Bound Entanglement with Negative Partial Transpose

Bound Entanglement with Negative Partial Transpose

Definitions and mathematical statement

A finite-dimensional bipartite state is a positive semidefinite trace-one operator $\rho$ on $H_A\otimes H_B$. It is entangled if it is not a convex combination of product states. Partial transpose $\rho^{T_B}$ transposes the second subsystem in an orthonormal basis; having a negative eigenvalue is basis independent. LOCC consists of local quantum operations with classical communication. The target asks for $\rho^{T_B}\not\succeq0$ but zero asymptotic rate of distilling maximally entangled pairs from $\rho^{\otimes n}$. Equivalently, in the standard finite-dimensional distillability criterion, seek an NPT state with \[ \langle\psi|(\rho^{T_B})^{\otimes n}|\psi\rangle\ge0 \] for every $n\ge1$ and every vector $\psi$ of Schmidt rank at most two across $A^n:B^n$. This is not merely failure of single-copy distillation.

Formulation and concept references: [S1].

Short English statement

Can a quantum state have a negative partial transpose yet yield no distillable entanglement, even when arbitrarily many copies and two-way classical communication are available?

Research attempt

Scope and recent-source check.

GPT-6 Astra Ultra prepared this unresolved attempt on 27 September 2026 using explicit matrix reductions and primary-source inspection. The target requires non-distillability for every number of copies, across the grouped bipartition $A^n:B^n$. Source [S1] also discusses a specific two-copy Werner-state problem. July 2026 manuscripts [E1, E2] report solutions to that two-copy question in arbitrary dimension. Their stated scope was inspected; their proofs are not independently verified or used here. Even accepting a two-copy theorem would not produce the all-copy state requested by the catalogue. The work below keeps that quantifier visible and proves elementary restricted cases.

A normalized Werner family and its negative eigenvalue.

Let $F$ be the swap on $\mathbb C^d\otimes\mathbb C^d$, and for $0\le\alpha\le1$ set \[ \rho_\alpha=\frac{I-\alpha F}{d^2-\alpha d}. \] Since $F$ has eigenvalues $+1,-1$, the numerator is positive semidefinite. Also $\operatorname{Tr}F=d$, so the denominator makes this a state; for $d\ge2$ it is strictly positive as a scalar. For $|\Phi_d\rangle=d^{-1/2}\sum_i|i,i\rangle$ and $P_d=|\Phi_d\rangle\langle\Phi_d|$, direct transposition gives \[ F^{T_B}=dP_d,\qquad \rho_\alpha^{T_B}=\frac{I-\alpha dP_d}{d^2-\alpha d}. \] Its eigenvalue on $\Phi_d$ is negative exactly when $\alpha>1/d$; all orthogonal directions have positive eigenvalue. This verifies NPT by an exact spectrum calculation rather than a rounded numerical test. Because a separable state stays positive under partial transpose, every one of these NPT states is entangled.

The one-copy Schmidt-rank threshold is exact.

Represent a normalized bipartite vector by its coefficient matrix $X$, so $\|X\|_F=1$, its Schmidt rank is $\operatorname{rank}X$, and \[ |\langle\Phi_d|\psi\rangle|^2=|\operatorname{Tr}X|^2/d. \] For rank at most $r$, the trace is bounded by the sum of the singular values, which is at most $\sqrt r\|X\|_F$. This proves the sharp bound $|\langle\Phi_d|\psi\rangle|^2\le r/d$. Equality is attained by the normalized sum of $r$ matching basis vectors. For Schmidt rank at most two, therefore, \[ \langle\psi|\rho_\alpha^{T_B}|\psi\rangle \ge\frac{1-2\alpha}{d^2-\alpha d}. \] It follows that the state is one-copy undistillable exactly for $\alpha\le1/2$, and one-copy distillable for $\alpha>1/2$. For $d\ge3$, the interval \[ 1/d<\alpha\le1/2 \] contains NPT states which pass the entire one-copy Schmidt-rank-two test. In dimension two that interval is empty. This is a fully checked partial result, but passing one copy is not the requested all-copy conclusion.

The two-copy condition becomes a rank-two matrix inequality.

For two copies, group the row indices as $A_1A_2$ and the column indices as $B_1B_2$. Write the coefficient matrix of a vector as $X\in M_{d^2}(\mathbb C)$, with entries $X_{ia,jb}$. The relevant restriction is $\operatorname{rank}X\le2$; assume $\|X\|_F=1$. Define the partial traces algebraically by \[ (\operatorname{Tr}_1X)_{ab}=\sum_iX_{ia,ib},\qquad (\operatorname{Tr}_2X)_{ij}=\sum_aX_{ia,ja}. \] These are partial traces of a coefficient matrix and do not require $X$ itself to be Hermitian or positive.

Expanding the rank-one maximally entangled projectors gives \[ \langle P_{A_1B_1}\otimes I\rangle =\frac1d\|\operatorname{Tr}_1X\|_F^2, \qquad \langle I\otimes P_{A_2B_2}\rangle =\frac1d\|\operatorname{Tr}_2X\|_F^2, \] and the expectation of both projectors is $|\operatorname{Tr}X|^2/d^2$. Consequently the sign of the two-copy expectation is the sign of \begin{equation}\label{wsref-0-61-32-37-36-3a-74-77-6f-63-6f-70-79} 1-\alpha\left(\|\operatorname{Tr}_1X\|_F^2+ \|\operatorname{Tr}_2X\|_F^2\right) +\alpha^2|\operatorname{Tr}X|^2. \end{equation} The omitted denominator is $(d^2-\alpha d)^2>0$. At $\alpha=1/2$, nonnegativity is equivalent to \[ \|\operatorname{Tr}_1X\|_F^2+ \|\operatorname{Tr}_2X\|_F^2 \le2\|X\|_F^2+\tfrac12|\operatorname{Tr}X|^2 \quad(\operatorname{rank}X\le2). \] This is an exact reformulation, not a proof of the displayed inequality for every rank-two matrix. It also explains why positivity tests only on Hermitian or diagonal matrices cannot exhaust the distillability criterion.

Two restricted tests that can be proved directly.

First, rank-one vectors across $A^n:B^n$ can never give a negative expectation for any physical state $\rho$, at any number of copies: \[ \langle a\otimes b|(\rho^{T_B})^{\otimes n}|a\otimes b\rangle =\langle a\otimes\overline b|\rho^{\otimes n} |a\otimes\overline b\rangle\ge0. \] The complex conjugation is in the same basis used for the partial transpose. This is block positivity; it does not extend from rank one to rank two by linearity, since the cross terms then matter.

Second, restrict the test vector to factor across copies. Its coefficient matrix has the form $X=X_1\otimes\cdots\otimes X_n$. The rank is $\prod_i\operatorname{rank}X_i$. If this is at most two, every factor has rank at most two and all but possibly one have rank one. For normalized factors and $\alpha\le1/2$, each single-copy expectation of $I-\alpha dP_d$ is nonnegative by the preceding threshold proof. Their product is nonnegative. Thus no copy-factorized eligible vector can distill a Werner state in the candidate interval, regardless of how many copies are supplied.

The restriction is strong. A rank-two matrix on the grouped spaces need not be a tensor product of single-copy matrices. Conversely, tensoring $n$ arbitrary single-copy rank-two witnesses generally gives rank $2^n$, which is not an allowed witness in the criterion. Both mistakes can lead to invalid inductions on copy number.

For a simple genuinely correlated rank-two test, take the diagonal matrix with entries $1/\sqrt2$ at $(00,00)$ and $(11,11)$ and zeros elsewhere, for $d\ge2$. Both partial traces have squared norm one, and $|\operatorname{Tr}X|^2=2$. Formula (a276:twocopy) becomes $1-2\alpha+2\alpha^2$, always positive. This verifies one family of correlated tests without pretending that it exhausts the compact rank-two variety.

Why tensor positivity is not a valid induction.

At an NPT parameter the operator $I-\alpha dP_d$ is not positive. Its negative eigenvector has Schmidt rank $d$, so the one-copy rank-two restriction hides that eigenvalue when $d>2$. Tensoring operators that are nonnegative only on a restricted class of vectors does not in general preserve that restricted positivity: the allowed vectors in the grouped tensor product have new cross-copy correlations. An induction would need a decomposition theorem expressing every eligible global vector through eligible local tests with nonnegative weights. No such theorem has been established here. Expanding a vector in a basis gives interference terms, not a convex combination of product expectations.

Similarly, the presence of a negative tensor-product eigenvalue says nothing unless a negative vector of Schmidt rank at most two is exhibited. Searching all eigenvectors of the full operator addresses ordinary positivity, a different question that NPT has already answered.

Compactness exposes the remaining quantifier issue.

Let $B_n\subset[1/d,1/2]$ be the set of parameters that pass every Schmidt-rank-two test at copy number $n$. It is closed: each fixed vector's expectation is continuous in $\alpha$, and the set is the intersection of their closed nonnegativity sets. Also $B_{n+1}\subseteq B_n$. To see this for the Werner family, tensor any negative $n$-copy witness with a product vector orthogonal to $\Phi_d$ in the new copy. The extra numerator expectation is one, so negativity and Schmidt rank at most two are preserved.

If there were a fixed $\delta>0$ such that every $B_n\cap[1/d+\delta,1/2]$ were nonempty, compactness and nesting would give a parameter in their intersection. Its state would be NPT and undistillable at every copy number, answering the problem. The uniform margin above $1/d$ is essential. Nonempty sets of NPT candidates for each finite $n$ could approach the PPT endpoint; the abstract nested intervals $[1/d,1/d+1/n]$ show the logical failure. Their intersection contains no NPT parameter, despite positive-width NPT portions at every finite stage.

Verification, gap, and outcome.

The partial-transpose spectrum, Schmidt-rank bound and coefficient-matrix contractions were checked independently, including nonsymmetric complex rank-two examples for the contraction identities. Finite numerical searches verify only individual vectors and cannot prove universal positivity on the rank-two variety. Recent two-copy claims do not remove the all-copy quantifier or establish the uniform-margin condition. The first remaining gap is an all-copy restricted-positivity mechanism for one fixed NPT state, or a proof every NPT state eventually fails it. The one-copy threshold and the stated restricted multi-copy tests are proved; the existence of NPT bound entanglement remains UNRESOLVED.

Sources

• S1 P. Horodecki, L. Rudnicki, and K. Zyczkowski, Five Open Problems in Quantum Information Theory, PRX Quantum 3 $2022$. https://arxiv.org/pdf/2002.03233. Formulation source.
• E1 Jinshi Fu, Li Gao and Sang-Jun Park, A solution to 2-copy distillability of Werner states, arXiv:2607.21367, submitted 23 July 2026. https://arxiv.org/abs/2607.21367. Abstract and scope inspected 27 September 2026; claimed two-copy result, not used as a proved all-copy assertion.
• E2 Zhiwei Song and Lin Chen, A partial-trace matrix inequality and Werner-state distillability, arXiv:2607.23416, submitted 26 July 2026. https://arxiv.org/abs/2607.23416. Abstract and scope inspected 27 September 2026; proof not independently certified in this attempt.