Foundations of quantum mechanics · Conceptual analysis
Probing the extreme quantum violation of the Leggett-Garg inequality
arXiv:2611.16401v1 ↗PDF ↗Submitted 12 November 2026
- Question
- The three-time Leggett–Garg functional \(K=\langle M_1M_2\rangle+\langle M_2M_3\rangle-\langle M_1M_3\rangle\) is at most 1 in any macrorealist model, and at most 3/2 in quantum mechanics when each measurement is a two-outcome projective (Lüders) measurement; 3/2 is called the temporal Tsirelson bound. Two Physical Review Letters papers report quantum values above it, up to the algebraic maximum 3: Budroni and Emary (2014) with degeneracy-breaking measurements, and Chatterjee and co-workers (2025) with superpositions of unitary evolutions. Do these values say anything about macrorealism?
- Result
- The authors argue that they do not, because both constructions change the functional being evaluated. A degeneracy-breaking measurement of \(M_1\) is equivalent to first measuring a commuting observable \(S\); the correlations involving \(M_1\) are then evaluated on the state \((\rho+S\rho S)/2\), while \(\langle M_2M_3\rangle\) needs a different prior observable and is evaluated on another state. With three mutually commuting two-qubit observables and a Bell state, for which ordinary sequential measurements give \(K=1\), this already gives \(K=2\). In the second construction, the evolution from \(t_1\) to \(t_3\) is not the product of the two intermediate evolutions, so the third correlation involves a fourth observable; a superposition of unitaries is not needed, and a qubit example reaches 3 directly. For both modified functionals the macrorealist bound can itself be 3.
The paper in detail
Leggett and Garg asked whether a macroscopic system can be in one definite state at every time and be measured without disturbance. A dichotomic observable \(M_1\) is measured at three times; in the Heisenberg picture these are observables \(M_1\), \(M_2\) and \(M_3\), with \(M_2\) and \(M_3\) the evolved forms of \(M_1\). Pairwise sequential correlations enter \(K\), and macrorealism with non-invasive measurability gives \(K\le1\). Quantum mechanics violates this, and for two-outcome Lüders measurements, which project onto the full eigenspaces, the maximum is 3/2.
The bound has a two-line proof. For Lüders measurements of \(\pm1\)-valued observables the sequential correlation is \(\tfrac12\mathrm{Tr}[\rho\{M_i,M_j\}]\), and expanding the positive operator \((M_1-M_2+M_3)^2\) bounds the combination of anticommutators by 3. Equality needs \(M_2=M_1+M_3\), which a qubit realises with three directions at \(60^\circ\) and \(120^\circ\). Nothing in the argument depends on the dimension or the state, so exceeding 3/2 requires leaving its assumptions.
Budroni and Emary leave them through the measurement: in dimension above two, a \(\pm1\) observable has degenerate eigenspaces, and a von Neumann measurement can resolve them into rank-one projectors. The paper shows this already matters for commuting observables. With \(M_1=\sigma_x\otimes\sigma_x\), \(M_2=\sigma_y\otimes\sigma_y\), \(M_3=\sigma_z\otimes\sigma_z\) and \(\ket\psi=(\ket{01}+\ket{10})/\sqrt2\), coarse-grained measurements disturb nothing and give \(K=1\), but splitting each eigenspace with an auxiliary commuting observable gives \(K=2\). The fine-grained projectors do not commute with the other observables, which is where the extra disturbance comes from.
Chatterjee and co-workers leave them through the dynamics. If the evolution from \(t_1\) to \(t_3\) is not the composition of the evolutions from \(t_1\) to \(t_2\) and from \(t_2\) to \(t_3\), then \(\langle M_1M_3\rangle\) involves an observable \(M_3^\prime\) different from the \(M_3\) in \(\langle M_2M_3\rangle\). The paper shows that this mismatch, not the superposition of unitaries, produces the value 3, and that a macrorealist assigning fixed values to four observables can also reach 3.
What remains open is which inequality an experiment with invasive fine-grained measurements actually tests. The authors offer two readings and do not choose between them: if the two effective post-measurement states correspond to the same distribution of hidden states, the classical bound is still 1; if not, it is 3.
Transferable methods
A sum-of-squares bound on sequential correlations
To show that 3/2 cannot be exceeded within the standard framework, the paper writes the quantum functional as an expectation of anticommutators and bounds it by a single positive operator. Any construction that goes beyond 3/2 must break one of the steps.
Editorial derivation of the correlation formula: measuring \(M_i\) with projectors \(\Pi_\pm=(\mathbb 1\pm M_i)/2\) and then \(M_j\) gives \(\langle M_iM_j\rangle=\mathrm{Tr}[M_j(\Pi_+\rho\Pi_+-\Pi_-\rho\Pi_-)]\). Expanding the projectors, \(\Pi_+\rho\Pi_+-\Pi_-\rho\Pi_-=\tfrac12(M_i\rho+\rho M_i)\), so \[\langle M_iM_j\rangle=\tfrac12\mathrm{Tr}\bigl[\rho\{M_i,M_j\}\bigr].\]
Since \(M_i^2=\mathbb 1\), \[0\le(M_1-M_2+M_3)^2=3\cdot\mathbb 1-\{M_1,M_2\}-\{M_2,M_3\}+\{M_1,M_3\},\] and taking the expectation in \(\rho\) gives \(2K\le3\). Equality on every state requires \(M_2=M_1+M_3\), and squaring both sides shows this needs \(\{M_1,M_3\}=-\mathbb 1\): directions at \(120^\circ\) on the Bloch sphere. The paper's example \(M_{1,3}=\tfrac12\sigma_z\pm\tfrac{\sqrt3}{2}\sigma_x\), \(M_2=\sigma_z\) gives \(K=3/2\) exactly (editorial check).
How to reuse it
For any linear combination of two-time correlators with Lüders measurements, look for a positive operator whose expansion produces the same anticommutators. It gives the quantum maximum and its equality conditions without optimising over states.
Fine-grained measurement as a hidden prior measurement
The key observation is operational: to resolve a degenerate eigenspace of \(M_1\) one must measure something else that commutes with it. Making that prior measurement explicit shows on which state each correlation is really evaluated.
Let \(S\) commute with \(M_1\) and put \(R=-M_1S\). The four rank-two projectors \(\tfrac14(\mathbb 1\pm M_1\pm R\pm S)\), with signs as in Eq. (14), refine the two eigenprojectors of \(M_1\). The resulting correlation is \[\langle M_1M_2\rangle_{db}=\tfrac12\langle M_1M_2\rangle_{dp}+\tfrac12\mathrm{Tr}[\rho\,S M_1M_2S]=\mathrm{Tr}[\rho^\prime M_1M_2],\qquad\rho^\prime=\tfrac12(\rho+S\rho S),\] for commuting \(M_1,M_2\): first dephasing in the eigenbasis of \(S\), then measuring as usual. \(\langle M_2M_3\rangle\) needs another observable \(F\) and the state \(\rho^{\prime\prime}=\tfrac12(\rho+F\rho F)\).
Editorial check of the two-qubit example: with \(S=\sigma_z\otimes\sigma_z\) and \(F=\sigma_y\otimes\mathbb 1\), the projectors are valid and rank one, the coarse-grained value is \(K_{dp}=1\), and the fine-grained value is \(K_{db}=2\). For the dynamical construction, \(\sigma_\xi=\cos\xi\,\sigma_x+\sin\xi\,\sigma_z\) with \(\theta=\pi/2\) makes the first two evolutions commute with \(\sigma_z\), so \(M_2=M_3=\sigma_z\), while \(U(t_1,t_3)=e^{-i\pi\sigma_x/2}\) gives \(M_3^\prime=-\sigma_z\). Then \(K=1+1+1=3\), and a macrorealist with fixed values \(+1,+1,+1,-1\) for \(M_1,M_2,M_3,M_3^\prime\) reaches the same 3. Eq. (26) agrees with direct evaluation at random parameters.
How to reuse it
When a reported quantum value exceeds a known bound, write every measurement as an instrument on a single state, including the auxiliary measurements that implement it. If different terms are evaluated on different effective states or observables, recompute the classical bound for that modified expression before comparing.
Context & limitations
Why it matters. The comment separates two readings of results above the temporal Tsirelson bound: stronger nonclassicality within the Leggett–Garg test, or a different inequality whose classical bound has not been recomputed. It argues for the second and gives minimal examples that make the change of functional explicit.
Limitations. This is a short conceptual comment. For degeneracy-breaking measurements the authors leave the conclusion open: if the two post-measurement states are treated as macrorealistically equivalent, the bound stays 1 and the violation stands, and they do not argue which reading is correct. They state rather than derive the macrorealist value 3 for the second construction, do not re-analyse the experiments reported as exceeding the bound, and do not engage with the arguments in the original papers beyond their examples. The paper has no AI-use statement.