Foundations of quantum mechanics · Theory

Measurement-order dependence bounds uncertainty, contextuality, steering, and Bell nonlocality

Shengjun Wu

arXiv:2611.24604v1 ↗PDF ↗Submitted 18 November 2026

Question
Uncertainty relations, contextuality, steering and Bell violations all need incompatible measurements, and with sharp measurements they all disappear when the observables commute. That is a qualitative statement. Is there one operational quantity, measurable in a simple sequential experiment, whose size bounds how strong each of these signatures can be? The paper takes the order of two measurements: when the measurements are applied with the standard state update, the joint record of “\(A\) then \(B\)” differs from that of “\(B\) then \(A\)” on some input exactly when they do not commute.
Result
Order scores are defined as the largest distance between the two order records over all inputs, measured by total variation, Jensen–Shannon divergence or squared Hellinger distance. For two binary sharp observables the total-variation score equals \(\|[A,B]\|/4\). Entropic and variance uncertainty relations bound the scores from below, and violations of noncontextuality, steering and Bell inequalities are bounded by order effects. For correlation Bell inequalities the squared Bell operator splits into a part without commutators and tensor products of local commutators, so that when one party has two settings the excess of the squared quantum value over the squared classical bound is bilinear in the two parties' order scores; for CHSH this is Landau's identity. The same bilinear form bounds Bell's original inequality on a maximally entangled state and the KCBS pentagon on a qutrit, with equality at the maximal KCBS violation \(\sqrt5-2\).

The paper in detail

If two classical quantities can be read without disturbing the system, the order of reading does not change their joint statistics. A quantum measurement updates the state, \(\rho\mapsto\sqrt{E}\rho\sqrt{E}\) for outcome effect \(E\) (the Lüders rule), so the second measurement sees a changed input. The paper's experiment is minimal: a coin chooses the order, both outcomes are recorded, and the task is to guess the order from the record. The records \(p_{AB}\) and \(p_{BA}\) are the statistics of two sequential POVMs, and the order scores \(\mathcal T\), \(\mathcal J\) and \(\mathcal H^2\) are their largest total-variation, Jensen–Shannon and squared Hellinger distances over inputs. By a theorem of Gudder and Nagy, the scores vanish exactly when the effects commute, and they bound one another.

For two \(\pm1\)-valued sharp observables, Jordan's lemma splits the space into qubit blocks where \(A=\sigma_z\) and \(B=\cos\theta_j\,\sigma_z+\sin\theta_j\,\sigma_x\), and \(\mathcal T=\max_j\tfrac12\sin\theta_j=\|[A,B]\|/4\). Even for an anticommuting pair one record identifies the order only with probability \(3/4\) and carries 0.311 bits. For projective measurements in general the two records always overlap: the affinity bound of Theorem 2 shows that the order can never be read off with certainty. Mutually unbiased bases, extremal for entropic uncertainty, are not extremal here; pairs of bases built so that a suitable input occupies about half of the outcomes of each measurement have a larger Hellinger score, although their minimal entropy sum is smaller.

The link to the signatures runs through one observation: measuring all observables of a scenario in sequence produces a single joint distribution, which is a classical model. The model is exact on whichever context is measured first, and its error on another context is the disturbance caused by the measurements made before it, which is a distance between two order records. That gives linear bounds for contextuality, steering and Bell inequalities. For correlation Bell inequalities a sharper, bilinear statement follows from squaring the Bell operator, which isolates the part of \(\mathfrak B^2\) that needs noncommutativity on both sides. Two further scenarios have the same form: Bell's original inequality with a maximally entangled state, where the violation is an order effect of three measurements on the maximally mixed state, and the rank-one KCBS pentagon, whose violation is bounded by the order scores of two crossing diagonals.

The main open problem is the conjectured bilinear bound for all correlation Bell inequalities, which on Clifford realisations would follow from a Hilbert–Schmidt variant of Grothendieck's inequality proved here for at most three settings on one side; its planar case would imply Krivine's bound \(K_G(2)\le\sqrt2\). Also open are the pentagon bound for projectors of higher rank, the monotonicity of the Jensen–Shannon function \(\Psi\), and the behaviour of the scores when outcomes are merged.

Transferable methods

  1. Order records as sequential POVMs and the commutator norm

    The two orders define two POVMs on outcome pairs, so the largest total-variation distance between their records is a finite family of eigenvalue problems. For \(\pm1\)-valued observables the optimum is attained by a single marginal, the disturbance of \(A\) by \(B\), and in each Jordan block that disturbance is half the sine of the angle between the two Bloch vectors.

    With Lüders instruments the order “\(A\) then \(B\)” has effects \(X_{ab}=\sqrt{E_a}F_b\sqrt{E_a}\) and the reverse order \(Y_{ab}=\sqrt{F_b}E_a\sqrt{F_b}\). Since total variation is the largest difference of the probabilities of an event, \[\mathcal T(A,B)=\max_S\lambda_{\max}\Bigl(\sum_{(a,b)\in S}(X_{ab}-Y_{ab})\Bigr),\] maximised over sets \(S\) of outcome pairs. If \(\sqrt EF\sqrt E=\sqrt FE\sqrt F\) for all outcomes, the effects commute (Gudder–Nagy), so the score is faithful to noncommutativity.

    Editorial derivation for one Jordan block, \(A=\sigma_z\), \(B=\cos\theta\,\sigma_z+\sin\theta\,\sigma_x\). Take \(S\) to be the event \(a=+1\). In the order \(AB\) its probability is \(\operatorname{Tr}\rho P_+\) with \(P_+=(1+\sigma_z)/2\). In the order \(BA\), measuring \(B\) first dephases in \(B\)'s eigenbasis, and \(\sum_bQ_bP_+Q_b=(1+\cos\theta\,\mathbf n\cdot\vec\sigma)/2\) with \(\mathbf n=(\sin\theta,0,\cos\theta)\). The difference is \[P_+-\sum_bQ_bP_+Q_b=\tfrac12\sin\theta\,\bigl(\sin\theta\,\sigma_z-\cos\theta\,\sigma_x\bigr),\] whose largest eigenvalue is \(\tfrac12\sin\theta\). Since \([A,B]=2i\sin\theta\,\sigma_y\) has norm \(2\sin\theta\), this equals \(\|[A,B]\|/4\), and the optimal input is the eigenstate along \(\mathbf a-(\mathbf a\cdot\mathbf b)\mathbf b\), as the paper states. Theorem 1 shows that no other event does better.

    Editorial check: computing Eq. (3) over all subsets of outcome pairs for random \(\pm1\)-valued observables in dimensions 3 and 5 reproduced \(\|[A,B]\|/4\) to six digits (0.4477 and 0.4972). Maximising the Jensen–Shannon divergence over qubit inputs gave \(\Psi(1/2)=0.311278\), equal to \(h(3/4)-\tfrac12\), and 0.1119 bits at \(\theta=0.6\). For unsharp measurements with sharpness \(\eta,\zeta\) the same eigenvalue computation reproduces Eq. (28), \(\mathcal T=\tfrac12\max\{\eta(1-\sqrt{1-\zeta^2}),\zeta(1-\sqrt{1-\eta^2})\}\), which is positive even when the pair is jointly measurable.

    How to reuse it

    Asking how well a record reveals the order of two operations is a general way to quantify their incompatibility from data alone. Wherever sequential measurements are available (contextuality tests, weak-measurement and Leggett–Garg experiments), order scores give directly measured, operationally interpreted lower bounds on commutator norms.

  2. Squaring the Bell operator: Jordan part and commutator part

    For \(\pm1\)-valued observables the square of a correlation Bell operator separates into products of symmetrised (Jordan) products and products of commutators, one from each party. The first part is bounded by the classical value whenever one party's Jordan products can be written as a mixture of products of classical signs; the second is bounded by order scores. The same algebra governs Bell's original inequality and the KCBS pentagon.

    Write \(\mathfrak B=\sum_{x,y}M_{xy}A_x\otimes B_y\) and \(A_xA_{x'}=A_x\circ A_{x'}+\tfrac12[A_x,A_{x'}]\) with \(X\circ Y=\tfrac12(XY+YX)\). In \(\mathfrak B^2=\sum_{x,x',y,y'}M_{xy}M_{x'y'}A_xA_{x'}\otimes B_yB_{y'}\) the mixed terms pair a part symmetric in \((x,x')\) with a part antisymmetric in \((x,x')\) and cancel in the sum, leaving \[\mathfrak B^2=\Gamma-\sum_{x,x',y,y'}M_{xy}M_{x'y'}\,\frac{[A_x,A_{x'}]}{2i}\otimes\frac{[B_y,B_{y'}]}{2i},\] with \(\Gamma\) the Jordan part. Each commutator has norm \(4\mathcal T\), so \(\beta_{\mathrm Q}^2\le\lambda_{\max}(\Gamma)+4\sum|M_{xy}M_{x'y'}|\,\mathcal T^A_{xx'}\mathcal T^B_{yy'}\). For CHSH, \(\Gamma=4\,\mathbb 1\) and \(\beta_{\mathrm Q}\le2\sqrt{1+4\mathcal T_A\mathcal T_B}\), which is Landau's identity.

    For three observables on a maximally entangled state, \(AC=(AB)(BC)\) gives \(\operatorname{Tr}(AC)=\operatorname{Tr}[(A\circ B)(B\circ C)]-\operatorname{Tr}(K_{AB}K_{BC})\) with \(K_{XY}=[X,Y]/2i\), and Theorem 9 turns it into \[E_{AB}+E_{BC}-E_{AC}=1+\kappa_B-4\,p_{(B)}(a\ne b\ne c),\] where \(\kappa_B=\operatorname{Tr}(K_{AB}K_{BC})/d\) is the change in the correlation of \(A\) and \(C\) caused by measuring \(B\) between them on \(\mathbb 1/d\). The last term is never positive, so the violation of Bell's original inequality is at most \(\kappa_B\le4\,\mathcal T(A,B)\,\mathcal T(B,C)\). For a rank-one pentagon on \(\mathbb C^3\) with chord overlaps \(z_i=|\langle v_i,v_{i+2}\rangle|^2\), Theorem 10 shows \(z_{i+3}=(1-z_i)(1-z_{i+1})\) and \(2-\sum_iz_i=V^2(1+V)/(1-V)\), from which \(V\le\max_i\mathcal T_i\mathcal T_{i+1}\) with \(\mathcal T_i=\sqrt{z_i(1-z_i)}\).

    Editorial check: optimising Bob's settings for Alice's pair at angle \(\theta\) gives \(\beta=2.2764,2.7140,2.8284\) at \(\theta=0.3,1,\pi/2\), equal to \(2\sqrt{1+2\mathcal T_A}\) with \(\mathcal T_A=\tfrac12\sin\theta\). The identity of Theorem 9 holds to ten digits for random observables in dimension 4, and at qubit angles of \(60^\circ\) the value \(\beta_3=3/2\) is reached with \(\mathcal T=\sqrt3/4\). Over 20,000 random orthogonality pentagons in \(\mathbb C^3\) the relations of Theorem 10 hold to \(5\times10^{-15}\), the ratio \(V/\max_i\mathcal T_i\mathcal T_{i+1}\) never exceeds 0.9988, and at the KCBS optimum \(z_i=(3-\sqrt5)/2\) one finds \(\mathcal T_i^2=\sqrt5-2=V\).

    How to reuse it

    Splitting a squared operator into a symmetrised part and a commutator part separates what classical, compatible observables could already produce from what needs incompatibility on both sides. The same move bounds other quadratic witnesses, such as Leggett–Garg combinations, uncertainty sums and steering functionals, by products of local commutator norms.

Source: Section II (Definition 1, Eq. (3), Theorem 1), Section IV.3 (Theorem 6), Section V.1 (Theorem 7), Section VI (Theorems 9 and 10). Based on the linked paper version. Derivations marked as editorial, and suggested uses in other problems, are editorial guidance rather than results of the paper.

Context & limitations

Why it matters. The commutator norm of two \(\pm1\)-valued observables gets an operational reading: \((1+\|[A,B]\|/4)/2\) is the best probability of guessing, from one record, in which order they were applied. Each signature of incompatibility then becomes a necessary condition with an explicit constant on quantities that sequential contextuality experiments already measure. The scores also rank pairs of measurements differently from entropic uncertainty: in every dimension \(n\ge3\) there are pairs of bases with a larger Hellinger score than mutually unbiased bases but a smaller minimal entropy sum.

Limitations. All bounds are necessary conditions: a given order effect guarantees no violation for a specified state and scenario. Order scores belong to instruments rather than measurement effects, so a classical device that disturbs what it reads also produces them, and for unsharp measurements a positive score does not imply incompatibility; the scores are not device-independent witnesses. The steering and Bell bounds of Section IV are linear and loose. The bilinear bound for general correlation Bell inequalities is a conjecture, proved when one party has two settings and, on Clifford realisations, when one party has at most three; the pentagon bound is proved only for rank-one projectors on \(\mathbb C^3\), with numerical evidence beyond. Monotonicity of the scores under coarse-graining is open. The paper contains no statement about AI use.

Description based on the full arXiv HTML of v1, read 2026-11-20: Sections I–VIII, Table 1 and the proof of Proposition 1 in Appendix A; the other appendix proofs were not checked line by line. Editorial checks: for random \(\pm1\)-valued observables in dimensions 3 and 5 the score from Eq. (3), a maximum over all subsets of outcome pairs, equals \(\|[A,B]\|/4\); \(\Psi(1/2)=h(3/4)-\tfrac12\approx0.3113\) bits and the \(\theta=0.6\) optimum (0.1119 bits, Bloch vector \(\pm(0.074,0,-0.997)\)) by numerical maximisation; the ideal-partner CHSH value \(2\sqrt{1+2\mathcal T_A}\) at three angles; Landau's identity for random qutrit observables; the identity (24) of Theorem 9 for random observables in dimension 4 to ten digits and \(\beta_3=3/2\) at \(\mathcal T=\sqrt3/4\); for 20,000 random orthogonality pentagons in \(\mathbb C^3\) the relations of Theorem 10(a,b) to \(5\times10^{-15}\), with the largest ratio of violation to bound 0.9988; and Eq. (28) for unsharp pairs.

Read the accessible explanation

Google tests quantum contextuality on up to 105 qubits