How much entanglement does it take to tell four twisted Bell states apart?

Two separated parties hold one of four entangled two-qubit states and have to work out which one. They also share a second, partially entangled pair to help. Primus for Quantum Mechanics derived the exact best-case success probability for this task under PPT measurements, proved it with matching upper and lower bounds, and found the precise condition under which the parties can succeed every time. A domain expert who had obtained the result privately confirmed it. This was the first result from Primus for Quantum Mechanics.

The problem

Alice and Bob sit in separate laboratories. Between them they share one of four orthogonal two-qubit states, chosen at random with equal probability, and their job is to identify which one they hold. A single party looking at their own qubit learns nothing. The four states are the twisted Bell states: the familiar Bell states, but with the two equal Schmidt coefficients replaced by an arbitrary pair $\alpha \ge \beta \ge 0$ with $\alpha^2 + \beta^2 = 1$.

$$|\psi_0\rangle = \alpha|00\rangle + \beta|11\rangle, \quad |\psi_1\rangle = \beta|00\rangle - \alpha|11\rangle, \quad |\psi_2\rangle = \alpha|01\rangle + \beta|10\rangle, \quad |\psi_3\rangle = \beta|01\rangle - \alpha|10\rangle.$$

When $\alpha = 1/\sqrt{2}$ these are the maximally entangled Bell states. When $\alpha = 1$ they are unentangled product states and the problem becomes trivial.

To help them, Alice and Bob share an extra pair of qubits in the resource state

$$|\tau_\varepsilon\rangle = \sqrt{\tfrac{1+\varepsilon}{2}}\,|00\rangle + \sqrt{\tfrac{1-\varepsilon}{2}}\,|11\rangle, \qquad \varepsilon \in [0, 1],$$

which is maximally entangled at $\varepsilon = 0$ and a product state at $\varepsilon = 1$. The question is how the best achievable success probability depends on both knobs: how entangled the signal states are ($\alpha$) and how much entanglement the resource carries ($\varepsilon$).

Why PPT measurements

What Alice and Bob can actually do is local operations and classical communication, or LOCC. LOCC protocols are notoriously hard to optimize over directly. The standard workaround is to optimize over a larger, better-behaved class: measurements whose every element has a positive partial transpose (PPT). Every LOCC measurement is PPT, so the best PPT success probability is an upper bound on what any local protocol can achieve. And because the PPT class is defined by linear and positive semidefinite constraints alone, the best PPT measurement is the solution of a semidefinite program (SDP).

The point of an SDP is that it comes with a dual. Any feasible measurement gives a lower bound on the optimum. Any feasible dual witness gives an upper bound. When the two values coincide, the optimum is known exactly. That is the standard the paper holds itself to: no numerics, only a closed-form measurement and a closed-form witness that agree.

The result

Define $\lambda_0 = (1 + \varepsilon)/2$ and $\lambda_1 = (1 - \varepsilon)/2$, the Schmidt weights of the resource, and set

$$v = 2\alpha\beta\lambda_0 - \lambda_1, \qquad v_+ = \max\{v, 0\}.$$

Then the optimal probability that a PPT measurement identifies the shared state is

$$P_{\text{succ}} = \frac{1}{2}\left(1 + \sqrt{1 - v_+^2}\right).$$

Three things follow directly.

• Perfect discrimination has a clean boundary. The parties succeed with certainty exactly when $v \le 0$, that is, when $2\alpha\beta(1 + \varepsilon) \le 1 - \varepsilon$. It is a trade-off rather than a threshold: the more entangled the four signal states are, the more entanglement the resource has to carry before perfect discrimination is possible. Outside that region no PPT measurement succeeds perfectly, so no LOCC protocol does either.
• Known results fall out as special cases. On the Bell line $\alpha = \beta$ the formula reduces to $\frac{1}{2}(1 + \sqrt{1 - \varepsilon^2})$, the value established by Bandyopadhyay, Cosentino, Johnston, Russo, Watrous and Yu for four Bell states. With no resource ($\varepsilon = 1$) it gives $1/2$, the classical result for four Bell states under LOCC.
• Teleportation stops being optimal. For the Bell states, the best protocol is for Alice to teleport her qubit through the resource and for Bob to measure. The paper shows this protocol attains only $\frac{1}{2}(1 + \sqrt{1 - (c\varepsilon)^2})$ with $c = 2\alpha\beta$, which is strictly below the PPT value whenever $0 < \beta < \alpha$ and $0 < \varepsilon < 1$. Twisting the basis breaks the protocol that worked for the untwisted one.

How the proof goes

The paper builds both halves of the certificate explicitly.

The primal side is a family of PPT measurements built from the symmetry of the problem. The four states are related by local Pauli operations, and the resource is invariant under a matching group, so the measurement can be chosen covariant under that group. This reduces the search to a small number of real parameters, and the paper maximizes the resulting success probability in closed form and verifies that every element of the measurement is PPT.

The dual side is an explicit dual witness whose value equals the primal maximum. Constructing it means exhibiting positive semidefinite operators satisfying the dual constraints sector by sector. The paper does this by hand, checking completeness, the trace conditions, and the positivity of each block, including the case $v \le 0$ where the witness certifies that perfect discrimination is possible.

Because the two values match, the formula is the exact optimum, not a numerical approximation. Demonstrating that an autonomous system can produce a certificate of this kind, where every step is checkable, is the part of this work we find most promising as a benchmark for computational physics.

What remains open

The paper closes with three questions it does not answer.

• Whether separable or LOCC measurements attain the PPT value when $\alpha \ne \beta$. The optimal PPT measurement exhibited here has no product structure, so no local protocol achieving the value is in hand. For three Bell states the separable and PPT values are known to differ, so either answer is possible.
• The doubly twisted bases studied by Bandyopadhyay, Halder and Nathanson, where the two pairs of states carry independent Schmidt coefficients. The symmetry argument here requires the two pairs to be related by a local unitary.
• The three-state problem. Even on the Bell line the PPT value is unknown for $\varepsilon > 1/3$, where perfect discrimination is known to fail. The transitive symmetry used here is absent, and a different construction is needed.

How this paper was made

This result was produced by Primus for Quantum Mechanics, a specialization of Primus with support for Physlib's QuantumInfo library for Lean 4, toqito, cvxpy semidefinite programming, sympy and QuTiP. Primus was given the problem at a high level, read the field's literature, set up the SDP, constructed the measurement and the witness, and wrote the paper. The derivation was reviewed by a domain expert in quantum information theory who had previously obtained the result privately and who confirmed the closed-form answer and the dual SDP framework.

The full Primus research log is public, and the key steps were also formalized in Lean.

• Browse the Primus research report →
• See the Lean formalization →

The paper, with the full construction of the measurement and the dual witness, can be downloaded below.