slug: simnacher-2021-two-qubit-separable-numerical-range-ratio
bibkey: simnacher2021separable
doi: 10.1103/PhysRevA.104.042420
url: https://arxiv.org/abs/2107.04365v1
triage: theorem
motivation_gids:
  - D5/S3/Quantum/Entanglement/SeparableNumericalRangeRatio.result

The minimal separable numerical range ratio of one two-qubit observable

Problem

T. Simnacher, J. Czartowski, K. Szymański, K. Życzkowski, Confident entanglement detection via the separable numerical range, Phys. Rev. A 104, 042420 (2021), arXiv:2107.04365v1, Conjecture 8:

For a single measurement on a two-qubit system, the minimal volume ratio is ${\mu }_{2,1}=\frac{1}{2}$.

Here ${\mu }_{2,1}$ is the minimum over Hermitian two-qubit observables $A$ of $\mathrm{vol}{L}_{\mathrm{Sep}}(A)/\mathrm{vol}L(A)$, where $L_X(A)=\{\mathrm{Tr}\rho A\mid \rho \in X\}$ for the set of all states ($L$) or of separable states ($L_{\mathrm{Sep}}$), and vol is the Euclidean volume (Definitions 1 and 2). The paper proves ${\mu }_{2,1}\ge \sqrt{2}-1$ (Proposition 7) and ${\mu }_{2,1}\le \frac{1}{2}$ with the projector onto $|{\varphi }^{+}⟩=(|00⟩+|11⟩)/\sqrt{2}$.

Motivation

An experiment detects entanglement with statistical confidence only when its confidence region misses the separable numerical range, so a small ratio $\mathrm{vol}{L}_{\mathrm{Sep}}/\mathrm{vol}L$ makes detection easier. The conjecture fixes the best possible ratio for one two-qubit observable. The frozen declaration D5/S3/Quantum/Entanglement/SeparableNumericalRangeRatio.result proves it.

Gap

Issue #12146 classifies the conjecture as Tier 1 and records the literature checks before any Lean:

• arXiv:2107.04365 has one version, and the published version states the conjecture;
• the seven citing works listed by Crossref (Sauer–Bernád 2022, Szymański’s 2022 thesis, Balanzó-Juandó–Studziński–Huber 2024, Wang–Bernád 2025, Pikul et al. arXiv:2510.27670, Li–Wang arXiv:2605.24360, Zhang–Xie arXiv:2601.01858) do not prove or refute it;
• an adversarial search of product numerical ranges, entanglement gaps, robustness and common noise, and base norms found no publication of the width inequality or of a noise operator shared by two states.

These are orchestrator-reported readings and seat-reported readings, not-found-in-searched-scope; they do not establish exhaustive worldwide novelty. The rank-two step (Route, step 3) also follows from the rank-two concurrence theorem of Hill–Wootters, PRL 78, 5022 (1997); the conjecture needs in addition the shared noise of step 4.

Route

Write two-qubit vectors in the magic basis: $u_{00}=(z_0+i{z}_1)/\sqrt{2}$, $u_{11}=(z_0-i{z}_1)/\sqrt{2}$, $u_{01}=(i{z}_2+z_3)/\sqrt{2}$, $u_{10}=(i{z}_2-z_3)/\sqrt{2}$. This is unitary, and $u_{00}{u}_{11}-u_{01}{u}_{10}=z^{T}z/2$.

1. $u$ is a product vector $a\otimes b$ exactly when $z^{T}z=0$, and the projector of $a\otimes b$ is $(a{a}^{\ast })\otimes (b{b}^{\ast })$.
2. After a phase, $z^{T}z=C$ with $0\le C\le \Vert z{\Vert }^2=1$.
3. If $C=0$, $z$ is a product vector by step 1. If $C>0$, for a real unit $r\perp \mathrm{Re}z$ put $\eta =(\mathrm{Im}z)\cdot r$, $\delta =\sqrt{{\eta }^2+C}>0$ and $t_{\pm }=-\eta \pm \delta$. The vectors $z+i{t}_{\pm }r$ are product vectors, and with weights $(\delta \pm \eta )/(2\delta )$ their projectors sum to $z{z}^{\ast }+Cr{r}^T$, which is therefore separable.
4. For unit $u,v$ take orthonormal real $r,s$ orthogonal to $\mathrm{Re}{z}_u$ and $\mathrm{Re}{z}_v$, and $N=(r{r}^T+s{s}^T)/2$, a separable state. With $c=\max (C_u,C_v)$ both $(P_u+cN)/(1+c)$ and $(P_v+cN)/(1+c)$ are separable states, and their difference is $(P_u-P_v)/(1+c)$.
5. So $\langle u,Au\rangle-\langle v,Av\rangle\le 2,(\sup L_{\mathrm{Sep}}(A)- \inf L_{\mathrm{Sep}}(A))$; by the spectral decomposition of states the same bound holds for any two states.
6. $L_{\mathrm{Sep}}(A)$ is a bounded interval, so $\mathrm{vol}L(A)\le 2\mathrm{vol}{L}_{\mathrm{Sep}}(A)$.
7. For $A=|{\varphi }^{+}⟩⟨{\varphi }^{+}|$, $L(A)\supseteq [0,1]$, and a product state $\sigma \otimes \tau$ has expectation $\frac{1}{2}{\sum }_{a,b}{\sigma }_{ab}{\tau }_{ab}\in [0,\frac{1}{2}]$; so the ratio is $\frac{1}{2}$.

Falsifier

The kernel-checked result states that $\frac{1}{2}$ is the least element of the set of ratios $\mathrm{vol}{L}_{\mathrm{Sep}}(A)/\mathrm{vol}L(A)$ over Hermitian $A$ with $\mathrm{vol}L(A)\ne 0$, with vol the Lebesgue measure on $\mathbb{R}$. A different definition of separability, of the volume, or of the set over which the minimum is taken changes the question.

Evidence

A NumPy recomputation of the construction (seed 20261002, 2000 random pairs of unit vectors, half orthonormal) reconstructs $P_u+cN$ from the product vectors of steps 3–4 with maximum entrywise residual $3.3\times 10^{-16}$; every product vector has normalized $|u_{00}u_{11}-u_{01}u_{10}|\le 5.3\times10^{-16}$,everyweightisnonnegative,and$1+c\le1.9990$. Dropping the hypothesis $r\perp \mathrm{Re}z$ makes all 2000 vectors $z+i{t}_{\pm }r$ non-product (issue #12146).

The canonical source is D5/S3/Quantum/Entanglement/SeparableNumericalRangeRatio.lean. Its public declarations are separableStates, numericalRange, claim and result; states are the frozen D5/S3/Quantum/Entanglement/StructuredNegativityCoincidenceRefutation.IsDensity at d = 2. The frozen module state has statement identity sha256:8d89638351ff9e36c9f931fce5e088ab93cc308c8d733623bc0d1c9c9a92ee1f. The result declaration has statement identity sha256:8c35814a5581b8dd85f6068c299608bb3ab6358ca90eb7b729ad3bb8ebd2a1e0. The Freeze event is sha256:efd937da6ba0de445d31603562f7d3eded02963f281f6bbf4a1a1254b47107c2; its project-level frozen prerequisites are the Freeze events of D5/S3/Resource/EntanglementWitness (through which the module imports CompositeCones and CompositeConeDuality) and of D5/S3/Quantum/Entanglement/StructuredNegativityCoincidenceRefutation. The proof uses only the standard axioms propext, Classical.choice and Quot.sound; no sorry, native_decide, or new axiom.

Triage

Tier 1 published conjecture; resolution Proved by D5/S3/Quantum/Entanglement/SeparableNumericalRangeRatio.result. proof_shape: content; admission_basis: open-problem-resolution (issue #12146). Utility kind none (a theorem about every two-qubit observable).

What the settlement shows

• Proved by result: $\operatorname{vol}L_{\mathrm{Sep}}(A)\ge\frac12 \operatorname{vol}L(A)$foreveryHermitiantwo-qubit$A$, with equality for the Bell projector; so ${\mu }_{2,1}=\frac{1}{2}$.
• Proved inside the proof of result: for any two unit vectors $u,v$ there is $c\in [0,1]$ and separable states $\sigma ,\tau$ with $P_u-P_v=(1+c)(\sigma -\tau )$, where $c$ is the larger of the concurrences $|z_u^T{z}_u|$ and $|z_v^T{z}_v|$.
• Follows from it (not stated in Lean): for extreme eigenvectors $u,v$ of $A$, $\operatorname{vol}L_{\mathrm{Sep}}(A)\ge(\lambda_{\max}- \lambda_{\min})/(1+\max(C_u,C_v))$,sotheratioexceeds$\frac12$ whenever $A$ has extreme eigenvectors $u,v$ neither of which is maximally entangled ($C_u,C_v<1$).
• Mechanism: the noise operator is chosen once for both states, from two real directions orthogonal to the real parts of both magic coordinate vectors, which exist because these real parts span at most two of the four dimensions of ${\mathbb{R}}^4$.
• Open here: the values of ${\mu }_{2,k}$ for $k\ge 2$ (the paper’s Table 1), of ${\mu }_{d,1}$ for larger local dimension, and whether the common-noise construction extends beyond $2\times 2$.

ASSUMED-UNVERIFIED

The literature checks do not establish worldwide priority or absence of an independent answer. The Lean kernel verifies the encoded statement and its axiom closure; its correspondence to the paper, including the reading of the minimum as the least element over observables with nonzero range and of separability as the frozen separableCone with unit trace, is checked by reading the source and the definitions.