Claimed solutionOpenContext 5.13
A proposed analytic proof of the sharp bound cₙ = 1/2 for every number of qubits
by Yujie Zhang · accepted 20 Nov 2026 · partially AI-assisted
Main result
Proposed theorem. Write . Let be the largest constant such that every -qubit density operator of purity at most is a convex mixture of stabilizer states. Then
The centered Hilbert–Schmidt inradius is . The proof also determines every equality witness. If a trace-one Hermitian is nonnegative on all pure stabilizer states and , then is Clifford-equivalent to
Which steps use a computer? The argument below is analytic throughout. This is a fully AI-generated proposed proof for independent review. Its external spectral input is Rafael Wagner’s upper eigenvalue bound in OpenContext 5.11, §2.1. All other estimates below follow from the displayed Pauli identities and inequalities. The earlier computer-assisted estimates in 5.10 and the grid checks in 5.12 are unnecessary for this proof. OpenContext 5.10 · OpenContext 5.11 · OpenContext 5.12
Proof strategy. Suppose an -qubit witness violates the bound. We show that some Pauli measurement outcome leaves a violating witness on qubits. To prove this, we classify the witnesses exactly on the boundary for which every outcome obeys the lower-dimensional bound. Wagner’s theorem forces large Pauli expectations to carry more than of a particular probability measure. Their commutation relations leave at most two logical qubits. The two remaining cases can then be settled explicitly.
1. The boundary statement and induction
Write for the normalized witnesses above. For a nonidentity Hermitian Pauli , let , and . The outcome has a stabilizer basis, so . If , identifying the code with qubits gives . If , extend each logical stabilizer vector to a stabilizer basis. Every expectation in that basis vanishes, since they are nonnegative and sum to zero. Logical stabilizer projectors span the Hermitian operators, hence .
Boundary lemma. If , , and every signed Pauli outcome satisfies , then is Clifford-equivalent to .
Here is why the lemma suffices. Interpolate a hypothetical witness with toward until its squared norm equals . Every stabilizer expectation of the interpolated witness is strictly positive. Each Clifford image of has a zero stabilizer expectation. The lemma therefore forces some outcome to violate . The same outcome violates the inequality on the original witness, since the inequality defines a convex cone containing the compression of . This gives the descent. On one qubit each Bloch coordinate of a witness lies in , giving squared norm at most , with equality precisely at the cube vertices. Induction excludes every violation. The lemma then gives the asserted equality classification at every .
2. The weight carried by large Pauli expectations
Assume the hypotheses of the boundary lemma. Put and . Wagner proves . For , summing gives . For , the remaining eigenvalues have squared sum , hence
The middle expression increases for , and . OpenContext 5.11, §2.1
Label phase-free Paulis by , with when their Hermitian representatives anticommute. Choose and put
The two nonnegative Pauli outcome probabilities give . Direct expansion with , , gives
Thus , and at both sides vanish. Define . Pauli orthogonality gives and . Hence , so is a probability distribution. Taking the Fourier transform again gives
Let and . By Pauli orthogonality and the bound on ,
Consequently .
3. The anticommutation graph leaves at most two logical qubits
Join two labels in when they anticommute. Three induced graphs on four vertices are impossible. Up to Clifford conjugation their four observables are
Here , and identities on additional qubits are implicit. Each graph has an odd number of perfect matchings, so its alternating adjacency matrix is nonsingular over . Its four Pauli labels therefore form a symplectically nondegenerate basis of their span, which proves the asserted equivalence to the displayed realizations. Independence also lets Pauli conjugation prescribe all four signs.
For each row the sum of the four absolute expectations is at most . To check this, choose signs giving positive expectations. For commuting signed Paulis the operator is four times a stabilizer code projector. For the path take the six triples below, two per line.
For the two disjoint edges, replace the first triple by . For the triangle with a pendant vertex use
In each case the six operators sum to minus twice the sum of the four observables. Stabilizer positivity proves the claimed upper bound, whereas four elements of have total absolute expectation greater than .
All edges now lie in one connected component, since two edges in different components would be disjoint and induced. The absence of an induced four-vertex path bounds the component’s diameter by two. If are adjacent and is adjacent to neither, take a common neighbor of in this component. When meets there is a triangle with pendant . Otherwise is an induced path. Thus nonadjacency, including equality, partitions the component into independent parts, with all edges between distinct parts.
Put and . Every maximal commuting label space has . Indeed, and , so by the Fourier formula. The latter sum divided by is the collision probability of the stabilizer basis distribution for , hence at most one. Since , the graph has an edge.
Choose one representative from each of the parts. They pairwise anticommute. Labels in a part have identical neighbors in , so their differences lie in , as do all isolated labels. Thus . The representatives span . Their Gram matrix has ones off the diagonal and zeros on it, so its rank is , where or . In the odd case their sum lies in . Since a nondegenerate complement of in has dimension , .
A label orthogonal to anticommutes with at most entire cosets of , by the sum relation when is odd. A label outside anticommutes with half of each coset. Summing (1) over gives , hence . This rules out . Also would give . Therefore and . A joint eigenspace of , called a syndrome space below, encodes logical qubits.
For , complete the six or seven representatives to seven by adjoining their sum if needed. Their only relation modulo is that all seven sum to zero. Each logical label is the sum of at most three of them. Let be and the seven representative cosets, and set . There are labels in . A sum of one, two or three generators anticommutes with respectively six, two or four generators. Thus a physical label in anticommutes with at most labels in . A label outside anticommutes with at most , including exactly half of each coset when it lies outside . Sum from (1) over . Then , so , a contradiction. Only remain.
4. Bounding the blocks outside one syndrome
Fix signed generators of . Write for their joint syndrome projectors, and . The are nonnegative and sum to one. Fourier orthogonality on and averaging (1) over give
Let have the largest weight . Since and , we get (at the last upper bound is ). Label by syndrome zero. In blocks on write , and .
For each nonzero syndrome parity , apply the outcome hypothesis to . Among all parities a fixed nonzero is selected half the time, and two distinct nonzero indices are both selected one quarter of the time. Summing the squared block norms on the left and the squared traces on the right gives
Let project onto either outcome of a nonidentity logical Pauli on , and put . Multiply that signed Pauli by a uniformly random syndrome parity. The resulting physical outcome selects on and selects a uniformly random one of the two complementary logical eigenspaces on each nonzero syndrome. Their total selected trace satisfies and . Keep the and blocks in the squared norm of the physical compression and average. Each nonzero-syndrome projector has mean , giving
The first term is at least . Hence
For a logical axis with a physical lift in , sign that lift to have expectation . Its opposite outcome has total trace . On its trace is at most , since each other syndrome contributes a nonnegative trace. The estimates also hold for a single syndrome, when the complementary blocks are empty.
5. The one-logical-qubit case
All three logical Pauli axes must occur in . Otherwise would lie in the union of two maximal commuting spaces, of total mass at most . For the opposite projectors on these signed axes write , and . The one-qubit Pauli expansion gives . The three projectors sum to an operator of least eigenvalue . Summing (2) thus gives , with . Substitution into yields
Both brackets are negative at and increase with . Hence . The complement and coherence bounds give , and all three signed logical coordinates are . These are precisely the Clifford images of .
6. The two-logical-qubit case
Complete the four or five anticommuting representatives to five axes , adjoining their sum when needed. The other ten nonzero logical labels are the pair sums . For each logical label choose a lift and let . Fourier orthogonality on the syndromes gives . For a set of logical labels write for its full preimage in . Summing the Fourier formula for over its cosets gives
Take to consist of zero and the five axes. Let sum the on the axes and sum them on the pairs. The character sums over are on zero, axes and pairs respectively. For they are . Thus
The second mass is nonnegative and the first contains , so .
Order decreasingly on , and set , . At least four axes have a lift in , so the opposite-outcome estimate in §4 gives on the top four. Every logical coefficient on any other syndrome has magnitude at most . Consequently
For each pair choose the index with smaller inside the pair and the two with smallest outside it, breaking ties to avoid index five. Let be the sum of the three chosen deficits. The chosen axes form a triangle and is adjacent only to the chosen index inside the pair. The four-observable inequality in §3 applied to gives
The chosen deficit inside the pair is bounded by half its sum, and the two outside by two thirds of their sum. Thus . Summing over pairs gives since every index is inside four pairs and outside six. Hence . Expanding in the fifteen nonidentity Paulis gives
For a lower bound take four signed axes with lifts in . Their signed sum squares to , so their opposite projectors sum to at least . Sum (2) with and use . This gives
To see the contradiction using exact numbers, weaken to , so . Let and . By (3), . The derivative of has the sign of , hence this ratio decreases. On we have and . The correction in (4) above is at most
On , , making the correction nonpositive, whereas . The lower bounds on follow by concavity and substitution at the two endpoints of each interval. Both cases contradict (5). This excludes two logical qubits and completes the boundary lemma.
7. Sharpness and the purity threshold
Postselecting a pure stabilizer state onto the defining code of gives zero or a multiple of a logical stabilizer state. Thus , and direct multiplication gives . By polarity the stabilizer polytope is the intersection of the half-spaces for . The squared distance from to such a hyperplane is . The minimum is , so adding gives the threshold .
The state reaches the hyperplane for and is positive definite. Moving slightly farther on the ray from through therefore gives physical magic states approaching the threshold from above. This establishes sharpness for every .
A trace-distance consequence. With , the centered inradius is . Both spectral bounds in 5.11 give . A traceless perturbation at trace distance from zero has positive and negative parts of trace , so it can lower by at most . Equality is attained along the direction , where are extremal eigenprojectors of . The touching state is positive definite. OpenContext 5.11
Attribution. Liu, Haug, Ye, Liu and Roth posed the purity conjecture and proved the upper bound on its threshold. Zurel and Davis related it to maximal witness norm. Pauli compression as an inradius recursion was developed by Liu and Liu and explored here in Yixuan Wang’s notes. Wagner’s §2.1 supplies the spectral input to §2 above. The additional step in this proof is to retain the commutation relations between large Pauli coefficients and bound each remaining logical block. Triangle Criterion, Conjecture 1 · Zurel and Davis · Liu and Liu, Appendix A · Wang, OpenContext 5.6 · Wagner, OpenContext 5.11
Permanent link:open-context.jan-bulling.com/contribution/5.13
Discussion
Sign in to join the discussion.
Rafael Wagner · 20 Nov 2026
I think that my Claude's solution is along similar lines