---
source: 'OpenContext: Open problems in generalized contextuality'
url: 'https://open-context.jan-bulling.com/problems/test-problem-hilbert-schmidt-inradius-of-the-multiqubit-stab'
number: 5
title: 'Test problem: Hilbert–Schmidt inradius of the multiqubit stabilizer polytope'
status: under-revision
statusSince: '2026-11-20'
published: true
posedBy:
  - Zhenhuan Liu
  - Tobias Haug
  - Qi Ye
  - Zi-Wen Liu
  - Ingo Roth
posed: '2026-11-17'
updated: '2026-11-20'
agenda:
  - classicality
  - framework
  - applications
tags:
  - Nonstabilizerness
  - Magic
  - Purity
  - Convex polytopes
  - Clifford group
  - Stabilizer states
  - Qubits
contributions: 17
comments: 19
license: 'CC BY 4.0, unless the entry says otherwise'
---

# Problem 5: Test problem: Hilbert–Schmidt inradius of the multiqubit stabilizer polytope

From [OpenContext: Open problems in generalized contextuality](https://open-context.jan-bulling.com/problems/test-problem-hilbert-schmidt-inradius-of-the-multiqubit-stab).

- **Status:** under revision since 20 Nov 2026
- **Posed by:** Zhenhuan Liu, Tobias Haug, Qi Ye, Zi-Wen Liu and Ingo Roth, 17 Nov 2026
- **Last updated:** 20 Nov 2026
- **Research agenda:** I. Reconceptualization of classicality; II. Towards a better framework; III. Useful applications
- **Tags:** Nonstabilizerness, Magic, Purity, Convex polytopes, Clifford group, Stabilizer states, Qubits

## Statement

Is every $n$-qubit state $\rho$ of purity $\mathrm{Tr}[\rho^2]$ at most $1/(2^n-1/2)$ a convex mixture of stabilizer states, i.e., not a magic state?

In other words, let $d=2^n$ and 

$$
\mathrm{STAB}_n = \mathrm{conv}\{ \vert s \rangle \langle s \vert \ : \ \vert s \rangle \text{ is an }n\text{-qubit stabililzer state}\}.
$$

Is the following implication valid for every $n\geq 1$?

$$
\mathrm{Tr}(\rho^2) \leq \frac{1}{d-1/2} \implies \rho \in \mathrm{STAB}_n.
$$

## Background
 
[Stabilizer states](https://en.wikipedia.org/wiki/Stabilizer_code), [Clifford gates](https://en.wikipedia.org/wiki/Clifford_gate) and Pauli measurements can be simulated efficiently on a classical computer ([Gottesman–Knill theorem](https://en.wikipedia.org/wiki/Gottesman%E2%80%93Knill_theorem)); states outside the stabilizer polytope $\mathrm{STAB}_n$ ("magic" states) are the resource that lifts this to universal quantum computation, for instance through [magic state distillation](https://en.wikipedia.org/wiki/Magic_state_distillation) [5]. A basic question is how noisy a state can be while still being magic. Since [purity](https://en.wikipedia.org/wiki/Purity_(quantum_mechanics)) is the simplest basis-independent measure of noise, this asks for the largest Hilbert–Schmidt ball centred at the maximally mixed state $\mathbb{I}/d$ that fits inside $\mathrm{STAB}_n$, i.e. the inradius of the polytope.
 
For one qubit the answer is elementary: the stabilizer octahedron contains the [Bloch ball](https://en.wikipedia.org/wiki/Bloch_sphere) of radius $1/\sqrt{3}$, which is purity $2/3 = 1/(d-1/2)$. The problem asks whether the same formula holds for every $n$.
 
What is known:
 
- The statement was posed as Conjecture 1 in [1], where it is verified for $n=1,2$ and numerically up to $n=4$. The authors prove the weaker threshold $1/(d-1/d)$, and they show that $1/(d-1/2)$ cannot be improved: there are magic states with purity arbitrarily close to it.
- In [2] the same inradius is obtained conditionally on an equivalent conjecture, and the analogous question is resolved unconditionally for odd prime local dimension, where the threshold is $1/(D-1/D)$ for total dimension $D$.
- The strongest proved result we are aware of is in [3]: purity at most $1/(d-a_*)$ with $a_* = 0.458327\ldots$ suffices for every $n$, which leaves a small gap to the conjectured constant $1/2$.
- By [polarity](https://en.wikipedia.org/wiki/Dual_cone_and_polar_cone), the statement is equivalent to a bound on the dual polytope $\Lambda_n = \{A : \mathrm{Tr}[A]=1,\ \langle s|A|s\rangle \ge 0 \text{ for all stabilizer states } s\}$ of [4], namely $\mathrm{Tr}[A^2]\le 2$ for all $A\in\Lambda_n$. The bound is attained by a one-qubit phase-point operator tensored with a stabilizer state. The two-qubit case follows from the enumeration of the facets of the two-qubit stabilizer polytope in [5].
So only one direction is open: that every state below the threshold is a stabilizer mixture, for $n\ge 3$.
 

## Relevance
 
The dual polytope $\Lambda_n$ is the state space of a [hidden-variable](https://en.wikipedia.org/wiki/Hidden-variable_theory) description of quantum computation with [magic](https://en.wikipedia.org/wiki/Magic_(quantum_information)) states [4]: its elements assign non-negative probabilities to every stabilizer measurement. The problem therefore asks how large such a "classical-looking" object can be, measured by $\mathrm{Tr}[A^2]$.
 
The link to [contextuality](https://en.wikipedia.org/wiki/Quantum_contextuality) is direct. Non-negativity on each stabilizer basis alone gives only the trivial bound $\mathrm{Tr}[A^2]\le d$, and that bound would be reached exactly by a deterministic, noncontextual $\pm1$ value assignment to all Pauli observables. Such an assignment is forbidden by [Mermin-square](https://en.wikipedia.org/wiki/Mermin-Peres_square) ([Kochen–Specker](https://en.wikipedia.org/wiki/Kochen%E2%80%93Specker_theorem)) arguments. The conjectured value $2$ is thus a quantitative statement of how strongly state-independent contextuality of the multiqubit Pauli group constrains these assignments.
 
This is also why qubits are the hard case. For odd prime dimension the stabilizer subtheory admits a noncontextual ([Wigner-function](https://en.wikipedia.org/wiki/Wigner_quasiprobability_distribution)) model and the corresponding inradius is already known [2]; for qubits the stabilizer subtheory is itself contextual, and no such model is available.
 
For the applications agenda, a proof would give a certified, basis-independent purity threshold: below it, no magic and hence no contextuality-based advantage can be present, whatever the state. The best threshold of this kind proved so far is that of [3]; the conjecture of [1] asserts the optimal one.
 
## Significance
 
A proof would:
 
- settle the exact Hilbert–Schmidt inradius of the multiqubit stabilizer polytope, and with it the minimal purity of a magic state, for all $n$, closing the constant gap left open by [3];
- turn the conditional results of [2] on absolutely stabilizer states into theorems;
- show that the largest-norm vertices of $\Lambda_n$ are of the simplest known (closed-noncontextual) type, which is structural information about a polytope whose vertices are unclassified for $n\ge 3$ (for $n=2$ they follow from the facet enumeration of [5]).
A counterexample would be equally informative: it would exhibit a new kind of vertex of $\Lambda_n$ and a magic state noisier than any currently known. By [3], such a state could only exist in a narrow window of purities, between $1/(d-a_*)$ with $a_*=0.458327\ldots$ and $1/(d-\frac{1}{2})$.
 
The scope is limited. The problem concerns one geometric constant of the stabilizer subtheory, and it bears on generalized noncontextuality only through the identification of magic with contextuality. It is best read as a sharp, well-posed test case rather than a step that would by itself advance the framework.

## Note

This problem can also be found at the [QIQCOP Zoo database](https://qiqc-op.com/problem/op_6d9a72b32070a900/). See Ref. [6].

## Research agenda

The long-term goals of the research agenda that this problem traces to.

### I. Reconceptualization of classicality

Determine whether generalized noncontextuality captures the appropriate boundary between classical and nonclassical explanation, the thesis being that it offers a broad, formal, and meaningful framework for what it means for a process in Nature to be “classical”.

### II. Towards a better framework

Develop a formalism that generalizes ontological models while cleanly separating causal (ontic) from inferential (epistemic) elements, in which a suitably adapted notion of noncontextuality can be formulated.

### III. Useful applications

Make the failure of classicality, as captured by generalized noncontextuality, a certified and quantified resource, witnessed in experiments and tied to the advantages that quantum devices offer.

## References

1. [Zhenhuan Liu, Tobias Haug, Qi Ye, Zi-Wen Liu, and Ingo Roth, Triangle Criterion: a mixed-state magic criterion with applications in distillation and detection, arXiv:2512.16777 (2025)](https://arxiv.org/pdf/2512.16777v2)
2. [Michael Zurel and Jack Davis, Basis-independent stabilizerness and maximally noisy magic states, arXiv:2602.22336 (2026)](https://arxiv.org/pdf/2602.22336v1)
3. [Zhenhuan Liu and Zi-Wen Liu, On the geometry, typicality, and detectability of quantum magic, arXiv:2609.03944 (2026)](https://arxiv.org/abs/2609.03944)
4. [Michael Zurel, Cihan Okay, and Robert Raussendorf, Hidden Variable Model for Universal Quantum Computation with Magic States on Qubits, Phys. Rev. Lett. 125, 260404 (2020)](https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.125.260404)
5. [Ben W. Reichardt, Quantum universality by state distillation, arXiv:quant-ph/0608085 (2006)](https://arxiv.org/abs/quant-ph/0608085)
6. [Hilbert–Schmidt inradius of the multiqubit stabilizer polytope, Quantum Information and Quantum Computation Open Problem Zoo (QIQCOP Zoo), (2026)](https://qiqc-op.com/problem/op_6d9a72b32070a900/)
7. [John Watrous, The Theory of Quantum Information, Cambridge University Press (2018)](https://jhwatrous.github.io/TQI.pdf)
8. [Cihan Okay, Michael Zurel, Robert Raussendorf, On the extremal points of the $\Lambda$-polytopes and classical simulation of quantum computation with magic states, 	Quantum Information and Computation Vol.21 No.13&14, 1533-7146 (2021)](https://doi.org/10.26421/QIC21.13-14-2)
9. [Michael Zurel, Cihan Okay, Robert Raussendorf, and Arne Heimendahl, Hidden variable model for quantum computation with magic states on qudits of any dimension, Quantum 8, 1323 (2024)](https://doi.org/10.22331/q-2024-04-30-1323)
10. [Michael Zurel, Cihan Okay, and Robert Raussendorf, Simulating Quantum Computation: How Many “Bits” for “It”?, PRX Quantum 5, 030343 (2024)](https://doi.org/10.1103/PRXQuantum.5.030343)
11. [Cihan Okay, Michael Zurel, Robert Raussendorf, ERRATUM: On the extremal points of the $\Lambda$-polytopes and classical simulation of quantum computation with magic states, 	Quantum Information and Computation Vol.22 No.7&8, 627-628 (2022)](https://www.rintonpress.com/xxqic22/qic-22-78/0627-0628.pdf)
12. [C Dankert, R Cleve, J Emerson, E Livine, Exact and approximate unitary 2-designs and their application to fidelity estimation, Phys. Rev. A 80, 012304 (2009)](https://doi.org/10.1103/PhysRevA.80.012304)
13. [J. Ernest Wilkins, “A Note on Skewness and Kurtosis,” The Annals of Mathematical Statistics 15(3), 333–335 (1944)](https://projecteuclid.org/journals/annals-of-mathematical-statistics/volume-15/issue-3/A-Note-on-Skewness-and-Kurtosis/10.1214/aoms/1177731243.full)
14. [Robert C. Thompson and Linda J. Freede, Eigenvalues of partitioned hermitian matrices, Bull. Austral. Math. Soc. 3 (1970)](https://www.cambridge.org/core/journals/bulletin-of-the-australian-mathematical-society/article/eigenvalues-of-partitioned-hermitian-matrices/E252E270CD3258755D98F19849ABB622)

## Progress and history

- 17 Nov 2026, Posed: Proposed by the community and accepted by the moderators.
- 17 Nov 2026, Posed: Improved the problem's statement
- 17 Nov 2026, Posed: Numerical [evidence](https://open-context.jan-bulling.com/contribution/5.3) for the cases $n=1$, $2$, and $3$
- 17 Nov 2026, Edit: Tested if edits are now credited ([version 6](https://open-context.jan-bulling.com/contribution/5v6), by Jan, Albert)
- 17 Nov 2026, Posed: Verification
- 17 Nov 2026, Note: Explanatory exposition: On the geometry of the inner ball ([OpenContext 5.1](https://open-context.jan-bulling.com/contribution/5.1), by Rafael Wagner)
- 17 Nov 2026, Advance: Partial result: The solution for N=1 ([OpenContext 5.2](https://open-context.jan-bulling.com/contribution/5.2), by Rudi B. P. Pietsch)
- 17 Nov 2026, Advance: Computational evidence: Numerical test of the stabilizer-polytope inradius conjecture for small values of $n$ ([OpenContext 5.3](https://open-context.jan-bulling.com/contribution/5.3), by Rafael Wagner)
- 17 Nov 2026, Advance: Independent verification: Test ([OpenContext 5.4](https://open-context.jan-bulling.com/contribution/5.4), by Jan)
- 18 Nov 2026, Edit: Improved the problem's background ([version 7](https://open-context.jan-bulling.com/contribution/5v7), by the editors)
- 18 Nov 2026, Posed: Improved the references
- 18 Nov 2026, Posed: Partial result claim
- 18 Nov 2026, Posed: Improved the references
- 18 Nov 2026, Edit: I've suggested improvements on the background of the problem ([version 8](https://open-context.jan-bulling.com/contribution/5v8), by Rafael Wagner)
- 18 Nov 2026, Edit: I think that the Relevance and Significance could be improved ([version 9](https://open-context.jan-bulling.com/contribution/5v9), by Rafael Wagner)
- 18 Nov 2026, Posed: Partial result claim
- 18 Nov 2026, Note: Challenge to a solution: I want to challenge your solution ([OpenContext 5.8](https://open-context.jan-bulling.com/contribution/5.8), by Jan)
- 18 Nov 2026, Advance: Partial result: A weaker bound on the stabilizer-inradius conjecture ([OpenContext 5.5](https://open-context.jan-bulling.com/contribution/5.5), by Rafael Wagner)
- 18 Nov 2026, Advance: Partial result: Partial Results on the Hilbert–Schmidt Inradius of the Multiqubit Stabilizer Polytope ([OpenContext 5.6](https://open-context.jan-bulling.com/contribution/5.6), by Yixuan Wang)
- 18 Nov 2026, Advance: Partial result: An improved bound relative to Liu and Liu's result: from approximately 0.4583 to 0.4589. ([OpenContext 5.7](https://open-context.jan-bulling.com/contribution/5.7), by Rafael Wagner)
- 19 Nov 2026, Advance: Partial result: An improved bound at 0.4962, and an exact bound for n=3 at 1/2
- 19 Nov 2026, Edit: Improvements on the references ([version 10](https://open-context.jan-bulling.com/contribution/5v10), by Rafael Wagner)
- 19 Nov 2026, Advance: Partial result: Proofs for OpenContext 5.9: the sharp three-qubit threshold and a uniform bound of 0.496222 ([OpenContext 5.10](https://open-context.jan-bulling.com/contribution/5.10), by Yujie Zhang)
- 19 Nov 2026, Edit: Improvements on the references ([version 11](https://open-context.jan-bulling.com/contribution/5v11), by Rafael Wagner)
- 19 Nov 2026, Note: OpenContext contribution 5.9 has been retracted
- 19 Nov 2026, Advance: Partial result: An improved computer-assisted bound at 0.4996 for all $n$, and a sharp 0.5 value for $n\leq 5$ (part I) ([OpenContext 5.11](https://open-context.jan-bulling.com/contribution/5.11), by Rafael Wagner)
- 19 Nov 2026, Advance: Partial result: An improved computer-assisted bound at 0.4996 for all $n$, and a sharp 0.5 value for $n\leq 5$ (part II) ([OpenContext 5.12](https://open-context.jan-bulling.com/contribution/5.12), by Rafael Wagner)
- 20 Nov 2026, Advance: Claimed solution: A proposed analytic proof of the sharp bound cₙ = 1/2 for every number of qubits ([OpenContext 5.13](https://open-context.jan-bulling.com/contribution/5.13), by Yujie Zhang)
- 20 Nov 2026, Status: under revision: Status changed to under revision.
- 20 Nov 2026, Edit: Improvements on the references ([version 12](https://open-context.jan-bulling.com/contribution/5v12), by Rafael Wagner)
- 20 Nov 2026, Advance: Independent verification: A solution to the problem which uses Aronszajn's inequality ([OpenContext 5.14](https://open-context.jan-bulling.com/contribution/5.14), by Rafael Wagner)

## Contributors

Everyone who posed or proposed the problem, edited its text, or added a result or reference, in the order they first contributed.

- **Ingo Roth:** Posed the problem
- **Qi Ye:** Posed the problem
- **Tobias Haug:** Posed the problem
- **Zhenhuan Liu:** Posed the problem
- **Zi-Wen Liu:** Posed the problem
- **Rafael Wagner:** Proposed it for the site; Edited the text: version 2 (5v2), Erased an earlier comment from the question posed, and added context to the question.; Explanatory exposition: On the geometry of the inner ball (5.1); Edited the text: version 5 (5v5), Included background and other things; Computational evidence: Numerical test of the stabilizer-polytope inradius conjecture for small values of $n$ (5.3); Partial result: A weaker bound on the stabilizer-inradius conjecture (5.5); Suggested a reference: Michael Zurel, Cihan Okay, and Robert Raussendorf, Hidden Variable Model for Universal Quantum Computation with Magic States on Qubits, Phys. Rev. Lett. 125, 260404 (2020); Suggested a reference: Ben W. Reichardt, Quantum universality by state distillation, arXiv preprint: quant-ph/0608085 (2006); Suggested a reference: Zhenhuan Liu and Zi-Wen Liu, On the geometry, typicality, and detectability of quantum magic, arXiv preprint:2609.03944 (2026); Edited the text: version 8 (5v8), I've suggested improvements on the background of the problem; Edited the text: version 9 (5v9), I think that the Relevance and Significance could be improved; Partial result: An improved bound relative to Liu and Liu's result: from approximately 0.4583 to 0.4589. (5.7); Edited the text: version 10 (5v10), Improvements on the references; Edited the text: version 11 (5v11), Improvements on the references; Partial result: An improved computer-assisted bound at 0.4996 for all $n$, and a sharp 0.5 value for $n\leq 5$ (part I) (5.11); Partial result: An improved computer-assisted bound at 0.4996 for all $n$, and a sharp 0.5 value for $n\leq 5$ (part II) (5.12); Edited the text: version 12 (5v12), Improvements on the references; Independent verification: A solution to the problem which uses Aronszajn's inequality (5.14)
- **Rudi B. P. Pietsch:** Suggested a reference: On the geometry, typicality, and detectability of quantum magic; Partial result: The solution for N=1 (5.2)
- **Yìlè Yīng:** Edited the text: version 3 (5v3), add the term magic state to the statement
- **Jan:** Independent verification: Test (5.4); Edited the text: version 6 (5v6), Tested if edits are now credited; Challenge to a solution: I want to challenge your solution (5.8)
- **Albert:** Edited the text: version 6 (5v6), Tested if edits are now credited
- **Yixuan Wang:** Partial result: Partial Results on the Hilbert–Schmidt Inradius of the Multiqubit Stabilizer Polytope (5.6)
- **Yujie Zhang:** Partial result: Proofs for OpenContext 5.9: the sharp three-qubit threshold and a uniform bound of 0.496222 (5.10); Claimed solution: A proposed analytic proof of the sharp bound cₙ = 1/2 for every number of qubits (5.13)

## Contributions

Material submitted by researchers and approved by the moderators, oldest first.

### 1. Relevant reference: On the geometry, typicality, and detectability of quantum magic

- **Date:** 17 Nov 2026
- **Credit:** Rudi B. P. Pietsch
- **AI use:** no AI assistance
- **Link:** https://arxiv.org/abs/2609.03944

This seems to be a relevant recent preprint that might include advancements towards the problem.

### 2. Explanatory exposition: On the geometry of the inner ball

- **Date:** 17 Nov 2026
- **Credit:** Rafael Wagner
- **AI use:** no AI assistance
- **Permanent link:** https://open-context.jan-bulling.com/contribution/5.1
- **Votes:** 1 (1 up, 0 down)

## Background 

If we consider Ref. [2], we can note that the set of all states which lie in the inner ball - i.e., the one inside the octahedron - are of a special kind of nonstabilizer states. The way in which they are special is as follows: take any arbitrary number of states $\rho_1,\ldots, \rho_n$. Regardless of how one rotates these states, if they lie inside the inner ball, they will certainly remain inside the set of stabilizer states of the qubit. 

The question is then what is the value of the purity for arbitrary number of qubits.

## Important considerations

The most crucial consideration one needs to take here is the fact that for arbitrary dimensions, the one-to-one correspondance between the Bloch ball and quantum states is lost. Specifically, while in two dimensions every point in the ball corresponds to one and only one state, and vice versa, the same does not holds for arbitrary dimensions. In arbitrary dimensions, the set of states is contained in a real ball of radius 1, but it is simple to find points in the ball which corresponds to no quantum state.

#### Comment by Rafael Wagner, 18 Nov 2026

I cannot revise my own contribution. I wonder whether we should warn people of that, or allow for small edits to be made.

### 3. Partial result: The solution for N=1

- **Date:** 17 Nov 2026
- **Credit:** Rudi B. P. Pietsch
- **AI use:** partially AI-assisted
- **Permanent link:** https://open-context.jan-bulling.com/contribution/5.2
- **Votes:** -1 (1 up, 2 down)

Let us just consider the case of $n=1$.

In that case, we have  $d=2^1=2$.

The state can be expressed as $\varrho = \frac{\mathbb{1} + r \sigma}{2}$. An octahedron with $|r| \leq 1$ defines the stabilizer polytope, and the purity is given by $\mathrm{Tr} (\varrho^2) = \frac{1 + |r|^2}{2}$.

For the implication $\mathrm{Tr}(\varrho^2) \leq \frac{1}{d-1/2} = \frac{2}{3}$ to hold, we need $|r|^2 \leq \frac{1}{3}$, which holds as $|r| \leq \sqrt{3} |r| \leq 1$.

#### Comment by Rafael Wagner, 18 Nov 2026

I think it might be possible to have advance these results to the cases (at least numerically) for up to n=5 using neat mathematical tricks.

#### Comment by Rafael Wagner, 18 Nov 2026

I dont understand the downvotes! this is clearly a good (albeit trivial) advance

### 4. Computational evidence: Numerical test of the stabilizer-polytope inradius conjecture for small values of $n$

- **Date:** 17 Nov 2026
- **Credit:** Rafael Wagner
- **AI use:** fully AI-generated
- **Permanent link:** https://open-context.jan-bulling.com/contribution/5.3
- **Votes:** 2 (2 up, 0 down)

## Numerical test of the stabilizer-polytope inradius conjecture, n = 1, 2, 3

**Claim tested.** For $d = 2^n$: $\mathrm{Tr}(\rho^2) \le 1/(d-\tfrac12) \implies \rho \in \mathrm{STAB}_n$.

**Outcome.** No violation was found for $n = 1, 2, 3$. In every case the search reaches the conjectured bound exactly and never exceeds it. This is numerical evidence, not a proof: the search is a multi-start local optimisation and cannot certify a global maximum for $n = 3$.

*AI-generated (Claude, Fable 5.2); not independently checked.*

## Method

A state is represented by its $d^2-1$ Pauli expectation values $b_P = \mathrm{Tr}(\rho P)$, so that $\mathrm{Tr}(\rho^2) = (1+|b|^2)/d$ and $\mathrm{STAB}_n$ is the convex hull of the stabilizer-state vectors (6, 60 and 1080 of them for $n=1,2,3$).

**Experiment 1 (dual search).** A witness $W = I - \sum_P w_P P$ is non-negative on all stabilizer states iff $Vw \le 1$, where the rows of $V$ are the stabilizer vectors. Then $A = W/d$ lies in the dual polytope $\Lambda_n$ and $\mathrm{Tr}(A^2) = (1+|w|^2)/d$. The conjecture is equivalent to $\max_{A\in\Lambda_n} \mathrm{Tr}(A^2) = 2$. Since this maximises a convex function over a polytope, I iterate linear programs (each step maximises $\langle w_k, w\rangle$) from random starting directions until a vertex is a fixed point, and record $\mathrm{Tr}(A^2)$ there.

**Experiment 2 (primal check).** For a unit direction $u$, a linear program gives the largest $t$ with $t\,u \in \mathrm{STAB}_n$, hence the purity at which the ray from $I/d$ leaves the polytope. This is done for 300 random directions per $n$ and for the worst direction found in Experiment 1. Along that worst ray I also compute the robustness of magic (minimal $\sum_s |q_s|$ over affine stabilizer decompositions; it equals 1 exactly on stabilizer mixtures) at 41 points, up to the purity where $\rho$ stops being positive semidefinite.

## Results

| | n = 1 | n = 2 | n = 3 |
|---|---|---|---|
| Stabilizer states | 6 | 60 | 1080 |
| Search starts | 200 | 300 | 300 |
| Largest $\mathrm{Tr}(A^2)$ found | 2.000000 | 2.000000 | 2.000000 |
| Starts that reach 2 | 200 (100%) | 241 (80%) | 160 (53%) |
| Conjectured threshold $1/(d-\tfrac12)$ | 0.666667 | 0.285714 | 0.133333 |
| Exit purity along worst ray found | 0.666667 | 0.285714 | 0.133333 |
| Lowest exit purity, 300 random rays | 0.666785 | 0.297227 | 0.147028 |
| Median exit purity, random rays | 0.7186 | 0.3321 | 0.1641 |
| Non-zero Pauli coefficients of best witness | 3 | 7 | 15 |



- The best witness for each $n$ has $2d-1$ Pauli coefficients equal to $\pm1$ and the rest zero. This is the known extremal type (a one-qubit phase-point operator tensored with a stabilizer state, up to Clifford).
- Other local maxima are strictly smaller. For $n=2$ they sit at $\mathrm{Tr}(A^2) = 4/3,\ 11/8,\ 3/2$; for $n=3$ they spread between about 1.09 and 1.5.
- Random directions leave the polytope at higher purity than the threshold, and the gap widens with $n$; the tight directions are rare and structured.

Along the worst ray, robustness stays at 1 up to the threshold and rises immediately after it. The grid points bracketing the transition are 0.6653/0.6800 ($n=1$), 0.2843/0.2871 ($n=2$) and 0.1330/0.1337 ($n=3$); the exact exit values from the linear program are in the table. The ray stays a valid quantum state up to purity 1.0, 0.338 and 0.144 respectively, so genuine magic states exist just above the threshold in each case.

## Limitations

- For $n=3$ the dual polytope's vertices are not enumerated; 300 starts sample its local maxima but do not exclude a rarer, larger one.
- Floating-point linear programming (HiGHS); values agree with the exact ones to about $10^{-12}$ but are not exact arithmetic.
- One random seed. The run takes a few minutes, almost all of it for $n=3$.

## Code

Requires Python 3 with `numpy`, `scipy` and `matplotlib`. Run `python3 stab_inradius_sim.py results.json`, then `python3 plot_results.py results.json`.

### `stab_inradius_sim.py`

```python
"""Numerical test of: Tr(rho^2) <= 1/(d - 1/2)  =>  rho in STAB_n   (d = 2^n), for n = 1, 2, 3.

Coordinates: a state is its vector of d^2-1 Pauli expectation values b_P = Tr(rho P);
purity = (1 + |b|^2)/d.  STAB_n = conv(rows of V), V = Pauli vectors of all stabilizer states.
"""
import itertools, json, sys
import numpy as np
from scipy.optimize import linprog

I2 = np.eye(2); X = np.array([[0, 1], [1, 0]]); Y = np.array([[0, -1j], [1j, 0]]); Z = np.diag([1.0, -1.0])
H = np.array([[1, 1], [1, -1]]) / np.sqrt(2); S = np.diag([1, 1j])
CNOT = np.array([[1, 0, 0, 0], [0, 1, 0, 0], [0, 0, 0, 1], [0, 0, 1, 0]])

def kron(*ops):
    out = np.eye(1)
    for o in ops:
        out = np.kron(out, o)
    return out

def paulis(n):
    return np.array([kron(*p) for p in itertools.product([I2, X, Y, Z], repeat=n)][1:])

def stabilizer_vectors(n, P):
    """All n-qubit stabilizer states as +-1/0 Pauli vectors (Clifford orbit of |0...0>)."""
    gens = [kron(*[G if j == i else I2 for j in range(n)]) for i in range(n) for G in (H, S)]
    gens += [kron(*([I2] * i + [CNOT] + [I2] * (n - i - 2))) for i in range(n - 1)]
    vec = lambda psi: tuple(np.rint(np.einsum("i,pij,j->p", psi.conj(), P, psi).real).astype(int))
    psi = np.zeros(2 ** n, complex); psi[0] = 1
    seen, frontier = {vec(psi)}, [psi]
    while frontier:
        new = []
        for p in frontier:
            for G in gens:
                q = G @ p; v = vec(q)
                if v not in seen:
                    seen.add(v); new.append(q)
        frontier = new
    return np.array(sorted(seen), float)

# ---- Experiment 1 (dual): largest witness -------------------------------------------------
# W = I - sum_P w_P P is non-negative on all stabilizer states  <=>  V w <= 1.
# A = W/d lies in the dual polytope Lambda_n and Tr(A^2) = (1 + |w|^2)/d.
# Conjecture  <=>  max |w|^2 = 2d - 1  <=>  max Tr(A^2) = 2.
def local_max(V, rng, iters=80):
    c = rng.normal(size=V.shape[1])
    for _ in range(iters):                       # Frank-Wolfe for a convex maximum: iterate LPs
        w = linprog(-c, A_ub=V, b_ub=np.ones(len(V)), bounds=(None, None), method="highs").x
        if np.allclose(w, c, atol=1e-9):
            break
        c = w
    return w

# ---- Experiment 2 (primal): where does a ray leave STAB_n? --------------------------------
def exit_radius(V, u):
    """Largest t with t*u in conv(V)."""
    N, D = V.shape
    A_eq = np.vstack([np.hstack([V.T, -u[:, None]]), np.r_[np.ones(N), 0][None]])
    r = linprog(np.r_[np.zeros(N), -1], A_eq=A_eq, b_eq=np.r_[np.zeros(D), 1],
                bounds=[(0, None)] * N + [(0, None)], method="highs")
    return r.x[-1]

def robustness(V, b):
    """min sum|q_s| over affine decompositions b = sum_s q_s v_s, sum q = 1 (equals 1 iff b in STAB)."""
    N, D = V.shape
    A_eq = np.vstack([np.hstack([V.T, -V.T]), np.r_[np.ones(N), -np.ones(N)][None]])
    r = linprog(np.ones(2 * N), A_eq=A_eq, b_eq=np.r_[b, 1], bounds=(0, None), method="highs")
    return r.fun

def run(n, starts, n_random, seed=0):
    d = 2 ** n; rng = np.random.default_rng(seed)
    P = paulis(n); V = stabilizer_vectors(n, P)
    ws = [local_max(V, rng) for _ in range(starts)]
    trA2 = np.array([(1 + w @ w) / d for w in ws])
    best = ws[int(np.argmax(trA2))]
    u = best / np.linalg.norm(best)
    # random directions
    R = rng.normal(size=(n_random, d * d - 1)); R /= np.linalg.norm(R, axis=1)[:, None]
    pur_rand = np.array([(1 + exit_radius(V, r) ** 2) / d for r in R])
    t_best = exit_radius(V, u)
    # largest t for which rho(t) = (I + t sum u_P P)/d is still a physical state
    M = np.einsum("p,pij->ij", u, P)
    t_phys = -1 / np.linalg.eigvalsh(M).min()
    ts = np.linspace(0, t_phys, 41)
    rob = np.array([robustness(V, t * u) for t in ts])
    return dict(n=n, d=d, n_stab=len(V), starts=starts, trA2=trA2.tolist(),
                threshold=1 / (d - 0.5), best_trA2=float(trA2.max()),
                frac_best=float(np.mean(trA2 > 2 - 1e-6)),
                best_exit_purity=float((1 + t_best ** 2) / d),
                rand_exit_purity=pur_rand.tolist(),
                ray_purity=((1 + ts ** 2) / d).tolist(), ray_robustness=rob.tolist(),
                ray_max_physical_purity=float((1 + t_phys ** 2) / d),
                best_w=np.round(best, 6).tolist())

if __name__ == "__main__":
    cfg = {1: (200, 300), 2: (300, 300), 3: (300, 300)}
    out = []
    for n, (starts, nr) in cfg.items():
        res = run(n, starts, nr); out.append(res)
        print(n, res["n_stab"], "max Tr A^2 =", res["best_trA2"], "frac at max", res["frac_best"],
              "| best exit purity", res["best_exit_purity"], "threshold", res["threshold"],
              "| min random exit purity", min(res["rand_exit_purity"]), flush=True)
    json.dump(out, open(sys.argv[1] if len(sys.argv) > 1 else "results.json", "w"))
```

### `plot_results.py`

```python
import json, sys
import numpy as np
import matplotlib; matplotlib.use("Agg")
import matplotlib.pyplot as plt

res = json.load(open(sys.argv[1]))
BLUE, ORANGE, INK, MUTED, GRID, SURF = "#2a78d6", "#eb6834", "#0b0b0b", "#52514e", "#e4e3df", "#fcfcfb"
plt.rcParams.update({"font.size": 10, "axes.edgecolor": GRID, "axes.labelcolor": MUTED, "text.color": INK,
                     "xtick.color": MUTED, "ytick.color": MUTED, "axes.spines.top": False,
                     "axes.spines.right": False, "figure.facecolor": SURF, "axes.facecolor": SURF,
                     "savefig.facecolor": SURF})

# Figure 1: local maxima of Tr(A^2) found by the search
fig, axs = plt.subplots(1, 3, figsize=(11, 3.4), sharey=False)
for ax, r in zip(axs, res):
    v = np.round(np.array(r["trA2"]), 6); bins = np.arange(0.975, 2.08, 0.05)
    ax.hist(v, bins=bins, color=BLUE, edgecolor=SURF, linewidth=1.5)
    ax.axvline(2, color=INK, lw=1, ls=(0, (4, 3)))
    ax.set_title(f"n = {r['n']}   ({r['starts']} starts, {r['n_stab']} stabilizer states)", fontsize=10, loc="left")
    ax.set_xlabel("Tr(A²) at local maximum"); ax.grid(axis="y", color=GRID, lw=0.6); ax.set_axisbelow(True)
    ax.set_xlim(0.95, 2.3); ax.set_xticks([1.0, 1.25, 1.5, 1.75, 2.0])
    ax.annotate("conjectured\nmax = 2", xy=(2.05, ax.get_ylim()[1] * 0.6), ha="left", va="center", fontsize=8.5, color=MUTED)
axs[0].set_ylabel("number of starts")
fig.suptitle("No search run exceeds Tr(A²) = 2 on the dual polytope Λₙ", x=0.01, ha="left", fontsize=12, fontweight="bold")
fig.tight_layout(rect=(0, 0, 1, 0.94)); fig.savefig("fig1_search_maxima.png", dpi=170); plt.close(fig)

# Figure 2: best-case ray, robustness of magic vs purity
fig, axs = plt.subplots(1, 3, figsize=(11, 3.4))
for ax, r in zip(axs, res):
    p, R = np.array(r["ray_purity"]), np.array(r["ray_robustness"]); th = r["threshold"]
    ax.plot(p, R, color=ORANGE, lw=2, marker="o", ms=3.5, mec=SURF, mew=0.6)
    ax.axvline(th, color=INK, lw=1, ls=(0, (4, 3)))
    ax.axhline(1, color=GRID, lw=1)
    d = r["d"]
    ax.annotate(f"1/(d−½) = {th:.4f}", xy=(th, R.max()), xytext=(6, 0), textcoords="offset points",
                ha="left", va="top", fontsize=8.5, color=MUTED)
    ax.set_title(f"n = {r['n']}   (d = {d})", fontsize=10, loc="left")
    ax.set_xlabel("purity Tr(ρ²) along the worst-case ray"); ax.grid(axis="y", color=GRID, lw=0.6); ax.set_axisbelow(True)
axs[0].set_ylabel("robustness of magic")
fig.suptitle("Along the worst direction found, magic appears exactly at purity 1/(d − ½)", x=0.01, ha="left", fontsize=12, fontweight="bold")
fig.text(0.01, 0.86, "Robustness = 1 means the state is a stabilizer mixture; dashed line is the conjectured threshold.", fontsize=9, color=MUTED)
fig.tight_layout(rect=(0, 0, 1, 0.86)); fig.savefig("fig2_best_case_ray.png", dpi=170); plt.close(fig)

for r in res:
    v = np.round(np.array(r["trA2"]), 4); u, c = np.unique(v, return_counts=True)
    rp = np.array(r["rand_exit_purity"]); R = np.array(r["ray_robustness"]); p = np.array(r["ray_purity"])
    print(r["n"], dict(zip(u.tolist(), c.tolist())), "rand min/med", rp.min(), np.median(rp),
          "ray phys max", r["ray_max_physical_purity"], "first magic purity", p[np.argmax(R > 1 + 1e-7)],
          "last stab", p[np.argmax(R > 1 + 1e-7) - 1], "Rmax", R.max(),
          "best w values", dict(zip(*[x.tolist() for x in np.unique(np.round(r["best_w"], 3), return_counts=True)])))
```

### 5. Independent verification: Test

- **Date:** 17 Nov 2026
- **Credit:** Jan
- **AI use:** partially AI-assisted
- **Permanent link:** https://open-context.jan-bulling.com/contribution/5.4
- **Votes:** -1 (0 up, 1 down)

Hello, this is a test

#### Comment by Jan, 17 Nov 2026

I think this is completely wrong

#### Comment by Rafael Wagner, 18 Nov 2026

Is this a reply to the verification part?

#### Reply by Jan, 18 Nov 2026

Yes! At least one can discuss the contribution.

### 6. Partial result: A weaker bound on the stabilizer-inradius conjecture

- **Date:** 18 Nov 2026
- **Credit:** Rafael Wagner
- **AI use:** partially AI-assisted
- **Permanent link:** https://open-context.jan-bulling.com/contribution/5.5
- **Votes:** 0 (0 up, 0 down)

## Result

Using Claude Fable 5.2 and a few prompts, it was possible to prove the following weaker statement: 

>*Proposition:* Let $\rho \in \mathcal{D}(\underbrace{\mathbb{C}^2 \otimes \cdots \otimes \mathbb{C}^2}_{n\text{ times}})$ be an 
>arbitrary $n$-qubit state, with $n\geq 2$. Denote, furthermore, $d_n = 2^n$ and the function $c_{(\cdot)}:\mathbb{N}\to\mathbb{Q}$ given by 
>$$
>c_n=\frac{2d_n+7}{7d_n+2}>\frac{2}{7}.
>$$ 
>Then, 
>$$
>\mathrm{Tr}[\rho^2] \leq \frac{1}{d_n-c_n} \implies \rho \in \mathrm{STAB}_n.
>$$ 
>In particular, for all $n \geq 2$ it holds that 
>$$
>\mathrm{Tr}[\rho^2] \leq \frac{1}{d_n-\frac{2}{7}} \implies \rho \in \mathrm{STAB}_n.
>$$ 

Note that this is only a partial result because the conjecture claims that the implication holds with $c_n=\frac{1}{2}$ for arbitrary $n$. 

## Background 

The key technical idea is to use the fact that one can investigate the easier properties associated with the [convex cones](https://en.wikipedia.org/wiki/Convex_cone) generated by the set of states and the stabilizer polytopes, as opposed to the sets themselves. In this case, one replaces rotations and the more complicated geometry of the inradius by properties of cone-preservation and related ideas.

**Isotropic Pauli subgroups**

*Definition (isotropic Pauli subgroup).* Let $\mathcal{P}_n$ be the set of $4^n$ Pauli strings, i.e. tensor products of [Pauli matrices](https://de.wikipedia.org/wiki/Pauli-Matrizen) $\mathbb{I}, X, Y, Z$ on $n$ qubits without phases. A subset $T \subseteq \mathcal{P}_n$ is an isotropic Pauli subgroup if
- (1) it contains the identity, i.e., if $\mathbb{I} \in T$;
- (2) it is [Abelian](https://en.wikipedia.org/wiki/Abelian_group), i.e. any two elements of $T$ commute; 
- (3) it is closed under multiplication up to a sign, i.e., $\forall R,R' \in T$, one has that either $RR' \in T$ or $-RR' \in T$.

Such a subgroup has $\vert T \vert = 2^k$ elements for some $0 \leq k \leq n$. A sign pattern $\tau$ assigns an eigenvalue $\tau(R) = \pm 1$ to each $R \in T$ consistently with multiplication (so, for example, $\tau(\mathbb{I})=1$ and $\tau(R)\tau(R') = \pm \tau(R'')$ if $RR'= \pm R''$, with the same sign on both sides). For any fixed $T$ there are finitely many sign patterns, namely $2^k$ of them. The operator 
$$
\Pi_\tau = \frac{1}{\vert T\vert} \sum_{R \in T}\tau(R)R
$$
is the projector onto the corresponding joint eigenspace, which has dimension $2^{n-k}$.  The centraliser $C(T)$ is the set of Pauli strings commuting with every element of $T$.

**Stabilizer codes** 

These constructions are very common in [quantum error correction (QEC)](https://en.wikipedia.org/wiki/Quantum_error_correction). An important family of quantum codes is that of [stabilizer codes](https://en.wikipedia.org/wiki/Stabilizer_code). Each pair $(T,\tau)$ defines a stabilizer group, and $\Pi_\tau$ projects onto its stabilizer code.

*Definition (stabilizer code space).* Given an isotropic Pauli subgroup $T$ with $\vert T \vert = 2^k$ and a sign pattern $\tau$, the stabilizer code space is the joint eigenspace given by 
$$
\mathcal{C}_{T,\tau} = \left\{ \vert \psi \rangle \ : \ R\vert \psi \rangle =\tau(R)\vert \psi \rangle \text{ for all }R \in T \right\} = \mathrm{range}[\Pi_\tau].
$$

## Mathematical preliminaries

**Lemma:** Let $T$ be an isotropic Pauli subgroup with $\vert T\vert = 2^k$, and $\Pi_\tau$ the projector onto its joint eigenspace with sign pattern $\tau$. **If** $A$ is non-negative on all $n$-qubit stabilizer states, i.e., $A$ is a Hermitian operator on $n$ qubits with
$$
\langle s|A|s\rangle \ge 0
$$ 
for every $n$-qubit stabilizer state $|s\rangle$, **then** there is a [Clifford unitary](https://en.wikipedia.org/wiki/Clifford_gate) $U$ and an operator $B$ on $n-k$ qubits such that
 
- (a) $U\,\Pi_\tau A\,\Pi_\tau\,U^\dagger = |0^k\rangle\langle 0^k|\otimes B$;
- (b) $\langle t|B|t\rangle \ge 0$ for every $(n-k)$-qubit stabilizer state $|t\rangle$.

In other words, the block of $A$ on the code space is, up to a Clifford change of basis, an $(n-k)$-qubit operator $B$ that satisfies the same kind of hypothesis as $A$. We can also gain some intuition for why such a result might be useful. From (a), we have that 
$$
\mathrm{Tr}[B] = \mathrm{Tr}[\Pi_\tau A]
$$
and 
$$
\mathrm{Tr}[B^2] = \mathrm{Tr}[(\Pi_\tau A\Pi_\tau)^2].
$$ 
From (b), any inequality $\mathrm{Tr}[B^2] \le M\,(\mathrm{Tr}[B])^2$ known for $(n-k)$-qubit operators satisfying the positivity constraint holds for $B$. Together they imply that 
$$
\mathrm{Tr}[(\Pi_\tau A\Pi_\tau)^2] \le M\,\big(\mathrm{Tr}[\Pi_\tau A]\big)^2.
$$ 
This is the way the lemma enters the final theorem.

*Proof.* We want to show that there is a Clifford unitary $U$ and an operator $B$ on $n-k$ qubits such that 
$$
U \Pi_\tau A \Pi_\tau U^\dagger =  \vert 0^k \rangle \langle 0^k \vert \otimes B,
$$
where $\langle t \vert B \vert t \rangle \geq 0$ for every $(n-k)$-qubit stabilizer state $\vert t \rangle$. The argument has four steps. First, we move the code space to a standard position. Pick independent generators $R_1,\ldots, R_k$ of $T$ and set $$
g_i=\tau(R_i)R_i.$$
These are $k$ independent, commuting, Hermitian Pauli operators, and they allow us to rewrite $\Pi_\tau$ as 
$$
\Pi_\tau=\prod_{i=1}^k\tfrac12(\mathbb{I}+g_i).
$$ 
It is a standard fact about the [Clifford group](https://en.wikipedia.org/wiki/Clifford_group) that any such list can be mapped to $Z_1,\ldots,Z_k$: there is a Clifford unitary $U$ with 
$$
U g_i U^\dagger = Z_i
$$ 
for all $i$. (In more mathematical terms, Cliffords realise every [symplectic transformation](https://arxiv.org/pdf/1803.06987) of the Pauli labels together with every choice of signs, and any $k$ independent commuting Paulis can be completed to a symplectic basis.) Consequently 
$$
U\,\Pi_\tau\,U^\dagger=\prod_{i=1}^k\tfrac12(\mathbb{I}+Z_i)=|0^k\rangle\langle0^k|\otimes \mathbb{I}_{n-k}.
$$

Now, let $A'=UAU^\dagger$. For any $n$-qubit stabilizer state $|s\rangle$,
$$
\langle s|A'|s\rangle=\langle U^\dagger s|A|U^\dagger s\rangle\ge0.
$$ 
Here the assumption on the positivity of $A$ is used: Cliffords map stabilizer states to stabilizer states, so $U^\dagger|s\rangle$ is an $n$-qubit stabilizer state, and $A$ is non-negative on it by assumption. So $A'$ satisfies the same positivity requirements as $A$. Let $B=(\langle0^k|\otimes \mathbb{I}_{n-k})\,A'\,(|0^k\rangle\otimes \mathbb{I}_{n-k})$, an operator on the last $n-k$ qubits. Then, using our findings above,
$$
U\,\Pi_\tau A\,\Pi_\tau\,U^\dagger=(|0^k\rangle\langle0^k|\otimes \mathbb{I}_{n-k})\,A'\,(|0^k\rangle\langle0^k|\otimes \mathbb{I}_{n-k})=|0^k\rangle\langle0^k|\otimes B.
$$
To conclude, let $|t\rangle$ be a stabilizer state of $n-k$ qubits, with stabilizer generators $h_1,\dots,h_{n-k}$. Then $|0^k\rangle\otimes|t\rangle$ is stabilized by the $n$ independent commuting Paulis $Z_1,\dots,Z_k,\ \mathbb{I}\otimes h_1,\dots,\mathbb{I}\otimes h_{n-k}$, so it is an $n$-qubit stabilizer state. Hence, by the above (that is, by the positivity requirements in the assumption of our lemma applied to this particular $n$-qubit stabilizer state),
 
$$
\langle t|B|t\rangle=\big(\langle0^k|\otimes\langle t|\big)\,A'\,\big(|0^k\rangle\otimes|t\rangle\big)\ \ge\ 0 .
$$
 
This concludes the proof.

## Proof of the main result

Let us write some definitions in order to state the main result differently. From now on we write $d := d_n = 2^n$. Write $A = \tfrac1d\sum_P a_P P$ over the $n$-qubit Paulis, with $a_{\mathbb{I}} := \mathrm{Tr}[A]$. Let $\Lambda_n$ be the set of Hermitian operators $A$ with $\mathrm{Tr}[A] = 1$ and $\langle s|A|s\rangle \ge 0$ for every stabilizer state $s$. This is known as the $\Lambda$-polytope. Furthermore, define the number
$$
M_n = \max_{A\in\Lambda_n}\mathrm{Tr}[A^2].
$$
 
We know that $\rho \in \mathrm{STAB}_n$ iff $\mathrm{Tr}[A\rho]\ge 0$ for all $A\in\Lambda_n$ (this is standard convex duality; the polytope $\Lambda_n$ was introduced in the work of [Zurel, Okay, and Raussendorf](https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.125.260404)). If we write $\rho = \mathbb{I}/d + Y$ with $\mathrm{Tr}[Y]=0$, we can take any $A\in\Lambda_n$ and then, since $\mathrm{Tr}[A]=1$ and $\mathrm{Tr}[Y]=0$, we find 
$$
\mathrm{Tr}[A\rho]=\frac{1}{d}+\mathrm{Tr}\big[(A-\tfrac{\mathbb{I}}{d})\,Y\big]\;\ge\;\frac{1}{d}-\big\|A-\tfrac{\mathbb{I}}{d}\big\|_2\,\|Y\|_2 \;\ge\; \frac1d-\sqrt{M_n-\tfrac1d}\;\|Y\|_2.
$$ 
Above we have used the [Cauchy–Schwarz](https://en.wikipedia.org/wiki/Cauchy%E2%80%93Schwarz_inequality) inequality and $\|A-\mathbb{I}/d\|_2^2=\mathrm{Tr}[A^2]-\tfrac{1}{d}\le M_n-\tfrac{1}{d}$. The right-hand side is non-negative as soon as $\|Y\|_2^2\le 1/\big(d(dM_n-1)\big)$. Because $\mathrm{Tr}[\rho^2]=\tfrac1d+\|Y\|_2^2$, this condition reads $\mathrm{Tr}[\rho^2]\le M_n/(dM_n-1)=1/(d-1/M_n)$.  
$$
\mathrm{Tr}[A\rho] \ge \tfrac1d - \sqrt{M_n - \tfrac1d}\,\|Y\|_2.
$$ 
Hence 
$$\mathrm{Tr}[\rho^2] \le \frac{1}{d - 1/M_n} \;\Longrightarrow\; \rho\in\mathrm{STAB}_n, $$
and the conjecture is exactly that $M_n = 2$ for all $n$. With the Lemma, we can now prove the result through a re-statement in terms of $M_n$, namely the bound $M_n \leq 1/c_n$. 

**Theorem 2.** For all $n \ge 2$,
$$
M_n \;\le\; \frac{7d+2}{2d+7} \;<\; \frac72 .
$$

*Proof*. Take $A\in\Lambda_n$ and an isotropic $T$ with $\vert T \vert = 2^{n-2} = d/4$, so that its centraliser has $|C(T)| = 4d$. By the Lemma, each block $\Pi_\tau A\Pi_\tau$ corresponds to a two-qubit operator $B$ that is non-negative on all two-qubit stabilizer states, with trace $q_\tau = \mathrm{Tr}[\Pi_\tau A]$. 

The two-qubit case of the conjecture, $M_2 = 2$, is known from the exhaustive enumeration of the facets of the two-qubit stabilizer polytope (see [Reichardt](https://arxiv.org/abs/quant-ph/0608085)); it gives $\mathrm{Tr}[B^2] \le 2\,\mathrm{Tr}[B]^2$ (if $\mathrm{Tr}[B] = 0$ then $B = 0$, and otherwise $B/\mathrm{Tr}[B] \in \Lambda_2$). Hence $\mathrm{Tr}[(\Pi_\tau A\Pi_\tau)^2] \le 2q_\tau^2$. Summing over $\tau$:
 
- Left side: $\sum_\tau \mathrm{Tr}[(\Pi_\tau A\Pi_\tau)^2] = \mathrm{Tr}\big[\big(\tfrac{1}{|T|}\sum_{R\in T}RAR\big)^2\big] = \tfrac1d\sum_{Q\in C(T)}a_Q^2$, because averaging over $T$ keeps exactly the Paulis commuting with $T$.
- Right side: $q_\tau = \tfrac{1}{|T|}\sum_{R\in T}\tau(R)a_R$, so $\sum_\tau q_\tau^2 = \tfrac{1}{|T|}\sum_{R\in T}a_R^2$.

Therefore
$$
\sum_{Q\in C(T)} a_Q^2 \;\le\; 8\sum_{R\in T} a_R^2 .
$$
 
Now average this inequality over all such $T$. The symplectic group is transitive on non-identity Paulis, so each $Q \ne \mathbb{I}$ lies in $C(T)$ for a fraction $(4d-1)/(d^2-1)$ of the $T$'s and in $T$ for a fraction $(d/4-1)/(d^2-1)$. With $\Sigma = \sum_{Q\ne \mathbb{I}}a_Q^2$ and $a_{\mathbb{I}} = 1$:
$$
1 + \frac{4d-1}{d^2-1}\Sigma \;\le\; 8 + \frac{2d-8}{d^2-1}\Sigma \quad\Longrightarrow\quad \Sigma \le \frac{7(d^2-1)}{2d+7}.
$$

Finally $\mathrm{Tr}[A^2] = (1+\Sigma)/d \le (7d+2)/(2d+7)$. Since this gives $1/M_n \ge (2d+7)/(7d+2) = c_n > 2/7$, the implication above yields the Proposition.

## Other results in the literature

The most notable recent result which is better than the above is the result shown recently by [Liu and Liu](https://arxiv.org/pdf/2609.03944) in ther arXiv preprint:2609.03944v3 titled "On the geometry, typicality, and detectability of quantum magic".

Their theorem 1 shows that one can get much better than the bound $c_n$ above. Their value is $\approx 0.458$.

### 7. Relevant reference: Michael Zurel, Cihan Okay, and Robert Raussendorf, Hidden Variable Model for Universal Quantum Computation with Magic States on Qubits, Phys. Rev. Lett. 125, 260404 (2020)

- **Date:** 18 Nov 2026
- **Credit:** Rafael Wagner
- **AI use:** no AI assistance
- **Link:** https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.125.260404

It introduces a notion of Lambda polytope which is crucial for the whole topic.

### 8. Relevant reference: Zhenhuan Liu and Zi-Wen Liu, On the geometry, typicality, and detectability of quantum magic, arXiv preprint:2609.03944 (2026)

- **Date:** 18 Nov 2026
- **Credit:** Rafael Wagner
- **AI use:** no AI assistance
- **Link:** https://arxiv.org/pdf/2609.03944

They have, as far as I am aware, the largest known value for the bound which gets closer to the conjectured 1/2.

### 9. Relevant reference: Ben W. Reichardt, Quantum universality by state distillation, arXiv preprint: quant-ph/0608085 (2006)

- **Date:** 18 Nov 2026
- **Credit:** Rafael Wagner
- **AI use:** no AI assistance
- **Link:** https://arxiv.org/abs/quant-ph/0608085

Motivated by quantum error correction and distillation, the author investigated properties of the geometry of stabilizer states which might be useful.

### 10. Partial result: Partial Results on the Hilbert–Schmidt Inradius of the Multiqubit Stabilizer Polytope

- **Date:** 18 Nov 2026
- **Credit:** Yixuan Wang
- **AI use:** fully AI-generated
- **Permanent link:** https://open-context.jan-bulling.com/contribution/5.6
- **Votes:** 1 (1 up, 0 down)

**Scope:** Analytical statements and proofs under additional spectral or structural hypotheses. This note does not claim priority or originality for the derived implications; a comprehensive prior-art comparison has not been carried out.

## Abstract

Let $\mathrm{STAB}_n$ be the convex hull of pure $n$-qubit stabilizer states and $d=2^n$. The open inradius conjecture asserts that $\operatorname{Tr}(\rho^2)\le(d-1/2)^{-1}$ implies $\rho\in\mathrm{STAB}_n$. **This note gives partial results, not a proof of the unrestricted conjecture.** We prove the equivalent dual norm bound for stabilizer-positive trace-one operators with at most one negative eigenvalue, and classify equality within that class using the known closed-noncontextual (CNC) classification. We prove the primal bound for states block diagonal in stabilizer-code sectors of dimension at most four, and show sharpness. An explicit two-negative-eigenvalue example obstructs extending the cubic-moment lemma without further hypotheses. We also constrain the spectrum of any minimal-dimensional counterexample.

## 1. Problem, notation, and dual formulation

Write
$$
\mathrm{STAB}_n=\operatorname{conv}\{\,|s\rangle\langle s|:\ |s\rangle\text{ is a pure }n\text{-qubit stabilizer state}\,\},
\quad d=2^n.
$$
The conjecture posed in [P] is
$$
\boxed{
\operatorname{Tr}(\rho^2)\le\frac{1}{d-\frac12}
\quad\Longrightarrow\quad \rho\in\mathrm{STAB}_n
\qquad\text{for all }n\ge1.
}\tag{C}
$$
It is established for $n=1,2$ and remains open in general. The best unconditional constant reported in [L2] is $a_*\simeq0.458327$, in place of the conjectured $1/2$.

Define the stabilizer-positive trace-one slice
$$
\Lambda_n=\left\{
A=A^\dagger:\ \operatorname{Tr}A=1,\quad
\langle s|A|s\rangle\ge0\ \forall\text{ pure stabilizer states }|s\rangle
\right\}.
$$
**An element of $\Lambda_n$ need not be positive semidefinite.** The conjecture is equivalent to
$$
\boxed{\quad A\in\Lambda_n\quad\Longrightarrow\quad
\operatorname{Tr}(A^2)\le2.\quad}\tag{D}
$$
Indeed, a nonzero Hermitian operator $B$ that is nonnegative on all pure stabilizer states has $\operatorname{Tr}B>0$. To see this, if $\operatorname{Tr}B=0$, the stabilizer mixtures $(I\pm P)/d$ for nonidentity Hermitian Paulis $P$ imply $\operatorname{Tr}(BP)=0$; Pauli completeness then forces $B=0$. Thus stabilizer-polytope separation may always be tested using trace-one operators in $\Lambda_n$.

For any $A\in\Lambda_n$, writing $q=\operatorname{Tr}(A^2)$, Hilbert–Schmidt Cauchy–Schwarz yields
$$
\operatorname{Tr}(A\rho)
\ge\frac1d-
\sqrt{q-\frac1d}\sqrt{\operatorname{Tr}(\rho^2)-\frac1d}.
$$
Consequently, (D) implies (C). Conversely, every state in the Hilbert–Schmidt ball of radius
$$
r_*=\frac1{\sqrt{d(2d-1)}}
$$
about $I/d$ is positive semidefinite (this radius is smaller than the inradius of the full density-operator set). If the whole ball were contained in $\mathrm{STAB}_n$, minimizing $\operatorname{Tr}(A\rho)$ over that ball for any $A\in\Lambda_n$ would give $q\le2$. This proves the equivalence.

We use the phase-free Hermitian Pauli basis
$$
\mathcal P_n=\{I,X,Y,Z\}^{\otimes n},\qquad
\operatorname{Tr}(PQ)=d\,\delta_{P,Q},\qquad
\sum_{P\in\mathcal P_n}PBP=d\operatorname{Tr}(B)I.
$$

## 2. A cubic-moment inequality for a single negative eigenvalue

**Lemma 1 (spectral cubic-moment bound).**  
Let $H=H^\dagger$, $\operatorname{Tr}H=1$, and suppose $H$ has at most one strictly negative eigenvalue, counting multiplicities. Set
$$
q=\operatorname{Tr}(H^2),\qquad t=\operatorname{Tr}(H^3).
$$
If $q\ge1$, then
$$
\boxed{\quad t\le\frac{3q-1}{2}.\quad}\tag{1}
$$
If $q>1$, equality in (1) requires exactly one positive and one negative eigenvalue, with all other eigenvalues zero.

*Proof.* If $H\ge0$, then $q\le1$. The case $q\ge1$ therefore forces $H$ to be pure and gives equality.

Otherwise, write the nonzero eigenvalues as
$$
u_1,\ldots,u_r,-b,\qquad u_i>0,\quad b>0,\quad
s:=\sum_i u_i=1+b.
$$
Let
$$
E_2=\sum_{i<j}u_i u_j,\qquad
E_3=\sum_{i<j<k}u_i u_j u_k.
$$
Since $q=s^2-2E_2+b^2\ge1$,
$$
E_2\le b(1+b)=bs.
$$
For nonnegative $u_i$, a direct expansion gives
$$
\begin{aligned}
E_2^2-sE_3
={}&\sum_{i<j}u_i^2u_j^2
+\sum_{i<j<k}u_i u_j u_k(u_i+u_j+u_k)\\
&+2\sum_{i<j<k<\ell}u_i u_j u_k u_\ell\ge0.
\end{aligned}
$$
The third elementary symmetric polynomial of *all* eigenvalues of $H$ is
$$
e_3(H)=E_3-bE_2
\le E_2\left(\frac{E_2}{s}-b\right)\le0.
$$
Newton's identities, using $\operatorname{Tr}H=1$, give
$$
\operatorname{Tr}(H^3)=\frac{3\operatorname{Tr}(H^2)-1}{2}+3e_3(H).
$$
This proves (1). For $r\ge2$, the displayed expansion has a strictly positive $\sum_{i<j}u_i^2u_j^2$ term, so equality is impossible. Thus, when $q>1$, equality forces $r=1$ and $\operatorname{rank}H=2$. $\square$

The restriction on negative inertia is essential; see Section 7.

## 3. Weighted Pauli compression

**Lemma 2 (compression-to-cubic-moment inequality).**  
Let $A\in\Lambda_n$. For each nonidentity $P\in\mathcal P_n$ and sign $\pm$, compress $A$ to the $\pm1$ eigenspace of $P$, and identify this eigenspace with $n-1$ logical qubits by a Clifford isometry; call the resulting Hermitian operator $B_{P,\pm}$. Assume
$$
\operatorname{Tr}(B_{P,\pm}^2)
\le2\bigl(\operatorname{Tr}B_{P,\pm}\bigr)^2
\qquad\text{for all }P\ne I\text{ and signs }\pm.
\tag{2}
$$
Then, with $q=\operatorname{Tr}(A^2)$ and $t=\operatorname{Tr}(A^3)$,
$$
\boxed{\quad(d+2)q\le2d+2t-1.\quad}\tag{3}
$$

*Proof.* Let
$$
\Pi_\pm=\frac{I\pm P}{2},\qquad
x_P=\operatorname{Tr}(AP),\quad
y_P=\operatorname{Tr}(A^2P),\quad
c_P=\operatorname{Tr}(APAP).
$$
Because $(I\pm P)/d$ is a stabilizer mixture, $|x_P|\le1$. Direct expansion gives
$$
\operatorname{Tr}B_{P,\pm}=\frac{1\pm x_P}{2},\qquad
\operatorname{Tr}(B_{P,\pm}^2)=\frac{q+c_P\pm2y_P}{4}.
$$
Thus (2) yields
$$
q+c_P\pm2y_P\le2(1\pm x_P)^2.
$$
Multiply the $+$ inequality by $(1-x_P)/2$ and the $-$ inequality by $(1+x_P)/2$. Both weights are nonnegative, and addition gives
$$
q+c_P-2x_Py_P\le2(1-x_P^2).
\tag{4}
$$
Pauli orthogonality and twirling imply
$$
\sum_{P\ne I}x_P^2=dq-1,\qquad
\sum_{P\ne I}c_P=d-q,\qquad
\sum_{P\ne I}x_Py_P=dt-q.
$$
Summing (4) over $P\ne I$ gives
$$
d^2q+d-2dt\le2d(d-q),
$$
which is (3). $\square$

A useful exact **slack identity** follows from the same calculation. Define
$$
\beta_{P,\pm}=\operatorname{Tr}B_{P,\pm},\qquad
g(B)=2(\operatorname{Tr}B)^2-\operatorname{Tr}(B^2).
$$
Then
$$
\begin{aligned}
&\sum_{P\ne I}\bigl[
\beta_{P,-}g(B_{P,+})+\beta_{P,+}g(B_{P,-})\bigr]\\
&\quad=\frac d4\bigl[2d-(d+2)q-1+2t\bigr]\\
&\quad=\frac d4\bigl[(d-1)(2-q)+6e_3(A)\bigr].
\end{aligned}
\tag{5}
$$
Under assumption (2), each term on the left-hand side is nonnegative.

The underlying Pauli-compression approach is related to the recursion in [L2]; the weighted form above is useful because it retains the signed cubic moment $t$ rather than bounding it by an absolute-value estimate.

## 4. Main conditional dual result

**Theorem 1 (dual bound for at most one negative eigenvalue).**  
For every $n\ge1$,
$$
\boxed{
A\in\Lambda_n,\quad \nu_-(A)\le1
\quad\Longrightarrow\quad
\operatorname{Tr}(A^2)\le2,
}\tag{6}
$$
where $\nu_-(A)$ is the number of negative eigenvalues counted with multiplicity. No rank constraint, Clifford symmetry, or stabilizer-code support is assumed.

*Proof.* We induct on $n$. For $n=1$, write
$$
A=\frac12(I+a_xX+a_yY+a_zZ).
$$
Nonnegativity on the six pure stabilizer states is equivalent to $|a_x|,|a_y|,|a_z|\le1$. Hence
$$
\operatorname{Tr}(A^2)=\frac12(1+a_x^2+a_y^2+a_z^2)\le2.
$$

Suppose the assertion holds for $n-1$ qubits. For every Pauli compression $B_{P,\pm}$:

1. It is nonnegative on every $(n-1)$-qubit pure logical stabilizer state, because the Clifford encoding maps that state to an $n$-qubit pure stabilizer state.
2. Hermitian compression cannot increase negative inertia, so $\nu_-(B_{P,\pm})\le\nu_-(A)\le1$.
3. Its trace is nonnegative. If it is positive, normalize and apply the induction hypothesis. If it is zero, stabilizer positivity forces $B_{P,\pm}=0$.

Therefore every compression satisfies (2), and Lemma 2 gives
$$
(d+2)q\le2d+2t-1.
$$
If $q<1$, the conclusion is automatic. If $q\ge1$, Lemma 1 implies
$$
(d+2)q\le2d+3q-2,
$$
so $(d-1)(q-2)\le0$. Because $d\ge2$, we obtain $q\le2$. $\square$

**Corollary 1 (arbitrary rank-two support).** Every $A\in\Lambda_n$ with $\operatorname{rank}A\le2$ satisfies $\operatorname{Tr}(A^2)\le2$. Its support need not be a stabilizer code.

**Corollary 2 (negative inertia required for a counterexample witness).** If a density matrix $\rho$ satisfies the purity threshold in (C), then any $A\in\Lambda_n$ with $\operatorname{Tr}(A\rho)<0$ has $\nu_-(A)\ge2$. Thus no violation of (C) can be certified by a single-negative-eigenvalue witness.

The corollaries do **not** prove (C), since $\Lambda_n$ also contains operators with multiple negative eigenvalues.

## 5. Equality within the one-negative-eigenvalue class

**Theorem 2 (conditional equality classification).** If $A\in\Lambda_n$ and $\nu_-(A)\le1$, then $\operatorname{Tr}(A^2)=2$ if and only if
$$
\boxed{A=U\left[|0^{n-1}\rangle\langle0^{n-1}|\otimes
\frac{I+\epsilon_xX+\epsilon_yY+\epsilon_zZ}{2}\right]U^\dagger,}\tag{7}
$$
where $U$ is Clifford and each $\epsilon_j\in\{\pm1\}$.

*Proof.* Sufficiency follows by stabilizer postselection and the one-qubit calculation. Necessity for $n=1$ follows from the Bloch representation. Let $n\ge2$ and $q=2$. Lemmas 1–2 force $t=5/2$, so equality holds in Lemma 1. Thus $A$ has rank two with eigenvalues $(1\pm\sqrt3)/2$. Let $\Pi$ be its support projection and put
$$
a_P=\operatorname{Tr}(AP),\qquad b_P=\operatorname{Tr}(\Pi P).
$$
For each $P$ with $|a_P|<1$, both compression traces $\beta_\pm=(1\pm a_P)/2$ are positive. The nonnegative slack identity (5) has zero sum; hence both compressions saturate (2).

Choose an isometry $V:\mathbb C^2\to\operatorname{Ran}\Pi$, set $D=V^\dagger AV$ and $H=V^\dagger PV=h_0I+\mathbf h\cdot\boldsymbol\sigma$, and write $K_\pm=(I\pm H)/2$. Since $\det D=-1/2$, the rank-two compression has determinant $-\det K_\pm/2$. Saturation and its trace $\beta_\pm$ imply $\det K_\pm=\beta_\pm^2$, or
$$
(1\pm h_0)^2-\|\mathbf h\|^2=(1\pm a_P)^2.
$$
Subtracting and then substituting yield $h_0=a_P$ and $\mathbf h=0$. As $b_P=2h_0$,
$$
\boxed{b_P=2a_P\quad\text{if }|a_P|<1.}\tag{8}
$$
Define $S=\{P\in\mathcal P_n:|a_P|=1\}$, $m=|S|$. Parseval gives $\sum_Pa_P^2=2d=\sum_Pb_P^2$. Using (8) outside $S$,
$$
\sum_{P\in S}b_P^2=4m-6d.
$$
As $I\in S$ and $b_I=2$, we get $m>3d/2$. Stabilizer positivity makes the saturated Pauli set $S$ closed-noncontextual: values are deterministic and multiplicative on commuting contexts (including Pauli product signs). By the known CNC-set classification [ZD], a CNC set of size greater than $3d/2$ must be the maximal one-logical-qubit type, of size $2d$: it consists of $n-1$ fixed stabilizer axes and all three logical Pauli axes, up to Clifford conjugation. Then $\sum_{P\notin S}a_P^2=2d-m=0$; its deterministic Pauli coefficients give exactly (7). $\square$

This classifies equality **only under** $\nu_-(A)\le1$; it does not classify unrestricted dual extremizers.

## 6. Block-diagonal primal results and sharpness

**Theorem 3 (commuting-Pauli symmetry).** Let $n\ge2$ and suppose $\rho$ commutes with $n-2$ independent, pairwise commuting Hermitian Paulis $G_1,\ldots,G_{n-2}$. Then
$$
\boxed{\operatorname{Tr}(\rho^2)\le\frac1{2^n-1/2}
\quad\Longrightarrow\quad\rho\in\mathrm{STAB}_n.}\tag{9}
$$

*Proof.* Clifford conjugation transforms $G_j$ into $Z_j$ on the first $n-2$ qubits, so
$$
U\rho U^\dagger=\sum_{z\in\{0,1\}^{n-2}}p_z|z\rangle\langle z|\otimes\rho_z,
\tag{10}
$$
with two-qubit density operators $\rho_z$ for $p_z>0$. The exact two-qubit result [ZD, R] states
$\operatorname{Tr}(\sigma^2)\le2/7\Rightarrow\sigma\in\mathrm{STAB}_2$.
For $n=2$ this proves the claim. If $n>2$ and some $\rho_z$ is nonstabilizer, then $\operatorname{Tr}(\rho_z^2)>1/(4-1/2)$. Writing $p=p_z>0$, the remaining blocks occupy $d-4$ dimensions. Cauchy–Schwarz implies
$$
\operatorname{Tr}(\rho^2)>
\frac{p^2}{4-1/2}+\frac{(1-p)^2}{d-4}
\ge\frac1{d-1/2},
$$
a contradiction. Every nonzero block is a two-qubit stabilizer mixture; (10) is therefore an $n$-qubit stabilizer mixture. $\square$

The condition is **commutation**, not fixed-code-space support: $\rho$ may be full rank. The affine dimension of this fixed-frame family is $4d-1$.

**Theorem 4 (orthogonal stabilizer-code blocks).** Let $\{\Pi_j\}$ be mutually orthogonal stabilizer-code projectors summing to $I$, with ranks $k_j\in\{1,2,4\}$. If $\rho=\sum_j\Pi_j\rho\Pi_j$, implication (9) holds.

*Proof.* Normalize each nonzero block and decode its stabilizer code to $0$, $1$, or $2$ logical qubits. The exact stabilizer inradius is known in all three cases. A nonstabilizer block of dimension $k\in\{2,4\}$ and weight $p>0$ would imply
$$
\operatorname{Tr}(\rho^2)>
\frac{p^2}{k-1/2}+\frac{(1-p)^2}{d-k}
\ge\frac1{d-1/2},
$$
again a contradiction. Rank-one blocks are stabilizer states by definition. Hence all blocks are stabilizer mixtures. $\square$

**Sharpness within the restricted family.** Put
$$
F=\tfrac12(I+X+Y+Z),\qquad
W=|0^{n-1}\rangle\langle0^{n-1}|\otimes F,
\qquad\rho_*=\frac{2I-W}{2d-1}.
$$
Stabilizer postselection gives $W\in\Lambda_n$, $\operatorname{Tr}W=1$ and $\operatorname{Tr}(W^2)=2$. Since $\lambda_{\max}(W)=(1+\sqrt3)/2<2$, $\rho_*>0$, and
$$
\operatorname{Tr}(\rho_*^2)=\frac1{d-1/2},\qquad
\operatorname{Tr}(W\rho_*)=0.
$$
It obeys the commuting-Pauli conditions of Theorem 3. Indeed, it is explicitly stabilizer-decomposable:
$$
\rho_*=\frac{\Pi_0\otimes(\tau_{-X}+\tau_{-Y}+\tau_{-Z})
+2(I-\Pi_0\otimes I_2)}{2d-1},
$$
where $\Pi_0=|0^{n-1}\rangle\langle0^{n-1}|$ and $\tau_{-P}=(I-P)/2$. For small $\varepsilon>0$, the state
$$
\rho_\varepsilon=\rho_*-\varepsilon(W-I/d)
$$
is still positive and block diagonal, but
$\operatorname{Tr}(W\rho_\varepsilon)=-\varepsilon(2-1/d)<0$.
Thus $\rho_\varepsilon\notin\mathrm{STAB}_n$ while its purity approaches $1/(d-1/2)$ from above. The constant is optimal even with the additional symmetry. Related boundary constructions are known [L1, ZD].

## 7. Obstruction to extending the cubic-moment bound

The single-negative-eigenvalue assumption in Lemma 1 cannot be discarded. Here is an explicit element of $\Lambda_2$ with two negative eigenvalues and $e_3(A)>0$. **It is not a counterexample to the inradius conjecture.**

Take orthonormal vectors
$$
|\psi\rangle=\frac{(1,2,3,5)^T}{\sqrt{39}},\quad
|v_1\rangle=\frac{(2,-1,0,0)^T}{\sqrt5},\quad
|v_2\rangle=\frac{(3,6,-5,0)^T}{\sqrt{70}},
$$
let $\varepsilon=1/200$, and set
$$
A=(1+2\varepsilon)|\psi\rangle\langle\psi|
-\varepsilon\sum_{j=1}^2|v_j\rangle\langle v_j|.
$$
Its spectrum is $(1+2\varepsilon,-\varepsilon,-\varepsilon,0)$ and $\operatorname{Tr}A=1$. Every two-qubit pure stabilizer state has computational-basis support of size $1,2$, or $4$, with equal-modulus nonzero entries and relative phases $\pm1,\pm i$. The unequal entries $1,2,3,5$ rule out cancellation for support $1$ or $2$; for support $4$, their sum is odd, so the associated Gaussian integer cannot vanish. In all cases $|\langle s|\psi\rangle|^2\ge1/156$. Orthogonality of $v_j$ to $\psi$ gives
$$
\langle s|A|s\rangle\ge
(1+3\varepsilon)|\langle s|\psi\rangle|^2-\varepsilon
\ge(1+3\varepsilon)/156-\varepsilon>0.
$$
Therefore $A\in\Lambda_2$. Yet
$$
q=\operatorname{Tr}(A^2)=1+4\varepsilon+6\varepsilon^2=1.02015,
\quad e_3(A)=(1+2\varepsilon)\varepsilon^2>0,
$$
and Newton's identity yields
$$
t-\frac{3q-1}{2}=3(1+2\varepsilon)\varepsilon^2=0.00007575>0.
$$
Thus Lemma 1 fails for arbitrary negative inertia even inside $\Lambda_2$; the witness still has $q<2$.

## 8. Necessary conditions on a minimal counterexample

Suppose Conjecture (D) fails and let $n$ be the smallest number of qubits for which there exists $A\in\Lambda_n$ with
$$
q=\operatorname{Tr}(A^2)>2.
$$
For this minimal $n$, every Pauli compression to $n-1$ logical qubits satisfies (2), since (D) is valid at lower dimensions. Lemma 2 then gives
$$
(d+2)q\le2d+2t-1.
$$
Using $t=(3q-1)/2+3e_3(A)$, we obtain
$$
\boxed{\quad 6e_3(A)\ge(d-1)(q-2)>0.\quad}\tag{12}
$$
Consequently a minimal counterexample must have **all** of the following features:

1. At least two negative eigenvalues, and therefore rank at least three (by Theorem 1).
2. A strictly positive third elementary symmetric polynomial $e_3(A)$, quantitatively constrained by (12).
3. Every normalized one-round Pauli postselection with nonzero probability has squared Hilbert–Schmidt norm at most two, even though the global witness has squared norm greater than two.

These conditions provide a target for further investigation, but **do not exclude such operators**. The analysis stops here, without resolving the unrestricted conjecture.

## 9. Scope and unresolved issue

**Proven here:** the dual bound for $\nu_-(A)\le1$; equality classification in this class (using the existing CNC classification); the exact purity threshold for the stated stabilizer-code block families; sharpness therein; an obstruction to a general cubic-moment inequality; and necessary conditions for a minimal counterexample.

**Not proven:** the unrestricted $n$-qubit inradius conjecture, the dual bound for witnesses with multiple negative eigenvalues, or a classification of all dual extremizers. The derivations are not claimed to be novel without a comprehensive literature comparison. The natural next target is to use stabilizer positivity to rule out $q>2$ in the multiple-negative-eigenvalue regime, beyond the necessary condition (12).

## References

- **[P]** OpenContext, Problem 5, *Test problem: Hilbert–Schmidt inradius of the multiqubit stabilizer polytope*. <https://open-context.jan-bulling.com/problems/test-problem-hilbert-schmidt-inradius-of-the-multiqubit-stab>
- **[L1]** Zhenhuan Liu, Tobias Haug, Qi Ye, Zi-Wen Liu, and Ingo Roth, *Triangle Criterion: a mixed-state magic criterion with applications in distillation and detection*, arXiv:2512.16777; *PRX Quantum* **7**, 033016 (2026). <https://arxiv.org/abs/2512.16777>
- **[L2]** Zhenhuan Liu and Zi-Wen Liu, *On the geometry, typicality, and detectability of quantum magic*, arXiv:2609.03944v3 (5 November 2026). <https://arxiv.org/abs/2609.03944>
- **[ZD]** Michael Zurel and Jack Davis, *Basis-independent stabilizerness and maximally noisy magic states*, arXiv:2602.22336v2 (30 September 2026). <https://arxiv.org/abs/2602.22336>
- **[R]** Ben W. Reichardt, *Quantum universality by state distillation*, arXiv:quant-ph/0608085. <https://arxiv.org/abs/quant-ph/0608085>

---

*Partial results only; the open unrestricted conjecture remains unresolved.*

### 11. Partial result: An improved bound relative to Liu and Liu's result: from approximately 0.4583 to 0.4589.

- **Date:** 18 Nov 2026
- **Credit:** Rafael Wagner
- **AI use:** fully AI-generated
- **Permanent link:** https://open-context.jan-bulling.com/contribution/5.7
- **Votes:** 0 (0 up, 0 down)

## Main result

The following is an output from Claude Fable 5.2. The main claim is that it has improved the best bounds known for $n=3,4,5$ and also the one valid for all $n$. The claimed bound is 

>$$
>\mathrm{Tr}[\rho^2] \leq \frac{1}{d_n-0.4859} \implies \rho \in \mathrm{STAB}_n.
>$$

Furthermore, the values for $n=3,4,5$ are all improved by numerical checks:

>- For $n=3$: $1/2.0082 \approx 0.4979583$.
>- For $n=4$: $1/2.0314 \approx 0.4922713$.
>- For $n=5$: $1/2.0444 \approx 0.48914106$.

The results below have not been polished nor revised by a human.

## 1. Setting
 
Throughout $d=2^n$, $\Lambda_n=\{A=A^\dagger:\ \mathrm{Tr}[A]=1,\ \langle s|A|s\rangle\ge 0\ \text{for all stabilizer states } s\}$, $L_n$ is the cone over $\Lambda_n$, $S_n$ is the cone generated by stabilizer states, and
$$M_n:=\max_{A\in\Lambda_n}\mathrm{Tr}[A^2].$$
Elements of $\Lambda_n$ are called (normalised stabilizer) witnesses below. By polarity, every state with $\mathrm{Tr}[\rho^2]\le 1/(d-1/M_n)$ lies in $\mathrm{STAB}_n$, and this threshold is sharp; the conjecture is $M_n=2$. It is known that $M_1=M_2=2$.
 
For a Pauli operator $P\neq\mathbb{I}$ write $\Pi_\pm=(\mathbb{I}\pm P)/2$, $a_P=\mathrm{Tr}[PA]$, $p_\pm=\mathrm{Tr}[\Pi_\pm A]=(1\pm a_P)/2$ and $B_\pm=\Pi_\pm A\Pi_\pm$.
 
**Block lemma (known).** If $A\in\Lambda_n$ then $B_\pm$, viewed as an operator on the $(n-1)$-qubit space $\mathrm{range}\,\Pi_\pm$, belongs to $L_{n-1}$ and has trace $p_\pm\ge 0$.
 
Let $A_1=(\mathbb{I}+X+Y+Z)/2$, with eigenvalues $\lambda_\pm=(1\pm\sqrt3)/2$, and $A^\star=A_1\otimes|0\dots0\rangle\langle 0\dots0|$, the conjectured maximiser ($\mathrm{Tr}[A^{\star2}]=2$). Two new constants are used:
$$v_n:=\max_{A\in\Lambda_n}\lambda_{\max}(A),\qquad c_n:=\max_{A\in\Lambda_n}\big(-\lambda_{\min}(A)\big).$$
Dually, $v_n$ and $c_n$ are the smallest numbers with $v_n\mathbb{I}-\rho\in S_n$ and $\rho+c_n\mathbb{I}\in S_n$ for every state $\rho$. Clearly $v_n\ge\lambda_+$ and $c_n\ge-\lambda_-$.
 
## 2. Lifting, and the spectral recursion
 
**Lifting principle.** Let $F$ be a homogeneous function with $F(B)\in\mathcal K$ for every $B\in L_{n-1}$, where $\mathcal K$ is $S_{n-1}$ or the positive-semidefinite cone. Then for any weights $w_{P\pm}\ge 0$ the operator $\sum_{P\ne\mathbb{I}}\sum_\pm w_{P\pm}F(B_{P\pm})$ lies in $S_n$, respectively is positive semidefinite. With $\mathcal K=S_{n-1}$, $F(B)=M_{n-1}\mathrm{Tr}[B]\,\mathbb{I}-B$ and $w_{P\pm}=p_\mp$ this is the operator form of the Liu–Liu recursion (Lemma A of the previous note). The choice $w_{P\pm}=p_\mp$ is the only linear weight for which nothing is lost at $A^\star$. The new step is to use the same weights with the positive-semidefinite cone.
 
**Lemma (two sums).** For Hermitian $A$ with $\mathrm{Tr}[A]=1$ and $q=\mathrm{Tr}[A^2]$,
$$\sum_{P\ne\mathbb{I}}\sum_\pm p_\mp\,p_\pm\Pi_\pm=\frac{d(d-q)}{4}\,\mathbb{I},\qquad \sum_{P\ne\mathbb{I}}\sum_\pm p_\mp\,\Pi_\pm A\Pi_\pm=\frac{d^2}{4}A+\frac d4\mathbb{I}-\frac d2A^2 .$$
 
*Proof.* The first sum is $\sum_P p_+p_-\mathbb{I}=\sum_P\frac{1-a_P^2}{4}\mathbb{I}$ and $\sum_{P\ne\mathbb{I}}a_P^2=dq-1$. For the second, $\Pi_+A\Pi_++\Pi_-A\Pi_-=\tfrac12(A+PAP)$ and $\Pi_+A\Pi_+-\Pi_-A\Pi_-=\tfrac12(PA+AP)$, so the inner sum equals $\tfrac14(A+PAP)-\tfrac{a_P}{4}(PA+AP)$. Summing with $\sum_{P}PAP=d\,\mathbb{I}$ and $\sum_Pa_PP=dA$ (all Paulis, identity included) gives $\tfrac14\big((d^2-2)A+d\,\mathbb{I}\big)-\tfrac14\big(2dA^2-2A\big)$. $\square$
 
**Theorem 1 (spectral recursion).** Let $n\ge 2$, $A\in\Lambda_n$, $q=\mathrm{Tr}[A^2]$. Then
$$A^2-\frac d2A+\frac12\big(v_{n-1}(d-q)-1\big)\mathbb{I}\ \succeq\ 0,\qquad A^2-\frac d2A-\frac12\big(c_{n-1}(d-q)+1\big)\mathbb{I}\ \preceq\ 0 .$$
 
*Proof.* By the block lemma and the definition of $v_{n-1},c_{n-1}$, for every $P$ and sign, $v_{n-1}p_\pm\Pi_\pm-B_\pm\succeq 0$ and $B_\pm+c_{n-1}p_\pm\Pi_\pm\succeq 0$. Multiply by $p_\mp\ge 0$, sum over $P\ne\mathbb{I}$ and both signs, and insert the two sums: the results are $\frac d2$ times the first operator and $-\frac d2$ times the second. $\square$
 
Both inequalities are equalities on the range of $A^\star$ when $v_{n-1}=\lambda_+$, $c_{n-1}=-\lambda_-$, because $\lambda_\pm^2=\lambda_\pm+\tfrac12$.
 
**Corollary 1 (eigenvalue cap).** Let $n\ge 3$ and $A\in\Lambda_n$. Then
$$\lambda_{\max}(A)\ \le\ r_n(q):=\frac14\Big(d-\sqrt{d^2-8\big(v_{n-1}(d-q)-1\big)}\Big)$$
whenever the radicand is non-negative, and $\lambda_{\min}(A)\ge-\frac14\big(\sqrt{d^2+8(c_{n-1}(d-q)+1)}-d\big)$. If $v_{n-1}=\lambda_+$, the value of the first quadratic form at $\lambda_+$ is $\tfrac{\lambda_+}{2}(2-q)$; hence $\mathrm{Tr}[A^2]\ge 2$ forces $\lambda_{\max}(A)\le\lambda_+$, with strict inequality if $\mathrm{Tr}[A^2]>2$. Likewise $c_{n-1}=-\lambda_-$ and $\mathrm{Tr}[A^2]\ge2$ force $\lambda_{\min}(A)\ge\lambda_-$.
 
*Proof.* Each eigenvalue $\lambda$ satisfies $\lambda^2-\frac d2\lambda+\frac12(v_{n-1}(d-q)-1)\ge 0$, so it lies outside the open interval between the two roots, which are symmetric about $d/4$. By Proposition 2(c) below all eigenvalues are at most $d/4$ when $d\ge 8$, so they are below the smaller root $r_n(q)$. The other claims are the same computation. $\square$
 
So a hypothetical counterexample at the first level where the conjecture fails would have its whole spectrum strictly inside $(\lambda_-,\lambda_+)$, provided the spectral constants are sharp one level below.
 
## 3. Input data
 
**Proposition 2.**
 
(a) $v_1=\lambda_+$, $c_1=-\lambda_-$ (the single-qubit witnesses form a cube).
 
(b) **[computed]** $M_2=2$, $v_2=\lambda_+$, $c_2=-\lambda_-$.
 
(c) For every $n$ and every $A\in\Lambda_n$: each eigenvalue $\lambda$ of $A$ satisfies
$$-\frac{\sqrt{2d(d+1)}-d}{d+2}\ \le\ \lambda\ \le\ \frac{\sqrt{2d(d+1)}+d}{d+2},$$
so $c_n<\sqrt2-1$ and $\mathbb{I}+2A\succ 0$; moreover
$$\mathrm{Tr}\big[(\mathbb{I}+2A)^{-1}\big]\ \le\ \mathrm{Tr}[\mathbb{I}+2A]=d+2,\qquad\text{equivalently}\qquad \sum_i\frac{\lambda_i^2}{1+2\lambda_i}\ \le\ 1 ,$$
with equality at $A^\star$ for every $n$.
 
(d) Consequently, with $\varphi(w)=w+1/w$, any two eigenvalues $\lambda_i,\lambda_j$ ($i\ne j$) of $A\in\Lambda_n$ satisfy $\varphi(1+2\lambda_i)+\varphi(1+2\lambda_j)\le 8$. Since $\varphi(1+2\lambda_\pm)=4$: if $\lambda_{\min}(A)\le\lambda_-$ then $\lambda_{\max}(A)\le\lambda_+$, and if $\lambda_{\max}(A)\ge\lambda_+$ then $\lambda_{\min}(A)\ge\lambda_-$.
 
*Proof of (b).* $\mathrm{Tr}[A^2]$, $\lambda_{\max}(A)$ and $-\lambda_{\min}(A)$ are convex, so their maxima over the polytope $\Lambda_2$ are attained at vertices. Vertex enumeration returns $22{,}320$ vertices (the number reported in the literature), in seven purity classes:
 
| $\mathrm{Tr}[A^2]$ | 2 | 3/2 | 11/8 | 43/32 | 4/3 | 5/4 | 7/8 |
|---|---|---|---|---|---|---|---|
| vertices | 240 | 192 | 5760 | 2304 | 5760 | 5760 | 2304 |
| largest eigenvalue | 1.3660 | 0.8090 | 1.1614 | 1.1520 | 1.1455 | 1.0557 | 0.7255 |
| smallest eigenvalue | −0.3660 | −0.3090 | −0.2050 | −0.0930 | −0.1455 | −0.2021 | −0.2255 |
 
The $240$ vertices of purity $2$ are the Clifford images of $A_1\otimes|0\rangle\langle0|$ and have eigenvalues exactly $\lambda_\pm,0,0$. The enumeration was run in floating point; every vertex was then converted to exact rationals (denominators at most $4$) and checked exactly to be feasible with a tight set of rank $15$. Completeness rests on the count matching the known one.
 
*Proof of (c).* Stabilizer states form a 3-design, so for any Hermitian $X$ the inequality $\sum_s\langle s|X|s\rangle^2\langle s|A|s\rangle\ge 0$ reads $(\mathrm{Tr}[X])^2+\mathrm{Tr}[X^2]+2\mathrm{Tr}[X]\mathrm{Tr}[XA]+2\mathrm{Tr}[X^2A]\ge 0$, that is, with $W=\mathbb{I}+2A$,
$$\mathrm{Tr}[X^2W]+\mathrm{Tr}[X]\,\mathrm{Tr}[XW]\ \ge\ 0\qquad\text{for all Hermitian } X .$$
For the eigenvalue bounds take $X=x\mathbb{I}+\psi$ with $A\psi=\lambda\psi$ and put $u=(d+2)x$: the inequality becomes $\frac{d+1}{d+2}u^2+u+(1+2\lambda)(u+2)\ge 0$ for all real $u$, and optimising over $u>-2$ and over $u<-2$ gives the two bounds. Hence $W\succ 0$, and $X=W^{-1}-\mathbb{I}$ is admissible: then $\mathrm{Tr}[X^2W]=\mathrm{Tr}[W^{-1}]-d+2$, $\mathrm{Tr}[X]=\mathrm{Tr}[W^{-1}]-d$ and $\mathrm{Tr}[XW]=-2$, so the inequality is $d+2-\mathrm{Tr}[W^{-1}]\ge 0$. The equivalent form uses $\frac{\lambda^2}{1+2\lambda}=\frac\lambda2-\frac14+\frac{1}{4(1+2\lambda)}$ and $\sum_i\lambda_i=1$. At $A^\star$ the eigenvalues of $W$ are $2+\sqrt3$, $2-\sqrt3$ and $1$, which are closed under $w\mapsto 1/w$. $\square$
 
*Proof of (d).* Let $w_k=1+2\lambda_k>0$, so $\sum_kw_k=d+2$ and $\sum_k1/w_k\le d+2$. By Cauchy–Schwarz, $\sum_{k\ne i,j}1/w_k\ge\frac{(d-2)^2}{d+2-w_i-w_j}\ge d-6+w_i+w_j$. $\square$
 
Statement (c) is an unconditional "weighted purity" bound that $A^\star$ saturates for every $n$; the conjecture is the unweighted version $\sum_i\lambda_i^2\le 2$. In dual form the eigenvalue bound says that $\rho+\frac{\sqrt{2d(d+1)}-d}{d+2}\mathbb{I}\in S_n$ for every state $\rho$: white noise with visibility about $2.41/d$ makes any state a stabilizer mixture. I have not found (c) in the literature, but it is a short consequence of the 3-design property and may well be known.
 
## 4. Improved bounds
 
Two elementary facts about a spectrum with $\sum_i\lambda_i=1$, $\sum_i\lambda_i^2=q$ and largest eigenvalue $\ell$:
$$\mathrm{Tr}[A^3]\le\ell q,\qquad \mathrm{Tr}[A^3]\le\ell^3+(q-\ell^2)^{3/2},$$
and $\ell\mapsto\min\{\ell q,\ \ell^3+(q-\ell^2)^{3/2}\}$ is non-decreasing on $(0,\sqrt q\,]$. The second bound can be sharpened with Wilkins' skewness inequality (1944) applied to the other $d-1$ eigenvalues: with $\mu=\frac{1-\ell}{d-1}$ and $S=q-\ell^2-\frac{(1-\ell)^2}{d-1}\ge 0$,
$$\mathrm{Tr}[A^3]\le\tau_d(\ell,q):=\ell^3+(d-1)\mu^3+3\mu S+\frac{d-3}{\sqrt{(d-1)(d-2)}}\,S^{3/2}.$$
 
**Theorem 3.** Put $\bar M_2=2$, $\bar v_2=\lambda_+$ and, for $n\ge 3$,
 
- $r_n(q)=\frac14\big(d-\sqrt{d^2-8(\bar v_{n-1}(d-q)-1)}\big)$,
- $T_n(q)=\max\big\{\min\{\ell q,\ \tau_d(\ell,q)\}:\ \tfrac1d\le\ell\le\min\{r_n(q),\sqrt q\},\ S\ge 0\big\}$,
- $\bar M_n=\sup\{q\ge\bar M_{n-1}:\ d(q-\bar M_{n-1})\le 2T_n(q)-\bar M_{n-1}q-1\}$,
- $\bar v_n=$ the solution $\ell$ of $\ell=r_n\big(\ell^2+\frac{(1-\ell)^2}{d-1}\big)$.
Then $M_n\le\bar M_n$ and $v_n\le\bar v_n$ for all $n$. **[computed]** Evaluating the recursion:
 
| $n$ | Liu–Liu bound on $M_n$ | this note, elementary | this note, with Wilkins | $\bar v_n$ |
|---|---|---|---|---|
| 3 | 2.0938 | 2.0150 | 2.0082 | 1.3906 |
| 4 | 2.1384 | 2.0380 | 2.0314 | 1.4020 |
| 5 | 2.1603 | 2.0506 | 2.0444 | 1.4076 |
| 6 | 2.1712 | 2.0570 | 2.0511 | 1.4103 |
| all $n$ | 2.1819 | 2.0635 | 2.0578 | 1.4130 |
 
("Elementary" replaces $\tau_d$ by $\ell^3+(q-\ell^2)^{3/2}$, in which case $T_n(q)=r_n(q)^3+(q-r_n(q)^2)^{3/2}$ in closed form.)
 
*Proof.* Induction on $n$. Let $A\in\Lambda_n$ with purity $q$ and largest eigenvalue $\ell$.
 
- (i) Corollary 1 gives $\ell\le r_n(q)$; replacing $v_{n-1}$ by the larger $\bar v_{n-1}$ only weakens Theorem 1.
- (ii) Hence $\mathrm{Tr}[A^3]\le T_n(q)$ by the spectral facts above.
- (iii) The Liu–Liu recursion states $d(q-M_{n-1})\le 2\mathrm{Tr}[A^3]-M_{n-1}q-1$, and it remains valid with $M_{n-1}$ replaced by $\bar M_{n-1}$ because $q\le d$. So either $q\le\bar M_{n-1}$ or $q$ satisfies the inequality defining $\bar M_n$.
- (iv) Always $q\ge\ell^2+\frac{(1-\ell)^2}{d-1}$, and $r_n$ is decreasing, so $\ell\le r_n\big(\ell^2+\frac{(1-\ell)^2}{d-1}\big)$, i.e. $\ell\le\bar v_n$. $\square$
**Corollary 2.** For every $n$: every state with $\mathrm{Tr}[\rho^2]\le 1/(d-0.4859)$ is a stabilizer mixture; the inradius of $\mathrm{STAB}_n-\mathbb{I}/d$ is at least $1/\sqrt{d(2.0578\,d-1)}$, within $1.5\%$ of the conjectured $1/\sqrt{d(2d-1)}$; and every eigenvalue of every $A\in\Lambda_n$ lies in $(-0.4143,\ 1.4130)$. For three qubits the purity threshold is at least $1/(8-0.4979)$.
 
Two remarks on robustness. If one does not want to rely on the two-qubit enumeration for $v_2$ and uses only $v_2\le\frac14+\sqrt{\tfrac34\cdot\tfrac74}$ (which follows from $M_2=2$), the same recursion still gives $M_n\le 2.1183$, already below Liu–Liu. And running the same recursion without the cap reproduces their values ($67/32$ for three qubits, $2.181847$ in the limit), which confirms that I am using their recursion as they do; the Liu–Liu column of the table is this reproduction.
 
**Proposition 4 (maximisers are pinned).** For every $A\in\Lambda_n$, $2\mathrm{Tr}[A^3]\ge M_nq+1-2d(M_n-q)$. In particular a maximiser ($q=M_n$) has $\mathrm{Tr}[A^3]\ge(M_n^2+1)/2\ge\tfrac52$, therefore $1.34\le\lambda_{\max}(A)\le 1.4130$, and then $\lambda_{\min}(A)\ge-0.368$.
 
*Proof.* $A\otimes|0\rangle\langle0|\in\Lambda_{n+1}$ has the same $q$ and $\mathrm{Tr}[A^3]$; apply the Liu–Liu recursion at level $n+1$. For the window of $\lambda_{\max}$ combine with $\mathrm{Tr}[A^3]\le\ell^3+(q-\ell^2)^{3/2}$ and Theorem 3; the bound on $\lambda_{\min}$ then follows from Proposition 2(d). $\square$
 
## Routes that do not work
 
- **A trace-norm bound.** $\|A\|_1\le\sqrt3$ on $\Lambda_n$ would imply both (E) and $M_n=2$, but it is false already for two qubits (the $192$ vertices of purity $3/2$ have $\|A\|_1=\sqrt5$).
- **Uniform (unweighted) lifting.** In the basis $(\mathbb{I},\,q\mathbb{I},\,A,\,A^2)$ the unweighted block sum acts as $\frac{d^2-4}{4}\big(\mathrm{id}+\frac{d}{d^2-4}\Psi\big)$, where $\Psi$ has eigenvalues $3,-3,0,0$. It reproduces $q\mathbb{I}-A$ and $\mathbb{I}-A^2$ exactly and distorts everything else at relative order $1/d$, so every unweighted lift loses $O(1/d)$ per level and the losses add up to $O(1)$.
- **Tight lifting of quadratic statements.** With weights $p_\mp$ a quadratic block statement lifts to a cubic one containing operators that are not polynomials in $A$ (for example $\frac1d\sum_Pa_P\,PAPAP$). The resulting chain of invariants $S_1\ge S_2/d\ge S_3/d^2$ is valid whenever the lower rung is non-negative on blocks, but $S_3$ is negative somewhere on $\Lambda_2$, so the chain does not close by positivity.
- **Moment methods for (E).** Weights that are polynomials of degree at most two in the overlaps give exactly Proposition 2(c). At degree three the failure of the 4-design property produces a term that cannot be cancelled (Section 6, item 3).
- **Lifting two-eigenvalue (Ky Fan) statements.** The two halves of a Pauli measurement can each host one unit of the top-two eigenspace, so only the top eigenvalue of each block enters and nothing is learned about $\lambda_2$.
## 6. Clifford-algebra and group-theoretic remarks
 
These are exact statements found while looking for structure; all were verified numerically on random instances.
 
1. **Two qubits are $\mathfrak{so}(6)$.** With five anticommuting Paulis $\gamma_1,\dots,\gamma_5$, set $\Gamma_{i6}=\gamma_i$ and $\Gamma_{ij}=-\mathrm{i}\gamma_i\gamma_j$. These are the $15$ non-identity Paulis, and $a_{ij}=\mathrm{Tr}[\Gamma_{ij}A]$ is a real antisymmetric $6\times6$ matrix. The $60$ stabilizer states correspond to the perfect matchings of six points with signs of fixed parity, the eigenvalues of $A$ are $(1\pm\mu_1\pm\mu_2\pm\mu_3)/4$ with a fixed sign parity, where $\pm\mathrm{i}\mu_k$ are the eigenvalues of $a$, and
$$\mathrm{Tr}[A^3]=\frac{12\,\mathrm{Tr}[A^2]-2}{16}+\frac38\,\mathrm{Pf}(a).$$
On each of the $15$ lines the three coefficients lie (after fixing signs) in a tetrahedron inscribed in the elliptope $x^2+y^2+z^2\le 1+2xyz$; summing over lines gives $\mathrm{Tr}[A^2]\le\frac32+\frac16\mathrm{Pf}(a)$, which is exactly the Liu–Liu recursion at $n=2$. So $M_2=2$ follows from $\mathrm{Pf}(a)\le 3$ on $\Lambda_2$, which holds numerically. I did not find a proof of this Pfaffian bound by hand.
2. **The uniform pinching operator** described in Section 5.3; its eigenvector for eigenvalue $3$ is $(1+q)\mathbb{I}+2A+2A^2$, the element of $S_n$ produced by the 3-design.
3. **Fourth moment of stabilizer states.** For a pure state $\psi$,
$$\sum_s\langle s|\psi|s\rangle^3\,|s\rangle\langle s|\ \propto\ \mathbb{I}+3\psi+\psi^{\circ3},\qquad \psi^{\circ 3}:=\frac1d\sum_P\langle\psi|P|\psi\rangle^3P=\frac1d\sum_P\langle\psi|P|\psi\rangle^2\,P\psi P .$$
The state $\psi^{\circ3}$ is the only trace of the "not a 4-design" property at this order.
4. **Closure properties.** $\Lambda_n$ is closed under the triple Pauli-coefficient product $(a_Pb_Pc_P)$ and under the $2^{2n}$ involutions $a_P\mapsto(-1)^{Q(P)}a_P$, $Q$ a quadratic form polarising to the symplectic form (transposition followed by a Pauli conjugation).
5. **Two-qubit pairings.** From the vertex list, $-1\le\mathrm{Tr}[AB]\le 2$ on $\Lambda_2\times\Lambda_2$, so $\mathbb{I}+A\in S_2$ as well as $2\mathbb{I}-A\in S_2$.

#### Comment by Jan, 18 Nov 2026

This looks nice. I can also btw now use \tag (this worked actually before, if one writes math in different lines. Now it always works $$2+2=4\tag{1}$$.

#### Reply by Jan, 18 Nov 2026

Replying to my own comment!

#### Reply by Rafael Wagner, 19 Nov 2026

Nice!

I have just found out another proof, which is improving the bound for all $n$ much more.

### 12. Challenge to a solution: I want to challenge your solution

- **Date:** 18 Nov 2026
- **Credit:** Jan
- **AI use:** partially AI-assisted
- **Permanent link:** https://open-context.jan-bulling.com/contribution/5.8
- **Votes:** 0 (0 up, 0 down)

This should appear automatically as an "edit" in the history section. Hope it does!

### 13. Partial result: Proofs for OpenContext 5.9: the sharp three-qubit threshold and a uniform bound of 0.496222

- **Date:** 19 Nov 2026
- **Credit:** Yujie Zhang
- **AI use:** fully AI-generated
- **Permanent link:** https://open-context.jan-bulling.com/contribution/5.10
- **Votes:** 1 (1 up, 0 down)

## Main result

This is the proof of the two bounds announced in [OpenContext 5.9](https://open-context.jan-bulling.com/contribution/5.9). Let $d=2^n$ and let $c_n$ be the largest constant such that every $n$-qubit density matrix of purity at most $1/(d-c_n)$ is a convex mixture of stabilizer states. Then

$$
\begin{aligned}
c_3&=\tfrac12,\\
c_n&\ge c_* =\frac{327659848669}{660308814208}\\
&=0.496222133672421\ldots \qquad(n\ge1).
\end{aligned}
$$

For three qubits the sharp purity threshold is $2/15$. The uniform bound applies to every state, with arbitrary spectrum and eigenbasis. The conjectured equality $c_n=1/2$ remains open for $n\ge4$. The main result of [OpenContext 5.7](https://open-context.jan-bulling.com/contribution/5.7) gave a uniform value of $0.4859$.

**Which steps use a computer?** Both proofs are computer-assisted. For $c_3=1/2$, exact polynomial certificates classify the witnesses on the sphere $\operatorname{Tr}(A^2)=2$. Interpolation then excludes witnesses outside that sphere, and an explicit contact state proves sharpness. These last two deductions are analytic. For the uniform bound, rigorous interval computations establish the finite estimates for $n=4,\ldots,15$. Once the bound on $c_{15}$ is in hand, **one analytic averaging identity covers every $n>15$**. No computation at 16 qubits or beyond is required. This contribution is AI-generated.

## 1. Dual formulation

A normalized witness is a trace-one Hermitian operator $A$ such that $\langle s|A|s\rangle\ge0$ for every pure stabilizer state $s$. Write $\Lambda_n$ for this polytope and $q=\operatorname{Tr}(A^2)$. By polarity,

$$
M_n=\max_{A\in\Lambda_n}\operatorname{Tr}(A^2),\qquad c_n=1/M_n.
$$

Indeed, the squared distance from $I/d$ to the witness hyperplane $\operatorname{Tr}(A\rho)=0$ is $1/[d(dq-1)]$. Minimizing over witnesses and adding $1/d$ gives the purity at first contact. Thus $c_n$ here is the purity parameter $1/M_n$ in the notation of OpenContext 5.7.

## 2. The sharp three-qubit bound

**Claim.** The only witnesses in $\Lambda_3$ with $q=2$ are the Clifford conjugates of

$$
V_0=|00\rangle\langle00|\otimes\frac{I+X+Y+Z}{2}.
$$

Postselecting the first two qubits of a stabilizer state onto $00$ leaves a one-qubit stabilizer state or zero. This proves feasibility of $V_0$, and direct calculation gives $\operatorname{Tr}(V_0^2)=2$. Consider any other candidate on the same sphere.

**Step 1. A large Pauli expectation.** Let $a_R=\operatorname{Tr}(AR)$ for each nonidentity Hermitian Pauli $R$. A signed Pauli outcome with projector $\Pi$ gives a two-qubit compression $B=\Pi A\Pi$ of trace $p$. If $p>0$, then $B/p\in\Lambda_2$. If $p=0$, the compression vanishes. Hence $\operatorname{Tr}(B^2)\le2p^2$ and $B\preceq Lp\Pi$, where $L=1.366026$ is a rational upper bound on the largest two-qubit witness eigenvalue. Suppose every $|a_R|\le r=49/50$. Sum the two compression inequalities over the two outcomes and the nonidentity Paulis, weighting the outcomes by $1\mp a_R/r$. With $t=\operatorname{Tr}(A^3)$, $x=\lambda_{\max}(A)$ and $h=1/r-1$, the sums give

$$
\begin{aligned}
0&\le J(t)=2t-5-2h(4-t),\\
0&\le W(x)=2x^2-8x+6L-1\\
&\hspace{19mm}+h[2x^2-L(2+x)+(L-x)/4].
\end{aligned}
$$

The convex quadratic $W$ is negative at both $1.359$ and $1.5$. Since $x\le\sqrt2$, this forces $x<1.359$. The fixed-mean, fixed-variance cubic bound on the seven remaining eigenvalues then gives $t<2.527$. But $J$ increases and $J(2.527)=-3/490$. Thus some signed Pauli has expectation above $49/50$. This step uses the Pauli compression and cubic moment estimate in [Liu and Liu, Appendix A](https://arxiv.org/html/2609.03944v3#A1).

**Step 2. Computer-assisted certificate for an overlap gap.** Fix a Clifford conjugate $V$ of $V_0$. Let $\delta$ be the difference of the 63 nonidentity Pauli coefficient vectors of $A$ and $V$, and put $D=8[2-\operatorname{Tr}(AV)]$. Both witnesses have $q=2$, so $\|\delta\|^2=2D$. The first rational certificate gives

$$
\frac{343}{1242}D^3-D\|\delta\|^2\ge0.
$$

The polynomial on the left is a positive sum of 38 symmetry-averaged triple products of stabilizer expectations. Each factor comes from a stabilizer projector on which $V$ has zero expectation. If $A\ne V$, then $D>0$. Substitution gives $D\ge2484/343$ and hence $\operatorname{Tr}(AV)\le T=751/686$.

**Step 3. Computer-assisted contradiction for the remaining witnesses.** Suppose $A$ differs from every conjugate of $V_0$. In any two-qubit block $B$ of trace $p$, the vertices of $\Lambda_2$ with squared norm 2 lift to conjugates of $V_0$. Every other vertex has squared norm at most $3/2$. Decomposing $B/p$ into vertices and using Cauchy–Schwarz gives the first inequality below. The second follows by polarizing the cubic and checking the resulting separately affine expression on triples of two-qubit vertices.

$$
\begin{aligned}
\operatorname{Tr}(B^2)&\le \max\{Tp,3p^2/2\},\\
\operatorname{Tr}(B^3)&\le p\operatorname{Tr}(B^2)+p^3/2.
\end{aligned}
$$

Rotate the signed Pauli of Step 1 to $Z$ on the first qubit. The second certificate combines 42 nonnegative polynomials from the compression bounds, stabilizer positivity, the overlap bounds and the large Pauli expectation. It adds six polynomial multiples of $q-2$, which vanish on the sphere. Exact expansion bounds this nonnegative sum strictly below $-19/100000$. This contradiction proves the claim. The enumeration used here is the published classification of 22,320 **two-qubit** vertices. [Reichardt, Table 2](https://arxiv.org/abs/quant-ph/0608085) · [Zurel and Davis, Appendix A](https://arxiv.org/html/2602.22336#A1)

**Analytic conclusion and sharpness.** If a feasible $A$ had $q>2$, mixing it with $I/8$ would reach $q=2$ with every nonidentity Pauli expectation strictly smaller than one in absolute value. Every conjugate of $V_0$ has a Pauli expectation of magnitude one. This excludes $q>2$ and proves $c_3=1/2$. The contact state $\rho_*=(2I-V_0)/15$ is positive definite, has purity $2/15$, and satisfies $\operatorname{Tr}(V_0\rho_*)=0$. Continuing a short distance along the ray from $I/8$ through $\rho_*$ preserves positivity and makes the witness expectation negative. Magic-state purities therefore approach $2/15$ from above.

## 3. A lower bound for every number of qubits

**Analytic bound from the off-diagonal block.** Choose a signed Pauli with maximal absolute expectation $\alpha\ge0$. In its eigenbasis write $A$ with diagonal blocks $B,C$ of size $k=d/2$ and upper-right block $H$, where $C$ has the smaller trace. Let $u$ minimize $m=\langle u|B|u\rangle$ over logical stabilizer states, and take $b\ge\|C\|_{\rm op}$. Complete $u$ to a Clifford image of the computational basis, with vectors $v_j$. For each $v_j$ the following is a stabilizer state for all four phases $\omega\in\{1,-1,i,-i\}$.

$$
\frac{|0\rangle|u\rangle+\omega|1\rangle|v_j\rangle}{\sqrt2}.
$$

Positivity at the four phases bounds the real and imaginary parts of $\langle u|H|v_j\rangle$ and gives $\|H^\dagger u\|\le\sqrt{k/2}(m+b)$. Compress to the entire $C$ block together with the single vector $|0\rangle|u\rangle$. With eigenvalues ordered decreasingly, min–max gives

$$
\lambda_{k+1}(A)\ge\frac{m-b-\sqrt{1+2k}\,(m+b)}{2}.
$$

The right side decreases in both $m$ and $b$. Let $x,\ell$ be the extreme eigenvalues of $A$, and let $L$ bound the operator norm of every normalized witness on $n-1$ qubits. Take $b=L(1-\alpha)/2$. The operator $B-mI$ is stabilizer-positive, so its top eigenvalue is at most $L$ times its trace. Also, compressing $A$ to the span of the two block components of a top eigenvector gives $x+\ell\le\lambda_{\max}(B)+\lambda_{\max}(C)$. Together these imply

$$
m\le\frac{L-x-\ell+b}{kL-1}.
$$

**Computer-assisted bounds for 4 through 15 qubits.** Apply this extra eigenvalue bound directly at $n=4,5,6$, alongside the weighted Pauli compression and spectral inequalities. At fixed trace and squared norm, maximizing the third moment of the remaining eigenvalues reduces to finitely many multiplicities. Rational interval bounds with outward-rounded square roots exclude each candidate sphere. The lower bounds used in the uniform argument are $c_4\ge1/2.0070400$, $c_5\ge1/2.0108900$ and $c_6\ge1/2.0129180$. Propagating the weighted bounds through $n=7,\ldots,15$ gives $c_{15}\ge10000000/20151331$. The extra off-diagonal constraint enters at $n=4,5,6$. The scalar compression and fixed-moment cubic estimates come from [Liu and Liu](https://arxiv.org/html/2609.03944v3#A1), and the operator-weighted compression is developed in [OpenContext 5.7](https://open-context.jan-bulling.com/contribution/5.7).

**Analytic extension from 15 to every larger number of qubits.** Set $K=2^{15}$ and $U=20151331/10000000$. In dimension $d\ge K$, choose a uniform Clifford image $\Pi$ of a fixed rank-$K$ stabilizer-code projector. Its range encodes 15 qubits. The certified 15-qubit bound thus gives $\operatorname{Tr}[(\Pi A\Pi)^2]\le U[\operatorname{Tr}(\Pi A)]^2$. The Clifford second moment gives $\mathbb E[\Pi\otimes\Pi]=aI+bF$, where $F$ swaps the two copies and $b/a=(d-K)/(Kd-1)$. Taking the expectation and solving yields

$$
a q+b\le U(a+bq)\quad\Longrightarrow\quad
c_n\ge\frac{d c_*+1}{d+c_*},\qquad c_*=\frac{K-U}{KU-1}.
$$

This lower bound exceeds $c_*$. Substituting $K,U$ gives the fraction in the main result. The finite estimates cover $d<K$. This last deduction uses only the [Clifford second-moment identity](https://arxiv.org/abs/quant-ph/0606161) and algebra.

## 4. Reproducing the finite checks

The verification package is named **unrestricted-stabilizer-bound-certificate.zip**. The three-qubit package under **prior_inputs** contains **verify_all.py**, which checks the two-qubit data and the polynomial certificates using integer or rational arithmetic. It requires Python and NumPy. For the uniform result, run the following scripts from the outer package with Python 3.10 or later and assertions enabled. These scripts use the standard library. The README files in **prior_inputs** give the commands for checking the preceding spectral estimates.

~~~sh
python3 verify_four.py
python3 verify_five_six.py
python3 verify_transfer.py
~~~

#### Comment by Rafael Wagner, 19 Nov 2026

This is quite interesting; I am reviewing and digesting the proof of 0.4996 and will upload it soon today. I really think that, given more time we could be closing the problem entirely?!

Crazy

#### Reply by Rafael Wagner, 19 Nov 2026

The result I have has the sharp proof that $c=1/2$ with the conjecture remaining open for $n\geq 6$.

I wonder whether we could simply use an insight that, nothing in principle seems to be stopping us from just making the bound tighter and tigther.

### 14. Partial result: An improved computer-assisted bound at 0.4996 for all $n$, and a sharp 0.5 value for $n\leq 5$ (part I)

- **Date:** 19 Nov 2026
- **Credit:** Rafael Wagner
- **AI use:** partially AI-assisted
- **Permanent link:** https://open-context.jan-bulling.com/contribution/5.11
- **Votes:** 0 (0 up, 0 down)

## Main result

After a couple of prompts, Claude Fable 5.2 found another advance:

>**Theorem ($0.4996$ bound):** Let $\rho \in \mathcal{D}(\underbrace{\mathbb{C}^2 \otimes \cdots \otimes \mathbb{C}^2}_{n\text{ times}})$ be an arbitrary $n$-qubit state, and denote $\mathrm{STAB}_n$ the convex polytope of $n$-qubit stabilizer states. For every integer $n \geq 1$, it holds that 
>$$
>\mathrm{Tr}[\rho^2] \leq \frac{1}{d_n-a^*} \implies \rho \in \mathrm{STAB}_n,
>$$
>where $d_n = 2^n$ is the dimension of the underlying Hilbert space and $a^*=0.4996$. Furthermore, for $1\leq n \leq 5$ the conjecture is shown to hold, i.e., $a^*=\tfrac{1}{2}$.  

The results have used the strategy present in the notes by Y. Wang. ["Partial Results on the Hilbert–Schmidt Inradius of the Multiqubit Stabilizer Polytope"](https://open-context.jan-bulling.com/contribution/5.6), in collaboration with ChatGPT 6 Astra. 

**How this contribution is organised.** Because of its length, the argument is divided into two parts. This first part contains the preliminaries and two new results on the spectrum of the operators in the $\Lambda$-polytope: the theorem on the spectrum of a witness and the star inequality. Both are proved analytically. The second part combines them with the Liu–Liu inequality of Ref. [3] to prove the theorem above.

## Preliminary mathematical ideas

Recall that a crucial ingredient in this discussion is the notion of the $\Lambda$-polytope from Ref. [4]. For more on $\Lambda$-polytopes we refer to Refs. [8, 9, 10, 11, 4]. It can be defined as follows:
$$
\Lambda_n=\left \{A=A^\dagger:\ \mathrm{Tr}[A]=1,\ \langle s|A|s\rangle\ge 0\ \text{for all stabilizer states } \vert s\rangle  \right\}.
$$ 
Above the relevant operators are always defined relative to $n$-qubit Hilbert spaces $\mathbb{C}^2 \otimes \cdots \otimes \mathbb{C}^2$ of dimension $d_n = 2^n$. Below, all sums taken over Paulis $P$ run over the $d^2-1$ Pauli operators which are different from the identity $\mathbb{I}$. For such Pauli $P$, we also denote  
$$
\Pi_\pm=(\mathbb{I}\pm P)/2,
$$ 
$a_P=\mathrm{Tr}[PA]$ as the trace of that Pauli with an operator $A$, and $B_\pm=\Pi_\pm A\Pi_\pm$. Let $A_1=(\mathbb{I}+X+Y+Z)/2$, with eigenvalues $\lambda_\pm=(1\pm\sqrt3)/2$, and $A^\star=A_1\otimes|0\dots0\rangle\langle0\dots0|$. Throughout this document we define 
$$
v:=\lambda_+,\qquad c:=-\lambda_-=v-1.
$$
With these definitions, it is simple to show that the following equations all hold:
$$
\qquad 2v^2=2v+1,\quad vc=\tfrac12,\quad v^2+c^2=2,\quad v^3-c^3=\frac52.
$$

Building on my [earlier contribution](https://open-context.jan-bulling.com/contribution/5.7), we recall that we are using ideas related to the stabilizer theory, which is tied to quantum error correction. Because of that, the language of "codes" is used below, without loss of generality.

>**Lemma (Block lemma).** For an arbitrary Pauli $P$ the range of $\Pi_\pm$ is an $(n-1)$-qubit code space whose stabilizer states are $n$-qubit stabilizer states. Hence $B_\pm$ is non-negative on them, 
>$$
>\mathrm{Tr}[B_\pm]=(1\pm a_P)/2\ge0,
>$$ 
>with $a_P=\mathrm{Tr}[PA]$, and 
>$$
>B_\pm/\mathrm{Tr}[B_\pm]\in\Lambda_{n-1}.
>$$
>As a specific case, if $\mathrm{Tr}[B_\pm]=0$ then $B_\pm=0$.

This lemma is shown in Ref. [4] (their Lemma 1, part (2)) and discussed also in Ref. [3], Appendix A (see the "Pauli-compression recursion" subsection). 

The following is another simple Lemma, which can be found, for example, in standard textbooks [7].

>**Lemma (a rule for Pauli sums).** For arbitrary Hermitian operators $X,Y$, with the identity included in the sums, we have that 
>$$
>\sum_P PXP = d\, \mathrm{Tr}[X]\,\mathbb{I}
>$$ 
>and 
>$$
>\sum_P\mathrm{Tr}[PX]\mathrm{Tr}[PY]=d\,\mathrm{Tr}[XY].
>$$

The first rule in the above lemma says that conjugation by a uniformly random Pauli operator is the completely depolarizing channel, i.e. that the Pauli group is a unitary 1-design: Watrous, Eq. (4.109); Ref. [12]; for one qubit it is the fully depolarizing case of the [quantum depolarizing channel](https://en.wikipedia.org/wiki/Quantum_depolarizing_channel) written with the Kraus operators $\mathbb{I},X,Y,Z$.

The second rule is [Parseval's identity](https://en.wikipedia.org/wiki/Parseval%27s_identity) for the Pauli operators, which form an orthogonal basis of the space of operators for the Hilbert–Schmidt inner product: see [Pauli matrices](https://en.wikipedia.org/wiki/Pauli_matrices) ("Algebraic properties") for one qubit, and J. Watrous [7], Section 4.1.2, Eq. (4.78) (equation numbers as in the author's online version), where it is stated for the discrete Weyl operators (for a qubit these are the Pauli matrices up to phases, Example 4.12). 

## The spectrum of a witness

In order to show the main result, we need to first show a few theorems. The first one is the following: 

>**Theorem (spectrum of a witness).** Let $A_1=(\mathbb{I}+X+Y+Z)/2$, with eigenvalues $\lambda_\pm$. For all $n\ge1$ and $A\in\Lambda_n$ it holds that $\lambda_-\le\lambda_{\min}(A)$ and $\lambda_{\max}(A)\le\lambda_+$. If $\lambda_{\max}(A)=\lambda_+$ or $\lambda_{\min}(A)=\lambda_-$, then $A$ has eigenvalues $\lambda_+,\lambda_-$ and $d-2$ zeros.

*Proof:* The proof is an induction on $n$. For $n=1$, $A=(\mathbb{I}+r\cdot\sigma)/2$ with $|r_i|\le1$, so the eigenvalues $(1\pm|r|)/2$ lie in $[\lambda_-,\lambda_+]$, and they are $\lambda_\pm$ when a bound is attained. Assume the theorem for $n-1$ and let $A\in\Lambda_n$, $n\ge2$. The upper bound is proved in Section 2.1 and the lower bound in Section 2.2.

### 2.1 The largest eigenvalue

In this subsection $n\ge2$ and $A\in\Lambda_n$ are fixed, and the induction hypothesis is used in the form $\lambda_{\max}(B)\le v$ for every $B\in\Lambda_{n-1}$. Let $\ell=\lambda_{\max}(A)$ with unit eigenvector $\varphi$, let $q=\mathrm{Tr}[A^2]$ and let $\kappa=\max(0,-\lambda_{\min}(A))$. For every Pauli $P\ne\mathbb{I}$ we define the real numbers
$$
x_P=\langle\varphi|P|\varphi\rangle,\qquad \alpha_P=\langle\varphi|PAP|\varphi\rangle,\qquad u_P=v\,a_P-(2\ell-v)\,x_P,\qquad \sigma_P=\mathrm{sign}(u_P),
$$
and we also define
$$
R'=A+\kappa\,\mathbb{I}-(\ell+\kappa)\,|\varphi\rangle\langle\varphi|,\qquad e_P=\langle\varphi|P\,R'\,P|\varphi\rangle,\qquad \omega=v-\ell+\kappa .
$$
The operator $R'$ is positive semidefinite: $A+\kappa\mathbb{I}\ge0$, and $R'$ is this operator with its eigenvalue $\ell+\kappa$ at $\varphi$ removed. Therefore $e_P=\alpha_P+\kappa-(\ell+\kappa)x_P^2\ge0$ for every $P$. We also note that $|a_P|\le1$, because by the block lemma $a_P=\mathrm{Tr}[B_+]-\mathrm{Tr}[B_-]$ is a difference of two non-negative numbers whose sum is $\mathrm{Tr}[A]=1$.

>**Lemma (per-Pauli inequality).** For every Pauli $P\ne\mathbb{I}$,
>$$
>0\ \le\ e_P\ \le\ \omega\,(1-x_P^2)-|u_P|\,(1-\sigma_Px_P).
>$$
>Consequently, $\omega\ge0$ and $|u_P|\le\omega\,(1+\sigma_Px_P)$ for every $P\ne\mathbb{I}$.

*Proof:* By the block lemma and the induction hypothesis, $B_\pm\le v\,\mathrm{Tr}[B_\pm]\,\Pi_\pm$. We take the expectation value in $\varphi$ and use $A\varphi=\ell\varphi$, $\mathrm{Tr}[B_\pm]=(1\pm a_P)/2$ and $\langle\varphi|\Pi_\pm|\varphi\rangle=(1\pm x_P)/2$. This gives
$$
\ell+\alpha_P\pm2\ell x_P\ \le\ v\,(1\pm a_P)(1\pm x_P).
$$
Now insert $\alpha_P=e_P-\kappa+(\ell+\kappa)x_P^2$. The left-hand side becomes $\ell(1\pm x_P)^2-\kappa(1-x_P^2)+e_P$, and collecting terms we find
$$
e_P\ \le\ (1\pm x_P)\,(\omega\pm\zeta_P),\qquad \zeta_P:=v\,a_P-(\ell+\kappa)\,x_P .
$$
Since $u_P=\zeta_P+\omega x_P$, the two bounds are $\omega(1-x_P^2)+(1+x_P)\,u_P$ and $\omega(1-x_P^2)-(1-x_P)\,u_P$. The smaller of the two is $\omega(1-x_P^2)-|u_P|(1-\sigma_Px_P)$, which is the claimed inequality.

For the consequences, note that there is a Pauli $P$ with $|x_P|<1$ (because $\sum_Px_P^2=d-1<d^2-1$ by the rule for Pauli sums), and for this $P$ the inequality gives $0\le\omega(1-x_P^2)$, so $\omega\ge0$. If $\sigma_Px_P\ne1$, we divide $|u_P|(1-\sigma_Px_P)\le\omega(1-x_P^2)$ by $1-\sigma_Px_P$ and obtain $|u_P|\le\omega(1+\sigma_Px_P)$. If $\sigma_Px_P=1$, the claim reads $|u_P|\le2\omega$, and it holds because $|a_P|\le1$ gives $|u_P|=\sigma_Pv\,a_P-(2\ell-v)\le2(v-\ell)\le2\omega$. $\square$

>**Theorem (star inequality).** Let $n\ge2$ and $A\in\Lambda_n$, with largest eigenvalue $\ell$, purity $q$ and $\kappa=\max(0,-\lambda_{\min}(A))$. Then
>$$
>v^2(dq-1)-2v(2\ell-v)(d\ell-1)+(2\ell-v)^2(d-1)\ \le\ (v-\ell+\kappa)\,G,
>$$
>where
>$$
>G=(v-\ell)\,d(d-1)+d(\ell-1)+2v(d\ell-1)-2(2\ell-v)(d-1).
>$$
>Equality holds for $A=A^\star$.

At this point of the argument the star inequality is proved under the induction hypothesis for $n-1$. Once the theorem on the spectrum of a witness is established, it holds for every $n\ge2$ without any assumption, and this is how it is used in the second part.

*Proof:* The rule for Pauli sums gives the four identities (all sums run over $P\ne\mathbb{I}$, as before)
$$
\sum_P x_P^2=d-1,\qquad \sum_P a_Px_P=d\ell-1,\qquad \sum_P a_P^2=dq-1,\qquad \sum_P\alpha_P=d-\ell .
$$
Therefore $\sum_Pe_P=d(1-\ell)+\kappa\,d(d-1)$, and summing the per-Pauli inequality over $P$ yields
$$
\sum_P|u_P|\,(1-\sigma_Px_P)\ \le\ \omega\,d(d-1)-\sum_Pe_P=(v-\ell)\,d(d-1)+d(\ell-1).
$$
Next we multiply $|u_P|\le\omega(1+\sigma_Px_P)$ by $|u_P|$ and sum over $P$. Since $|u_P|\,\sigma_P=u_P$ and $\omega\ge0$, we get
$$
\sum_Pu_P^2\ \le\ \omega\Big(\sum_P|u_P|\,(1-\sigma_Px_P)+2\sum_Pu_Px_P\Big)\ \le\ \omega\,G ,
$$
where we used $\sum_Pu_Px_P=v(d\ell-1)-(2\ell-v)(d-1)$. Finally, the four identities give
$$
\sum_Pu_P^2=v^2(dq-1)-2v(2\ell-v)(d\ell-1)+(2\ell-v)^2(d-1),
$$
which is the left-hand side of the star inequality. For $A^\star$ one has $\ell=v$, $\kappa=c$ and $q=2$, and both sides are equal to $d/2$. $\square$

*Proof of the upper bound $\lambda_{\max}(A)\le\lambda_+$:* Suppose that $\ell=v+\varepsilon$ with $\varepsilon\ge0$. Then $\omega=\kappa-\varepsilon\ge0$. A direct expansion, using $2v^2=2v+1$, gives
$$
G=dv-\varepsilon\,\big(d(d+2-2v)-4\big)\ \le\ dv
$$
and
$$
\sum_Pu_P^2=d\Big[v^2(q-\ell^2)+\big(\tfrac12-(2-v)\,\varepsilon\big)^2\Big]-4\varepsilon^2 .
$$
Since $\omega\ge0$ and $G\le dv$, the star inequality implies $\sum_Pu_P^2\le(\kappa-\varepsilon)\,dv$. Dividing by $d$ and rearranging, this reads
$$
v^2\,(q-\ell^2-\kappa^2)+\big(v\kappa-\tfrac12\big)^2+(\sqrt3-1)\,\varepsilon+\Big((2-v)^2-\frac4d\Big)\varepsilon^2\ \le\ 0 .
$$
The first term is non-negative because $q\ge\ell^2+\kappa^2$, and the second term is a square. If $\varepsilon>0$, the last two terms are positive. Indeed, the coefficient of $\varepsilon^2$ is positive for $d\ge16$. For $d=4$ and $d=8$ it equals $-0.60$ and $-0.10$, but $\varepsilon\le\sqrt d-v$ (because $\ell^2\le q\le d$, the last inequality following from $|a_P|\le1$), that is, $\varepsilon<0.64$ and $\varepsilon<1.47$ respectively, and then $(\sqrt3-1)+\big((2-v)^2-\tfrac4d\big)\varepsilon>0$. This contradiction shows that $\varepsilon=0$, i.e. $\ell\le\lambda_+$. If $\ell=\lambda_+$, the displayed inequality forces $v\kappa=\tfrac12$, that is $\kappa=c=-\lambda_-$, and $q=\ell^2+\kappa^2$, so that all the other eigenvalues of $A$ vanish. This proves the upper bound and its equality case.

### 2.2 The smallest eigenvalue

The argument is the mirror image of the previous one, and the induction hypothesis is now used in the form $\lambda_{\min}(B)\ge-c$ for every $B\in\Lambda_{n-1}$. If $\lambda_{\min}(A)\ge0$ there is nothing to prove. Otherwise let $-\kappa=\lambda_{\min}(A)<0$ with unit eigenvector $\chi$, and let $\ell=\lambda_{\max}(A)$. For every Pauli $P\ne\mathbb{I}$ we define
$$
y_P=\langle\chi|P|\chi\rangle,\qquad \alpha'_P=\langle\chi|PAP|\chi\rangle,\qquad u'_P=c\,a_P-(2\kappa-c)\,y_P,\qquad \sigma'_P=\mathrm{sign}(u'_P),
$$
and
$$
R''=\ell\,\mathbb{I}-A-(\ell+\kappa)\,|\chi\rangle\langle\chi|,\qquad f_P=\langle\chi|P\,R''\,P|\chi\rangle,\qquad \omega'=c-\kappa+\ell .
$$
The operator $R''$ is positive semidefinite, since $\ell\,\mathbb{I}-A\ge0$ and $R''$ is this operator with its eigenvalue $\ell+\kappa$ at $\chi$ removed. Therefore $f_P=\ell-\alpha'_P-(\ell+\kappa)y_P^2\ge0$.

>**Lemma (mirror inequalities).** For every Pauli $P\ne\mathbb{I}$,
>$$
>0\ \le\ f_P\ \le\ \omega'\,(1-y_P^2)-|u'_P|\,(1-\sigma'_Py_P),
>$$
>and, as a consequence,
>$$
>\sum_Pu_P'^{\,2}\ \le\ (c-\kappa+\ell)\,G',\qquad G'=(c-\kappa)\,d(d-1)+d(1+\kappa)-2c(d\kappa+1)-2(2\kappa-c)(d-1).
>$$

*Proof:* By the block lemma and the induction hypothesis, $B_\pm\ge-c\,\mathrm{Tr}[B_\pm]\,\Pi_\pm$. Taking the expectation value in $\chi$ and using $A\chi=-\kappa\chi$ gives
$$
-\kappa+\alpha'_P\mp2\kappa y_P\ \ge\ -c\,(1\pm a_P)(1\pm y_P).
$$
Inserting $\alpha'_P=\ell-(\ell+\kappa)y_P^2-f_P$ and collecting terms, we find
$$
f_P\ \le\ (1\pm y_P)\,(\omega'\pm\zeta'_P),\qquad \zeta'_P:=c\,a_P-(\kappa+\ell)\,y_P .
$$
Since $u'_P=\zeta'_P+\omega'y_P$, the smaller of the two bounds is $\omega'(1-y_P^2)-|u'_P|(1-\sigma'_Py_P)$, exactly as in the per-Pauli inequality. In the same way we obtain $\omega'\ge0$ and $|u'_P|\le\omega'(1+\sigma'_Py_P)$ for every $P$ (if $\sigma'_Py_P=1$, then $|u'_P|=\sigma'_Pc\,a_P-(2\kappa-c)\le2(c-\kappa)\le2\omega'$).

The rule for Pauli sums now gives $\sum_Py_P^2=d-1$, $\sum_Pa_Py_P=-d\kappa-1$, $\sum_Pa_P^2=dq-1$ and $\sum_P\alpha'_P=d+\kappa$. Hence $\sum_Pf_P=\ell\,d(d-1)-d(1+\kappa)$, and summing the first inequality over $P$ yields
$$
\sum_P|u'_P|\,(1-\sigma'_Py_P)\ \le\ (c-\kappa)\,d(d-1)+d(1+\kappa).
$$
Multiplying $|u'_P|\le\omega'(1+\sigma'_Py_P)$ by $|u'_P|$, summing over $P$, and using $\sum_Pu'_Py_P=-c(d\kappa+1)-(2\kappa-c)(d-1)$, we arrive at $\sum_Pu_P'^{\,2}\le\omega'\,G'$. $\square$

*Proof of the lower bound $\lambda_{\min}(A)\ge\lambda_-$:* Suppose that $\kappa=c+\varepsilon$ with $\varepsilon\ge0$. A direct expansion, using $c^2+c=vc=\tfrac12$, gives
$$
G'=dc-\varepsilon\,\big(d(d+2+2c)-4\big)\ \le\ dc
$$
and
$$
\sum_Pu_P'^{\,2}=d\Big[c^2(q-\kappa^2)+\big(\tfrac12+(2+c)\,\varepsilon\big)^2\Big]-4\varepsilon^2 .
$$
Since $\omega'=\ell-\varepsilon\ge0$ and $G'\le dc$, the mirror inequalities imply $\sum_Pu_P'^{\,2}\le(\ell-\varepsilon)\,dc$. Dividing by $d$ and rearranging, this reads
$$
c^2\,(q-\ell^2-\kappa^2)+\big(c\,\ell-\tfrac12\big)^2+(1+\sqrt3)\,\varepsilon+\Big((2+c)^2-\frac4d\Big)\varepsilon^2\ \le\ 0 .
$$
The first term is non-negative because $q\ge\ell^2+\kappa^2$, the second is a square, and the coefficients of $\varepsilon$ and $\varepsilon^2$ are positive. Hence $\varepsilon=0$, i.e. $\lambda_{\min}(A)\ge\lambda_-$. If $\lambda_{\min}(A)=\lambda_-$, the displayed inequality forces $c\,\ell=\tfrac12$, that is $\ell=v=\lambda_+$, and $q=\ell^2+\kappa^2$, so that all the other eigenvalues of $A$ vanish. This completes the induction and the proof of the theorem on the spectrum of a witness. $\square$

The theorem has a direct translation into a statement about states.

>**Corollary (dual form of the theorem on the spectrum).** For every $n$-qubit state $\rho$,
>$$
>\frac{\lambda_+\mathbb{I}-\rho}{\lambda_+d-1}\in\mathrm{STAB}_n
>$$
>and
>$$
>p\,\rho+(1-p)\,\frac{\mathbb{I}}{d}\in\mathrm{STAB}_n\qquad\text{whenever}\qquad p\le\frac{1}{1+(\sqrt3-1)\,2^{\,n-1}} .
>$$
>Both constants are optimal for every $n$.

*Proof:* A state $\sigma$ belongs to $\mathrm{STAB}_n$ if and only if $\mathrm{Tr}[A\sigma]\ge0$ for all $A\in\Lambda_n$ [4]. By the theorem, $\lambda_-\le\mathrm{Tr}[A\rho]\le\lambda_+$ for every state $\rho$ and every $A\in\Lambda_n$. Hence $\mathrm{Tr}[A(\lambda_+\mathbb{I}-\rho)]=\lambda_+-\mathrm{Tr}[A\rho]\ge0$, and $\mathrm{Tr}[A(p\rho+(1-p)\mathbb{I}/d)]\ge-p\,c+(1-p)/d$, which is non-negative exactly when $p\le1/(1+cd)$, and $c\,d=(\sqrt3-1)\,2^{n-1}$. For the optimality, take $A=A^\star$ and let $\rho$ be its eigenvector with eigenvalue $\lambda_+$ in the first statement and with eigenvalue $\lambda_-$ in the second. $\square$

### 2.3 The star inequality near $A^\star$

For the application to purity in the second part, it is convenient to rewrite the star inequality in variables that measure the distance of a witness from $A^\star$. For $A\in\Lambda_n$ with a negative eigenvalue we write
$$
\ell=\lambda_{\max}(A)=v-\delta,\qquad \kappa=-\lambda_{\min}(A)=c-k,\qquad q=\mathrm{Tr}[A^2]=2+\theta .
$$
By the theorem on the spectrum of a witness, $\delta\ge0$ and $k\ge0$, and for $A^\star$ we have $\delta=k=\theta=0$.

>**Lemma (the star inequality near $A^\star$).** Let $n\ge2$ and let $A\in\Lambda_n$ have a negative eigenvalue. Then
>$$
>d\,v^2\,\theta+d\,v\,k\ \le\ \delta\,\big[(c-k)\,D-d\,(v+3)\big]+\delta^2\,\big[D+4d(v-1)+4\big],
>$$
>where $D:=d(d+2-2v)-4$.

*Proof:* We insert $\ell=v-\delta$, $\kappa=c-k$ and $q=2+\theta$ in the star inequality. Using $2v^2=2v+1$, $vc=\tfrac12$ and $v^2+c^2=2$, one finds $G=dv+\delta D$ and
$$
v^2(dq-1)-2v(2\ell-v)(d\ell-1)+(2\ell-v)^2(d-1)=d\Big[\tfrac12+v^2\theta+(2v+3)\,\delta-4(v-1)\,\delta^2\Big]-4\delta^2 ,
$$
while
$$
(v-\ell+\kappa)\,G=(c-k+\delta)(dv+\delta D)=\tfrac d2-d\,v\,k+d\,v\,\delta+(c-k)\,D\,\delta+D\,\delta^2 .
$$
Subtracting the two expressions gives the claim. $\square$

The second part combines this lemma with the Liu–Liu inequality of Ref. [3] to bound the purity $\mathrm{Tr}[A^2]$ of every witness, which is what the theorem stated at the beginning requires.

**Remark.** All the inequalities used in the proofs above were also tested numerically on elements of $\Lambda_n$ for $n=2,3,4$. The argument is AI-generated and has not been peer reviewed.

### 15. Partial result: An improved computer-assisted bound at 0.4996 for all $n$, and a sharp 0.5 value for $n\leq 5$ (part II)

- **Date:** 19 Nov 2026
- **Credit:** Rafael Wagner
- **AI use:** partially AI-assisted
- **Permanent link:** https://open-context.jan-bulling.com/contribution/5.12
- **Votes:** 0 (0 up, 0 down)

## Main result

After a couple of prompts, Claude Fable 5.2 found another advance:

>**Theorem ($0.4996$ bound):** Let $\rho \in \mathcal{D}(\underbrace{\mathbb{C}^2 \otimes \cdots \otimes \mathbb{C}^2}_{n\text{ times}})$ be an arbitrary $n$-qubit state, and denote $\mathrm{STAB}_n$ the convex polytope of $n$-qubit stabilizer states. For every integer $n \geq 1$, it holds that 
>$$
>\mathrm{Tr}[\rho^2] \leq \frac{1}{d_n-a^*} \implies \rho \in \mathrm{STAB}_n,
>$$
>where $d_n = 2^n$ is the dimension of the underlying Hilbert space and $a^*=0.4996$. Furthermore, for $1\leq n \leq 5$ the conjecture is shown to hold, i.e., $a^*=\tfrac{1}{2}$.  

The results have used the strategy present in the notes by Y. Wang. ["Partial Results on the Hilbert–Schmidt Inradius of the Multiqubit Stabilizer Polytope"](https://open-context.jan-bulling.com/contribution/5.6), in collaboration with ChatGPT 6 Astra. 

**How this contribution is organised.** This is the second of two parts. The first part proves two results on the spectrum of the operators in the $\Lambda$-polytope, which are recalled below without proof. Here they are combined with the Liu–Liu inequality of Ref. [3] to bound the purity of these operators, which proves the theorem above. The numbering of the sections continues that of the first part.

**What "[computed]" means.** Two steps of the argument are *computer-assisted* and are marked **[computed]** below: the case $n=5$, and the constant $0.4996$ for all $n$. In these steps the proof is carried out analytically until the claim is reduced to an explicit inequality between elementary functions of two real variables, and this last inequality is then checked by evaluating the functions numerically on a computer (floating-point arithmetic on a fine grid). This is strong numerical evidence rather than a fully rigorous proof, which would require error-controlled (interval) arithmetic. Everything that is not marked **[computed]** is proved analytically, in particular the conjecture for $n\le4$ and the weaker constant $a^*=0.4926$ for all $n$. All the inequalities used in the two parts were also tested numerically on elements of $\Lambda_n$ for $n=2,3,4$. The argument is AI-generated and has not been peer reviewed.

## Preliminary mathematical ideas

Recall that a crucial ingredient in this discussion is the notion of the $\Lambda$-polytope from Ref. [4]. For more on $\Lambda$-polytopes we refer to Refs. [8, 9, 10, 11, 4]. It can be defined as follows:
$$
\Lambda_n=\left \{A=A^\dagger:\ \mathrm{Tr}[A]=1,\ \langle s|A|s\rangle\ge 0\ \text{for all stabilizer states } \vert s\rangle  \right\}.
$$ 
Above the relevant operators are always defined relative to $n$-qubit Hilbert spaces $\mathbb{C}^2 \otimes \cdots \otimes \mathbb{C}^2$ of dimension $d_n = 2^n$. Let $A_1=(\mathbb{I}+X+Y+Z)/2$, with eigenvalues $\lambda_\pm=(1\pm\sqrt3)/2$, and $A^\star=A_1\otimes|0\dots0\rangle\langle0\dots0|$. Throughout this document we define 
$$
v:=\lambda_+,\qquad c:=-\lambda_-=v-1.
$$
With these definitions, it is simple to show that the following equations all hold:
$$
\qquad 2v^2=2v+1,\quad vc=\tfrac12,\quad v^2+c^2=2,\quad v^3-c^3=\frac52.
$$

Finally, we recall a Lemma that can be found in Ref. [3].  

>**Lemma (Liu–Liu inequality).** For an arbitrary $A\in\Lambda_n$, define $q=\mathrm{Tr}[A^2]$, $M_n=\max\{\mathrm{Tr}[A^2]:A\in\Lambda_n\}$, and $t=\mathrm{Tr}[A^3]$. Then, 
>$$
>dq+1-2t\ \le\ M_{n-1}\,(d-q).
>$$

Note that in various of these advances we have been using that we are interested in asking when 
$$
\mathrm{Tr}[\rho^2]\le \frac{1}{d-\frac{1}{M_n}}
$$
is sufficient as a condition for $\rho$ to be a stabilizer mixture and the conjecture is $M_n=2$ for all $n\in \mathbb{N}$.

## Results from the first part

The first part proves the following two statements, which we use without repeating their proofs.

>**Theorem (spectrum of a witness).** Let $A_1=(\mathbb{I}+X+Y+Z)/2$, with eigenvalues $\lambda_\pm$. For all $n\ge1$ and $A\in\Lambda_n$ it holds that $\lambda_-\le\lambda_{\min}(A)$ and $\lambda_{\max}(A)\le\lambda_+$. If $\lambda_{\max}(A)=\lambda_+$ or $\lambda_{\min}(A)=\lambda_-$, then $A$ has eigenvalues $\lambda_+,\lambda_-$ and $d-2$ zeros.

For a witness $A\in\Lambda_n$ we write
$$
\ell=\lambda_{\max}(A)=v-\delta,\qquad \kappa=-\lambda_{\min}(A)=c-k,\qquad q=\mathrm{Tr}[A^2]=2+\theta,\qquad t=\mathrm{Tr}[A^3].
$$
By the theorem, $\delta\ge0$ and $k\ge0$, and for $A^\star$ we have $\delta=k=\theta=0$.

>**Lemma (the star inequality near $A^\star$).** Let $n\ge2$ and let $A\in\Lambda_n$ have a negative eigenvalue. Then
>$$
>d\,v^2\,\theta+d\,v\,k\ \le\ \delta\,\big[(c-k)\,D-d\,(v+3)\big]+\delta^2\,\big[D+4d(v-1)+4\big],
>$$
>where $D:=d(d+2-2v)-4$.

## Purity

We now turn the results of the previous section into bounds on $M_n=\max\{\mathrm{Tr}[A^2]:A\in\Lambda_n\}$. Let us first recall why this is the quantity we need. For a state $\rho$ and $A\in\Lambda_n$, the Cauchy–Schwarz inequality gives
$$
\mathrm{Tr}[A\rho]=\frac1d+\mathrm{Tr}\Big[\Big(A-\frac{\mathbb{I}}{d}\Big)\Big(\rho-\frac{\mathbb{I}}{d}\Big)\Big]\ \ge\ \frac1d-\sqrt{\Big(\mathrm{Tr}[A^2]-\frac1d\Big)\Big(\mathrm{Tr}[\rho^2]-\frac1d\Big)} ,
$$
and the right-hand side is non-negative for every $A\in\Lambda_n$ as soon as $\mathrm{Tr}[\rho^2]\le1/(d-1/M_n)$. Since $\rho\in\mathrm{STAB}_n$ if and only if $\mathrm{Tr}[A\rho]\ge0$ for all $A\in\Lambda_n$, we conclude that **every upper bound $M_n\le m$ proves the main theorem with $a^*=1/m$**. In particular, $M_n=2$ gives the conjectured constant $a^*=\tfrac12$. On the other hand, $M_n\ge2$ for every $n$, because $A^\star\in\Lambda_n$ and $\mathrm{Tr}[(A^\star)^2]=2$.

### 3.1 Three bounds on the third moment

The following lemma collects three elementary upper bounds on the third moment $t$ in terms of the extreme eigenvalues and the purity. The third one uses Wilkins' inequality for the skewness of a finite set of real numbers [13].

>**Lemma (three bounds on the third moment).** Let $n\ge2$ and $A\in\Lambda_n$. Then
>$$
>t\ \le\ \ell\,q .
>$$
>If, moreover, $A$ has a negative eigenvalue, put
>$$
>s_1=1-\ell+\kappa,\qquad s_2=q-\ell^2-\kappa^2,\qquad S=s_2-\frac{s_1^2}{d-2}.
>$$
>Then $S\ge0$ and
>$$
>t\ \le\ \ell^3-\kappa^3+s_2^{3/2},\qquad t\ \le\ \ell^3-\kappa^3+W,
>$$
>where
>$$
>W=\frac{s_1^3}{(d-2)^2}+\frac{3\,s_1}{d-2}\,S+\frac{d-4}{\sqrt{(d-2)(d-3)}}\,S^{3/2}.
>$$

*Proof:* For the first bound, every eigenvalue $\lambda$ of $A$ satisfies $\lambda^3\le\ell\,\lambda^2$, and we sum over the eigenvalues. For the other two, let $\mu_1,\dots,\mu_{d-2}$ be the eigenvalues of $A$ other than $\ell$ and $-\kappa$. Since $\mathrm{Tr}[A]=1$ and $\mathrm{Tr}[A^2]=q$, we have $\sum_i\mu_i=s_1$ and $\sum_i\mu_i^2=s_2$, so that $t=\ell^3-\kappa^3+\sum_i\mu_i^3$ and $S=\sum_i(\mu_i-\bar\mu)^2\ge0$ with $\bar\mu=s_1/(d-2)$. The second bound follows from $\sum_i\mu_i^3\le\max_i|\mu_i|\,\sum_i\mu_i^2\le s_2^{3/2}$. For the third, Wilkins' inequality states that $N$ real numbers with mean $\bar\mu$ satisfy $\sum_i(\mu_i-\bar\mu)^3\le\frac{N-2}{\sqrt{N(N-1)}}\,S^{3/2}$; with $N=d-2$ and $\sum_i\mu_i^3=N\bar\mu^3+3\bar\mu\,S+\sum_i(\mu_i-\bar\mu)^3$ this is the bound with $W$. $\square$

### 3.2 The conjecture for $n\le5$

>**Theorem (the conjecture for $n\le5$).** For $1\le n\le5$ we have $M_n=2$. Consequently, for these $n$ every state with $\mathrm{Tr}[\rho^2]\le1/(d-\tfrac12)$ belongs to $\mathrm{STAB}_n$, and the constant $\tfrac12$ cannot be improved [1]. The cases $n\le4$ are proved analytically; the case $n=5$ is computer-assisted.

*Proof:* For $n=1$ we have $\mathrm{Tr}[A^2]=(1+|r|^2)/2\le2$. Let now $2\le n\le5$, assume that $M_{n-1}=2$, and suppose that some $A\in\Lambda_n$ has purity $q=2+\theta$ with $\theta>0$. We show that this is impossible. With $M_{n-1}=2$ the Liu–Liu inequality reads
$$
t\ \ge\ \frac52+\frac{d+2}{2}\,\theta .
$$
We combine it with the first two bounds on the third moment.

**(a)** From $t\le\ell q$ we get $\theta\,(d+2-2\ell)\le4\ell-5$. Hence $\ell>\tfrac54$, that is $\delta<v-\tfrac54<0.1161$, and $\theta\le(4v-5)/(d+2-2v)$. In particular $\ell>1=\mathrm{Tr}[A]$, so $A$ has a negative eigenvalue, and both the lemma of Section 3.1 and the star inequality near $A^\star$ apply.

**(b)** From $t\le\ell^3-\kappa^3+s_2^{3/2}$ and $v^3-c^3=\tfrac52$ we get
$$
\frac{d+2}{2}\,\theta\ \le\ -(v^3-\ell^3)+(c^3-\kappa^3)+s_2^{3/2},\qquad s_2=\theta+(v^2-\ell^2)+(c^2-\kappa^2).
$$

**The cases $n=2$ and $n=3$.** Since $k\ge0$ and $\theta>0$, the star inequality near $A^\star$ requires $\delta\,[c\,D-d(v+3)]+\delta^2\,[D+4d(v-1)+4]>0$. For $d=4$ the two brackets are $-14.14$ and $18.93$, so $\delta>0.74$. For $d=8$ they are $-15.11$ and $69.86$, so $\delta>0.21$. Both contradict (a). Hence $M_2=2$ (from $M_1=2$), and then $M_3=2$.

**The case $n=4$.** Here $D=240.29$, $c\,D-d(v+3)=18.09$ and $D+4d(v-1)+4=267.71$. By (a), $267.71\,\delta^2<31.1\,\delta$, so the star inequality near $A^\star$ (dropping the term $-kD\delta\le0$) gives $29.86\,\theta+21.86\,k<49.2\,\delta$, that is,
$$
\delta\ >\ 0.606\,\theta+0.444\,k .
$$
In (b) we use $v^3-\ell^3\ge3\ell^2\delta\ge4.6875\,\delta$, $c^3-\kappa^3\le3c^2k\le0.402\,k$ and $s_2\le\theta+2v\delta+2ck\le0.616$ (by (a), $\theta\le0.0304$ and $\delta<0.1161$), so that $s_2^{3/2}\le0.785\,(\theta+2.733\,\delta+0.733\,k)$. With $\tfrac{d+2}{2}=9$ this gives $2.54\,\delta\le0.978\,k-8.21\,\theta$, that is,
$$
\delta\ \le\ 0.386\,k-3.2\,\theta .
$$
The two bounds on $\delta$ are incompatible for $\theta>0$ and $k\ge0$. Hence $M_4=2$.

**The case $n=5$** **[computed]**. For $d=32$ the crude estimates of the previous case are not sufficient, and we use the third bound on the third moment instead of the second. For a fixed pair $(\ell,\kappa)$ we have three conditions on $\theta$. The star inequality near $A^\star$ reads $\theta\le\theta_T(\ell,\kappa)$, where
$$
\theta_T(\ell,\kappa):=\frac{\delta\,\big[(c-k)\,D-d\,(v+3)\big]+\delta^2\,\big[D+4d(v-1)+4\big]-d\,v\,k}{d\,v^2},\qquad \delta=v-\ell,\quad k=c-\kappa .
$$
Item (a) reads $\theta\le\theta_{\rm lin}(\ell):=(4\ell-5)/(d+2-2\ell)$. The Liu–Liu inequality with the bound $t\le\ell^3-\kappa^3+W$ reads $h(\theta)\ge0$, where
$$
h(\theta):=2\,\big(\ell^3-\kappa^3+W\big)-5-(d+2)\,\theta
$$
is a convex function of $\theta$ (because $S$ is linear in $\theta$). Therefore $\theta>0$ is possible only at pairs $(\ell,\kappa)$ with $\theta_T>0$, $\theta_{\rm lin}>0$ and
$$
\max\big\{h(0),\ h\big(\min(\theta_T,\theta_{\rm lin})\big)\big\}\ \ge\ 0 .
$$
We evaluated the left-hand side on a grid of $3001\times3001$ points of the box $\tfrac54\le\ell\le v$, $0\le\kappa\le c$, refined towards $A^\star$. Wherever $\theta_T>0$ and $\theta_{\rm lin}>0$, it is at most $-0.97\,(\delta+k)$ for $d=32$ (and at most $-3.1\,(\delta+k)$ for $d=16$, in agreement with the analytical proof). Hence no witness in $\Lambda_5$ has purity above $2$, and $M_5=2$. This step is a floating-point evaluation of an explicit function with a uniform margin; it is not an interval-arithmetic proof. $\square$

### 3.3 The bound for all $n$

For $n\ge6$ the argument of Section 3.2 stops working: the conditions used there are no longer contradictory, so we cannot conclude that $M_n=2$, and we cannot use $M_{n-1}=2$ as an input for the next level. The way out is to carry an upper bound from one level to the next instead of an exact value. The Liu–Liu inequality only needs an upper bound on $M_{n-1}$: if $M_{n-1}\le m$, then, because $d-q\ge0$,
$$
dq+1-2t\ \le\ m\,(d-q)\qquad\text{for every }A\in\Lambda_n .
$$
So from a number $m_{n-1}\ge M_{n-1}$ we can compute a number $m_n\ge M_n$, from it a number $m_{n+1}\ge M_{n+1}$, and so on. Each step can increase the bound only by an amount of order $1/d$, and $\sum_n2^{-n}$ converges, so the bounds stay close to $2$ for all $n$. We first make this precise in an elementary form, and then in the sharper form that gives the constant $0.4996$.

>**Lemma (one level of the recursion, elementary form).** Let $n\ge2$ and let $m\ge2$ be any number with $M_{n-1}\le m$. Then
>$$
>M_n\ \le\ g_d(m):=\frac{m\,d-1}{d+m-2\lambda_+},
>$$
>and
>$$
>g_d(m)-m=\frac{m\,(2\lambda_+-m)-1}{d+m-2\lambda_+}\ \le\ \frac{0.4642}{d-0.74}.
>$$

*Proof:* Let $A\in\Lambda_n$. By the first bound on the third moment and the theorem on the spectrum, $t\le\ell\,q\le\lambda_+q$. Inserting this in the Liu–Liu inequality with $M_{n-1}\le m$ gives $dq+1-2\lambda_+q\le m\,(d-q)$, that is, $q\,(d+m-2\lambda_+)\le m\,d-1$. Maximising over $A$ proves the first claim. For the second, the numerator $m(2\lambda_+-m)-1$ is decreasing in $m$ for $m\ge\lambda_+$ and equals $2\sqrt3-3<0.4642$ at $m=2$, while the denominator is at least $d+2-2\lambda_+>d-0.74$. $\square$

>**Corollary (a bound for all $n$ without computer assistance).** For every $n\ge1$ we have $M_n<2.03$. Hence the main theorem holds with $a^*=0.4926$ by an argument that uses no numerical computation.

*Proof:* By the cases $n\le4$ of the theorem in Section 3.2, which are proved analytically, $M_n=2$ for $n\le4$. For $n\ge5$ we apply the lemma repeatedly, starting from $M_4=2$:
$$
M_n\ \le\ 2+0.4642\sum_{j=5}^{n}\frac{1}{2^j-0.74}\ <\ 2+0.4642\cdot\frac{32}{31.26}\cdot\sum_{j\ge5}2^{-j}\ <\ 2.03 ,
$$
and $1/2.03>0.4926$. $\square$

If one also uses $M_5=2$, the same estimate starting from $j=6$ gives $M_n<2.0147$ and $a^*=0.4963$. To reach $0.4996$ we replace the elementary step by a sharper one, which uses the star inequality and the Wilkins bound.

>**Theorem (the bound for all $n$)** **[computed]**. For every $n\ge1$ we have $M_n\le2.0015$. Hence the main theorem holds with $a^*=1/2.0015>0.4996$.

*Proof:* **Step 1 (conditions that a witness with large purity must satisfy).** Let $n\ge2$, let $m\ge2$ be a number with $M_{n-1}\le m$, and suppose that $A\in\Lambda_n$ has purity $q>m$. The Liu–Liu inequality gives $2t\ge dq+1-m(d-q)$. We combine it with the star inequality near $A^\star$ and with the first and the third bound on the third moment. This gives three conditions on the triple $(\ell,\kappa,q)$:

**(i)** the star inequality near $A^\star$, with $\theta_T$ as in Section 3.2:
$$
q\ \le\ q_T(\ell,\kappa):=2+\theta_T(\ell,\kappa);
$$
**(ii)** the Liu–Liu inequality with $t\le\ell\,q$:
$$
q\ \le\ q_L(\ell):=\frac{m\,d-1}{d+m-2\ell};
$$
**(iii)** the Liu–Liu inequality with $t\le\ell^3-\kappa^3+W$:
$$
h(q):=2\,\big(\ell^3-\kappa^3+W\big)-\big(dq+1-m\,(d-q)\big)\ \ge\ 0 ,
$$
where $W$ depends on $q$ through $S=q-\ell^2-\kappa^2-(1-\ell+\kappa)^2/(d-2)$.

Moreover, (ii) and $q>m$ force $\ell>(m^2+1)/(2m)\ge\tfrac54$. So $A$ has a negative eigenvalue, and by the theorem on the spectrum $\tfrac54<\ell\le v$ and $0<\kappa\le c$.

**Step 2 (the function $F_d$).** For $m\ge2$ we say that a number $q>m$ is *admissible* if there is a pair $(\ell,\kappa)$ with $1\le\ell\le v$ and $0\le\kappa\le c$ such that (i), (ii) and (iii) hold. We define
$$
F_d(m):=\sup\{\,q>m:\ q\ \text{is admissible}\,\},
$$
and $F_d(m):=m$ if there is no admissible $q$. By Step 1, the purity of every witness in $\Lambda_n$ is either at most $m$ or admissible, so
$$
M_{n-1}\le m\qquad\Longrightarrow\qquad M_n\le F_d(m),\qquad d=2^n .
$$

**Step 3 (how $F_d(m)$ is evaluated).** Fix a pair $(\ell,\kappa)$. Conditions (i) and (ii) say that $q\le q_U:=\min\{q_T(\ell,\kappa),q_L(\ell)\}$, and $h$ is a convex function of $q$ (on this box and for $q\ge2$ one has $S\ge0$, so $W$ is well defined, and $S$ is linear in $q$). Hence the largest admissible $q$ for this pair is $q_U$ if $h(q_U)\ge0$; it is the unique zero of $h$ in $[m,q_U)$ if $h(q_U)<0\le h(m)$; and there is no admissible $q$ for this pair if $h(m)<0$ and $h(q_U)<0$. The number $F_d(m)$ is the maximum of this quantity over the box $1\le\ell\le v$, $0\le\kappa\le c$. We computed it numerically: a grid of $601\times601$ points of the box, followed by six rounds of local refinement around the maximiser. The values in the table below are stable to about $10^{-8}$ under changes of the grid.

**Step 4 (the recursion).** We put $m_5:=2$, which is the value of $M_5$, and
$$
m_n:=F_{2^n}(m_{n-1})\qquad\text{for }n\ge6 .
$$
By Step 2, $M_n\le m_n$ for every $n\ge5$. For $d\le32$ no purity above $2$ is admissible, in agreement with Section 3.2; the first level at which an admissible purity above the input appears is $d=64$, where $F_{64}(2)=2.00009$. The results are collected in the following table. The bounds on $M_n$ are rounded up, and the resulting constants $a^*=1/M_n$ are rounded down.

| $n$ | $d=2^n$ | bound on $M_n$ | increase from level $n-1$ | resulting constant $a^*$ | how the bound is obtained |
|---|---|---|---|---|---|
| 1 | 2 | 2 | – | 0.5 | direct computation |
| 2 | 4 | 2 | 0 | 0.5 | analytical proof |
| 3 | 8 | 2 | 0 | 0.5 | analytical proof |
| 4 | 16 | 2 | 0 | 0.5 | analytical proof |
| 5 | 32 | 2 | 0 | 0.5 | computed: $F_{32}(2)=2$ |
| 6 | 64 | 2.00009 | $8.98\cdot10^{-5}$ | 0.49997 | computed: $m_6=F_{64}(2)$ |
| 7 | 128 | 2.00052 | $4.29\cdot10^{-4}$ | 0.49987 | computed: $m_7=F_{128}(m_6)$ |
| 8 | 256 | 2.00091 | $3.83\cdot10^{-4}$ | 0.49977 | computed: $m_8=F_{256}(m_7)$ |
| 9 | 512 | 2.00116 | $2.52\cdot10^{-4}$ | 0.49971 | computed |
| 10 | 1024 | 2.00131 | $1.47\cdot10^{-4}$ | 0.49967 | computed |
| 11 | 2048 | 2.00139 | $8.05\cdot10^{-5}$ | 0.49965 | computed |
| 12 | 4096 | 2.00143 | $4.26\cdot10^{-5}$ | 0.49964 | computed |
| 13 to 20 | up to 1048576 | 2.00147 | less than $2.3\cdot10^{-5}$ per level | 0.49963 | computed |
| all $n\ge21$ | – | 2.0015 | less than $5\cdot10^{-7}$ in total | 0.4996 | tail estimate (Step 5) |

**Step 5 (the tail).** The computed value at level $20$ is $m_{20}=2.00146925\ldots<2.0014693$. For $n>20$ we use the elementary lemma, which bounds the increase at level $j$ by $0.4642/(2^j-0.74)$:
$$
M_n\ \le\ m_{20}+0.4642\sum_{j\ge21}\frac{1}{2^j-0.74}\ <\ 2.0014693+5\cdot10^{-7}\ <\ 2.0015 .
$$
Together with the table this proves $M_n\le2.0015$ for every $n$, and $1/2.0015>0.4996$. $\square$

*What the numbers mean.* The bound $2.00009$ at $n=6$ does not say that a witness with this purity exists. It says that conditions (i), (ii) and (iii) no longer exclude a hypothetical witness in $\Lambda_6$ with $\lambda_{\max}\approx1.361$, $\lambda_{\min}\approx-0.311$ and purity $2.00009$. At the next level the Liu–Liu inequality is used with $m_6$ instead of $2$, which leaves a little more room, and so on. The increases soon shrink by a factor of about two per level, and the bounds converge to a number slightly below $2.00147$.

*Why the exact value $2$ is obtained only up to $n=5$.* Near $A^\star$ the star inequality needs $\delta\ge d\,v\,k/(c\,D-d(v+3))$, while item (b) needs $v^2\delta<c^2k$. These two requirements are incompatible exactly as long as $c\,D-d(v+3)<v^3d/c^2$, that is, for $d<64.6$. So the induction step closes for $d\le32$ and fails, by a very small amount, from $d=64$ on. Closing this gap for all $n$ would prove the conjecture.

### 16. Claimed solution: A proposed analytic proof of the sharp bound cₙ = 1/2 for every number of qubits

- **Date:** 20 Nov 2026
- **Credit:** Yujie Zhang
- **AI use:** partially AI-assisted
- **Permanent link:** https://open-context.jan-bulling.com/contribution/5.13
- **Votes:** 1 (1 up, 0 down)

## Main result

**Proposed theorem.** Write $d=2^n$. Let $c_n$ be the largest constant such that every $n$-qubit density operator of purity at most $1/(d-c_n)$ is a convex mixture of stabilizer states. Then

$$
\boxed{c_n=\tfrac12\qquad\text{for every }n\ge1.}
$$

The centered Hilbert–Schmidt inradius is $1/\sqrt{d(2d-1)}$. The proof also determines every equality witness. If a trace-one Hermitian $A$ is nonnegative on all pure stabilizer states and $\operatorname{Tr}(A^2)=2$, then $A$ is Clifford-equivalent to

$$
V_n=|0^{n-1}\rangle\langle0^{n-1}|\otimes\frac{I+X+Y+Z}{2}.
$$

**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](https://open-context.jan-bulling.com/contribution/5.10) · [OpenContext 5.11](https://open-context.jan-bulling.com/contribution/5.11) · [OpenContext 5.12](https://open-context.jan-bulling.com/contribution/5.12)

**Proof strategy.** Suppose an $n$-qubit witness violates the bound. We show that some Pauli measurement outcome leaves a violating witness on $n-1$ 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 $73.8\%$ 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 $\Lambda_n$ for the normalized witnesses above. For a nonidentity Hermitian Pauli $P$, let $\Pi=(I\pm P)/2$, $B=\Pi A\Pi$ and $p=\operatorname{Tr}B$. The outcome has a stabilizer basis, so $p\ge0$. If $p>0$, identifying the code with $n-1$ qubits gives $B/p\in\Lambda_{n-1}$. If $p=0$, 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 $B=0$.

**Boundary lemma.** If $A\in\Lambda_n$, $\operatorname{Tr}(A^2)=2$, and every signed Pauli outcome satisfies $\operatorname{Tr}(B^2)\le2p^2$, then $A$ is Clifford-equivalent to $V_n$.

Here is why the lemma suffices. Interpolate a hypothetical witness with $\operatorname{Tr}(A^2)>2$ toward $I/d$ until its squared norm equals $2$. Every stabilizer expectation of the interpolated witness is strictly positive. Each Clifford image of $V_n$ has a zero stabilizer expectation. The lemma therefore forces some outcome to violate $\|B\|_{\mathrm{HS}}\le\sqrt2\,p$. The same outcome violates the inequality on the original witness, since the inequality defines a convex cone containing the compression of $I/d$. This gives the descent. On one qubit each Bloch coordinate of a witness lies in $[-1,1]$, giving squared norm at most $2$, with equality precisely at the cube vertices. Induction excludes every violation. The lemma then gives the asserted equality classification at every $n$.

## 2. The weight carried by large Pauli expectations

Assume the hypotheses of the boundary lemma. Put $t=\operatorname{Tr}(A^3)$ and $\ell=\lambda_{\max}(A)$. Wagner proves $\ell\le u=(1+\sqrt3)/2$. For $\ell\le1$, summing $\lambda^3\le\ell\lambda^2$ gives $t\le2$. For $\ell>1$, the remaining eigenvalues have squared sum $2-\ell^2$, hence

$$
t\le\ell^3+(2-\ell^2)^{3/2}
\le u^3+(u-1)^3=\frac{3\sqrt3}{2}.
$$

The middle expression increases for $1\le\ell\le\sqrt2$, and $u^2+(u-1)^2=2$. [OpenContext 5.11, §2.1](https://open-context.jan-bulling.com/contribution/5.11)

Label phase-free Paulis by $v\in\mathbb F_2^{2n}$, with $[v,w]=1$ when their Hermitian representatives anticommute. Choose $P_0=I$ and put

$$
a_v=\operatorname{Tr}(AP_v),\quad f_v=a_v^2,\quad k_v=\operatorname{Tr}(AP_vAP_v),\quad z_v=\operatorname{Tr}((2A-A^2)P_v).
$$

The two nonnegative Pauli outcome probabilities give $|a_v|\le1$. Direct expansion with $\Pi_\pm=(I\pm P_v)/2$, $B_\pm=\Pi_\pm A\Pi_\pm$, gives

$$
4(2p_\pm^2-\operatorname{Tr}(B_\pm^2))=2f_v-k_v\pm2z_v.
$$

Thus $2f_v-k_v\ge2|z_v|$, and at $v=0$ both sides vanish. Define $\mu_v=(2f_v-k_v)/(3d)$. Pauli orthogonality gives $\sum_v f_v=2d$ and $k_v=d^{-1}\sum_w(-1)^{[v,w]}f_w$. Hence $\sum_v k_v=d$, so $\mu$ is a probability distribution. Taking the Fourier transform again gives

$$
\Pr_{w\sim\mu}([v,w]=1)=d\mu_v+\frac{1-f_v}{2}.\tag{1}
$$

Let $\mathcal H=\{v\ne0:|a_v|>3/4\}$ and $\beta=\mu(\mathcal H)$. By Pauli orthogonality and the bound on $|z_v|$,

$$
4-t=\frac1d\sum_v a_vz_v
\le\frac32\sum_v\mu_v|a_v|
\le\frac32\left(\frac34+\frac{\beta}{4}\right).
$$

Consequently $\beta\ge(23-8t)/3\ge(23-12\sqrt3)/3>0.738$.

## 3. The anticommutation graph leaves at most two logical qubits

Join two labels in $\mathcal H$ when they anticommute. Three induced graphs on four vertices are impossible. Up to Clifford conjugation their four observables are

$$
\begin{array}{c|c}
\text{two disjoint edges}&XI,ZI,IX,IZ\\
\text{four-vertex path}&XI,ZI,XX,IZ\\
\text{triangle with a pendant vertex}&XI,ZI,YX,IZ
\end{array}
$$

Here $XI=X\otimes I$, and identities on additional qubits are implicit. Each graph has an odd number of perfect matchings, so its alternating adjacency matrix is nonsingular over $\mathbb F_2$. 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 $3$. To check this, choose signs giving positive expectations. For commuting signed Paulis $U,V$ the operator $I+U+V+UV$ is four times a stabilizer code projector. For the path take the six triples $(U,V,UV)$ below, two per line.

$$
\begin{array}{cc}
(IX,-XI,-XX)&(-IX,-ZI,ZX)\\
(-IZ,-XI,XZ)&(-IZ,-ZI,ZZ)\\
(-XX,-YY,-ZZ)&(-XZ,YY,-ZX)
\end{array}
$$

For the two disjoint edges, replace the first triple by $(-IX,-XI,XX)$. For the triangle with a pendant vertex use

$$
\begin{array}{cc}
(IY,-XI,-XY)&(-IY,-ZI,ZY)\\
(-IZ,-XI,XZ)&(-IZ,-ZI,ZZ)\\
(XY,-YX,-ZZ)&(-XZ,-YX,-ZY).
\end{array}
$$

In each case the six operators sum to $6I$ minus twice the sum of the four observables. Stabilizer positivity proves the claimed upper bound, whereas four elements of $\mathcal H$ have total absolute expectation greater than $3$.

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 $a,b$ are adjacent and $c$ is adjacent to neither, take a common neighbor $e$ of $a,c$ in this component. When $e$ meets $b$ there is a triangle with pendant $c$. Otherwise $c,e,a,b$ is an induced path. Thus nonadjacency, including equality, partitions the component into independent parts, with all edges between distinct parts.

Put $W=\operatorname{span}(\mathcal H)$ and $L=W\cap W^\perp$. Every maximal commuting label space $F$ has $\mu(F)\le1/3$. Indeed, $F=F^\perp$ and $|F|=d$, so $\sum_{v\in F}k_v=\sum_{v\in F}f_v$ by the Fourier formula. The latter sum divided by $d$ is the collision probability of the stabilizer basis distribution for $F$, hence at most one. Since $\beta>1/3$, the graph has an edge.

Choose one representative $v_i$ from each of the $m\ge2$ parts. They pairwise anticommute. Labels in a part have identical neighbors in $\mathcal H$, so their differences lie in $L$, as do all isolated labels. Thus $\mathcal H\subset\Omega=L\cup\bigcup_i(v_i+L)$. The representatives span $W/L$. Their Gram matrix has ones off the diagonal and zeros on it, so its rank is $2r$, where $m=2r$ or $2r+1$. In the odd case their sum lies in $L$. Since a nondegenerate complement of $L$ in $W$ has dimension $2r$, $|L|\le2^{n-r}$.

A label orthogonal to $L$ anticommutes with at most $2r$ entire cosets of $\Omega$, by the sum relation when $m$ is odd. A label outside $L^\perp$ anticommutes with half of each coset. Summing (1) over $\Omega$ gives $d\beta\le2r|L|$, hence $\beta\le2r/2^r$. This rules out $r\ge4$. Also $|L|\le2^{n-r-1}$ would give $\beta\le r/2^r\le1/2$. Therefore $\dim L=n-r$ and $W=L^\perp$. A joint eigenspace of $L$, called a syndrome space below, encodes $r$ logical qubits.

For $r=3$, complete the six or seven representatives to seven by adjoining their sum if needed. Their only relation modulo $L$ is that all seven sum to zero. Each logical label is the sum of at most three of them. Let $\Omega'$ be $L$ and the seven representative cosets, and set $M=\mu(\Omega')$. There are $8|L|=d$ labels in $\Omega'$. A sum of one, two or three generators anticommutes with respectively six, two or four generators. Thus a physical label in $\Omega'$ anticommutes with at most $3d/4$ labels in $\Omega'$. A label outside anticommutes with at most $d/2$, including exactly half of each coset when it lies outside $L^\perp$. Sum $d\mu_v\le\Pr([v,w]=1)$ from (1) over $\Omega'$. Then $M\le3M/4+(1-M)/2$, so $\beta\le M\le2/3$, a contradiction. Only $r=1,2$ remain.

## 4. Bounding the blocks outside one syndrome

Fix signed generators of $L$. Write $E_s$ for their joint syndrome projectors, $B_s=E_sAE_s$ and $p_s=\operatorname{Tr}B_s$. The $p_s$ are nonnegative and sum to one. Fourier orthogonality on $L$ and averaging (1) over $v\in L$ give

$$
S_0:=\sum_s p_s^2=\frac1{|L|}\sum_{v\in L} f_v
=\mu(L^\perp)+\frac{2d}{|L|}\mu(L)\ge\beta.
$$

Let $E$ have the largest weight $p=1-\eta$. Since $p\ge S_0\ge\beta$ and $S_0\le p^2+\eta^2$, we get $\eta<0.156$ (at $\eta=0.156$ the last upper bound is $0.736672$). Label $E$ by syndrome zero. In blocks on $E\oplus E^\perp$ write $A=\left(\begin{smallmatrix}B&H\\H^\dagger&C\end{smallmatrix}\right)$, $h=\operatorname{Tr}(HH^\dagger)$ and $q_B=\operatorname{Tr}(B^2)$.

For each nonzero syndrome parity $a$, apply the outcome hypothesis to $a\cdot i=1$. Among all parities a fixed nonzero $i$ is selected half the time, and two distinct nonzero indices $i,j$ are both selected one quarter of the time. Summing the squared block norms on the left and the squared traces on the right gives

$$
\|C\|_{\mathrm{HS}}^2+\sum_{i\ne0}\|C_{ii}\|_{\mathrm{HS}}^2
\le 2\eta^2+2\sum_{i\ne0}p_i^2\le4\eta^2.
$$

Let $\Pi$ project onto either outcome of a nonidentity logical Pauli on $E$, and put $b=\operatorname{Tr}(\Pi B)$. Multiply that signed Pauli by a uniformly random syndrome parity. The resulting physical outcome selects $\Pi$ on $E$ and selects a uniformly random one of the two complementary logical eigenspaces on each nonzero syndrome. Their total selected trace $T$ satisfies $0\le T\le\eta$ and $\mathbb ET=\eta/2$. Keep the $00$ and $0i,i0$ blocks in the squared norm of the physical compression and average. Each nonzero-syndrome projector has mean $I/2$, giving

$$
\operatorname{Tr}[(\Pi B\Pi)^2]+\operatorname{Tr}(\Pi HH^\dagger)
\le2\mathbb E(b+T)^2\le2b^2+2b\eta+\eta^2 .
$$

The first term is at least $b^2/\operatorname{rank}\Pi$. Hence

$$
\operatorname{Tr}(\Pi HH^\dagger)\le
\begin{cases}
(b+\eta)^2,&r=1,\\
\tfrac32b^2+2b\eta+\eta^2,&r=2 .
\end{cases}\tag{2}
$$

For a logical axis with a physical lift in $\mathcal H$, sign that lift to have expectation $>3/4$. Its opposite outcome has total trace $<1/8$. On $E$ its trace $b$ is at most $1/8$, 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 $\mathcal H$. Otherwise $\mathcal H$ would lie in the union of two maximal commuting spaces, of total mass at most $2/3$. For the opposite projectors on these signed axes write $b_j\le1/8$, $S=\sum_j b_j$ and $S_2=\sum_j b_j^2\le S/8$. The one-qubit Pauli expansion gives $q_B=2p^2-2pS+2S_2$. The three projectors sum to an operator of least eigenvalue $(3-\sqrt3)/2$. Summing (2) thus gives $2h\le K(S_2+2\eta S+3\eta^2)$, with $K=4/(3-\sqrt3)<3.16$. Substitution into $2=q_B+2h+\|C\|_{\mathrm{HS}}^2$ yields

$$
0\le\eta[-4+(6+3K)\eta]+S[-2+(2+2K)\eta+(2+K)/8].
$$

Both brackets are negative at $\eta=0.156$ and increase with $\eta$. Hence $\eta=S=0$. The complement and coherence bounds give $C=H=0$, and all three signed logical coordinates are $+1$. These are precisely the Clifford images of $V_n$.

## 6. The two-logical-qubit case

Complete the four or five anticommuting representatives to five axes $R_1,\ldots,R_5$, adjoining their sum when needed. The other ten nonzero logical labels are the pair sums $R_i+R_j$. For each logical label $j$ choose a lift $T_j$ and let $S_j=\sum_s(\operatorname{Tr}(B_sT_j))^2$. Fourier orthogonality on the syndromes gives $S_j=|L|^{-1}\sum_{v\in j+L}f_v$. For a set $J$ of logical labels write $J+L$ for its full preimage in $L^\perp$. Summing the Fourier formula for $k_v$ over its cosets gives

$$
12\mu(J+L)=2\sum_{j\in J}S_j-
\frac14\sum_{v\in L^\perp/L} S_v\sum_{j\in J}(-1)^{[j,v]} .
$$

Take $J$ to consist of zero and the five axes. Let $U$ sum the $S_j$ on the axes and $V$ sum them on the pairs. The character sums over $J$ are $6,-2,2$ on zero, axes and pairs respectively. For $J^c$ they are $10,2,-2$. Thus

$$
24\mu(J+L)=S_0+5U-V,\qquad
24\mu(J^c+L)=-5S_0-U+5V .
$$

The second mass is nonnegative and the first contains $\mathcal H$, so $\beta\le\mu(J+L)\le U/5$.

Order $\alpha_i=|\operatorname{Tr}(BR_i)|$ decreasingly on $E$, and set $s_i=p^2-\alpha_i^2$, $\Delta=\sum_i s_i$. At least four axes have a lift in $\mathcal H$, so the opposite-outcome estimate in §4 gives $\alpha_i\ge p-1/4$ on the top four. Every logical coefficient on any other syndrome has magnitude at most $p_s$. Consequently

$$
0\le\Delta\le5S_0-U\le5(p^2+\eta^2-\beta).\tag{3}
$$

For each pair $Q=R_i+R_j$ choose the index with smaller $s$ inside the pair and the two with smallest $s$ outside it, breaking ties to avoid index five. Let $m_Q$ be the sum of the three chosen deficits. The chosen axes form a triangle and $Q$ is adjacent only to the chosen index inside the pair. The four-observable inequality in §3 applied to $B/p$ gives

$$
|\operatorname{Tr}(BQ)|\le\sum_{\text{chosen }k}(p-\alpha_k)
\le \frac{m_Q}{2p-1/4}.
$$

The chosen deficit inside the pair is bounded by half its sum, and the two outside by two thirds of their sum. Thus $m_Q\le2\Delta/3$. Summing over pairs gives $\sum_Qm_Q\le6\Delta$ since every index is inside four pairs and outside six. Hence $\sum_Qm_Q^2\le4\Delta^2$. Expanding $B$ in the fifteen nonidentity Paulis gives

$$
q_B\le\tfrac32p^2+\Delta\left(\frac{\Delta}{(2p-1/4)^2}-\tfrac14\right).\tag{4}
$$

For a lower bound take four signed axes with lifts in $\mathcal H$. Their signed sum squares to $4I_E$, so their opposite projectors sum to at least $I_E$. Sum (2) with $b_j\le1/8$ and use $\|C\|_{\mathrm{HS}}^2\le4\eta^2$. This gives

$$
q_B\ge\frac{29}{16}-2\eta-12\eta^2
=\tfrac32p^2+g(\eta),\qquad
g(\eta)=\tfrac5{16}+\eta-\tfrac{27}{2}\eta^2.\tag{5}
$$

To see the contradiction using exact numbers, weaken $\beta$ to $13/18$, so $\eta\le1/6$. Let $D=5[(1-\eta)^2+\eta^2-13/18]$ and $a=7/4-2\eta$. By (3), $\Delta\le D$. The derivative of $D/a^2$ has the sign of $-43/18+3\eta$, hence this ratio decreases. On $[0,1/12]$ we have $D\le25/18$ and $D/a^2\le200/441$. The correction in (4) above $3p^2/2$ is at most

$$
\frac{25}{18}\left(\frac{200}{441}-\frac14\right)<\frac{29}{96}\le g(\eta).
$$

On $[1/12,1/6]$, $D/a^2\le90/361<1/4$, making the correction nonpositive, whereas $g(\eta)\ge5/48$. The lower bounds on $g$ 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 $V_n$ gives zero or a multiple of a logical stabilizer state. Thus $V_n\in\Lambda_n$, and direct multiplication gives $\operatorname{Tr}(V_n^2)=2$. By polarity the stabilizer polytope is the intersection of the half-spaces $\operatorname{Tr}(A\rho)\ge0$ for $A\in\Lambda_n$. The squared distance from $I/d$ to such a hyperplane is $(1/d)^2/(\operatorname{Tr}(A^2)-1/d)$. The minimum is $1/[d(2d-1)]$, so adding $1/d$ gives the threshold $1/(d-1/2)$.

The state $\rho_*=(2I-V_n)/(2d-1)$ reaches the hyperplane for $V_n$ and is positive definite. Moving slightly farther on the ray from $I/d$ through $\rho_*$ therefore gives physical magic states approaching the threshold from above. This establishes sharpness for every $n$.

**A trace-distance consequence.** With $D_{\mathrm{tr}}(\rho,\sigma)=\|\rho-\sigma\|_1/2$, the centered inradius is $1/(d\sqrt3)$. Both spectral bounds in 5.11 give $\lambda_{\max}(A)-\lambda_{\min}(A)\le\sqrt3$. A traceless perturbation at trace distance $r$ from zero has positive and negative parts of trace $r$, so it can lower $\operatorname{Tr}(AI/d)=1/d$ by at most $r\sqrt3$. Equality is attained along the direction $P_--P_+$, where $P_\pm$ are extremal eigenprojectors of $V_n$. The touching state $I/d+(P_--P_+)/(d\sqrt3)$ is positive definite. [OpenContext 5.11](https://open-context.jan-bulling.com/contribution/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](https://arxiv.org/abs/2512.16777) · [Zurel and Davis](https://arxiv.org/abs/2602.22336) · [Liu and Liu, Appendix A](https://arxiv.org/html/2609.03944v3#A1) · [Wang, OpenContext 5.6](https://open-context.jan-bulling.com/contribution/5.6) · [Wagner, OpenContext 5.11](https://open-context.jan-bulling.com/contribution/5.11)

#### Comment by Rafael Wagner, 20 Nov 2026

I think that my Claude's solution is along similar lines

#### Comment by Rafael Wagner, 20 Nov 2026

It is just super cool that we are really solving the problem. Now we need people to second the result's correctness somehow.

### 17. Independent verification: A solution to the problem which uses Aronszajn's inequality

- **Date:** 20 Nov 2026
- **Credit:** Rafael Wagner
- **AI use:** partially AI-assisted
- **Permanent link:** https://open-context.jan-bulling.com/contribution/5.14
- **Votes:** 0 (0 up, 0 down)

## Main result

Below, I present an independent claim from the one by Yujie where it seems that Fable 5.2 found a solution to the problem. The interesting aspects is that it clearly has succeeded by building on the theory of $\Lambda$-polytopes. Below, I mark all the points where I believe further exposition is needed with a triple star $(\star\star\star)$. 

This is the main output:

>**Theorem (inradius of the stabilizer polytope).** Let $\Lambda_n$ be the $\Lambda$-polytope of $n$ qubits. Denote by $d=2^n$ the Hilbert space dimension, and $$M_n:=\max_{A\in\Lambda_n}\,\mathrm{Tr}[A^2].$$ For every integer $n\ge1$,
>$$
>M_n=2 .
>$$
>Consequently, for every $n$-qubit state $\rho$,
>$$
>\mathrm{Tr}[\rho^2]\le\frac{1}{d-\frac12}\ \implies\ \rho\in\mathrm{STAB}_n ,
>$$
>and the constant $\tfrac12$ cannot be improved [1]. Moreover, $A\in\Lambda_n$ has $\mathrm{Tr}[A^2]=2$ if and only if $A=UA^\star U^\dagger$ for a Clifford unitary $U$, where $A^\star=\tfrac12(\mathbb{I}+X+Y+Z)\otimes|0\ldots0\rangle\langle0\ldots0|$.

This is the first contribution of two parts. First part proves $M_n=2$ and shows that every maximiser has largest eigenvalue $\lambda_+=\tfrac{1+\sqrt3}{2}$. The second part proves that such an operator lies in the Clifford orbit of $A^\star$ and reports numerical experiments on sub-polytopes of $\mathrm{STAB}_n$.

The proof is analytical; no step is computer-assisted. It relies on three earlier results, recalled below: the block lemma [4], the Liu–Liu inequality [3], and the theorem on the spectrum of a witness proved in a previous contribution. The full strategy is close to the one by [Yujie Zhang](https://open-context.jan-bulling.com/contribution/5.13). 

## Preliminary mathematical ideas

We keep the notation of the previous contributions. $\Lambda_n$ is the set of Hermitian operators $A$ with $\mathrm{Tr}[A]=1$ and $\langle s|A|s\rangle\ge0$ for all stabilizer states, known as the $\Lambda$-polytope [4, 2, 9, 10, 8, 11]. A state belongs to $\mathrm{STAB}_n$ if and only if $\mathrm{Tr}[A\rho]\ge0$ for all $A\in\Lambda_n$ [4]. By the [Cauchy–Schwarz inequality](https://en.wikipedia.org/wiki/Cauchy%E2%80%93Schwarz_inequality) every bound $M_n\le m$ implies that $\mathrm{Tr}[\rho^2]\le1/(d-\tfrac1m)$ is sufficient for $\rho\in\mathrm{STAB}_n$ (see [OpenContext §5.12](https://open-context.jan-bulling.com/contribution/5.12), the purity section). Define
$$
v:=\lambda_+=\frac{1+\sqrt3}{2},\qquad c:=-\lambda_-=v-1,\qquad r:=\frac{\sqrt3}{2}.
$$
It is simple to show that $(\star \star\star)$ $$vc=\tfrac12, \quad v^2=v+\tfrac12, \quad c^2=\tfrac12-c, \quad v^2+c^2=2, \quad r=v-\tfrac12=1-c^2.$$ 
For $A\in\Lambda_n$ we write
$$
\ell=\lambda_{\max}(A),\qquad \kappa=-\lambda_{\min}(A),\qquad q=\mathrm{Tr}[A^2]=2+\theta,\qquad a_P=\mathrm{Tr}[PA].
$$
These quantities are instrumental for the proof. For a Pauli operator $P\ne\mathbb{I}$ let $\Pi_\pm=\tfrac12(\mathbb{I}\pm P)$, $B_\pm=\Pi_\pm A\Pi_\pm$ and $t_\pm=\mathrm{Tr}[B_\pm]=\tfrac12(1\pm a_P)$. All sums over $P$ run over the $d^2-1$ Pauli operators $P\ne\mathbb{I}$.

We use the following results discussed previously in [OpenContext §5.12](https://open-context.jan-bulling.com/contribution/5.12).

>**Lemma (block lemma [4]).** The range of $\Pi_\pm$ is an $(n-1)$-qubit code space whose stabilizer states are $n$-qubit stabilizer states. Hence $t_\pm\ge0$, and $B_\pm$ is $t_\pm$ times an element of $\Lambda_{n-1}$ (and $B_\pm=0$ if $t_\pm=0$). In particular
>$$
>\mathrm{Tr}[B_\pm^2]\le M_{n-1}\,t_\pm^2 .
>$$

>**Theorem (spectrum of a witness, previous contribution).** For all $n\ge1$ and $A\in\Lambda_n$: $-c\le\lambda_{\min}(A)$ and $\lambda_{\max}(A)\le v$.

>**Lemma (Liu–Liu inequality [3]).** For $n\ge2$ and $A\in\Lambda_n$: $\ d\,q+1-2\,\mathrm{Tr}[A^3]\le M_{n-1}\,(d-q)$.

We also use the rule for Pauli sums [7]: $\sum_{P}\mathrm{Tr}[PX]\mathrm{Tr}[PY]=d\,\mathrm{Tr}[XY]$ and $\sum_PPXP=d\,\mathrm{Tr}[X]\,\mathbb{I}$, with the identity included in the sums. The last ingredient is a classical inequality for partitioned Hermitian matrices, attributed to Aronszajn [14]. We include its short proof (\star \star \star).

>**Lemma (Aronszajn's inequality).** Let $H$ be a Hermitian operator on $V_1\oplus V_2$ and let $H_1,H_2$ be its compressions to $V_1$ and $V_2$. Then
>$$
>\lambda_{\max}(H)+\lambda_{\min}(H)\le\lambda_{\max}(H_1)+\lambda_{\max}(H_2).
>$$

*Proof:* $K:=H-\lambda_{\min}(H)\mathbb{I}\ge0$. For a unit vector $\psi=\psi_1+\psi_2$ with $\psi_j\in V_j$, the triangle inequality for $K^{1/2}$ gives
$$
\langle\psi|K|\psi\rangle\le\big(\|\psi_1\|\sqrt{k_1}+\|\psi_2\|\sqrt{k_2}\big)^2\le k_1+k_2,\qquad k_j:=\lambda_{\max}(H_j)-\lambda_{\min}(H)\ge0 ,
$$
where the last step is the Cauchy–Schwarz inequality. Taking the supremum over $\psi$ gives $\lambda_{\max}(H)-\lambda_{\min}(H)\le k_1+k_2$. $\square$

## Block slacks

In this section and the next, $n\ge2$. Also, we assume $M_{n-1}=2$, and $A\in\Lambda_n$ is fixed in order to structure an induction step later on. For $P\ne\mathbb{I}$ we define the *slacks* as 
>$$
>s^P_\pm:=2\,t_\pm^2-\mathrm{Tr}[B_\pm^2] \tag{slacks}.
>$$
We also define the operator $W$ by
$$
W:=2A-A^2.
$$
Additionally, we denote $w_P:=\mathrm{Tr}[PW]$ the expectation value of $P$ on $W$. With these ingredients at hand, we can state and prove the following lemma $(\star \star \star)$.

>**Lemma (block slacks).** For every $P\ne\mathbb{I}$ we have $s^P_\pm\ge0$ and $s^P_+-s^P_-=w_P$. Moreover,
>$$
>\sum_{P}\big(s^P_++s^P_-\big)=\frac32\,d-\frac{d^2-2d-2}{2}\,\theta,\qquad \sum_{P}w_P^2=d\,\mathrm{Tr}[W^2]-\theta^2 .
>$$

*Proof:* $s^P_\pm\ge0$ is the block lemma with $M_{n-1}=2$. Let $b_P=\mathrm{Tr}[PA^2]$ and $c_P=\mathrm{Tr}[PAPA]$. Expanding $\Pi_\pm$,
$$
\mathrm{Tr}[B_\pm^2]=\mathrm{Tr}[\Pi_\pm A\Pi_\pm A]=\tfrac14\,(q+c_P\pm2b_P).
$$
Since $t_+^2-t_-^2=a_P$ and $t_+^2+t_-^2=\tfrac12(1+a_P^2)$, we get
$$
s^P_+-s^P_-=2a_P-b_P=w_P,\qquad s^P_++s^P_-=1+a_P^2-\tfrac12\,(q+c_P).
$$
The rule for Pauli sums gives $\sum_Pa_P^2=dq-1$ and $\sum_Pc_P=d-q$. Hence $\sum_P(s^P_++s^P_-)=d^2-2+dq-\tfrac12(d^2q-2q+d)$, which is the claimed expression for $q=2+\theta$. Finally, $\sum_Pw_P^2=d\,\mathrm{Tr}[W^2]-\mathrm{Tr}[W]^2$ and $\mathrm{Tr}[W]=2-q=-\theta$. $\square$

>*Remark.* The same computation gives $\tfrac4d\sum_P\big(t_-s^P_++t_+s^P_-\big)=2d-1-(d+2)\,q+2\,\mathrm{Tr}[A^3]$ $(\star \star \star)$.  Since the left-hand side is non-negative, this is the Liu–Liu inequality for $M_{n-1}=2$, so the proof below is self-contained once the block lemma and the theorem on the spectrum of a witness are granted.

## The key inequality

The proposition above shows how one can bound the sum of the largest and smallest eigenvalues of $A$ in the $\Lambda$-polytope $(\star \star \star)$.

>**Proposition.** Let $n\ge2$, $M_{n-1}=2$ and $A\in\Lambda_n$. For every $P\ne\mathbb{I}$,
>$$
>\lambda_{\max}(A)+\lambda_{\min}(A)\ \le\ \sqrt{2-|w_P|}\,.
>$$
>If moreover $\theta\ge0$, there is a Pauli operator $P\ne\mathbb{I}$ with
>$$
>|w_P|\ \ge\ \frac23\Big(\mathrm{Tr}[W^2]-\frac{\theta^2}{d}\Big).
>$$

*Proof:* Replacing $P$ by $-P$ exchanges the two blocks and changes the sign of $w_P$, so we may assume $w_P\ge0$. Then $s^P_+=w_P+s^P_-\ge w_P$, that is, $\mathrm{Tr}[B_+^2]\le2t_+^2-w_P$. The largest eigenvalue of a Hermitian operator is at most its Hilbert–Schmidt norm, so the compressions of $A$ to the ranges of $\Pi_+$ and $\Pi_-$ have largest eigenvalues at most $\sqrt{2t_+^2-w_P}$ and $\sqrt{\mathrm{Tr}[B_-^2]}\le\sqrt2\,t_-=\sqrt2\,(1-t_+)$. By Aronszajn's inequality,
$$
\lambda_{\max}(A)+\lambda_{\min}(A)\le g(t_+),\qquad g(t):=\sqrt{2t^2-w_P}+\sqrt2\,(1-t).
$$
Since $g'(t)=2t/\sqrt{2t^2-w_P}-\sqrt2\ge0$ and $t_+\le1$, we have $g(t_+)\le g(1)=\sqrt{2-w_P}$.

For the second statement, $|w_P|=|s^P_+-s^P_-|\le s^P_++s^P_-$, so by the lemma $\sum_P|w_P|\le\tfrac32d$ when $\theta\ge0$. Therefore
$$
d\,\mathrm{Tr}[W^2]-\theta^2=\sum_Pw_P^2\le\max_P|w_P|\,\sum_P|w_P|\le\tfrac32\,d\,\max_P|w_P| ,
$$
which is the claim. $\square$

For $A^\star$ (see definition at the top of this document) all inequalities are *equalities*. To see this, note that $W=r\,(|\varphi\rangle\langle\varphi|-|\chi\rangle\langle\chi|)$ for the two eigenvectors of $A^\star$, $\mathrm{Tr}[W^2]=\tfrac32$, and $|w_P|=1$ exactly for the $\tfrac32d$ Pauli operators that act as logical Pauli operators on the code space of $A^\star$; these carry the whole slack $\tfrac32d$, and $\lambda_{\max}+\lambda_{\min}=1$.

## A lemma on spectra

Let $f(\lambda):=(2\lambda-\lambda^2)^2$, so that $\mathrm{Tr}[W^2]=\sum_if(\lambda_i)$ for the eigenvalues $\lambda_i$ of $A$ $(\star \star \star)$.

>**Lemma (spectral inequality).** Let $\lambda_1\ge\dots\ge\lambda_d$ be real numbers, $d\ge4$, with $\tfrac54\le\lambda_1\le v$ and $-c\le\lambda_d\le0$. Put $\delta=v-\lambda_1$, $k=c+\lambda_d$, $s=2-\lambda_1^2-\lambda_d^2$, and assume that $\sigma:=\sum_{1<i<d}\lambda_i^2=s+\theta$ with $\theta\ge0$ and $\sigma\le\tfrac7{16}$. Then
>$$
>(\lambda_1+\lambda_d)^2+\frac23\sum_{i}f(\lambda_i)-2\ \ge\ Q(\delta,k)+\frac\theta4,
>$$
>where
>$$
>Q(\delta,k)=\Big(2+\tfrac{4}{\sqrt3}\Big)\delta+\Big(\tfrac{4}{\sqrt3}-2\Big)k-\Big(9+\tfrac{14}{\sqrt3}\Big)\delta^2-10\,\delta k+\Big(\tfrac{17}{\sqrt3}-7\Big)k^2 ,
>$$
>and $Q(\delta,k)>0$ for $0\le\delta\le v-\tfrac54$, $0\le k\le c$, $(\delta,k)\ne(0,0)$.

*Proof:* Let $a=\delta(2c-\delta)\ge0$, $b=k(2v-k)\ge0$ and $e=k-\delta$. Then $\lambda_1+\lambda_d=1+e$ and, using $v^2+c^2=2$ and $v-c=1$, $s=\delta(2v-\delta)+k(2c-k)=a+b-2e$.

*The two extreme eigenvalues.* $2\lambda_1-\lambda_1^2=1-(\lambda_1-1)^2=r+a$, because $\lambda_1-1=c-\delta$ and $1-c^2=r$. Similarly $2\lambda_d-\lambda_d^2=1-(v-k)^2=-(r-b)$, because $1-\lambda_d=v-k$ and $v^2-1=r$. Hence $f(\lambda_1)=(r+a)^2\ge\tfrac34+\sqrt3\,a$ and $f(\lambda_d)=(r-b)^2=\tfrac34-\sqrt3\,b+b^2$.

*The other eigenvalues.* For $1<i<d$ we have $\lambda_i\le\max(\lambda_2,0)\le\sqrt\sigma<2$, hence $(2-\lambda_i)^2\ge(2-\sqrt\sigma)^2\ge3(1-\sigma)$, the last step because $(2-\sqrt\sigma)^2-3(1-\sigma)=(1-2\sqrt\sigma)^2$. Therefore $\sum_{1<i<d}f(\lambda_i)\ge3\sigma(1-\sigma)$. The function $u\mapsto3u(1-u)$ is concave with derivative at least $\tfrac38$ on $[0,\tfrac7{16}]$, so $3\sigma(1-\sigma)\ge3s(1-s)+\tfrac38\theta$.

*Collecting.* The left-hand side of the lemma is at least
$$
(1+e)^2-2+\tfrac23\Big[\tfrac32+\sqrt3\,(a-b)+b^2+3s-3s^2\Big]+\tfrac\theta4
=-2e+e^2+\Big(2+\tfrac2{\sqrt3}\Big)a+\Big(2-\tfrac2{\sqrt3}\Big)b+\tfrac23b^2-2s^2+\tfrac\theta4 ,
$$
where we used $s=a+b-2e$ in the linear term. Now $a=2c\delta-\delta^2$, $b=2vk-k^2$ and $(2+\tfrac2{\sqrt3})\,2c=(2-\tfrac2{\sqrt3})\,2v=\tfrac4{\sqrt3}$. Moreover $b\ge(v+1)k$ gives $\tfrac23b^2\ge(2+\sqrt3)k^2$, and $0\le s\le2v\delta+2ck$ gives $2s^2\le(8+4\sqrt3)\delta^2+8\delta k+(8-4\sqrt3)k^2$. Inserting these and $e^2=\delta^2-2\delta k+k^2$ yields $Q(\delta,k)+\tfrac\theta4$.

*Positivity of $Q$.* Rounding all coefficients against us, $Q\ge\delta\,(4.309-17.083\,\delta-10k)+k\,(0.309+2.814\,k)$. Since $\delta\le v-\tfrac54<0.1161$, the first bracket exceeds $2.325-10k$. If $k\le0.2325$, both terms are non-negative, and they vanish only for $k=\delta=0$. If $0.2325<k\le c<0.3661$, the first term is at least $-0.1161\cdot1.336>-0.156$ and the second is at least $0.2325\cdot0.963>0.223$. $\square$

## Proof of the theorem

We prove by induction on $n$ that $M_n=2$, and that every $A\in\Lambda_n$ with $\mathrm{Tr}[A^2]=2$ has $\lambda_{\max}(A)=\lambda_+$. For $n=1$, $A=\tfrac12(\mathbb{I}+\vec r\cdot\vec\sigma)$ with $|r_i|\le1$, so $\mathrm{Tr}[A^2]=\tfrac12(1+|\vec r|^2)\le2$, with equality only if $|\vec r|=\sqrt3$, i.e. $\lambda_{\max}(A)=\lambda_+$. Since $A^\star\in\Lambda_n$ has purity $2$, we have $M_n\ge2$ for all $n$.

Let $n\ge2$, assume $M_{n-1}=2$, and let $A\in\Lambda_n$ have $q=2+\theta$ with $\theta\ge0$. We show that $\theta=0$ and $\ell=v$.

**Step 1 (location of the spectrum).** With $M_{n-1}=2$ the Liu–Liu inequality reads $2\,\mathrm{Tr}[A^3]\ge5+(d+2)\theta$. Since every eigenvalue satisfies $\lambda^3\le\ell\lambda^2$, also $\mathrm{Tr}[A^3]\le\ell\,(2+\theta)$. Hence
$$
\theta\,(d+2-2\ell)\le4\ell-5 ,
$$
so $\ell\ge\tfrac54>1=\mathrm{Tr}[A]$ and $A$ has a negative eigenvalue. By the theorem on the spectrum of a witness, $\tfrac54\le\ell\le v$ and $0<\kappa\le c$, so $\delta:=v-\ell\in[0,v-\tfrac54]$ and $k:=c-\kappa\in[0,c)$. Moreover $\theta\le(4\ell-5)/(6-2\ell)<\tfrac16$, and the sum $\sigma$ of the squares of the other $d-2$ eigenvalues satisfies
$$
\sigma=2+\theta-\ell^2-\kappa^2\le2-\ell^2+\frac{4\ell-5}{6-2\ell}\le\frac7{16},
$$
because the function of $\ell$ on the right is decreasing on $[\tfrac54,v]$ (its derivative is $-2\ell+14/(6-2\ell)^2<0$) and equals $\tfrac7{16}$ at $\ell=\tfrac54$.

**Step 2 (the key inequality).** By Step 1, $\ell-\kappa\ge\tfrac54-c>0$. We square the first inequality of the proposition for the Pauli operator of its second statement:
$$
(\ell-\kappa)^2\le2-\frac23\Big(\mathrm{Tr}[W^2]-\frac{\theta^2}{d}\Big).
$$

**Step 3 (conclusion).** The eigenvalues of $A$ satisfy the assumptions of the lemma on spectra, with $\lambda_1=\ell$ and $\lambda_d=-\kappa$. Together with Step 2,
$$
Q(\delta,k)+\frac\theta4\ \le\ \frac23\,\frac{\theta^2}{d}\ \le\ \frac{\theta}{36},
$$
where we used $\theta<\tfrac16$ and $d\ge4$. Since $Q(\delta,k)\ge0$, this forces $\theta=0$ and $Q(\delta,k)=0$, hence $\delta=k=0$. So no element of $\Lambda_n$ has purity larger than $2$, i.e. $M_n=2$, and every $A\in\Lambda_n$ with $\mathrm{Tr}[A^2]=2$ has $\lambda_{\max}(A)=\lambda_+$ (and $\lambda_{\min}(A)=\lambda_-$). This completes the induction.

The statement about states follows from $M_n=2$ as recalled in the preliminaries. The second part shows that $\lambda_{\max}(A)=\lambda_+$ implies $A=UA^\star U^\dagger$. $\square$

#### Comment by Jan, 20 Nov 2026

Wow, nice that this was solved! Amazing!

#### Comment by Rafael Wagner, 20 Nov 2026

Below I'll work out the various parts in which I haven't fully checked in order to improve my understanding of it. 

The first bit is simple, and it is related to the first batch of claims. In particular, I wanted to understand better the following simple bit first:

>Define
>$$
>v:=\lambda_+=\frac{1+\sqrt3}{2},\qquad c:=-\lambda_-=v-1,\qquad r:=\frac{\sqrt3}{2}.
>$$
>It is simple to show that
>$$
>vc=\tfrac12,\qquad v^2=v+\tfrac12,\qquad c^2=\tfrac12-c,\qquad v^2+c^2=2,\qquad r=v-\tfrac12=1-c^2.
>$$

Indeed, $\lambda_\pm$ are the two roots of $\lambda^2-\lambda-\tfrac12=0$. They are the eigenvalues of $$\tfrac12(\mathbb{I}+X+Y+Z),$$ and since we have that $$(X+Y+Z)^2=3\,\mathbb{I}$$ we get these values for the eigenvalues. Here, these are simply the Pauli matrices. 

The sum and the product of the roots give $v-c=1$ and $vc=\tfrac12$. The first relation is trivial, the product follows from 
$$
vc = -\lambda_+\lambda_- = \frac{1+\sqrt{3}}{2} \left(\frac{1+\sqrt{3}}{2}-1\right) = \frac{1+\sqrt{3}}{2} \left(\frac{\sqrt{3}-1}{2}\right) = \frac{(\sqrt{3})^1-1^2}{4} = \frac{1}{2}.
$$

Now, for  $v^2=v+\tfrac12$ and $c^2=\tfrac12-c$. Inserting each root in the equation gives

$$v^2 = \left ( \frac{1+\sqrt{3}}{2} \right)^2 = \frac{4+2\sqrt{3}}{4} = 1 + \frac{\sqrt{3}}{2} = \frac{1}{2} + \left(\underbrace{\frac{1+\sqrt{3}}{2}}_{=v}\right).$$

Then $$v^2+c^2=(v-c)^2+2vc=1^2+2\underbrace{vc}_{=\frac{1}{2}} = 2$$ and $1-c^2=\tfrac12+c=v-\tfrac12=r$. So, these are simply aspects.

#### Comment by Rafael Wagner, 20 Nov 2026

I would be nice to have Lean certificates of these proofs.

## Discussion

Comments approved by the moderators, oldest first, each followed by its replies.

### Rudi B. P. Pietsch, 17 Nov 2026

Hey I am joining the discussion now. I don't believe in magic so everybody should stop working on this problem ASAP.

#### Reply by Rafael Wagner, 17 Nov 2026

Hi, are you sure you’re understanding magic well?

#### Reply by Rafael Wagner, 18 Nov 2026

Just to make a more concrete comment about it. It is important to note that magic is not the "subjective" term commonly used. Here, it has a rather precise meaning: magic is the property that a quantum state can have. 

Different communities tend to use the term distinctively, but the most common use is that magic is just the property that states have whenever they are not elements of the stabilizer polytope $\mathrm{STAB}_n$. 

To construct one such polytope, what one does is the following: take $\vert 0 \rangle$ a reference stabilizer state, for example, the +1 eigenvalue of $Z \cdots Z$. Then, the vertices of the stabilizer polytope are found by applying all possible Clifford unitaries on $\vert 0 \rangle$. In this way, one ends up with all vertices. The polytope is then the convex-hull (in the space of quantum states) of these vertices.

### Rafael Wagner, 17 Nov 2026

I would like to respond to a comment made by Rudi above. 

Note that here we are treating a *formal* notion of magic, best put it as nonstabilizerness. This is related to the failure of having a state to be a stabilizer state.

### Rudi B. P. Pietsch, 17 Nov 2026

I would like to comment that the partial result is already known in the literature: "Basis-independent stabilizerness and maximally noisy magic states" https://arxiv.org/abs/2602.22336

#### Reply by Jan, 17 Nov 2026

Hi Rudi, that sounds nice! This is a reply comment!

#### Reply by Jan, 17 Nov 2026

Yes, even [2] citations and math $2+3^x=1$ works well!

#### Reply by Jan, 17 Nov 2026

Different commment!

#### Reply by Rafael Wagner, 17 Nov 2026

Test

#### Reply by Rafael Wagner, 18 Nov 2026

Nice, I really like how comments are shown here.

#### Reply by Rafael Wagner, 18 Nov 2026

Btw, we cite this work in the problem statement

### Jan, 17 Nov 2026

Another test comment

### Rafael Wagner, 17 Nov 2026

Nice

### Rafael Wagner, 18 Nov 2026

I have recently found an interesting advance to this problem; a comment section is too short for it, and I am revising the generated data from Claude

#### Reply by Rafael Wagner, 18 Nov 2026

It is taking time to do it, but the thing is that I found a bound $c_n$ which is distinct from 1/2 int he denominator.

### yujie4phy, 18 Nov 2026

Ref 1 was published and could be updated

https://journals.aps.org/prxquantum/abstract/10.1103/rcpf-8nh9

#### Reply by Rafael Wagner, 18 Nov 2026

Nice! Well do so

### Rafael Wagner, 18 Nov 2026

This is another test

### Yixuan Wang, 18 Nov 2026

**Test**
$$
Tr(A^2)=?
$$
🙌

## How to cite

```bibtex
@misc{opencontext-problem-5,
  author        = {Zhenhuan Liu and Tobias Haug and Qi Ye and Zi-Wen Liu and Ingo Roth},
  title         = {{Test problem: Hilbert–Schmidt inradius of the multiqubit stabilizer polytope}},
  howpublished  = {OpenContext: Open problems in generalized contextuality, Problem 5, version 12},
  year          = {2026},
  month         = nov,
  date          = {2026-11-17},
  url           = {https://open-context.jan-bulling.com/contribution/5},
  urldate       = {2026-11-20},
  note          = {Problem posed 17 November 2026; status: under revision, last updated 20 November 2026}
}
```

## Working on this problem

- A solution must resolve the problem **as stated**, by the criterion fixed in the statement. A proof of a weakened or modified statement is a partial result, not a solution.
- Check the contributions above before starting, so as not to repeat existing work.
- Submit results at https://open-context.jan-bulling.com/problems/test-problem-hilbert-schmidt-inradius-of-the-multiqubit-stab, with the extent of AI use and the names of the humans who take responsibility for them. AI-generated material counts as progress only after human scrutiny.
- Cite the problem and the people who posed it. The full rules: https://open-context.jan-bulling.com/guidelines
