2}\] because each of \(A(\vec{n}_2)^2,B(\vec{n}_2)^2,C(\vec{n}_2)^2\) equals \(+1\). Hence whenever the three product-correlation factors in \(I\) are simultaneously deterministic at the LHV level, eq. 1 forces \[\label{eq:LHV-zero} I(\vec{n}_1,\vec{n}_2)\,\big|_{\textrm{deterministic LHV}}=0,\tag{3}\] which is the standard GHZ–Mermin equality [2], [3]. For the special case \(\vec{n}_1=\hat{x}\), \(\vec{n}_2=\hat{y}\) on the GHZ state \(\ket{\mathrm{GHZ}}=(\ket{000}+\ket{111})/\sqrt{2}\), quantum mechanics gives \(\langle\sigma(\vec{n}_{1},\vec{n}_{2},\vec{n}_{2})\rangle=\langle\sigma(\vec{n}_{2},\vec{n}_{1},\vec{n}_{2})\rangle=\langle\sigma(\vec{n}_{2},\vec{n}_{2},\vec{n}_{1})\rangle=-1\) and \(\langle\sigma(\vec{n}_{1},\vec{n}_{1},\vec{n}_{1})\rangle=+1\), so \[\label{eq:Ighzminus2} I_{\mathrm{GHZ}}(\hat{x},\hat{y})=1- (-1)^3= 2,\tag{4}\] manifestly violating 3 . The maximal violation \(|I|=2\) is the algebraic essence of the GHZ paradox.

The remainder of this paper studies the functional \(I\) quantitatively: who else can saturate the bound (Sec. 3), what value it takes on a generic three-qubit pure state (Sec. 4), how it gives rise to a genuine tripartite entanglement measure (Sec. 5), and how it lifts to qudits (Sec. 6).

2 The four stabiliser-like operators↩︎

It is convenient to package \(I(\vec{n}_1,\vec{n}_2)\) in terms of four Hermitian observables on \(\mathcal{H}\): \[\label{eq:OOOO} \begin{align} O_1 &= \sigma_{\vec{n}_1}\otimes\sigma_{\vec{n}_2}\otimes\sigma_{\vec{n}_2},\\ O_2 &= \sigma_{\vec{n}_2}\otimes\sigma_{\vec{n}_1}\otimes\sigma_{\vec{n}_2},\\ O_3 &= \sigma_{\vec{n}_2}\otimes\sigma_{\vec{n}_2}\otimes\sigma_{\vec{n}_1},\\ O_4 &= \sigma_{\vec{n}_1}\otimes\sigma_{\vec{n}_1}\otimes\sigma_{\vec{n}_1}. \end{align}\tag{5}\] With \(e_i\equiv\langle O_i\rangle_\rho\) (\(i=1,\dots,4\)), \[I(\vec{n}_1,\vec{n}_2)\;=\;e_4 - e_1 e_2 e_3 .\] Each \(O_i\) is unitary and Hermitian and obeys \(O_i^2=\openone\), hence \(|e_i|\le 1\) and \(|I|\le 2\) trivially.

The algebraic structure of these operators is most transparent in terms of the inner and cross products of the two direction vectors, \(c\equiv\vec{n}_1\cdot\vec{n}_2\) and \(\boldsymbol{m}\equiv\vec{n}_1\times\vec{n}_2\). Using \(\sigma_{\vec{n}_1}\sigma_{\vec{n}_2}=c\,\openone+\mathrm{i}\,\boldsymbol{m}\cdot\boldsymbol{\sigma}\) one finds the following two key identities (proved in App. 8): \[\begin{align} [O_1,O_2] &= 2\mathrm{i}\,c\, \bigl(\boldsymbol{m}\cdot\boldsymbol{\sigma}\!\otimes\!\openone-\openone\!\otimes\!\boldsymbol{m}\cdot\boldsymbol{\sigma}\bigr) \otimes\openone,\tag{6}\\ O_1 O_2 O_3 + O_4 &= 2c\,\sigma_{\vec{n}_1}\otimes\sigma_{\vec{n}_2}\otimes\sigma_{\vec{n}_1}. \tag{7} \end{align}\] Equation 7 is the operator avatar of Mermin’s identity 2 : the right-hand side vanishes exactly when \(\vec{n}_1\perp\vec{n}_2\), in which case \(\{O_1,O_2,O_3,O_4\}\) commute pairwise and satisfy the abelian relation \[\label{eq:GHZ-stabiliser} O_1\,O_2\,O_3\,O_4=-\openone\qquad(\text{when }\vec{n}_1\perp\vec{n}_2).\tag{8}\] This is precisely the defining relation of the three-qubit GHZ stabiliser group [18], [19] written in the locally rotated Pauli basis \((\sigma_{\vec{n}_1},\sigma_{\vec{n}_2})\).

3 Maximum violation: a sharp bound↩︎

Our first main result completely characterises the states and directions saturating \(|I|=2\).

Bell-type interpretation. Proposition 1 is more than a Cirel’son-type bound [20]: the LHV identity 3 predicts \(I=0\) whereas quantum mechanics achieves \(|I|=2\) exactly on the GHZ orbit. In other words, \(I(\vec{n}_1,\vec{n}_2)\) is an equality-form GHZ witness whose violation measures how close \(\rho\) is, in the sense of stabiliser correlations, to a GHZ-type state. We emphasise that the saturation condition is sharper than what is available for the linear Mermin operator: the closed-form upper bound of Siddiqui and Sazim [21] for the linear functional \(M_3=\langle\sigma_{xxx}\rangle-\langle\sigma_{xyy}\rangle-\langle\sigma_{yxy}\rangle-\langle\sigma_{yyx}\rangle\) involves diagonalisation of state-dependent correlation matrices, whereas our multiplicative bound factorises and is achieved on the pure GHZ orbit alone. The complementary “inequality-free” viewpoint of Cabello [13], which produces an all-versus-nothing separator of GHZ and W via incompatible composite observables, parallels but does not subsume the present quantitative witness.

4 Evaluation on canonical pure states↩︎

We now evaluate \(I\) on prototypical three-qubit pure states.

4.1 The Acín canonical family↩︎

Up to local unitaries, every pure three-qubit state can be brought to the Acín canonical form [22], [23] \[\label{eq:acin} \ket{\Psi}=\lambda_0\ket{000}+\lambda_1\mathrm{e}^{\mathrm{i}\varphi}\ket{100} +\lambda_2\ket{101}+\lambda_3\ket{110}+\lambda_4\ket{111},\tag{11}\] with \(\lambda_i\ge 0\), \(\sum_i\lambda_i^2=1\), and \(\varphi\in[0,\pi]\). This parametrisation has five independent real parameters.

A direct calculation (App. 9) yields \[\begin{align} \langle\sigma_x^{\otimes3}\rangle_{\ket\Psi}&=2\lambda_0\lambda_4,\\ \langle\sigma_x\sigma_y\sigma_y\rangle_{\ket\Psi} =\langle\sigma_y\sigma_x\sigma_y\rangle_{\ket\Psi} =\langle\sigma_y\sigma_y\sigma_x\rangle_{\ket\Psi} &=-2\lambda_0\lambda_4, \end{align}\] independently of \(\lambda_1,\lambda_2,\lambda_3,\varphi\). Substituting in 1 we obtain the closed form \[\label{eq:I-acin} \;I(\hat{x},\hat{y};\ket\Psi)= 2\lambda_0\lambda_4\bigl[4(\lambda_0\lambda_4)^2+1\bigr].\;\tag{12}\] Maximising \(|I|\) over the simplex \(\{\lambda_i^2:\sum_i\lambda_i^2=1,\, \lambda_i\ge0\}\) reduces to maximising the single product \(\mu\equiv\lambda_0\lambda_4\), which by Cauchy–Schwarz is bounded by \(\mu\le 1/2\) with equality iff \(\lambda_0=\lambda_4=1/\sqrt{2}\) and all other \(\lambda_i=0\). This is precisely the GHZ state, and \[|I_{\mathrm{GHZ}}(\hat{x},\hat{y})|=8\,(\tfrac12)^3+2\cdot\tfrac12=1+1=2,\] in agreement with Proposition 1. The function \(|I(\hat{x},\hat{y})|=8\mu^3+2\mu\) is monotonically increasing in \(\mu\) on \([0,1/2]\) [Fig. 1 (a)]. In particular every state with \(\lambda_0\lambda_4=0\) — including all states whose Acín form has no \(\ket{111}\) component — gives \(I(\hat{x},\hat{y})=0\). This is the case for the W state, see Sec. 4.2 below.

A further consequence of 12 : on the GHZ subfamily \(\lambda_1=\lambda_2=\lambda_3=0\), \(\varphi=0\), namely \(\ket{\Psi(\beta)}=\cos\beta\ket{000}+\sin\beta\ket{111}\), one has \(\lambda_0\lambda_4=\tfrac12\sin(2\beta)\) and hence \[I(\hat{x},\hat{y};\ket{\Psi(\beta)})= \sin^3(2\beta) + \sin(2\beta),\] which is determined solely by the Schmidt angle and reaches the GHZ extremum \(|I|=2\) at \(\beta=\pi/4\).

Relation to the three-tangle. On the Acín canonical family the Coffman–Kundu–Wootters three-tangle equals \(\tau_3=4\lambda_0^2\lambda_4^2\) [6]. Equation 12 therefore admits the equivalent form \[\label{eq:I-tau3} I(\hat{x},\hat{y};\ket\Psi)= \sqrt{\tau_3}\,(\tau_3+1),\tag{13}\] exhibiting \(|I(\hat{x},\hat{y})|\) as a strictly increasing degree-3 polynomial in \(\sqrt{\tau_3}\) on the canonical pure-state slice. However, this identification is specific to the canonical form and to the direction pair \((\hat{x},\hat{y})\): on a generic state, neither representative of an LU orbit need be in Acín form, and \(I\) depends on \((\vec{n}_1,\vec{n}_2)\). Accordingly, \(\mathcal{E}_{\mathrm{GHZ}}\) in Sec. 5 is not a function of \(\tau_3\) on arbitrary states: most strikingly, \(\tau_3(\mathrm{W})=0\) [6] while \(\mathcal{E}_{\mathrm{GHZ}}(\mathrm{W})=35/54\) (see Sec. 4.2 below). This separation reflects that \(\mathcal{E}_{\mathrm{GHZ}}\) couples to the entire stabiliser-correlation structure of the GHZ class, not merely to the residual hyperdeterminant captured by \(\tau_3\) [24].

4.2 The W state↩︎

The W state \(\ket\mathrm{W}=(\ket{001}+\ket{010}+\ket{100})/\sqrt{3}\) is the representative of the W class, which is SLOCC-inequivalent to the GHZ class [25]. Direct computation of all 27 three-qubit Pauli correlators yields \[\label{eq:Wcorr} \begin{align} \langle\sigma_z^{\otimes3}\rangle_\mathrm{W}&=-1,\\ \langle\sigma_a \otimes \sigma_a \otimes \sigma_z\rangle_\mathrm{W}&=\langle\sigma_a \otimes \sigma_z \otimes \sigma_a\rangle_\mathrm{W}\\ &=\langle\sigma_z\otimes \sigma_a\otimes \sigma_a\rangle_\mathrm{W}=\tfrac23, \end{align}\tag{14}\] for \(a\in\{x,y\}\), all other components vanishing.

Special directions. With \(\vec{n}_1=\hat{x}\), \(\vec{n}_2=\hat{y}\), every correlator appearing in \(I\) vanishes by 14 ; therefore \[\label{eq:I-W-xy} I_\mathrm{W}(\hat{x},\hat{y})=0.\tag{15}\] By contrast, with \(\vec{n}_1=\hat{z}\), \(\vec{n}_2=\hat{x}\) (or any \(\vec{n}_2\) in the \(\hat{x}\!\text{-}\hat{y}\) plane, by the \(\hat{z}\)-rotation symmetry of \(\ket\mathrm{W}\)) one obtains \[\label{eq:I-W-zx} I_\mathrm{W}(\hat{z},\hat{x})=(-1)-\bigl(\tfrac23\bigr)^3=-1-\frac{8}{27}=-\frac{35}{27}.\tag{16}\]

Maximum over directions. For arbitrary orthogonal pairs we set \(a_3\equiv\hat{z}\cdot\vec{n}_1\), \(b_3\equiv\hat{z}\cdot\vec{n}_2\). Using 14 together with \(\vec{n}_1\cdot\vec{n}_2=0\) and \(\vec{n}_2\cdot\vec{n}_2=1\) one obtains the explicit formula (derived in App. 10) \[\label{eq:I-W-formula} I_\mathrm{W}(\vec{n}_1,\vec{n}_2)= a_3\bigl(2-3a_3^2\bigr) -a_3^3\Bigl(\tfrac23-3b_3^2\Bigr)^{\!3}.\tag{17}\] A short Lagrangian analysis (App. 10) shows that the global extremum of \(|I_\mathrm{W}|\) on the constraint surface \(a_3^2+b_3^2\le 1\) is attained at \((a_3,b_3)=(\pm1,0)\), giving \[\label{eq:I-W-max} \max_{\vec{n}_1,\vec{n}_2}|I_\mathrm{W}(\vec{n}_1,\vec{n}_2)|=\frac{35}{27}\approx 1.296.\tag{18}\] This is strictly less than the GHZ value \(2\), reflecting the well-known fact that the W state lies outside the GHZ class [25]. Equation 18 can therefore be regarded as a sharp quantitative separator of the two SLOCC classes via \(I\). To the best of our knowledge the value \(35/27\) has not previously appeared in the Mermin-functional literature: under the linear Mermin operator, recent analytical work [21] reports \(\max\langle M_3\rangle_\mathrm{W}\approx 3.046\) (out of an algebraic ceiling of \(4\)), while the present multiplicative functional gives the simpler closed form 18 .

4.3 Biseparable and product states↩︎

For an \(A|BC\) biseparable state \(\rho=\rho_A\otimes\rho_{BC}\), all four expectations factorise as \(\langle\sigma(\vec{n}_{a},\vec{n}_{b},\vec{n}_{c})\rangle=\langle\sigma_{\vec{n}_a}\rangle_A\cdot \langle\sigma_{\vec{n}_b}\otimes\sigma_{\vec{n}_c}\rangle_{BC}\). When \(\rho_A=\ket{0}\bra{0}\) and \(\rho_{BC}=\ket{\Phi^+}\bra{\Phi^+}\) with \(\ket{\Phi^+}=(\ket{00}+\ket{11})/\sqrt2\), taking \(\vec{n}_1=\hat{z},\vec{n}_2=\hat{x}\) yields \(e_1=\langle\sigma_z\rangle\langle\sigma_{xx}\rangle=1\cdot 1=1\), \(e_2=e_3=0\), \(e_4=\langle\sigma_z\rangle\langle\sigma_{zz}\rangle =1\cdot 1=1\), so \(I=1\). Numerical optimisation over all directions (Fig. 1 (b)) gives \(\sup|I_{\textrm{bisep}}|=1\) when restricted to orthogonal \(\vec{n}_1\perp\vec{n}_2\), in agreement with Proposition 1 forbidding saturation by biseparable states. Fully separable (product) states likewise reach at most \(1\).

Random pure states. Sampling Haar-random three-qubit pure states and numerically maximising \(|I|\) over directions, we find \(\sup|I|<2\) in every case; random pure states cluster around values \(0.4\)–\(1.6\), consistent with the fact that the GHZ orbit is a measure-zero subset of pure-state space.

5 A genuine GHZ-type entanglement measure↩︎

The quantitative content of Proposition 1 naturally suggests the following measure on three-qubit states.

By Proposition 1, \(\mathcal{E}_{\mathrm{GHZ}}\in[0,1]\) with \(\mathcal{E}_{\mathrm{GHZ}}(\rho)=1\) if and only if \(\rho\) is locally unitarily equivalent to \(\ket{\mathrm{GHZ}}\). We collect its main properties in the following proposition.

Comparison with other measures. Unlike the three-tangle \(\tau_3\) [6] or the GHZ\(\oplus\)W-restricted closed forms of Eltschka et al. [24], both of which vanish on the W class, \(\mathcal{E}_{\mathrm{GHZ}}\) takes nontrivial values on the W class and on biseparable states: it tells GHZ from W (with \(\mathcal{E}_{\mathrm{GHZ}}(\mathrm{GHZ})=1>\mathcal{E}_{\mathrm{GHZ}}(\mathrm{W})=35/54\)) and from biseparable states (with \(\mathcal{E}_{\mathrm{GHZ}}(\textrm{bisep})\le 1/2<\mathcal{E}_{\mathrm{GHZ}}(\mathrm{W})\)). In contrast to GHZ-fidelity-based witnesses \(F_\mathrm{GHZ}(\rho)=\bra{\mathrm{GHZ}}\rho\ket{\mathrm{GHZ}}\) [14], [15], \(\mathcal{E}_{\mathrm{GHZ}}\) is constructed from local Pauli measurements alone and is invariant under arbitrary local unitaries, so it does not depend on a choice of GHZ representative. Compared to the genuine multipartite negativity [16] or the GME-concurrence of Ma et al. [17], both of which involve semidefinite-programming or convex-roof constructions, our \(\mathcal{E}_{\mathrm{GHZ}}\) has a closed form on every pure Acín state. This makes \(\mathcal{E}_{\mathrm{GHZ}}\) device-independent in the operational sense: only the spin directions need be agreed upon.

Strict-monotonicity caveat. We do not claim \(\mathcal{E}_{\mathrm{GHZ}}\) is a strict entanglement monotone (i.e. monotone under all local operations and classical communication). Constructing a LOCC-monotone GHZ-class measure that agrees with \(\mathcal{E}_{\mathrm{GHZ}}\) on the GHZ orbit and vanishes on the W class is an open problem; one natural candidate is the convex roof \(\mathcal{E}_{\mathrm{GHZ}}^{\textrm{cr}}(\rho)=\inf_{\{p_i,\ket{\psi_i}\}}\sum_i p_i\,\mathcal{E}_{\mathrm{GHZ}}(\ket{\psi_i})\), which inherits convexity but loses the closed-form simplicity of ?? .

6 Extension to three-qudit systems↩︎

We now sketch a generalisation of \(I\) from qubits (\(d=2\)) to qudits (\(d\ge 2\)). Two natural strategies exist: (a) replace the spin observables \(\sigma_{\vec{n}}\) by the higher-spin generalisations \(S_{\vec{n}}=\vec{n}\cdot\boldsymbol{S}\) acting on the symmetric subspace, or (b) replace them by the unitary Heisenberg–Weyl operators [26]–[29]. We adopt the latter, which preserves the algebraic clarity of Sec. 2 and is directly compatible with the qudit GHZ paradox of Cerf, Massar, and Pironio [27] and its multisetting refinement by Ryu et al. [29].

Heisenberg–Weyl operators. On a single qudit \(\mathbb{C}^d\), with \(\omega=\mathrm{e}^{2\pi\mathrm{i}/d}\), define \[X\ket j=\ket{j+1\bmod d},\qquad Z\ket j=\omega^j\ket j.\] For \(\boldsymbol{p}=(p,q)\in\mathbb{Z}_d^2\), the Weyl operator is \[W(\boldsymbol{p})=\omega^{-pq/2}\,X^p Z^q.\] The \(W(\boldsymbol{p})\) are unitary (not Hermitian for \(d>2\)) and obey \[\begin{align} W(\boldsymbol{p})W(\boldsymbol{q})&=\omega^{\langle\boldsymbol{p},\boldsymbol{q}\rangle}W(\boldsymbol{q})W(\boldsymbol{p}),\\ \langle\boldsymbol{p},\boldsymbol{q}\rangle&\equiv p_1 q_2-p_2 q_1\pmod d, \end{align}\] together with \(W(\boldsymbol{p})^d=\openone\) for odd \(d\). For \(d=2\), \(W(0,0)=\openone\), \(W(1,0)=\sigma_x\), \(W(0,1)=\sigma_z\), and \(W(1,1)\propto\sigma_y\), so the Pauli operators are recovered.

Three-qudit GHZ state. The qudit GHZ state is \(\ket{\mathrm{GHZ}_d}=d^{-1/2}\sum_{j=0}^{d-1}\ket{jjj}\). It satisfies the perfect-correlation identities \(X^{\otimes 3}\ket{\mathrm{GHZ}_d}=\ket{\mathrm{GHZ}_d}\) and \((Z^{q_1}\otimes Z^{q_2}\otimes Z^{q_3})\ket{\mathrm{GHZ}_d}=\ket{\mathrm{GHZ}_d}\) when \(q_1+q_2+q_3\equiv 0\pmod d\). The qudit Mermin paradox [27], [28] exploits these algebraic relations to produce a complex-valued analogue of the dichotomic GHZ identity.

Qudit functional. Fix two pairs \(\boldsymbol{g}_1,\boldsymbol{g}_2\in\mathbb{Z}_d^2\) and define \[\begin{align} G_1&=W(\boldsymbol{g}_1)\!\otimes\!W(\boldsymbol{g}_2)\!\otimes\!W(\boldsymbol{g}_2),\nonumber\\ G_2&=W(\boldsymbol{g}_2)\!\otimes\!W(\boldsymbol{g}_1)\!\otimes\!W(\boldsymbol{g}_2),\nonumber\\ G_3&=W(\boldsymbol{g}_2)\!\otimes\!W(\boldsymbol{g}_2)\!\otimes\!W(\boldsymbol{g}_1),\nonumber\\ G_4&=W(\boldsymbol{g}_1)\!\otimes\!W(\boldsymbol{g}_1)\!\otimes\!W(\boldsymbol{g}_1). \end{align}\] A short calculation analogous to 7 gives \[G_1 G_2 G_3 = \omega^{2\langle\boldsymbol{g}_1,\boldsymbol{g}_2\rangle}G_4\] when \(\boldsymbol{g}_2\) is twice-applied (using \(W(\boldsymbol{p})^2=\omega^{-pq}W(2\boldsymbol{p})\) for the Heisenberg–Weyl group; here \(W(\boldsymbol{g}_2)^2=\openone\) in the case \(2\boldsymbol{g}_2\equiv 0\bmod d\) that arises naturally for \(d=2\)). Define the qudit functional \[\label{eq:I-qudit} I_d(\boldsymbol{g}_1,\boldsymbol{g}_2)\;=\; \langle G_4 \rangle - \omega^{2\langle\boldsymbol{g}_1,\boldsymbol{g}_2\rangle} \langle G_1\rangle\langle G_2\rangle\langle G_3\rangle.\tag{19}\] For \(d=2\), with \(\boldsymbol{g}_1=(1,0)\) (so \(W(\boldsymbol{g}_1)=\sigma_x\)) and \(\boldsymbol{g}_2=(0,1)\) (so \(W(\boldsymbol{g}_2)=\sigma_z\)), the symplectic form is \(\langle\boldsymbol{g}_1,\boldsymbol{g}_2\rangle=1\), \(\omega=-1\), and 19 reduces (after relabelling \(\sigma_z\leftrightarrow\sigma_y\)) to the original \(I(\hat{x},\hat{y})\).

Bound and saturation. Each \(\langle G_i\rangle\) is a complex number with \(|\langle G_i\rangle|\le1\), hence \(|I_d|\le 2\). Saturation occurs at the qudit GHZ state when \(\boldsymbol{g}_1\) and \(\boldsymbol{g}_2\) generate a maximally non-commutative subgroup of the single-qudit Heisenberg–Weyl group; the proof is structurally identical to that of Proposition 1 once one reads “orthogonality” as “maximal symplectic non-commutation” \(\langle\boldsymbol{g}_1,\boldsymbol{g}_2\rangle\ne 0\) \(\bmod \; d\). We omit the detailed argument; see [27], [28] for parallel constructions.

Spin-\(S\) alternative. For experimental settings in which spin-\(S\) measurements are natural (e.g.atomic ensembles), one may instead replace \(\sigma_{\vec{n}}/2\) by the projection \(S_{\vec{n}}=\vec{n}\cdot\boldsymbol{S}\) rescaled to have spectrum in \([-1,1]\) via a normalisation factor \(1/S\). All qualitative statements survive, but eigenvalues are no longer dichotomic and Proposition 1 becomes a strict inequality with saturation only in the \(S\to\infty\) classical limit.

7 Conclusion↩︎

We have introduced a real-valued functional \(I(\vec{n}_1,\vec{n}_2)\) on three-qubit states, built from four locally measurable Pauli correlations, that simultaneously

• reproduces the GHZ algebraic paradox at \(\vec{n}_1\perp\vec{n}_2\),

• obeys a tight quantum bound \(|I|\le 2\) that is saturated only on the GHZ orbit (Proposition 1),

• admits a closed-form expression on the Acín canonical family showing that \(I(\hat{x},\hat{y})\) depends only on \(\lambda_0\lambda_4\) [Eq. 12 ],

• gives a sharp distinction between the GHZ and W classes, with \(\max|I_\mathrm{W}|=35/27\) [Eq. 18 ],

• induces an LU-invariant measure \(\mathcal{E}_{\mathrm{GHZ}}\in[0,1]\) that vanishes smoothly between the GHZ orbit (where \(\mathcal{E}_{\mathrm{GHZ}}=1\)) and biseparable states (where \(\mathcal{E}_{\mathrm{GHZ}}\le 1/2\)),

• generalises naturally to three qudits via the Heisenberg–Weyl operators [Eq. 19 ].

A number of natural extensions remain open. First, while \(\mathcal{E}_{\mathrm{GHZ}}\) is LU-invariant and convex, its monotonicity under the full SLOCC hierarchy is unclear; a positive answer would promote \(\mathcal{E}_{\mathrm{GHZ}}\) to a genuine entanglement monotone. Second, the qudit construction in 19 merits independent quantitative study: in particular, the qudit analogue of the closed-form Acín evaluation in 12 would shed light on the high-dimensional analogue of the \(\lambda_0\lambda_4\) “GHZ slot.” Third, \(I(\vec{n}_1,\vec{n}_2)\) generalises immediately to \(N\)-qubit GHZ-Mermin functionals [4], [5], [8]; an \(N\)-party version of the entanglement measure \(\mathcal{E}_{\mathrm{GHZ}}\) would furnish a uniform indicator of GHZ-type genuine multipartite entanglement, with maximal value \(2^{N-1}/2\) on the \(N\)-qubit GHZ state. Experimental implementation requires only Pauli-string correlation measurements, well within reach of current photonic and trapped-ion platforms [30], [31].

Acknowledgments↩︎

This work is supported by the National Natural Science Foundation of China (12475020 and 92565111), Quantum Science and Technology-National Science and Technology Major Project (2021ZD0301701), and the National Key Research and development Program of China (2023YFC2205802).

8 Operator identities of Section 2↩︎

Using \(\sigma_{\vec{n}_1}\sigma_{\vec{n}_2}=c\,\openone+\mathrm{i}\,\boldsymbol{m}\cdot\boldsymbol{\sigma}\) with \(c=\vec{n}_1\cdot\vec{n}_2\), \(\boldsymbol{m}=\vec{n}_1\times\vec{n}_2\), \[\begin{align} O_1 O_2&=\sigma_{\vec{n}_1}\sigma_{\vec{n}_2}\otimes\sigma_{\vec{n}_2}\sigma_{\vec{n}_1}\otimes\openone\\ &=(c+\mathrm{i}\boldsymbol{m}\cdot\boldsymbol{\sigma})\otimes(c-\mathrm{i}\boldsymbol{m}\cdot\boldsymbol{\sigma})\otimes\openone,\\ O_2 O_1&=(c-\mathrm{i}\boldsymbol{m}\cdot\boldsymbol{\sigma})\otimes(c+\mathrm{i}\boldsymbol{m}\cdot\boldsymbol{\sigma})\otimes\openone, \end{align}\] so that \([O_1,O_2]=2\mathrm{i}c\,(\boldsymbol{m}\cdot\boldsymbol{\sigma}\otimes\openone-\openone\otimes\boldsymbol{m}\cdot\boldsymbol{\sigma})\otimes\openone\), which is 6 .

For 7 , one computes \[\begin{align} O_1 O_2 O_3 &=\sigma_{\vec{n}_1}\sigma_{\vec{n}_2}^2\otimes\sigma_{\vec{n}_2}\sigma_{\vec{n}_1}\sigma_{\vec{n}_2} \otimes\sigma_{\vec{n}_2}^2\sigma_{\vec{n}_1}\\ &=\sigma_{\vec{n}_1}\otimes\bigl(\sigma_{\vec{n}_2}\sigma_{\vec{n}_1}\sigma_{\vec{n}_2}\bigr)\otimes\sigma_{\vec{n}_1}. \end{align}\] The middle factor satisfies \(\sigma_{\vec{n}_2}\sigma_{\vec{n}_1}\sigma_{\vec{n}_2}=2c\,\sigma_{\vec{n}_2}-\sigma_{\vec{n}_1}\), since \(\{\sigma_{\vec{n}_1},\sigma_{\vec{n}_2}\}=2c\openone\). Substituting, \[O_1 O_2 O_3=2c\,\sigma_{\vec{n}_1}\otimes\sigma_{\vec{n}_2}\otimes\sigma_{\vec{n}_1}-O_4,\] which is 7 .

9 Closed form for the Acín family↩︎

For \(\ket\Psi\) as in 11 , the only \(\sigma_x^{\otimes3}\) matrix elements between basis states present in \(\ket\Psi\) that survive are \(\bra{000}\sigma_x^{\otimes3}\ket{111}=1\) and its conjugate. Hence \(\langle\sigma_x^{\otimes3}\rangle_\Psi=2\lambda_0\lambda_4\). For \(\sigma_x\sigma_y\sigma_y\), \(\sigma_x\otimes\sigma_y\otimes\sigma_y\ket{000} =-\ket{111}\) and \(\sigma_x\otimes\sigma_y\otimes\sigma_y\ket{111}=-\ket{000}\); the only contributing matrix elements are again \(\bra{000}\cdots\ket{111}=-1\) and \(\bra{111}\cdots\ket{000}=-1\), so \(\langle\sigma_x\sigma_y\sigma_y\rangle_\Psi=-2\lambda_0\lambda_4\). The remaining basis states \(\ket{100},\ket{101},\ket{110}\) contribute zero to all four expectations because the action of any tensor product of \(\sigma_x,\sigma_y\) that flips the first qubit cannot return any of \(\{\ket{000},\ket{100},\ket{101},\ket{110},\ket{111}\}\) to itself with nonzero coefficient. \(\blacksquare\)

10 Maximum of \(|I_\mathrm{W}|\) over directions↩︎

Reduction to two scalar variables. Using 14 , \[\begin{align} \langle\sigma_{\vec{n}_1}\otimes \sigma_{\vec{n}_2}\otimes \sigma_{\vec{n}_2}\rangle_\mathrm{W} &=\sum_{i,j,k}n^{(1)}_i n^{(2)}_j n^{(2)}_k\langle\sigma_i \otimes \sigma_j \otimes \sigma_k\rangle_\mathrm{W}\\ &=\tfrac23\bigl[2 b_3(\vec{n}_1\cdot\vec{n}_2) -2 a_3 b_3^2+a_3(1-b_3^2)\bigr]\nonumber\\ &\quad{}-a_3 b_3^2, \end{align}\] where \(a_3=\hat{z}\cdot\vec{n}_1\), \(b_3=\hat{z}\cdot\vec{n}_2\), and \(n^{(1)}, n^{(2)}\) are the cartesian components of \(\vec{n}_1,\vec{n}_2\). When \(\vec{n}_1\cdot\vec{n}_2=0\), the first square-bracket term simplifies giving \[\langle\sigma_{\vec{n}_1}\otimes \sigma_{\vec{n}_2}\otimes \sigma_{\vec{n}_2}\rangle_\mathrm{W} =a_3\bigl(\tfrac23-3b_3^2\bigr).\] By the permutation symmetry of \(\ket\mathrm{W}\), the other two “one-\(\vec{n}_1\), two-\(\vec{n}_2\)” correlators take the same value. Similarly, \[\langle\sigma_{\vec{n}_1}^{\otimes 3}\rangle_\mathrm{W}=a_3(2-3 a_3^2),\] yielding 17 .

Optimisation. Let \(f(u,v)=u^3(\tfrac23-3v^2)^3-u(2-3u^2)\) with constraints \(u^2+v^2\le 1\), \(u,v\in[-1,1]\). Critical points in the interior: \(\partial_v f=-18 u^3 v(\tfrac23-3v^2)^2\) vanishes at \(v=0\), \(u=0\), or \(v^2=2/9\). The branch \(v=0\) reduces \(f\) to \(\tfrac{89}{27}u^3-2u\), maximised at the boundary \(u=\pm 1\) giving \(|f|=35/27\). The branch \(v^2=2/9\) gives \(f=-u(2-3u^2)\), maximised under \(u^2\le 7/9\) at \(u^2=2/9\) with \(|f|=4\sqrt{2}/9\approx 0.629<35/27\). On the boundary \(u^2+v^2=1\), \(f(u)=u^3(3u^2-7/3)^3-u(2-3u^2)\) is again maximised at \(u=\pm1\), \(v=0\). Hence \(\max|f|=35/27\), attained at \(\vec{n}_1=\pm\hat{z}\), \(\vec{n}_2\) in the \(\hat{x}\!\text{-}\hat{y}\) plane.\(\blacksquare\)

11 Convexity of \(\mathcal{E}_{\mathrm{GHZ}}\)↩︎

Write \(e_i(\rho)=\mathrm{Tr}(O_i\rho)\), linear in \(\rho\). For each fixed \((\vec{n}_1,\vec{n}_2)\), the function \(g(\rho)\equiv|e_4 - e_1 e_2 e_3 |\) is convex on the convex set of states because it is the absolute value of a linear function of \(\rho\) once three of \(\{e_1,e_2,e_3,e_4\}\) are held fixed, and the absolute-value composition with multilinear forms is convex on each variable. Combined with the inequality \(|d- abc|\le |d|+ |a||bc|\), one verifies \(g(p\rho_1+(1-p)\rho_2)\le p g(\rho_1)+(1-p)g(\rho_2)\) for all \(\rho_1,\rho_2\) and \(p\in[0,1]\). The supremum \(\mathcal{E}_{\mathrm{GHZ}}(\rho)=\frac{1}{2}\sup_{\vec{n}_1\perp\vec{n}_2}g(\rho)\) is convex because the supremum of convex functions is convex.\(\blacksquare\)

References↩︎