Some quantum states are entangled, meaning their parts are correlated in ways no classical mixture can reproduce, yet no procedure using local operations and classical messages can extract even one clean unit of usable entanglement from them. These are called bound entangled states. They are the sharpest example of irreversibility in quantum information: making them costs entanglement, and nothing can be recovered afterwards. Finding such states is hard because the target is a conjunction of properties that pull against each other. The state must be a valid density matrix. It must stay positive after one subsystem is transposed, which is the "positive partial transpose" or PPT test and is what forbids distillation. It must still be entangled. Often it must also meet further requirements, such as a fixed rank or a nonzero lower bound on its entanglement cost. Today such states come either from hand-built algebraic constructions or from blind search.
We propose to treat this as a compositional generation problem, in the sense text-to-image researchers use the phrase. Their image generators draw each requested object well and then drop or mis-bind objects when a prompt asks for several at once. Our bet is that direct optimizers and code-writing language models fail on conjunctions of entanglement properties in the same way: they satisfy the dominant constraint and silently neglect the rarer one. The fix from image generation also transfers. Plan a discrete structure first (there, a layout; here, a set of product vectors whose orthogonal complement contains no product state), then fill in continuous parameters under verifier guidance. The deliverables are three. The first is a generator that produces exactly certified bound entangled states for specifications that known constructions do not reach. The second is a checklist benchmark for this task. The third is a first empirical map between the dimension of a trapping subspace and a computable lower bound on its entanglement cost.
Core Claim
The mathematics fixes the claim. A target specification $s$ consists of local dimensions $d \times d'$, a rank $r$, and a cost threshold $\tau$. A state $\rho$ satisfies $s$ when three conditions hold: $\rho \ge 0$ and $\mathrm{tr}\,\rho = 1$; $\rho^{\Gamma} \ge 0$, where $\Gamma$ denotes partial transpose on the second subsystem; and $\mathrm{rank}\,\rho = r$. In addition, entanglement must be certified by a certificate $c(\rho)$ satisfying
$$ c(\rho) \;\ge\; \tau > 0 . $$
Here $c$ is the logarithmic fidelity of $k$-negativity. Computable and Faithful Lower Bound on Entanglement Cost introduced this quantity as a semidefinite-programming lower bound on entanglement cost under PPT operations. Because PPT forces zero distillable entanglement, any state meeting the specification has a certified irreversibility gap of at least $\tau$.
We make two claims. The first is that generators without an explicit subspace plan show the text-to-image failure modes. "Neglect" means the output is PPT but separable, or entangled but not PPT. "Numeracy failure" means the rank drifts away from $r$. These failures should become rarer-constraint neglect as conjunctions grow. The second claim is that a planner that first commits to an unextendible product basis and then fills its complement will raise certified success on multi-constraint specifications at matched verifier budget.
Where This Sits
Direction 1 supplies the constructions that make planning possible. Robustness of entangled states that are positive under partial transposition formalizes the key fact: the normalized projector onto the complement of an unextendible product basis is PPT and entangled, because that complement contains no product vector. The same paper proves that almost all PPT entangled states keep both properties under small perturbation. Non-full-rank bound entangled states satisfying the range criterion mixes complement states with separable states inside the basis span to reach non-full-rank states that pass a classic separability test. Construction of bound entangled states based on permutation operators applies local scaled-permutation maps to existing bound states and obtains states that are PPT, violate the range criterion, and are not locally-unitarily equivalent to their seeds.
We reuse all three as moves available to the planner. None of these papers searches over specifications, and none relates the geometry of the construction to entanglement cost. The file lists that relation as open, and Zhao et al. leave it uncharacterized.
Direction 2 supplies the verifier and the scientific question. The Zhu–Jing–Wang semidefinite program is our reward and our certificate. We only use it for small bipartite systems, where the exponential scaling the file flags is manageable. Irreversibility of Asymptotic Entanglement Manipulation Under Quantum Operations Completely Preserving Positivity of Partial Transpose proves that the irreversibility we are certifying is real under physical PPT operations. Reversible Entanglement Beyond Quantum Operations shows the gap closes to the logarithmic negativity under non-physical quasi-operations. That result constrains our design in one specific way. Logarithmic negativity is zero on PPT states, so it cannot be the reward. Only a cost-type certificate distinguishes the states we want.
Direction 3 supplies the methodology. The planning stage borrows its logic from LayoutGPT, which plans bounding boxes before pixels, and from StruVis, which plans structured intermediates before generation. The continuous fill borrows from reward-based noise optimization in ReNO and CARINOX. Hallucination Early Detection in Diffusion Models detects object omission early in generation and restarts, and we test whether constraint neglect is similarly predictable from early optimization iterates. The evaluation design copies the checklist style of GenEval, which reports roughly 98% success on single objects falling to roughly 15% on spatial position. The analogy we test is that single properties are easy and conjunctions are hard.
QuanBench motivates the work by showing a gap. Its state-preparation tasks are pure textbook states, and it has no subspace-constrained or bound-entanglement tasks.
Research Program
We first build a benchmark we call BoundCheck. Its specifications span 3×3, 2×4, 3×4, 2×8 and 4×4 systems, where the semidefinite program is tractable. They are organized like GenEval categories. Single-property prompts ask only for "entangled" or only for "PPT". Binding prompts ask for "PPT and entangled". Counting prompts add an exact rank. Hard compositions add a cost threshold $\tau$, a range-criterion status (satisfied or violated), and a requirement of local-unitary inequivalence to a catalogue of known constructions.
Certification uses exact arithmetic wherever possible, because non-full-rank PPT states have zero eigenvalues in their partial transpose, and floating-point tests misclassify them. The planner therefore outputs rational or low-degree algebraic parameters. PPT and rank are checked in exact arithmetic. Unextendibility of the planned product set is checked exactly, by showing that the system of polynomial equations asking for a product vector orthogonal to the set has no solution. The cost certificate is a floating-point semidefinite program with a rigorous dual-feasibility margin. Local-unitary inequivalence is tested with spectra of reduced states and of partial transposes. Equal spectra are inconclusive, but different spectra prove inequivalence.
The generator has two stages. The planner is a language model prompted with the specification and a small library of basis families. It proposes an orthogonal product set: its size, its local vectors (restricted to real coefficients when the range criterion must be satisfied), and an optional local permutation-type transform. The exact unextendibility check gates each plan. Plans that fail return the offending product vector as feedback, much as RePrompt-style systems feed back verifier signals.
The filler then optimizes continuous parameters against the certificate using a fixed budget of calls to the semidefinite program. These parameters are the mixture weights over the complement and the optional separable admixture inside the basis span, and they are chosen so that the rank constraint holds by construction. We also train a small classifier that predicts final failure from early iterates. It triggers restarts, and we measure the saved verifier calls.
The scientific analysis regresses certified cost lower bounds on the complement dimension, rank, and the number and type of local transforms, across all certified outputs. This directly addresses the open relationship between subspace geometry and cost that the file identifies in Directions 1 and 2.
Baselines
The naive reference point is rejection sampling. We draw random density matrices at the specified rank, keep those that are PPT, and run the same certificate on them, with the same number of certificate calls as our method. It measures how rare certifiable PPT entanglement is by default. Beating it means higher certified success per verifier call.
The strongest existing method is the algebraic toolkit itself. We enumerate known basis complements, run the Bandyopadhyay et al. range-criterion mixtures, and apply the Zhao et al. permutation transforms, followed by small random perturbations justified by the robustness theorem. We give it the same certificate budget and the same exact checks. No reported result found on Lacuna for success rates of these constructions against a specification checklist. The papers establish existence, not coverage.
The third baseline is direct reward optimization without a plan. We parameterize the full density matrix as $\rho = AA^\dagger/\mathrm{tr}(AA^\dagger)$, where $A$ is a $dd' \times r$ complex matrix. We optimize a penalized reward for PPT plus the cost certificate at the same budget. This is the ReNO/CARINOX analogue and doubles as the ablation that removes subspace planning.
The fourth baseline is a strong language model writing Qiskit or NumPy code for the specification, with and without verifier feedback. The closest reported numbers come from QuanBench on standard tasks. There no model exceeded 40% Pass@1, the best Pass@5 was 50%, and "correct" code averaged only about 51.8% process fidelity. We expect lower success here and report it for comparison.
We claim success if plan-then-fill beats the algebraic toolkit on rank-constrained and cost-threshold specifications and beats direct optimization on every multi-constraint category at matched budget. It must do so while matching both on single-property prompts. A single-property win would suggest a difference in tuning rather than an effect of composition.
Why not just…?
Why not just perturb known constructions, since the robustness theorem says PPT entanglement survives small perturbations? This is the real threat. It will likely match us on full-rank "PPT and entangled" prompts. The weakness is that generic perturbations raise rank to full, so it cannot satisfy exact-rank or range-criterion specifications without re-deriving structure. It also stays near its seeds, so it should fail the inequivalence and high-$\tau$ prompts. If perturbation matches us on those categories too, the planning stage adds nothing and we will report that.
Why not just sample randomly and certify? Certified PPT entangled states occupy a small, hard-to-detect region at low rank. The certificate is expensive, so yield per call is what matters, and the rejection baseline measures exactly this.
Why not just ask a language model to write the state directly, with verifier feedback? QuanBench already shows weak fidelity on far easier pure-state tasks. More to the point, our ablation gives the same model the same feedback but removes the exact unextendibility gate. That isolates whether the gain comes from the model or from committing to a checked discrete plan.
Proposed Experiments and Decisive Checks
The first check tests whether neglect exists. For direct optimization and code generation, we tabulate which constraint fails as conjunctions grow. The claim is supported if failures concentrate on the rarer constraint, entanglement or rank, rather than spreading uniformly across constraints. The check also asks whether an early-iterate classifier predicts final failure well enough to save verifier calls, as HEaD+ reports for images, where restarts cut generation time by up to 32%.
The second check reports certified success per category at matched budget against all baselines, over several random seeds. It includes the perturbation baseline on exact-rank and inequivalence prompts, which is where the method should win or fail visibly.
The third check is the geometry-to-cost map. If certified lower bounds vary systematically with complement dimension and rank across hundreds of certified states, that pattern is a concrete conjecture for theorists. A null result, meaning no relation, is also informative. It would indicate that cost depends on finer structure than dimension counts.
References
- Computable and Faithful Lower Bound on Entanglement Cost
- Robustness of entangled states that are positive under partial transposition
- Non-full-rank bound entangled states satisfying the range criterion
- Construction of bound entangled states based on permutation operators
- Irreversibility of Asymptotic Entanglement Manipulation Under Quantum Operations Completely Preserving Positivity of Partial Transpose
- Reversible entanglement beyond quantum operations
- LayoutGPT: Compositional Visual Planning and Generation with Large Language Models
- StruVis: Enhancing Reasoning-based Text-to-Image Generation via Thinking with Structured Vision
- ReNO: Enhancing One-step Text-to-Image Models through Reward-based Noise Optimization
- CARINOX: Inference-time Scaling with Category-Aware Reward-based Initial Noise Optimization and Exploration
- Hallucination Early Detection in Diffusion Models
- GenEval: An object-focused framework for evaluating text-to-image alignment
- QuanBench: Benchmarking Quantum Code Generation with Large Language Models