A solution to the problem which uses Aronszajn's inequality
by Rafael Wagner · accepted 20 Nov 2026 · partially AI-assisted
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 Λ-polytopes. Below, I mark all the points where I believe further exposition is needed with a triple star (⋆⋆⋆).
This is the main output:
Theorem (inradius of the stabilizer polytope). Let Λn be the Λ-polytope of n qubits. Denote by d=2n the Hilbert space dimension, and Mn:=A∈ΛnmaxTr[A2]. For every integer n≥1,
Mn=2.
Consequently, for every n-qubit state ρ,
Tr[ρ2]≤d−211⟹ρ∈STABn,
and the constant 21 cannot be improved [1]. Moreover, A∈Λn has Tr[A2]=2 if and only if A=UA⋆U† for a Clifford unitary U, where A⋆=21(I+X+Y+Z)⊗∣0…0⟩⟨0…0∣.
This is the first contribution of two parts. First part proves Mn=2 and shows that every maximiser has largest eigenvalue λ+=21+3. The second part proves that such an operator lies in the Clifford orbit of A⋆ and reports numerical experiments on sub-polytopes of STABn.
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.
Preliminary mathematical ideas
We keep the notation of the previous contributions. Λn is the set of Hermitian operators A with Tr[A]=1 and ⟨s∣A∣s⟩≥0 for all stabilizer states, known as the Λ-polytope [4, 2, 9, 10, 8, 11]. A state belongs to STABn if and only if Tr[Aρ]≥0 for all A∈Λn[4]. By the Cauchy–Schwarz inequality every bound Mn≤m implies that Tr[ρ2]≤1/(d−m1) is sufficient for ρ∈STABn (see OpenContext §5.12, the purity section). Define
v:=λ+=21+3,c:=−λ−=v−1,r:=23.
It is simple to show that (⋆⋆⋆)vc=21,v2=v+21,c2=21−c,v2+c2=2,r=v−21=1−c2.
For A∈Λn we write
ℓ=λmax(A),κ=−λmin(A),q=Tr[A2]=2+θ,aP=Tr[PA].
These quantities are instrumental for the proof. For a Pauli operator P=I let Π±=21(I±P), B±=Π±AΠ± and t±=Tr[B±]=21(1±aP). All sums over P run over the d2−1 Pauli operators P=I.
We use the following results discussed previously in OpenContext §5.12.
Lemma (block lemma [4]). The range of Π± is an (n−1)-qubit code space whose stabilizer states are n-qubit stabilizer states. Hence t±≥0, and B± is t± times an element of Λn−1 (and B±=0 if t±=0). In particular
Tr[B±2]≤Mn−1t±2.
Theorem (spectrum of a witness, previous contribution). For all n≥1 and A∈Λn: −c≤λmin(A) and λmax(A)≤v.
Lemma (Liu–Liu inequality [3]). For n≥2 and A∈Λn: dq+1−2Tr[A3]≤Mn−1(d−q).
We also use the rule for Pauli sums [7]: ∑PTr[PX]Tr[PY]=dTr[XY] and ∑PPXP=dTr[X]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 V1⊕V2 and let H1,H2 be its compressions to V1 and V2. Then
λmax(H)+λmin(H)≤λmax(H1)+λmax(H2).
Proof:K:=H−λmin(H)I≥0. For a unit vector ψ=ψ1+ψ2 with ψj∈Vj, the triangle inequality for K1/2 gives
where the last step is the Cauchy–Schwarz inequality. Taking the supremum over ψ gives λmax(H)−λmin(H)≤k1+k2. □
Block slacks
In this section and the next, n≥2. Also, we assume Mn−1=2, and A∈Λn is fixed in order to structure an induction step later on. For P=I we define the slacks as
s±P:=2t±2−Tr[B±2].(slacks)
We also define the operator W by
W:=2A−A2.
Additionally, we denote wP:=Tr[PW] the expectation value of P on W. With these ingredients at hand, we can state and prove the following lemma (⋆⋆⋆).
Lemma (block slacks). For every P=I we have s±P≥0 and s+P−s−P=wP. Moreover,
The rule for Pauli sums gives ∑PaP2=dq−1 and ∑PcP=d−q. Hence ∑P(s+P+s−P)=d2−2+dq−21(d2q−2q+d), which is the claimed expression for q=2+θ. Finally, ∑PwP2=dTr[W2]−Tr[W]2 and Tr[W]=2−q=−θ. □
Remark. The same computation gives d4∑P(t−s+P+t+s−P)=2d−1−(d+2)q+2Tr[A3](⋆⋆⋆). Since the left-hand side is non-negative, this is the Liu–Liu inequality for Mn−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 Λ-polytope (⋆⋆⋆).
Proposition. Let n≥2, Mn−1=2 and A∈Λn. For every P=I,
λmax(A)+λmin(A)≤2−∣wP∣.
If moreover θ≥0, there is a Pauli operator P=I with
∣wP∣≥32(Tr[W2]−dθ2).
Proof: Replacing P by −P exchanges the two blocks and changes the sign of wP, so we may assume wP≥0. Then s+P=wP+s−P≥wP, that is, Tr[B+2]≤2t+2−wP. The largest eigenvalue of a Hermitian operator is at most its Hilbert–Schmidt norm, so the compressions of A to the ranges of Π+ and Π− have largest eigenvalues at most 2t+2−wP and Tr[B−2]≤2t−=2(1−t+). By Aronszajn's inequality,
λmax(A)+λmin(A)≤g(t+),g(t):=2t2−wP+2(1−t).
Since g′(t)=2t/2t2−wP−2≥0 and t+≤1, we have g(t+)≤g(1)=2−wP.
For the second statement, ∣wP∣=∣s+P−s−P∣≤s+P+s−P, so by the lemma ∑P∣wP∣≤23d when θ≥0. Therefore
For A⋆ (see definition at the top of this document) all inequalities are equalities. To see this, note that W=r(∣φ⟩⟨φ∣−∣χ⟩⟨χ∣) for the two eigenvectors of A⋆, Tr[W2]=23, and ∣wP∣=1 exactly for the 23d Pauli operators that act as logical Pauli operators on the code space of A⋆; these carry the whole slack 23d, and λmax+λmin=1.
A lemma on spectra
Let f(λ):=(2λ−λ2)2, so that Tr[W2]=∑if(λi) for the eigenvalues λi of A(⋆⋆⋆).
Lemma (spectral inequality). Let λ1≥⋯≥λd be real numbers, d≥4, with 45≤λ1≤v and −c≤λd≤0. Put δ=v−λ1, k=c+λd, s=2−λ12−λd2, and assume that σ:=∑1<i<dλi2=s+θ with θ≥0 and σ≤167. Then
Proof: Let a=δ(2c−δ)≥0, b=k(2v−k)≥0 and e=k−δ. Then λ1+λd=1+e and, using v2+c2=2 and v−c=1, s=δ(2v−δ)+k(2c−k)=a+b−2e.
The two extreme eigenvalues.2λ1−λ12=1−(λ1−1)2=r+a, because λ1−1=c−δ and 1−c2=r. Similarly 2λd−λd2=1−(v−k)2=−(r−b), because 1−λd=v−k and v2−1=r. Hence f(λ1)=(r+a)2≥43+3a and f(λd)=(r−b)2=43−3b+b2.
The other eigenvalues. For 1<i<d we have λi≤max(λ2,0)≤σ<2, hence (2−λi)2≥(2−σ)2≥3(1−σ), the last step because (2−σ)2−3(1−σ)=(1−2σ)2. Therefore ∑1<i<df(λi)≥3σ(1−σ). The function u↦3u(1−u) is concave with derivative at least 83 on [0,167], so 3σ(1−σ)≥3s(1−s)+83θ.
Collecting. The left-hand side of the lemma is at least
where we used s=a+b−2e in the linear term. Now a=2cδ−δ2, b=2vk−k2 and (2+32)2c=(2−32)2v=34. Moreover b≥(v+1)k gives 32b2≥(2+3)k2, and 0≤s≤2vδ+2ck gives 2s2≤(8+43)δ2+8δk+(8−43)k2. Inserting these and e2=δ2−2δk+k2 yields Q(δ,k)+4θ.
Positivity of Q. Rounding all coefficients against us, Q≥δ(4.309−17.083δ−10k)+k(0.309+2.814k). Since δ≤v−45<0.1161, the first bracket exceeds 2.325−10k. If k≤0.2325, both terms are non-negative, and they vanish only for k=δ=0. If 0.2325<k≤c<0.3661, the first term is at least −0.1161⋅1.336>−0.156 and the second is at least 0.2325⋅0.963>0.223. □
Proof of the theorem
We prove by induction on n that Mn=2, and that every A∈Λn with Tr[A2]=2 has λmax(A)=λ+. For n=1, A=21(I+r⋅σ) with ∣ri∣≤1, so Tr[A2]=21(1+∣r∣2)≤2, with equality only if ∣r∣=3, i.e. λmax(A)=λ+. Since A⋆∈Λn has purity 2, we have Mn≥2 for all n.
Let n≥2, assume Mn−1=2, and let A∈Λn have q=2+θ with θ≥0. We show that θ=0 and ℓ=v.
Step 1 (location of the spectrum). With Mn−1=2 the Liu–Liu inequality reads 2Tr[A3]≥5+(d+2)θ. Since every eigenvalue satisfies λ3≤ℓλ2, also Tr[A3]≤ℓ(2+θ). Hence
θ(d+2−2ℓ)≤4ℓ−5,
so ℓ≥45>1=Tr[A] and A has a negative eigenvalue. By the theorem on the spectrum of a witness, 45≤ℓ≤v and 0<κ≤c, so δ:=v−ℓ∈[0,v−45] and k:=c−κ∈[0,c). Moreover θ≤(4ℓ−5)/(6−2ℓ)<61, and the sum σ of the squares of the other d−2 eigenvalues satisfies
σ=2+θ−ℓ2−κ2≤2−ℓ2+6−2ℓ4ℓ−5≤167,
because the function of ℓ on the right is decreasing on [45,v] (its derivative is −2ℓ+14/(6−2ℓ)2<0) and equals 167 at ℓ=45.
Step 2 (the key inequality). By Step 1, ℓ−κ≥45−c>0. We square the first inequality of the proposition for the Pauli operator of its second statement:
(ℓ−κ)2≤2−32(Tr[W2]−dθ2).
Step 3 (conclusion). The eigenvalues of A satisfy the assumptions of the lemma on spectra, with λ1=ℓ and λd=−κ. Together with Step 2,
Q(δ,k)+4θ≤32dθ2≤36θ,
where we used θ<61 and d≥4. Since Q(δ,k)≥0, this forces θ=0 and Q(δ,k)=0, hence δ=k=0. So no element of Λn has purity larger than 2, i.e. Mn=2, and every A∈Λn with Tr[A2]=2 has λmax(A)=λ+ (and λmin(A)=λ−). This completes the induction.
The statement about states follows from Mn=2 as recalled in the preliminaries. The second part shows that λmax(A)=λ+ implies A=UA⋆U†. □
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)
Michael Zurel and Jack Davis, Basis-independent stabilizerness and maximally noisy magic states, arXiv:2602.22336 (2026)
Zhenhuan Liu and Zi-Wen Liu, On the geometry, typicality, and detectability of quantum magic, arXiv:2609.03944 (2026)
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)
Ben W. Reichardt, Quantum universality by state distillation, arXiv:quant-ph/0608085 (2006)
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) doi:10.26421/QIC21.13-14-2
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)
Michael Zurel, Cihan Okay, and Robert Raussendorf, Simulating Quantum Computation: How Many “Bits” for “It”?, PRX Quantum 5, 030343 (2024)