An improved bound relative to Liu and Liu's result: from approximately 0.4583 to 0.4589.
by Rafael Wagner · accepted 18 Nov 2026 · fully AI-generated
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
Tr[ρ2]≤dn−0.48591⟹ρ∈STABn.
Furthermore, the values for n=3,4,5 are all improved by numerical checks:
For n=3: 1/2.0082≈0.4979583.
For n=4: 1/2.0314≈0.4922713.
For n=5: 1/2.0444≈0.48914106.
The results below have not been polished nor revised by a human.
1. Setting
Throughout d=2n, Λn={A=A†:Tr[A]=1,⟨s∣A∣s⟩≥0for all stabilizer states s}, Ln is the cone over Λn, Sn is the cone generated by stabilizer states, and
Mn:=A∈ΛnmaxTr[A2].
Elements of Λn are called (normalised stabilizer) witnesses below. By polarity, every state with Tr[ρ2]≤1/(d−1/Mn) lies in STABn, and this threshold is sharp; the conjecture is Mn=2. It is known that M1=M2=2.
For a Pauli operator P=I write Π±=(I±P)/2, aP=Tr[PA], p±=Tr[Π±A]=(1±aP)/2 and B±=Π±AΠ±.
Block lemma (known). If A∈Λn then B±, viewed as an operator on the (n−1)-qubit space rangeΠ±, belongs to Ln−1 and has trace p±≥0.
Let A1=(I+X+Y+Z)/2, with eigenvalues λ±=(1±3)/2, and A⋆=A1⊗∣0…0⟩⟨0…0∣, the conjectured maximiser (Tr[A⋆2]=2). Two new constants are used:
vn:=A∈Λnmaxλmax(A),cn:=A∈Λnmax(−λmin(A)).
Dually, vn and cn are the smallest numbers with vnI−ρ∈Sn and ρ+cnI∈Sn for every state ρ. Clearly vn≥λ+ and cn≥−λ−.
2. Lifting, and the spectral recursion
Lifting principle. Let F be a homogeneous function with F(B)∈K for every B∈Ln−1, where K is Sn−1 or the positive-semidefinite cone. Then for any weights wP±≥0 the operator ∑P=I∑±wP±F(BP±) lies in Sn, respectively is positive semidefinite. With K=Sn−1, F(B)=Mn−1Tr[B]I−B and wP±=p∓ this is the operator form of the Liu–Liu recursion (Lemma A of the previous note). The choice wP±=p∓ is the only linear weight for which nothing is lost at A⋆. The new step is to use the same weights with the positive-semidefinite cone.
Lemma (two sums). For Hermitian A with Tr[A]=1 and q=Tr[A2],
P=I∑±∑p∓p±Π±=4d(d−q)I,P=I∑±∑p∓Π±AΠ±=4d2A+4dI−2dA2.
Proof. The first sum is ∑Pp+p−I=∑P41−aP2I and ∑P=IaP2=dq−1. For the second, Π+AΠ++Π−AΠ−=21(A+PAP) and Π+AΠ+−Π−AΠ−=21(PA+AP), so the inner sum equals 41(A+PAP)−4aP(PA+AP). Summing with ∑PPAP=dI and ∑PaPP=dA (all Paulis, identity included) gives 41((d2−2)A+dI)−41(2dA2−2A). □
Theorem 1 (spectral recursion). Let n≥2, A∈Λn, q=Tr[A2]. Then
A2−2dA+21(vn−1(d−q)−1)I⪰0,A2−2dA−21(cn−1(d−q)+1)I⪯0.
Proof. By the block lemma and the definition of vn−1,cn−1, for every P and sign, vn−1p±Π±−B±⪰0 and B±+cn−1p±Π±⪰0. Multiply by p∓≥0, sum over P=I and both signs, and insert the two sums: the results are 2d times the first operator and −2d times the second. □
Both inequalities are equalities on the range of A⋆ when vn−1=λ+, cn−1=−λ−, because λ±2=λ±+21.
Corollary 1 (eigenvalue cap). Let n≥3 and A∈Λn. Then
λmax(A)≤rn(q):=41(d−d2−8(vn−1(d−q)−1))
whenever the radicand is non-negative, and λmin(A)≥−41(d2+8(cn−1(d−q)+1)−d). If vn−1=λ+, the value of the first quadratic form at λ+ is 2λ+(2−q); hence Tr[A2]≥2 forces λmax(A)≤λ+, with strict inequality if Tr[A2]>2. Likewise cn−1=−λ− and Tr[A2]≥2 force λmin(A)≥λ−.
Proof. Each eigenvalue λ satisfies λ2−2dλ+21(vn−1(d−q)−1)≥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≥8, so they are below the smaller root rn(q). The other claims are the same computation. □
So a hypothetical counterexample at the first level where the conjecture fails would have its whole spectrum strictly inside (λ−,λ+), provided the spectral constants are sharp one level below.
3. Input data
Proposition 2.
(a) v1=λ+, c1=−λ− (the single-qubit witnesses form a cube).
(b) [computed]M2=2, v2=λ+, c2=−λ−.
(c) For every n and every A∈Λn: each eigenvalue λ of A satisfies
−d+22d(d+1)−d≤λ≤d+22d(d+1)+d,
so cn<2−1 and I+2A≻0; moreover
Tr[(I+2A)−1]≤Tr[I+2A]=d+2,equivalentlyi∑1+2λiλi2≤1,
with equality at A⋆ for every n.
(d) Consequently, with φ(w)=w+1/w, any two eigenvalues λi,λj (i=j) of A∈Λn satisfy φ(1+2λi)+φ(1+2λj)≤8. Since φ(1+2λ±)=4: if λmin(A)≤λ− then λmax(A)≤λ+, and if λmax(A)≥λ+ then λmin(A)≥λ−.
Proof of (b).Tr[A2], λmax(A) and −λmin(A) are convex, so their maxima over the polytope Λ2 are attained at vertices. Vertex enumeration returns 22,320 vertices (the number reported in the literature), in seven purity classes:
Tr[A2]
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 A1⊗∣0⟩⟨0∣ and have eigenvalues exactly λ±,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 ∑s⟨s∣X∣s⟩2⟨s∣A∣s⟩≥0 reads (Tr[X])2+Tr[X2]+2Tr[X]Tr[XA]+2Tr[X2A]≥0, that is, with W=I+2A,
Tr[X2W]+Tr[X]Tr[XW]≥0for all Hermitian X.
For the eigenvalue bounds take X=xI+ψ with Aψ=λψ and put u=(d+2)x: the inequality becomes d+2d+1u2+u+(1+2λ)(u+2)≥0 for all real u, and optimising over u>−2 and over u<−2 gives the two bounds. Hence W≻0, and X=W−1−I is admissible: then Tr[X2W]=Tr[W−1]−d+2, Tr[X]=Tr[W−1]−d and Tr[XW]=−2, so the inequality is d+2−Tr[W−1]≥0. The equivalent form uses 1+2λλ2=2λ−41+4(1+2λ)1 and ∑iλi=1. At A⋆ the eigenvalues of W are 2+3, 2−3 and 1, which are closed under w↦1/w. □
Proof of (d). Let wk=1+2λk>0, so ∑kwk=d+2 and ∑k1/wk≤d+2. By Cauchy–Schwarz, ∑k=i,j1/wk≥d+2−wi−wj(d−2)2≥d−6+wi+wj. □
Statement (c) is an unconditional "weighted purity" bound that A⋆ saturates for every n; the conjecture is the unweighted version ∑iλi2≤2. In dual form the eigenvalue bound says that ρ+d+22d(d+1)−dI∈Sn for every state ρ: 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 ∑iλi=1, ∑iλi2=q and largest eigenvalue ℓ:
Tr[A3]≤ℓq,Tr[A3]≤ℓ3+(q−ℓ2)3/2,
and ℓ↦min{ℓq,ℓ3+(q−ℓ2)3/2} is non-decreasing on (0,q]. The second bound can be sharpened with Wilkins' skewness inequality (1944) applied to the other d−1 eigenvalues: with μ=d−11−ℓ and S=q−ℓ2−d−1(1−ℓ)2≥0,
Tr[A3]≤τd(ℓ,q):=ℓ3+(d−1)μ3+3μS+(d−1)(d−2)d−3S3/2.
vˉn= the solution ℓ of ℓ=rn(ℓ2+d−1(1−ℓ)2).
Then Mn≤Mˉn and vn≤vˉn for all n. [computed] Evaluating the recursion:
n
Liu–Liu bound on Mn
this note, elementary
this note, with Wilkins
vˉn
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 τd by ℓ3+(q−ℓ2)3/2, in which case Tn(q)=rn(q)3+(q−rn(q)2)3/2 in closed form.)
Proof. Induction on n. Let A∈Λn with purity q and largest eigenvalue ℓ.
(i) Corollary 1 gives ℓ≤rn(q); replacing vn−1 by the larger vˉn−1 only weakens Theorem 1.
(ii) Hence Tr[A3]≤Tn(q) by the spectral facts above.
(iii) The Liu–Liu recursion states d(q−Mn−1)≤2Tr[A3]−Mn−1q−1, and it remains valid with Mn−1 replaced by Mˉn−1 because q≤d. So either q≤Mˉn−1 or q satisfies the inequality defining Mˉn.
(iv) Always q≥ℓ2+d−1(1−ℓ)2, and rn is decreasing, so ℓ≤rn(ℓ2+d−1(1−ℓ)2), i.e. ℓ≤vˉn. □Corollary 2. For every n: every state with Tr[ρ2]≤1/(d−0.4859) is a stabilizer mixture; the inradius of STABn−I/d is at least 1/d(2.0578d−1), within 1.5% of the conjectured 1/d(2d−1); and every eigenvalue of every A∈Λn 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 v2 and uses only v2≤41+43⋅47 (which follows from M2=2), the same recursion still gives Mn≤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∈Λn, 2Tr[A3]≥Mnq+1−2d(Mn−q). In particular a maximiser (q=Mn) has Tr[A3]≥(Mn2+1)/2≥25, therefore 1.34≤λmax(A)≤1.4130, and then λmin(A)≥−0.368.
Proof.A⊗∣0⟩⟨0∣∈Λn+1 has the same q and Tr[A3]; apply the Liu–Liu recursion at level n+1. For the window of λmax combine with Tr[A3]≤ℓ3+(q−ℓ2)3/2 and Theorem 3; the bound on λmin then follows from Proposition 2(d). □
Routes that do not work
A trace-norm bound.∥A∥1≤3 on Λn would imply both (E) and Mn=2, but it is false already for two qubits (the 192 vertices of purity 3/2 have ∥A∥1=5).
Uniform (unweighted) lifting. In the basis (I,qI,A,A2) the unweighted block sum acts as 4d2−4(id+d2−4dΨ), where Ψ has eigenvalues 3,−3,0,0. It reproduces qI−A and I−A2 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∓ a quadratic block statement lifts to a cubic one containing operators that are not polynomials in A (for example d1∑PaPPAPAP). The resulting chain of invariants S1≥S2/d≥S3/d2 is valid whenever the lower rung is non-negative on blocks, but S3 is negative somewhere on Λ2, 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 λ2.
6. Clifford-algebra and group-theoretic remarks
These are exact statements found while looking for structure; all were verified numerically on random instances.
Two qubits are so(6). With five anticommuting Paulis γ1,…,γ5, set Γi6=γi and Γij=−iγiγj. These are the 15 non-identity Paulis, and aij=Tr[ΓijA] is a real antisymmetric 6×6 matrix. The 60 stabilizer states correspond to the perfect matchings of six points with signs of fixed parity, the eigenvalues of A are (1±μ1±μ2±μ3)/4 with a fixed sign parity, where ±iμk are the eigenvalues of a, and
Tr[A3]=1612Tr[A2]−2+83Pf(a).
On each of the 15 lines the three coefficients lie (after fixing signs) in a tetrahedron inscribed in the elliptope x2+y2+z2≤1+2xyz; summing over lines gives Tr[A2]≤23+61Pf(a), which is exactly the Liu–Liu recursion at n=2. So M2=2 follows from Pf(a)≤3 on Λ2, which holds numerically. I did not find a proof of this Pfaffian bound by hand.
The uniform pinching operator described in Section 5.3; its eigenvector for eigenvalue 3 is (1+q)I+2A+2A2, the element of Sn produced by the 3-design.
Fourth moment of stabilizer states. For a pure state ψ,
s∑⟨s∣ψ∣s⟩3∣s⟩⟨s∣∝I+3ψ+ψ∘3,ψ∘3:=d1P∑⟨ψ∣P∣ψ⟩3P=d1P∑⟨ψ∣P∣ψ⟩2PψP.
The state ψ∘3 is the only trace of the "not a 4-design" property at this order.
Closure properties.Λn is closed under the triple Pauli-coefficient product (aPbPcP) and under the 22n involutions aP↦(−1)Q(P)aP, Q a quadratic form polarising to the symplectic form (transposition followed by a Pauli conjugation).
Two-qubit pairings. From the vertex list, −1≤Tr[AB]≤2 on Λ2×Λ2, so I+A∈S2 as well as 2I−A∈S2.