Quantum information · Theory (AI-assisted)

Bound entanglement as a trigonometric moment problem

Lilong Qian, Teng Zhang

arXiv:2611.24473v1 ↗PDF ↗Submitted 18 November 2026

Question
States that stay positive under partial transposition (PPT) but are entangled cannot be distilled, and they defeat the simplest separability test. Even for a few qubits three basic questions about them are open. Which bound entangled states exist? How much white noise does their entanglement survive? How small can their rank be? For the standard three-qubit examples, the states built from unextendible product bases (UPBs), only the noise tolerated by particular witnesses was known, and Wan and Guo had asked whether any witness does better than the UPB projector.
Result
The paper defines Toeplitz states, mixtures with real but possibly negative weights of products of equatorial spin coherent states. Every such state is PPT across every bipartition, and it is fully separable exactly when an associated multi-level Toeplitz matrix is the moment matrix of a positive measure on a torus. Within this class: every three-qubit rank-four PPT entangled state with nonzero Lorentz invariant is a Toeplitz state up to invertible local operators, which proves a working hypothesis of Garberg, Irgens and Myrheim; every three-qubit UPB state mixed with white noise is fully separable exactly when the noise fraction reaches \(\lambda^*=1-2/(1+\sqrt{1+s})\), with \(s\) a local-unitary invariant in \((0,1]\), so \(\lambda^*=3-2\sqrt2\approx0.1716\) for the Shifts UPB; and a closed-form witness detects the whole entangled range with eight local Pauli settings, four for Shifts. Minimal ranks of entangled Toeplitz states are determined for several group sizes.

The paper in detail

The partial transpose test detects most entanglement in small systems, but not all: PPT entangled states exist from two qutrits or three qubits on, and they cannot be distilled into Bell pairs, which is why they are called bound entangled. Deciding separability in general is hard, so progress comes from families where it can be decided exactly. The standard examples are built from unextendible product bases, sets of orthogonal product vectors with no further product vector orthogonal to all of them; the normalised projector onto the complement is PPT and entangled. How such states sit among all PPT entangled states, and how robust they are, were open even for three qubits.

Consider \(k\) groups of qubits and products of equatorial states \(\ket\theta^{\otimes m_j}\) with \(\ket\theta=(\ket0+e^{i\theta}\ket1)/\sqrt2\). Mixing them with a real, possibly negative weight function gives operators whose matrix elements depend only on differences of Hamming weights, \(\bra x\rho_T\ket{x'}=t(w(x)-w(x'))\), so the coefficients form a multi-level Toeplitz matrix \(T\). The authors show that every positive one is PPT across every cut, because each partial transpose is a bit flip of the original, and that it is fully separable exactly when the weights can be replaced by a probability measure, the truncated trigonometric moment problem on the torus. Duality supplies witnesses: a nonnegative trigonometric polynomial that is negative on \(T\) gives an operator measurable with equatorial settings.

Three results answer the three questions inside this class. For three qubits, rank-four PPT entangled states come in two types distinguished by a Lorentz invariant; every state of the first type is a Toeplitz state up to local operators, which yields a complete set of invariants for a seven-dimensional family, and of the second type exactly a one-parameter family is. Every three-qubit UPB state is Toeplitz up to local unitaries, which makes its white-noise threshold a moment problem that the authors solve exactly, finding a two-orbit, eight-point separable decomposition at the threshold and a matching witness. For rank, positive Toeplitz matrices below \(2(m_1+m_2)-2\) are moment matrices for two groups, and explicit integer extreme rays of that rank give extremal PPT states whose range contains no product vector.

The comparison in Table 3 puts the UPB states in context: they are not the most robust rank-four states. A GHZ-diagonal state tolerates \(1-1/\sqrt2\approx0.293\), Gabardo's state 0.1933 by a computer-assisted proof, and a local search found a type-I state at 0.2635. Which rank-four state tolerates the most noise, and how the method extends beyond the Toeplitz class or to more qubits, where the separable cone has no semidefinite description, are left open.

Transferable methods

  1. Partial transposition as a bit flip, separability as a moment problem

    Because a Toeplitz state's matrix elements depend only on differences of Hamming weights, transposing a set of qubits has the same effect as flipping them, so positivity under every partial transpose is automatic and the PPT test sees nothing. Separability then has to be decided another way, and it becomes the question of whether the Toeplitz coefficients are Fourier moments of a positive measure, a classical problem with a dual in nonnegative trigonometric polynomials.

    Let \(\bar x\) be \(x\) with the bits in \(S\) flipped. Then \(\bra x\rho_T^{\Gamma_S}\ket{x'}=t\bigl(w(x'')-w(x''')\bigr)\), where \(x''\) agrees with \(x\) outside \(S\) and with \(x'\) on \(S\), and \(x'''\) the other way round, while \(\bra xX_S\rho_TX_S\ket{x'}=t\bigl(w(\bar x)-w(\bar x')\bigr)\). Counting ones per group, both differences equal \(w(x|_{S^c})-w(x'|_{S^c})+w(x'|_S)-w(x|_S)\), so \[\rho_T^{\Gamma_S}=X_S\,\rho_T\,X_S\] and every partial transpose has the spectrum of \(\rho_T\).

    Full separability: if \(T=\sum_lp_l\,a(\theta_l)a(\theta_l)^*\) with \(a(\theta)_\alpha=e^{i\langle\alpha,\theta\rangle}\), then \(\rho_T\) is a mixture of products of equatorial states. Conversely, product vectors in the range lie in each group's symmetric subspace, so they have the form \(\bigotimes_ju_j^{\otimes m_j}\); equal diagonal entries force \(|a_{l,j}|=|b_{l,j}|\) for groups of two or more, and for single qubits the positive map \(\mathcal E(A)=\tfrac12(A+XA^{\mathsf T}X)\), which fixes \(\mathbb I,X,Y\) and kills \(Z\), turns any qubit state into a mixture of equatorial ones. If \(T\) is not a moment matrix, a separating hyperplane gives a trigonometric polynomial \(q\ge0\) on the torus with \(\langle q,T\rangle<0\), and \(W=\sum_d2^{|d|-n}q_d\bigotimes_i\Sigma_{d_i}\), with \(\Sigma_{\pm1}\) built from \(X\) and \(Y\), is a witness: \(\bra\theta W\ket\theta=2^{-n}q(\theta)\ge0\) on equatorial products and, since \(W\) has no \(Z\), on all products.

    Editorial check: for the Shifts UPB, \(\ket{0,1,+},\ket{1,+,0},\ket{+,0,1},\ket{-,-,-}\), the state \((\mathbb I-\Pi)/4\) has all three partial transposes positive, and the paper's eight product states with Bloch vectors \((\pm\alpha,\pm\beta)\) and \((\pm\beta,\pm\alpha)\) in the \(xz\) plane, \(\alpha=\sqrt{\sqrt2-1/2}\) and \(\beta=1-1/\sqrt2\), average exactly to the state with noise \(3-2\sqrt2\). That is an explicit separable model at the threshold, as Corollary 27 states.

    How to reuse it

    When a family of states has a phase-space representation with a symmetry that makes partial transposes trivial, separability becomes a classical moment problem. Positive polynomials then give witnesses restricted to the measurement bases of the representation, here equatorial ones, which is useful for spin ensembles with collective phase control.

  2. An exact noise threshold by symmetry reduction and a concave upper bound

    The threshold is the largest \(\tau=1-\lambda\) for which the scaled moments of the UPB state are moments of a probability measure. A Klein four-group of symmetries lets the measure be taken invariant, which reduces the moments to a point in five dimensions and the question to whether that point lies in the convex hull of an explicit surface. Two points on the surface show it does at \(\tau^*\); a supporting hyperplane, found through a concave function, shows it does not beyond.

    In Bravyi's normal form a three-qubit UPB is fixed by angles \(\theta_A,\theta_B,\theta_C\), and with \(U_B=\operatorname{diag}(1,i)e^{-i\pi X/4}\) on each qubit the state becomes \(T/8\) with \(t(d)=2\delta_{d,0}-\tfrac14\sum_{i=1}^4e^{i\langle d,\varphi_i\rangle}\) for four points \(\varphi_i\) of the torus. Averaging over the symmetry group, the moment conditions for \(T_\lambda=(1-\lambda)T+\lambda\mathbb I\) become \(\int\mathbf m\,d\mu=-\tau\mathbf m_0\) for a map \(\mathbf m:\mathbb T^3\to\mathbb R^5\) whose image is \[\mathcal M=\{m\in\mathbb R^5:\ |m|=1,\ \mu_4\mu_5=\mu_1\mu_2\mu_3\},\] and \(\mathbf m_0=(A_0,B_0,C_0,-e_1,e_2)\) has \(A_0B_0C_0=-s/8\), \(e_1e_2=s/8\), \(s=\sin2\theta_A\sin2\theta_B\sin2\theta_C\). So \(T_\lambda\) is separable exactly when \(-\tau\mathbf m_0\in\operatorname{conv}\mathcal M\).

    Achievability: with \((A,B,C)=-\tau(A_0,B_0,C_0)\), \(\delta=\tau(e_1+e_2)\), and \(u,v>0\) with \(v-u=\delta\), \(uv=\tau^3s/8\), the points \(P_\pm=(A,B,C,\cdot,\cdot)\) satisfy the product constraint and \[|P_\pm|^2=\tau^2\bigl(1+\tfrac s4\bigr)+\tau^3\tfrac s4.\] Editorial step: this equals 1 exactly when \(s\tau^2=4(1-\tau)\), since then \(\tau^2s/4\,(1+\tau)=(1-\tau)(1+\tau)\); the positive root is \(\tau^*=2/(1+\sqrt{1+s})\), and a convex combination of \(P_\pm\) hits \(-\tau^*\mathbf m_0\). Optimality: on \(\mathcal M\), \((\mu_4-\mu_5)^2=\det\Omega(\mu)\) for a \(3\times3\) correlation matrix, \(\sqrt{\det\Omega}\) is concave, so the linear functional \(\ell=(A+BC,B+AC,C+AB,\delta,-\delta)\) has maximum \(1+ABC\) on \(\mathcal M\), attained at an interior critical point, and \(-\tau\,\ell\cdot\mathbf m_0\le1+ABC\) forces \(\tau\le\tau^*\). The same \(\ell\) gives the witness polynomial \(q=1+ABC-\ell\cdot\mathbf m\ge0\).

    Editorial check: for three angle triples, including a random one, the witness of Eq. (18) has expectation zero on the threshold state to \(10^{-17}\), is negative on the noiseless UPB state, and has minimum zero over product states to \(10^{-11}\). Independently of the paper's construction, a linear program over probability measures supported on grids of \(24^3\) and \(48^3\) torus points finds separable decompositions at \(\lambda=0.1783\) and \(0.1738\) for Shifts and at \(0.1015\) and \(0.0976\) for \((0.4,0.9,1.2)\), decreasing towards the exact values 0.1716 and 0.0963, as grid restrictions must.

    How to reuse it

    For noise thresholds of symmetric entangled states: average the candidate separable measure over the symmetry group, describe the image of the reduced moment map algebraically, and look for a concave function, here the square root of a determinant, that turns the support function of its convex hull into a smooth optimisation. The optimal witness comes out of the same calculation.

Source: Section 2.1 (Definition 1, Proposition 2), Section 2.2 (Theorem 4), Section 4.1 (Proposition 23), Section 4.2 (Theorem 24, Lemmas 25–26), Section 4.3 (Shifts recipe, Table 2). 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. Exact white-noise thresholds were previously available mainly for highly symmetric families such as GHZ-diagonal states; here the whole three-qubit UPB family gets one, and Wan and Guo's question is answered: the new witness tolerates noise up to 0.1716 for Shifts against 0.1629 for the projector witness, and for random UPBs it detects a median of 6.7 times as wide a noise range. Since bound entanglement is fragile in experiments, a witness that is provably optimal against white noise, with a recipe of four Pauli settings for Shifts, is directly useful. The moment-problem reformulation brings classical tools, positive trigonometric polynomials and Hankel indices, to multipartite entanglement.

Limitations. The results are confined to the Toeplitz class, which for three qubits means states with no Pauli factor \(Z\) up to local unitaries, and mostly to three qubits; general multiqubit separability is untouched. The cone of separable Toeplitz matrices has no semidefinite description, so thresholds outside special families need computation: some values (Gabardo's state, 0.1933) rest on computer-assisted bounds, four-qubit values are numerical, one rank threshold is computer-assisted, and which rank-four state tolerates the most noise is open (a local search reached 0.2635 without converging). That the new witness is essentially unique is proved only near the Shifts UPB and tested numerically elsewhere. The authors state that Claude (Anthropic) assisted with literature searches, development and checking of proofs, numerical and symbolic computations and drafting, and that all results were verified by the authors.

Description based on the full arXiv HTML of v1, read 2026-11-20: Sections 1 and 2 (definitions, Theorem 4, Table 1) and Section 4 (Theorem 24, Corollaries 27–29, the Shifts recipe and Table 3); Sections 3 and 5 and the appendices were read only in their statements. Editorial checks: for the Shifts UPB in its standard frame, \(\operatorname{Tr}(W'_S\rho)=3\sqrt2-5\), all sixteen correlators equal \(\pm\tfrac14\) as stated, 300 local searches over product states found \(\min\langle W'_S\rangle\approx5\times10^{-15}\), and the eight product states of Table 2 average to \((1-\lambda^*)\rho+\lambda^*\mathbb I/8\) to \(6\times10^{-17}\). For Bravyi angles \((\pi/4,\pi/4,\pi/4)\), \((0.4,0.9,1.2)\) and a random triple, \(|\mathbf m_0|=1\), \(A_0B_0C_0=-s/8\), \(e_1e_2=s/8\), the witness of Eq. (18) vanishes on the threshold state and is nonnegative on product states to \(10^{-11}\). An independent linear program over measures on a grid of the torus gives separable states at \(\lambda=0.1783\) and \(0.1738\) (grids of \(24^3\) and \(48^3\) points) for Shifts, approaching 0.1716 from above, and 0.1015 and 0.0976 against the exact 0.0963 for \((0.4,0.9,1.2)\).