Skip to content
OpenContext
Contribution toProblem 5: Test problem: Hilbert–Schmidt inradius of the multiqubit stabilizer polytopeunder revision
0

Partial resultOpenContext 5.7

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=3,4,5n=3,4,5 and also the one valid for all nn. The claimed bound is

Tr[ρ2]≤1dn−0.4859  ⟹  ρ∈STABn.\mathrm{Tr}[\rho^2] \leq \frac{1}{d_n-0.4859} \implies \rho \in \mathrm{STAB}_n.

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

  • For n=3n=3: 1/2.0082≈0.49795831/2.0082 \approx 0.4979583.
  • For n=4n=4: 1/2.0314≈0.49227131/2.0314 \approx 0.4922713.
  • For n=5n=5: 1/2.0444≈0.489141061/2.0444 \approx 0.48914106.

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

1. Setting

Throughout d=2nd=2^n, Λn={A=A†: Tr[A]=1, ⟨s∣A∣s⟩≥0 for all stabilizer states s}\Lambda_n=\{A=A^\dagger:\ \mathrm{Tr}[A]=1,\ \langle s|A|s\rangle\ge 0\ \text{for all stabilizer states } s\}, LnL_n is the cone over Λn\Lambda_n, SnS_n is the cone generated by stabilizer states, and Mn:=max⁡A∈ΛnTr[A2].M_n:=\max_{A\in\Lambda_n}\mathrm{Tr}[A^2]. Elements of Λn\Lambda_n are called (normalised stabilizer) witnesses below. By polarity, every state with Tr[ρ2]≤1/(d−1/Mn)\mathrm{Tr}[\rho^2]\le 1/(d-1/M_n) lies in STABn\mathrm{STAB}_n, and this threshold is sharp; the conjecture is Mn=2M_n=2. It is known that M1=M2=2M_1=M_2=2.

For a Pauli operator P≠IP\neq\mathbb{I} write Π±=(I±P)/2\Pi_\pm=(\mathbb{I}\pm P)/2, aP=Tr[PA]a_P=\mathrm{Tr}[PA], p±=Tr[Π±A]=(1±aP)/2p_\pm=\mathrm{Tr}[\Pi_\pm A]=(1\pm a_P)/2 and B±=Π±AΠ±B_\pm=\Pi_\pm A\Pi_\pm.

Block lemma (known). If A∈ΛnA\in\Lambda_n then B±B_\pm, viewed as an operator on the (n−1)(n-1)-qubit space range Π±\mathrm{range}\,\Pi_\pm, belongs to Ln−1L_{n-1} and has trace p±≥0p_\pm\ge 0.

Let A1=(I+X+Y+Z)/2A_1=(\mathbb{I}+X+Y+Z)/2, with eigenvalues λ±=(1±3)/2\lambda_\pm=(1\pm\sqrt3)/2, and A⋆=A1⊗∣0…0⟩⟨0…0∣A^\star=A_1\otimes|0\dots0\rangle\langle 0\dots0|, the conjectured maximiser (Tr[A⋆2]=2\mathrm{Tr}[A^{\star2}]=2). Two new constants are used: vn:=max⁡A∈Λnλmax⁡(A),cn:=max⁡A∈Λn(−λmin⁡(A)).v_n:=\max_{A\in\Lambda_n}\lambda_{\max}(A),\qquad c_n:=\max_{A\in\Lambda_n}\big(-\lambda_{\min}(A)\big). Dually, vnv_n and cnc_n are the smallest numbers with vnI−ρ∈Snv_n\mathbb{I}-\rho\in S_n and ρ+cnI∈Sn\rho+c_n\mathbb{I}\in S_n for every state ρ\rho. Clearly vn≥λ+v_n\ge\lambda_+ and cn≥−λ−c_n\ge-\lambda_-.

2. Lifting, and the spectral recursion

Lifting principle. Let FF be a homogeneous function with F(B)∈KF(B)\in\mathcal K for every B∈Ln−1B\in L_{n-1}, where K\mathcal K is Sn−1S_{n-1} or the positive-semidefinite cone. Then for any weights wP±≥0w_{P\pm}\ge 0 the operator ∑P≠I∑±wP±F(BP±)\sum_{P\ne\mathbb{I}}\sum_\pm w_{P\pm}F(B_{P\pm}) lies in SnS_n, respectively is positive semidefinite. With K=Sn−1\mathcal K=S_{n-1}, F(B)=Mn−1Tr[B] I−BF(B)=M_{n-1}\mathrm{Tr}[B]\,\mathbb{I}-B and wP±=p∓w_{P\pm}=p_\mp this is the operator form of the Liu–Liu recursion (Lemma A of the previous note). The choice wP±=p∓w_{P\pm}=p_\mp is the only linear weight for which nothing is lost at A⋆A^\star. The new step is to use the same weights with the positive-semidefinite cone.

Lemma (two sums). For Hermitian AA with Tr[A]=1\mathrm{Tr}[A]=1 and q=Tr[A2]q=\mathrm{Tr}[A^2], ∑P≠I∑±p∓ p±Π±=d(d−q)4 I,∑P≠I∑±p∓ Π±AΠ±=d24A+d4I−d2A2.\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 ∑Pp+p−I=∑P1−aP24I\sum_P p_+p_-\mathbb{I}=\sum_P\frac{1-a_P^2}{4}\mathbb{I} and ∑P≠IaP2=dq−1\sum_{P\ne\mathbb{I}}a_P^2=dq-1. For the second, Π+AΠ++Π−AΠ−=12(A+PAP)\Pi_+A\Pi_++\Pi_-A\Pi_-=\tfrac12(A+PAP) and Π+AΠ+−Π−AΠ−=12(PA+AP)\Pi_+A\Pi_+-\Pi_-A\Pi_-=\tfrac12(PA+AP), so the inner sum equals 14(A+PAP)−aP4(PA+AP)\tfrac14(A+PAP)-\tfrac{a_P}{4}(PA+AP). Summing with ∑PPAP=d I\sum_{P}PAP=d\,\mathbb{I} and ∑PaPP=dA\sum_Pa_PP=dA (all Paulis, identity included) gives 14((d2−2)A+d I)−14(2dA2−2A)\tfrac14\big((d^2-2)A+d\,\mathbb{I}\big)-\tfrac14\big(2dA^2-2A\big). □\square

Theorem 1 (spectral recursion). Let n≥2n\ge 2, A∈ΛnA\in\Lambda_n, q=Tr[A2]q=\mathrm{Tr}[A^2]. Then A2−d2A+12(vn−1(d−q)−1)I ⪰ 0,A2−d2A−12(cn−1(d−q)+1)I ⪯ 0.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 vn−1,cn−1v_{n-1},c_{n-1}, for every PP and sign, vn−1p±Π±−B±⪰0v_{n-1}p_\pm\Pi_\pm-B_\pm\succeq 0 and B±+cn−1p±Π±⪰0B_\pm+c_{n-1}p_\pm\Pi_\pm\succeq 0. Multiply by p∓≥0p_\mp\ge 0, sum over P≠IP\ne\mathbb{I} and both signs, and insert the two sums: the results are d2\frac d2 times the first operator and −d2-\frac d2 times the second. □\square

Both inequalities are equalities on the range of A⋆A^\star when vn−1=λ+v_{n-1}=\lambda_+, cn−1=−λ−c_{n-1}=-\lambda_-, because λ±2=λ±+12\lambda_\pm^2=\lambda_\pm+\tfrac12.

Corollary 1 (eigenvalue cap). Let n≥3n\ge 3 and A∈ΛnA\in\Lambda_n. Then λmax⁡(A) ≤ rn(q):=14(d−d2−8(vn−1(d−q)−1))\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 λmin⁡(A)≥−14(d2+8(cn−1(d−q)+1)−d)\lambda_{\min}(A)\ge-\frac14\big(\sqrt{d^2+8(c_{n-1}(d-q)+1)}-d\big). If vn−1=λ+v_{n-1}=\lambda_+, the value of the first quadratic form at λ+\lambda_+ is λ+2(2−q)\tfrac{\lambda_+}{2}(2-q); hence Tr[A2]≥2\mathrm{Tr}[A^2]\ge 2 forces λmax⁡(A)≤λ+\lambda_{\max}(A)\le\lambda_+, with strict inequality if Tr[A2]>2\mathrm{Tr}[A^2]>2. Likewise cn−1=−λ−c_{n-1}=-\lambda_- and Tr[A2]≥2\mathrm{Tr}[A^2]\ge2 force λmin⁡(A)≥λ−\lambda_{\min}(A)\ge\lambda_-.

Proof. Each eigenvalue λ\lambda satisfies λ2−d2λ+12(vn−1(d−q)−1)≥0\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/4d/4. By Proposition 2(c) below all eigenvalues are at most d/4d/4 when d≥8d\ge 8, so they are below the smaller root rn(q)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 (λ−,λ+)(\lambda_-,\lambda_+), provided the spectral constants are sharp one level below.

3. Input data

Proposition 2.

(a) v1=λ+v_1=\lambda_+, c1=−λ−c_1=-\lambda_- (the single-qubit witnesses form a cube).

(b) [computed] M2=2M_2=2, v2=λ+v_2=\lambda_+, c2=−λ−c_2=-\lambda_-.

(c) For every nn and every A∈ΛnA\in\Lambda_n: each eigenvalue λ\lambda of AA satisfies −2d(d+1)−dd+2 ≤ λ ≤ 2d(d+1)+dd+2,-\frac{\sqrt{2d(d+1)}-d}{d+2}\ \le\ \lambda\ \le\ \frac{\sqrt{2d(d+1)}+d}{d+2}, so cn<2−1c_n<\sqrt2-1 and I+2A≻0\mathbb{I}+2A\succ 0; moreover Tr[(I+2A)−1] ≤ Tr[I+2A]=d+2,equivalently∑iλi21+2λi ≤ 1,\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⋆A^\star for every nn.

(d) Consequently, with φ(w)=w+1/w\varphi(w)=w+1/w, any two eigenvalues λi,λj\lambda_i,\lambda_j (i≠ji\ne j) of A∈ΛnA\in\Lambda_n satisfy φ(1+2λi)+φ(1+2λj)≤8\varphi(1+2\lambda_i)+\varphi(1+2\lambda_j)\le 8. Since φ(1+2λ±)=4\varphi(1+2\lambda_\pm)=4: if λmin⁡(A)≤λ−\lambda_{\min}(A)\le\lambda_- then λmax⁡(A)≤λ+\lambda_{\max}(A)\le\lambda_+, and if λmax⁡(A)≥λ+\lambda_{\max}(A)\ge\lambda_+ then λmin⁡(A)≥λ−\lambda_{\min}(A)\ge\lambda_-.

Proof of (b). Tr[A2]\mathrm{Tr}[A^2], λmax⁡(A)\lambda_{\max}(A) and −λmin⁡(A)-\lambda_{\min}(A) are convex, so their maxima over the polytope Λ2\Lambda_2 are attained at vertices. Vertex enumeration returns 22,32022{,}320 vertices (the number reported in the literature), in seven purity classes:

Tr[A2]\mathrm{Tr}[A^2]23/211/843/324/35/47/8
vertices24019257602304576057602304
largest eigenvalue1.36600.80901.16141.15201.14551.05570.7255
smallest eigenvalue−0.3660−0.3090−0.2050−0.0930−0.1455−0.2021−0.2255

The 240240 vertices of purity 22 are the Clifford images of A1⊗∣0⟩⟨0∣A_1\otimes|0\rangle\langle0| and have eigenvalues exactly λ±,0,0\lambda_\pm,0,0. The enumeration was run in floating point; every vertex was then converted to exact rationals (denominators at most 44) and checked exactly to be feasible with a tight set of rank 1515. Completeness rests on the count matching the known one.

Proof of (c). Stabilizer states form a 3-design, so for any Hermitian XX the inequality ∑s⟨s∣X∣s⟩2⟨s∣A∣s⟩≥0\sum_s\langle s|X|s\rangle^2\langle s|A|s\rangle\ge 0 reads (Tr[X])2+Tr[X2]+2Tr[X]Tr[XA]+2Tr[X2A]≥0(\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=I+2AW=\mathbb{I}+2A, Tr[X2W]+Tr[X] Tr[XW] ≥ 0for all Hermitian X.\mathrm{Tr}[X^2W]+\mathrm{Tr}[X]\,\mathrm{Tr}[XW]\ \ge\ 0\qquad\text{for all Hermitian } X . For the eigenvalue bounds take X=xI+ψX=x\mathbb{I}+\psi with Aψ=λψA\psi=\lambda\psi and put u=(d+2)xu=(d+2)x: the inequality becomes d+1d+2u2+u+(1+2λ)(u+2)≥0\frac{d+1}{d+2}u^2+u+(1+2\lambda)(u+2)\ge 0 for all real uu, and optimising over u>−2u>-2 and over u<−2u<-2 gives the two bounds. Hence W≻0W\succ 0, and X=W−1−IX=W^{-1}-\mathbb{I} is admissible: then Tr[X2W]=Tr[W−1]−d+2\mathrm{Tr}[X^2W]=\mathrm{Tr}[W^{-1}]-d+2, Tr[X]=Tr[W−1]−d\mathrm{Tr}[X]=\mathrm{Tr}[W^{-1}]-d and Tr[XW]=−2\mathrm{Tr}[XW]=-2, so the inequality is d+2−Tr[W−1]≥0d+2-\mathrm{Tr}[W^{-1}]\ge 0. The equivalent form uses λ21+2λ=λ2−14+14(1+2λ)\frac{\lambda^2}{1+2\lambda}=\frac\lambda2-\frac14+\frac{1}{4(1+2\lambda)} and ∑iλi=1\sum_i\lambda_i=1. At A⋆A^\star the eigenvalues of WW are 2+32+\sqrt3, 2−32-\sqrt3 and 11, which are closed under w↦1/ww\mapsto 1/w. □\square

Proof of (d). Let wk=1+2λk>0w_k=1+2\lambda_k>0, so ∑kwk=d+2\sum_kw_k=d+2 and ∑k1/wk≤d+2\sum_k1/w_k\le d+2. By Cauchy–Schwarz, ∑k≠i,j1/wk≥(d−2)2d+2−wi−wj≥d−6+wi+wj\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⋆A^\star saturates for every nn; the conjecture is the unweighted version ∑iλi2≤2\sum_i\lambda_i^2\le 2. In dual form the eigenvalue bound says that ρ+2d(d+1)−dd+2I∈Sn\rho+\frac{\sqrt{2d(d+1)}-d}{d+2}\mathbb{I}\in S_n for every state ρ\rho: white noise with visibility about 2.41/d2.41/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\sum_i\lambda_i=1, ∑iλi2=q\sum_i\lambda_i^2=q and largest eigenvalue ℓ\ell: Tr[A3]≤ℓq,Tr[A3]≤ℓ3+(q−ℓ2)3/2,\mathrm{Tr}[A^3]\le\ell q,\qquad \mathrm{Tr}[A^3]\le\ell^3+(q-\ell^2)^{3/2}, and ℓ↦min⁡{ℓq, ℓ3+(q−ℓ2)3/2}\ell\mapsto\min\{\ell q,\ \ell^3+(q-\ell^2)^{3/2}\} is non-decreasing on (0,q ](0,\sqrt q\,]. The second bound can be sharpened with Wilkins' skewness inequality (1944) applied to the other d−1d-1 eigenvalues: with μ=1−ℓd−1\mu=\frac{1-\ell}{d-1} and S=q−ℓ2−(1−ℓ)2d−1≥0S=q-\ell^2-\frac{(1-\ell)^2}{d-1}\ge 0, Tr[A3]≤τd(ℓ,q):=ℓ3+(d−1)μ3+3μS+d−3(d−1)(d−2) S3/2.\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 Mˉ2=2\bar M_2=2, vˉ2=λ+\bar v_2=\lambda_+ and, for n≥3n\ge 3,

  • rn(q)=14(d−d2−8(vˉn−1(d−q)−1))r_n(q)=\frac14\big(d-\sqrt{d^2-8(\bar v_{n-1}(d-q)-1)}\big),
  • Tn(q)=max⁡{min⁡{ℓq, τd(ℓ,q)}: 1d≤ℓ≤min⁡{rn(q),q}, S≥0}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\},
  • Mˉn=sup⁡{q≥Mˉn−1: d(q−Mˉn−1)≤2Tn(q)−Mˉn−1q−1}\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\},
  • vˉn=\bar v_n= the solution ℓ\ell of ℓ=rn(ℓ2+(1−ℓ)2d−1)\ell=r_n\big(\ell^2+\frac{(1-\ell)^2}{d-1}\big). Then Mn≤MˉnM_n\le\bar M_n and vn≤vˉnv_n\le\bar v_n for all nn. [computed] Evaluating the recursion:
nnLiu–Liu bound on MnM_nthis note, elementarythis note, with Wilkinsvˉn\bar v_n
32.09382.01502.00821.3906
42.13842.03802.03141.4020
52.16032.05062.04441.4076
62.17122.05702.05111.4103
all nn2.18192.06352.05781.4130

("Elementary" replaces τd\tau_d by ℓ3+(q−ℓ2)3/2\ell^3+(q-\ell^2)^{3/2}, in which case Tn(q)=rn(q)3+(q−rn(q)2)3/2T_n(q)=r_n(q)^3+(q-r_n(q)^2)^{3/2} in closed form.)

Proof. Induction on nn. Let A∈ΛnA\in\Lambda_n with purity qq and largest eigenvalue ℓ\ell.

  • (i) Corollary 1 gives ℓ≤rn(q)\ell\le r_n(q); replacing vn−1v_{n-1} by the larger vˉn−1\bar v_{n-1} only weakens Theorem 1.
  • (ii) Hence Tr[A3]≤Tn(q)\mathrm{Tr}[A^3]\le T_n(q) by the spectral facts above.
  • (iii) The Liu–Liu recursion states d(q−Mn−1)≤2Tr[A3]−Mn−1q−1d(q-M_{n-1})\le 2\mathrm{Tr}[A^3]-M_{n-1}q-1, and it remains valid with Mn−1M_{n-1} replaced by Mˉn−1\bar M_{n-1} because q≤dq\le d. So either q≤Mˉn−1q\le\bar M_{n-1} or qq satisfies the inequality defining Mˉn\bar M_n.
  • (iv) Always q≥ℓ2+(1−ℓ)2d−1q\ge\ell^2+\frac{(1-\ell)^2}{d-1}, and rnr_n is decreasing, so ℓ≤rn(ℓ2+(1−ℓ)2d−1)\ell\le r_n\big(\ell^2+\frac{(1-\ell)^2}{d-1}\big), i.e. ℓ≤vˉn\ell\le\bar v_n. □\square Corollary 2. For every nn: every state with Tr[ρ2]≤1/(d−0.4859)\mathrm{Tr}[\rho^2]\le 1/(d-0.4859) is a stabilizer mixture; the inradius of STABn−I/d\mathrm{STAB}_n-\mathbb{I}/d is at least 1/d(2.0578 d−1)1/\sqrt{d(2.0578\,d-1)}, within 1.5%1.5\% of the conjectured 1/d(2d−1)1/\sqrt{d(2d-1)}; and every eigenvalue of every A∈ΛnA\in\Lambda_n lies in (−0.4143, 1.4130)(-0.4143,\ 1.4130). For three qubits the purity threshold is at least 1/(8−0.4979)1/(8-0.4979).

Two remarks on robustness. If one does not want to rely on the two-qubit enumeration for v2v_2 and uses only v2≤14+34⋅74v_2\le\frac14+\sqrt{\tfrac34\cdot\tfrac74} (which follows from M2=2M_2=2), the same recursion still gives Mn≤2.1183M_n\le 2.1183, already below Liu–Liu. And running the same recursion without the cap reproduces their values (67/3267/32 for three qubits, 2.1818472.181847 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∈ΛnA\in\Lambda_n, 2Tr[A3]≥Mnq+1−2d(Mn−q)2\mathrm{Tr}[A^3]\ge M_nq+1-2d(M_n-q). In particular a maximiser (q=Mnq=M_n) has Tr[A3]≥(Mn2+1)/2≥52\mathrm{Tr}[A^3]\ge(M_n^2+1)/2\ge\tfrac52, therefore 1.34≤λmax⁡(A)≤1.41301.34\le\lambda_{\max}(A)\le 1.4130, and then λmin⁡(A)≥−0.368\lambda_{\min}(A)\ge-0.368.

Proof. A⊗∣0⟩⟨0∣∈Λn+1A\otimes|0\rangle\langle0|\in\Lambda_{n+1} has the same qq and Tr[A3]\mathrm{Tr}[A^3]; apply the Liu–Liu recursion at level n+1n+1. For the window of λmax⁡\lambda_{\max} combine with Tr[A3]≤ℓ3+(q−ℓ2)3/2\mathrm{Tr}[A^3]\le\ell^3+(q-\ell^2)^{3/2} and Theorem 3; the bound on λmin⁡\lambda_{\min} then follows from Proposition 2(d). □\square

Routes that do not work

  • A trace-norm bound. ∥A∥1≤3\|A\|_1\le\sqrt3 on Λn\Lambda_n would imply both (E) and Mn=2M_n=2, but it is false already for two qubits (the 192192 vertices of purity 3/23/2 have ∥A∥1=5\|A\|_1=\sqrt5).
  • Uniform (unweighted) lifting. In the basis (I, qI, A, A2)(\mathbb{I},\,q\mathbb{I},\,A,\,A^2) the unweighted block sum acts as d2−44(id+dd2−4Ψ)\frac{d^2-4}{4}\big(\mathrm{id}+\frac{d}{d^2-4}\Psi\big), where Ψ\Psi has eigenvalues 3,−3,0,03,-3,0,0. It reproduces qI−Aq\mathbb{I}-A and I−A2\mathbb{I}-A^2 exactly and distorts everything else at relative order 1/d1/d, so every unweighted lift loses O(1/d)O(1/d) per level and the losses add up to O(1)O(1).
  • Tight lifting of quadratic statements. With weights p∓p_\mp a quadratic block statement lifts to a cubic one containing operators that are not polynomials in AA (for example 1d∑PaP PAPAP\frac1d\sum_Pa_P\,PAPAP). The resulting chain of invariants S1≥S2/d≥S3/d2S_1\ge S_2/d\ge S_3/d^2 is valid whenever the lower rung is non-negative on blocks, but S3S_3 is negative somewhere on Λ2\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 λ2\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 so(6)\mathfrak{so}(6). With five anticommuting Paulis γ1,…,γ5\gamma_1,\dots,\gamma_5, set Γi6=γi\Gamma_{i6}=\gamma_i and Γij=−iγiγj\Gamma_{ij}=-\mathrm{i}\gamma_i\gamma_j. These are the 1515 non-identity Paulis, and aij=Tr[ΓijA]a_{ij}=\mathrm{Tr}[\Gamma_{ij}A] is a real antisymmetric 6×66\times6 matrix. The 6060 stabilizer states correspond to the perfect matchings of six points with signs of fixed parity, the eigenvalues of AA are (1±μ1±μ2±μ3)/4(1\pm\mu_1\pm\mu_2\pm\mu_3)/4 with a fixed sign parity, where ±iμk\pm\mathrm{i}\mu_k are the eigenvalues of aa, and Tr[A3]=12 Tr[A2]−216+38 Pf(a).\mathrm{Tr}[A^3]=\frac{12\,\mathrm{Tr}[A^2]-2}{16}+\frac38\,\mathrm{Pf}(a). On each of the 1515 lines the three coefficients lie (after fixing signs) in a tetrahedron inscribed in the elliptope x2+y2+z2≤1+2xyzx^2+y^2+z^2\le 1+2xyz; summing over lines gives Tr[A2]≤32+16Pf(a)\mathrm{Tr}[A^2]\le\frac32+\frac16\mathrm{Pf}(a), which is exactly the Liu–Liu recursion at n=2n=2. So M2=2M_2=2 follows from Pf(a)≤3\mathrm{Pf}(a)\le 3 on Λ2\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 33 is (1+q)I+2A+2A2(1+q)\mathbb{I}+2A+2A^2, the element of SnS_n produced by the 3-design.
  3. Fourth moment of stabilizer states. For a pure state ψ\psi, ∑s⟨s∣ψ∣s⟩3 ∣s⟩⟨s∣ ∝ I+3ψ+ψ∘3,ψ∘3:=1d∑P⟨ψ∣P∣ψ⟩3P=1d∑P⟨ψ∣P∣ψ⟩2 PψP.\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 ψ∘3\psi^{\circ3} is the only trace of the "not a 4-design" property at this order.
  4. Closure properties. Λn\Lambda_n is closed under the triple Pauli-coefficient product (aPbPcP)(a_Pb_Pc_P) and under the 22n2^{2n} involutions aP↦(−1)Q(P)aPa_P\mapsto(-1)^{Q(P)}a_P, QQ a quadratic form polarising to the symplectic form (transposition followed by a Pauli conjugation).
  5. Two-qubit pairings. From the vertex list, −1≤Tr[AB]≤2-1\le\mathrm{Tr}[AB]\le 2 on Λ2×Λ2\Lambda_2\times\Lambda_2, so I+A∈S2\mathbb{I}+A\in S_2 as well as 2I−A∈S22\mathbb{I}-A\in S_2.

Permanent link:open-context.jan-bulling.com/contribution/5.7

Discussion

Sign in to join the discussion.

  • 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(1)2+2=4\tag{1}.

    1. Jan · 18 Nov 2026

      Replying to my own comment!

    2. Rafael Wagner · 19 Nov 2026

      Nice!

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