Skip to content
OpenContext
Contribution toProblem 5: Test problem: Hilbert–Schmidt inradius of the multiqubit stabilizer polytopeunder revision
+1

Partial resultOpenContext 5.10

Proofs for OpenContext 5.9: the sharp three-qubit threshold and a uniform bound of 0.496222

by Yujie Zhang · accepted 19 Nov 2026 · fully AI-generated

Main result

This is the proof of the two bounds announced in OpenContext 5.9. Let d=2nd=2^n and let cnc_n be the largest constant such that every nn-qubit density matrix of purity at most 1/(d−cn)1/(d-c_n) is a convex mixture of stabilizer states. Then

c3=12,cn≥c∗=327659848669660308814208=0.496222133672421…(n≥1).\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 2/152/15. The uniform bound applies to every state, with arbitrary spectrum and eigenbasis. The conjectured equality cn=1/2c_n=1/2 remains open for n≥4n\ge4. The main result of OpenContext 5.7 gave a uniform value of 0.48590.4859.

Which steps use a computer? Both proofs are computer-assisted. For c3=1/2c_3=1/2, exact polynomial certificates classify the witnesses on the sphere Tr⁡(A2)=2\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,…,15n=4,\ldots,15. Once the bound on c15c_{15} is in hand, one analytic averaging identity covers every n>15n>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 AA such that ⟨s∣A∣s⟩≥0\langle s|A|s\rangle\ge0 for every pure stabilizer state ss. Write Λn\Lambda_n for this polytope and q=Tr⁡(A2)q=\operatorname{Tr}(A^2). By polarity,

Mn=max⁡A∈ΛnTr⁡(A2),cn=1/Mn.M_n=\max_{A\in\Lambda_n}\operatorname{Tr}(A^2),\qquad c_n=1/M_n.

Indeed, the squared distance from I/dI/d to the witness hyperplane Tr⁡(Aρ)=0\operatorname{Tr}(A\rho)=0 is 1/[d(dq−1)]1/[d(dq-1)]. Minimizing over witnesses and adding 1/d1/d gives the purity at first contact. Thus cnc_n here is the purity parameter 1/Mn1/M_n in the notation of OpenContext 5.7.

2. The sharp three-qubit bound

Claim. The only witnesses in Λ3\Lambda_3 with q=2q=2 are the Clifford conjugates of

V0=∣00⟩⟨00∣⊗I+X+Y+Z2.V_0=|00\rangle\langle00|\otimes\frac{I+X+Y+Z}{2}.

Postselecting the first two qubits of a stabilizer state onto 0000 leaves a one-qubit stabilizer state or zero. This proves feasibility of V0V_0, and direct calculation gives Tr⁡(V02)=2\operatorname{Tr}(V_0^2)=2. Consider any other candidate on the same sphere.

Step 1. A large Pauli expectation. Let aR=Tr⁡(AR)a_R=\operatorname{Tr}(AR) for each nonidentity Hermitian Pauli RR. A signed Pauli outcome with projector Π\Pi gives a two-qubit compression B=ΠAΠB=\Pi A\Pi of trace pp. If p>0p>0, then B/p∈Λ2B/p\in\Lambda_2. If p=0p=0, the compression vanishes. Hence Tr⁡(B2)≤2p2\operatorname{Tr}(B^2)\le2p^2 and B⪯LpΠB\preceq Lp\Pi, where L=1.366026L=1.366026 is a rational upper bound on the largest two-qubit witness eigenvalue. Suppose every ∣aR∣≤r=49/50|a_R|\le r=49/50. Sum the two compression inequalities over the two outcomes and the nonidentity Paulis, weighting the outcomes by 1∓aR/r1\mp a_R/r. With t=Tr⁡(A3)t=\operatorname{Tr}(A^3), x=λmax⁡(A)x=\lambda_{\max}(A) and h=1/r−1h=1/r-1, the sums give

0≤J(t)=2t−5−2h(4−t),0≤W(x)=2x2−8x+6L−1+h[2x2−L(2+x)+(L−x)/4].\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 WW is negative at both 1.3591.359 and 1.51.5. Since x≤2x\le\sqrt2, this forces x<1.359x<1.359. The fixed-mean, fixed-variance cubic bound on the seven remaining eigenvalues then gives t<2.527t<2.527. But JJ increases and J(2.527)=−3/490J(2.527)=-3/490. Thus some signed Pauli has expectation above 49/5049/50. This step uses the Pauli compression and cubic moment estimate in Liu and Liu, Appendix A.

Step 2. Computer-assisted certificate for an overlap gap. Fix a Clifford conjugate VV of V0V_0. Let δ\delta be the difference of the 63 nonidentity Pauli coefficient vectors of AA and VV, and put D=8[2−Tr⁡(AV)]D=8[2-\operatorname{Tr}(AV)]. Both witnesses have q=2q=2, so ∥δ∥2=2D\|\delta\|^2=2D. The first rational certificate gives

3431242D3−D∥δ∥2≥0.\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 VV has zero expectation. If A≠VA\ne V, then D>0D>0. Substitution gives D≥2484/343D\ge2484/343 and hence Tr⁡(AV)≤T=751/686\operatorname{Tr}(AV)\le T=751/686.

Step 3. Computer-assisted contradiction for the remaining witnesses. Suppose AA differs from every conjugate of V0V_0. In any two-qubit block BB of trace pp, the vertices of Λ2\Lambda_2 with squared norm 2 lift to conjugates of V0V_0. Every other vertex has squared norm at most 3/23/2. Decomposing B/pB/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.

Tr⁡(B2)≤max⁡{Tp,3p2/2},Tr⁡(B3)≤pTr⁡(B2)+p3/2.\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 ZZ 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−2q-2, which vanish on the sphere. Exact expansion bounds this nonnegative sum strictly below −19/100000-19/100000. This contradiction proves the claim. The enumeration used here is the published classification of 22,320 two-qubit vertices. Reichardt, Table 2 · Zurel and Davis, Appendix A

Analytic conclusion and sharpness. If a feasible AA had q>2q>2, mixing it with I/8I/8 would reach q=2q=2 with every nonidentity Pauli expectation strictly smaller than one in absolute value. Every conjugate of V0V_0 has a Pauli expectation of magnitude one. This excludes q>2q>2 and proves c3=1/2c_3=1/2. The contact state ρ∗=(2I−V0)/15\rho_*=(2I-V_0)/15 is positive definite, has purity 2/152/15, and satisfies Tr⁡(V0ρ∗)=0\operatorname{Tr}(V_0\rho_*)=0. Continuing a short distance along the ray from I/8I/8 through ρ∗\rho_* preserves positivity and makes the witness expectation negative. Magic-state purities therefore approach 2/152/15 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 α≥0\alpha\ge0. In its eigenbasis write AA with diagonal blocks B,CB,C of size k=d/2k=d/2 and upper-right block HH, where CC has the smaller trace. Let uu minimize m=⟨u∣B∣u⟩m=\langle u|B|u\rangle over logical stabilizer states, and take b≥∥C∥opb\ge\|C\|_{\rm op}. Complete uu to a Clifford image of the computational basis, with vectors vjv_j. For each vjv_j the following is a stabilizer state for all four phases ω∈{1,−1,i,−i}\omega\in\{1,-1,i,-i\}.

∣0⟩∣u⟩+ω∣1⟩∣vj⟩2.\frac{|0\rangle|u\rangle+\omega|1\rangle|v_j\rangle}{\sqrt2}.

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

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

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

m≤L−x−ℓ+bkL−1.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=4,5,6n=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 c4≥1/2.0070400c_4\ge1/2.0070400, c5≥1/2.0108900c_5\ge1/2.0108900 and c6≥1/2.0129180c_6\ge1/2.0129180. Propagating the weighted bounds through n=7,…,15n=7,\ldots,15 gives c15≥10000000/20151331c_{15}\ge10000000/20151331. The extra off-diagonal constraint enters at n=4,5,6n=4,5,6. The scalar compression and fixed-moment cubic estimates come from Liu and Liu, and the operator-weighted compression is developed in OpenContext 5.7.

Analytic extension from 15 to every larger number of qubits. Set K=215K=2^{15} and U=20151331/10000000U=20151331/10000000. In dimension d≥Kd\ge K, choose a uniform Clifford image Π\Pi of a fixed rank-KK stabilizer-code projector. Its range encodes 15 qubits. The certified 15-qubit bound thus gives Tr⁡[(ΠAΠ)2]≤U[Tr⁡(ΠA)]2\operatorname{Tr}[(\Pi A\Pi)^2]\le U[\operatorname{Tr}(\Pi A)]^2. The Clifford second moment gives E[Π⊗Π]=aI+bF\mathbb E[\Pi\otimes\Pi]=aI+bF, where FF swaps the two copies and b/a=(d−K)/(Kd−1)b/a=(d-K)/(Kd-1). Taking the expectation and solving yields

aq+b≤U(a+bq)⟹cn≥dc∗+1d+c∗,c∗=K−UKU−1.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∗c_*. Substituting K,UK,U gives the fraction in the main result. The finite estimates cover d<Kd<K. This last deduction uses only the Clifford second-moment identity 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.

python3 verify_four.py
python3 verify_five_six.py
python3 verify_transfer.py

Permanent link:open-context.jan-bulling.com/contribution/5.10

Discussion

Sign in to join the discussion.

  • 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

    1. Rafael Wagner · 19 Nov 2026

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