For C an operator on (C^d)^⊗3 of rank at most 2, let q(C) = sum over subsets S of {1,2,3} of (-1/2)^|S| times ||Tr_S C||_F^2. The inequality q(C) >= 0 is the three-copy Werner endpoint. I prove nothing new about Werner states in this note: the endpoint was already proved for every d >= 3 by Tomar, whose Lean formalization covers 3 <= d <= 8 only. What I contribute is a complete kernel-only Lean 4 proof of the endpoint for every d, by an independent, self-contained route: one exact rational certificate of the t = 1/d -> 0 limit of a strengthened moment relaxation, carried to every dimension by a lemma chain, with the certificate's exact identity proved in the Lean kernel. The final theorem depends only on the standard axioms propext, Classical.choice, and Quot.sound, with no sorry and no native_decide, and I had its statement independently checked to confirm it says what I claim here.
For C an operator on (C^d)^⊗3 of rank at most 2, let q(C) = sum over subsets S of {1,2,3} of (-1/2)^|S| times ||Tr_S C||_F^2. The inequality q(C) >= 0 is the three-copy Werner endpoint. I prove nothing new about Werner states in this note: the endpoint was already proved for every d >= 3 by Tomar, whose Lean formalization covers 3 <= d <= 8 only. What I contribute is a complete kernel-only Lean 4 proof of the endpoint for every d, by an independent, self-contained route: one exact rational certificate of the t = 1/d -> 0 limit of a strengthened moment relaxation, carried to every dimension by a lemma chain, with the certificate's exact identity proved in the Lean kernel. The final theorem depends only on the standard axioms propext, Classical.choice, and Quot.sound, with no sorry and no native_decide, and I had its statement independently checked to confirm it says what I claim here.
A Werner state is three-copy distillable exactly when it is one-copy distillable. For every local dimension d ≥ 3, the Werner state ρ(d,α) = (I + αF)/(d² + αd) is three-copy distillable if and only if α < −1/2. The same threshold for two-copy distillability was proved in July 2026, and the qutrit case of the three-copy problem had been settled by an archived certificate. In every dimension we prove four-copy undistillability for α ≥ −79/160 and five-copy undistillability for α ≥ −9/20; for qutrits (d = 3) the bounds improve to −0.4979 and −0.47. Applied to disjoint groups of copies, the results for up to five copies give k-copy undistillability for α ≥ −11/(5k) when k ≥ 6, again in all dimensions. The endpoint α = −1/2 remains open from four copies on, and numerically the four-copy relaxation is not exact there. By a criterion of P. Costa Rico, each statement is equivalent to an inequality for complex matrices of rank at most two. We prove these inequalities with quartic moment relaxations and exact rational certificates. For three copies there is one certificate for each dimension up to 16, and one certificate whose entries are polynomials in 1/d covers every dimension d ≥ 9. For four and for five copies, one certificate each covers all dimensions from 1.25·10⁹ and from 10¹⁰ on, respectively, and zero padding passes both results to all smaller dimensions. The bound for many copies comes from a linear program over averaged partial-trace norms, with exact rational certificates. Version 3 adds results on the zeros of the endpoint form. For two copies, the range and co-range of any matrix at which the form vanishes lie in tensor products of two-dimensional subspaces, and each such plane carries a zero; in dimension four this gives all equality cases of a singular-value bound for Kronecker sums. For any number of copies the form is nonnegative near its generic zeros of this kind and, when it is nonnegative for one copy fewer, near product zeros. If k − 1 copies are settled, a failure at k copies would first appear below the endpoint, at a minimiser whose range and co-range have equal local supports of dimension at least three. A quantitative conjecture that would imply undistillability for all numbers of copies is stated and proved for two copies on one class of planes. The record contains the paper, seven verification packages with the certificates, verifiers and checks, and two Lean 4 formalizations: of the three-copy result for d = 3, …, 8, checked by Lean's kernel alone, and of the four-copy qutrit bound, in which the checks of the certificate identity are evaluated by compiled code.
We solve the two-copy distillability problem for Werner states in every local dimension. Our main matrix result is a sharp, dimension-free inequality: for every rank-at-most-two operator, the sum of the squared Hilbert--Schmidt norms of its two partial traces is bounded by twice its squared Hilbert--Schmidt norm plus one half of the squared modulus of its trace. This implies that a Werner state ρα is two-copy distillable if and only if α<−1/2. In particular, the two-ququart state ρ−1/2(4) is two-copy undistillable, resolving Problem 5 of Horodecki, Rudnicki, and Życzkowski. For an arbitrary finite number k of copies, we give three exact formulations of the remaining problem. At the endpoint α=−1/2, undistillability is equivalent to nonnegativity of the endpoint partial-trace form on every rank-at-most-two operator. We also derive an equivalent hierarchy of operator inequalities Hk(ψ)⪰0 for pure states with a maximally mixed qubit marginal. The two-copy proof does not formally induct, because partial trace can increase rank and 2-positivity is not generally preserved by tensor products. We prove two rigorous many-copy extensions. First, the quadratic form factorizes exactly on tensor-factorized witnesses; for any such decomposition, the endpoint inequality holds if its possible rank-two factor is supported on a block containing at most two copies. Second, we construct explicit constants γk>0 such that α≥−γk implies k-copy undistillability in every dimension. The initial proofs were generated by ChatGPT 5.6 Sol; the authors have verified and rewritten them to enhance readability and provide additional context.
Whether every entangled state with a non-positive partial transpose (NPT) is distillable reduces to the case of the Werner states ρ(d,α) = (I + αF)/(d² + αd), where F is the swap. For d ≥ 3 and −1/2 ≤ α < −1/d these states are NPT but one-copy undistillable. Write a vector on k copies as a matrix C, and let Tr_J trace out the copies in J. By the Schmidt-rank criterion of DiVincenzo et al. and Dür et al. (2000), as rewritten by P. Costa Rico (2025), ρ(d,α) is k-copy undistillable exactly when Σ_J α^|J| ‖Tr_J C‖² ≥ 0 for every vector of Schmidt rank at most two on k copies. At α = −1/2 we call this the endpoint inequality. It was proved independently for two copies by Fu et al., Fraser et al., Song and Chen, and Bharti et al. (2026), and for three copies by a computer-assisted proof in the companion record doi:10.5281/zenodo.23604343. If it held for every k, NPT bound entanglement would exist. We describe its equality cases for two copies and where it could first fail, and we prove it on two classes of matrices, one for every number of copies and one for four copies. For two copies we prove that every equality case, that is, every nonzero C of rank at most two with equality in the endpoint inequality, has rank two and that the ranges of C and C* lie in products of two-dimensional single-copy subspaces. We obtain this by following the equality cases through Song and Chen's elementary proof of the two-copy inequality. Every known equality case, for any number of copies, has a range with a reduced state of rank at most two on some copy. We also show that a first failure would come with larger reduced states: if the endpoint inequality holds for k − 1 copies but fails for k, then for some α > −1/2 the inequality still holds, with equality at a nonzero C whose range has reduced states of rank at least three on every copy. This rests on a copy-removal lemma built on the separable ball of Gurvits and Barnum. We prove the endpoint inequality, for every k and d ≥ 3, for the matrices C whose range is spanned by a vector f_0 of V_0 and a vector f_1 of V_1 and whose co-range (the range of C*) is arbitrary. Here V_c is spanned by |c⟩^⊗k and the k vectors obtained from it by replacing one factor with |2⟩, for orthonormal |0⟩, |1⟩, |2⟩; for generic f_0 and f_1 the projector onto the range has reduced states of rank three on every copy, as a first failure would require. So no vector of this kind witnesses distillability of ρ(d,−1/2). Schur complements reduce the claim to a quadratic form; an induction on the number of copies then completes the argument. For four copies we establish the endpoint inequality in every dimension for all C whose range and co-range lie in one two-dimensional subspace. This class contains all normal matrices of rank two, for which Wu and Zou had proved the three-copy case (arXiv:2608.02647). Here we restrict the quartic moment relaxation used in the companion record to this class, where it involves the moments of a single vector; one exact certificate, together with an embedding into larger dimensions, then covers every dimension. For d = 3 the theorem is also proved in Lean 4 with Mathlib, in the package release_lean_commonplane. Some data checks in that proof are evaluated by compiled code (native_decide). For general C the endpoint inequality remains open from four copies on. This record contains the paper, the Lean package named above, and two packages of Python programs, each with a README that says how to run its checks. The package release_endpoint covers the computer-verified steps of the proofs, among them a symbolic check of five rational identities; these do not involve k, and checking them is the only computer step in the proof for every k. The package release_commonplane contains the four-copy certificates and two implementations of their exact checks; the second was written independently, from the definitions.
The best known lower bound for the minimum Kochen-Specker vector system in R3 -- 24 vectors -- rests on a computational proof whose combinatorial half emits DRAT proofs but whose geometric half does not: the non-embeddability of thousands of candidate graphs is established by Z3's nonlinear real arithmetic, which produces no checkable proof objects. We close this gap for the proof's blocking database. We introduce exact rational case-tree certificates of real non-embeddability, whose splits are polynomial factorizations and rational sum-of-squares decompositions and whose leaves are discharged by injectivity, ideal-membership, or Positivstellensatz-shaped positivity arguments, and we certify all 291 source lines (180 distinct graphs) of the published pipeline's order-10 to order-13 blocking lists. Certificates are replayed by two independent checkers that share no code with the generator: a pure-Python replay over exact fractions, and a total checker implemented and proved sound in Lean 4. The soundness theorem -- acceptance implies that no injective-on-rays, orthogonality-respecting assignment of nonzero real vectors realizes the graph -- is kernel-checked with axiom closure {propext, Classical.choice, Quot.sound}, and a gcd-free rational arithmetic layer makes the entire verdict computation kernel-reducible, so each per-graph non-embeddability result is a closed kernel theorem proved by decide. The formalization surfaced findings about the published pipeline, including a load-bearing injectivity side condition in its embeddability notion, hidden WLOG case obligations invisible to Z3-based workflows, and an unreproducible candidate count that we resolve against the published artifacts. All certificates, checkers, and proofs are available and replayable from a single build.
Preprint and verification scripts for the paper "Grouped endpoint forms and a sharp interior law for the Werner distillability inequalities". For an operator C on H_1 ⊗ … ⊗ H_k, let q_k(−t,C) = Σ_J (−t)^{|J|} ‖Tr_J C‖², over sets J of factors. For rank at most two, nonnegativity of q_k(−1/2,C) is the Schmidt-rank-two test of undistillability of a Werner state from k copies; for k ≥ 4 it is open. We prove an identity, for every k, t and operator, writing q_k(−t,C) − (1−2t)(1−t)^{k−1} ‖C‖² as a combination, with explicit weights nonnegative on [0,1/2], of the endpoint forms of all factor groupings; if a factor has dimension at least two, no other constant admits such an identity with C-independent weights. So, over all local dimensions and for rank at most two, the sharp bound q_k(−t,C) ≥ (1−2t)(1−t)^{k−1} ‖C‖² on [0,1/2] is equivalent to the endpoint inequality; for k = 4, given the computer-assisted three-copy endpoint, it holds for a fixed C if and only if q_4(−1/2,C) ≥ 0. The sharp bound holds for every k when the range or co-range of C contains a fully product vector, and the identity extends to every real m ≥ 1, with 1/m replacing 1/2. For operators with range and co-range in the strings with exactly one digit 2, we give an identity and pair bounds for every k, prove the endpoint inequality for every k on symmetric sectors and, for k = 4, on further classes, one of them given the three-copy endpoint. The inequality also holds when the range lies in the span of the vectors |j⟩^{⊗k} (for even k at every rank), and for k = 4 at every rank when the range lies in a three-dimensional span of W-type vectors. Files: the preprint (PDF); ten Python scripts that check the explicit weights, identities and examples of Sections 3, 4 and 6–8 (exact rational arithmetic, except the checks of Section 7 and of the operator C_0 in Proposition 8.3, which use floating point); README.md, which says what each script checks and how to run it; MANIFEST.tsv with the size and SHA-256 of every file.
Whether negative-partial-transpose entanglement can remain undistillable is a longstanding problem in quantum information theory. We analyze the first unresolved three-copy Werner endpoint using a sharp dimension-free inequality for complementary partial traces and an optimal exterior-square inequality for orthonormal tripartite vectors. The latter identifies local SWAP- parity statistics with metric data of a decomposable Plucker bivector. Together these inequalities prove endpoint nonnegativity for every positive semidefinite rank-two coefficient operator and for the complete normal rank-two sector in arbitrary finite local dimensions. For genuinely nonnormal operators, an exact crossed-Gram criterion proves nonnegativity when one local outpu-input support overlap is at most two, including every system with a qubit-sized factor, and when either support plane contains a product ray. Explicit anti-state and rank-boundary families establish optimality of the constants and the rank restriction.