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

Partial resultOpenContext 5.12

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

by Rafael Wagner · accepted 19 Nov 2026 · partially AI-assisted

Main result

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

Theorem (0.49960.4996 bound): Let ρ∈D(C2⊗⋯⊗C2⏟n times)\rho \in \mathcal{D}(\underbrace{\mathbb{C}^2 \otimes \cdots \otimes \mathbb{C}^2}_{n\text{ times}}) be an arbitrary nn-qubit state, and denote STABn\mathrm{STAB}_n the convex polytope of nn-qubit stabilizer states. For every integer n≥1n \geq 1, it holds that

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

where dn=2nd_n = 2^n is the dimension of the underlying Hilbert space and a∗=0.4996a^*=0.4996. Furthermore, for 1≤n≤51\leq n \leq 5 the conjecture is shown to hold, i.e., a∗=12a^*=\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", 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 Λ\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=5n=5, and the constant 0.49960.4996 for all nn. 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≤4n\le4 and the weaker constant a∗=0.4926a^*=0.4926 for all nn. All the inequalities used in the two parts were also tested numerically on elements of Λn\Lambda_n for n=2,3,4n=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 Λ\Lambda-polytope from Ref. [4]. For more on Λ\Lambda-polytopes we refer to Refs. [8, 9, 10, 11, 4]. It can be defined as follows:

Λn={A=A†: Tr[A]=1, ⟨s∣A∣s⟩≥0 for all stabilizer states ∣s⟩}.\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 nn-qubit Hilbert spaces C2⊗⋯⊗C2\mathbb{C}^2 \otimes \cdots \otimes \mathbb{C}^2 of dimension dn=2nd_n = 2^n. 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\langle0\dots0|. Throughout this document we define

v:=λ+,c:=−λ−=v−1.v:=\lambda_+,\qquad c:=-\lambda_-=v-1.

With these definitions, it is simple to show that the following equations all hold:

2v2=2v+1,vc=12,v2+c2=2,v3−c3=52.\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∈ΛnA\in\Lambda_n, define q=Tr[A2]q=\mathrm{Tr}[A^2], Mn=max⁡{Tr[A2]:A∈Λn}M_n=\max\{\mathrm{Tr}[A^2]:A\in\Lambda_n\}, and t=Tr[A3]t=\mathrm{Tr}[A^3]. Then,

dq+1−2t ≤ Mn−1 (d−q).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

Tr[ρ2]≤1d−1Mn\mathrm{Tr}[\rho^2]\le \frac{1}{d-\frac{1}{M_n}}

is sufficient as a condition for ρ\rho to be a stabilizer mixture and the conjecture is Mn=2M_n=2 for all n∈Nn\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 A1=(I+X+Y+Z)/2A_1=(\mathbb{I}+X+Y+Z)/2, with eigenvalues λ±\lambda_\pm. For all n≥1n\ge1 and A∈ΛnA\in\Lambda_n it holds that λ−≤λmin⁡(A)\lambda_-\le\lambda_{\min}(A) and λmax⁡(A)≤λ+\lambda_{\max}(A)\le\lambda_+. If λmax⁡(A)=λ+\lambda_{\max}(A)=\lambda_+ or λmin⁡(A)=λ−\lambda_{\min}(A)=\lambda_-, then AA has eigenvalues λ+,λ−\lambda_+,\lambda_- and d−2d-2 zeros.

For a witness A∈ΛnA\in\Lambda_n we write

ℓ=λmax⁡(A)=v−δ,κ=−λmin⁡(A)=c−k,q=Tr[A2]=2+θ,t=Tr[A3].\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, δ≥0\delta\ge0 and k≥0k\ge0, and for A⋆A^\star we have δ=k=θ=0\delta=k=\theta=0.

Lemma (the star inequality near A⋆A^\star). Let n≥2n\ge2 and let A∈ΛnA\in\Lambda_n have a negative eigenvalue. Then

d v2 θ+d v k ≤ δ [(c−k) D−d (v+3)]+δ2 [D+4d(v−1)+4],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)−4D:=d(d+2-2v)-4.

Purity

We now turn the results of the previous section into bounds on Mn=max⁡{Tr[A2]:A∈Λn}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 ρ\rho and A∈ΛnA\in\Lambda_n, the Cauchy–Schwarz inequality gives

Tr[Aρ]=1d+Tr[(A−Id)(ρ−Id)] ≥ 1d−(Tr[A2]−1d)(Tr[ρ2]−1d),\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∈ΛnA\in\Lambda_n as soon as Tr[ρ2]≤1/(d−1/Mn)\mathrm{Tr}[\rho^2]\le1/(d-1/M_n). Since ρ∈STABn\rho\in\mathrm{STAB}_n if and only if Tr[Aρ]≥0\mathrm{Tr}[A\rho]\ge0 for all A∈ΛnA\in\Lambda_n, we conclude that every upper bound Mn≤mM_n\le m proves the main theorem with a∗=1/ma^*=1/m. In particular, Mn=2M_n=2 gives the conjectured constant a∗=12a^*=\tfrac12. On the other hand, Mn≥2M_n\ge2 for every nn, because A⋆∈ΛnA^\star\in\Lambda_n and Tr[(A⋆)2]=2\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 tt 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≥2n\ge2 and A∈ΛnA\in\Lambda_n. Then

t ≤ ℓ q.t\ \le\ \ell\,q .

If, moreover, AA has a negative eigenvalue, put

s1=1−ℓ+κ,s2=q−ℓ2−κ2,S=s2−s12d−2.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≥0S\ge0 and

t ≤ ℓ3−κ3+s23/2,t ≤ ℓ3−κ3+W,t\ \le\ \ell^3-\kappa^3+s_2^{3/2},\qquad t\ \le\ \ell^3-\kappa^3+W,

where

W=s13(d−2)2+3 s1d−2 S+d−4(d−2)(d−3) S3/2.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 λ\lambda of AA satisfies λ3≤ℓ λ2\lambda^3\le\ell\,\lambda^2, and we sum over the eigenvalues. For the other two, let μ1,…,μd−2\mu_1,\dots,\mu_{d-2} be the eigenvalues of AA other than ℓ\ell and −κ-\kappa. Since Tr[A]=1\mathrm{Tr}[A]=1 and Tr[A2]=q\mathrm{Tr}[A^2]=q, we have ∑iμi=s1\sum_i\mu_i=s_1 and ∑iμi2=s2\sum_i\mu_i^2=s_2, so that t=ℓ3−κ3+∑iμi3t=\ell^3-\kappa^3+\sum_i\mu_i^3 and S=∑i(μi−μˉ)2≥0S=\sum_i(\mu_i-\bar\mu)^2\ge0 with μˉ=s1/(d−2)\bar\mu=s_1/(d-2). The second bound follows from ∑iμi3≤max⁡i∣μi∣ ∑iμi2≤s23/2\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 NN real numbers with mean μˉ\bar\mu satisfy ∑i(μi−μˉ)3≤N−2N(N−1) S3/2\sum_i(\mu_i-\bar\mu)^3\le\frac{N-2}{\sqrt{N(N-1)}}\,S^{3/2}; with N=d−2N=d-2 and ∑iμi3=Nμˉ3+3μˉ S+∑i(μi−μˉ)3\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 WW. □\square

3.2 The conjecture for n≤5n\le5

Theorem (the conjecture for n≤5n\le5). For 1≤n≤51\le n\le5 we have Mn=2M_n=2. Consequently, for these nn every state with Tr[ρ2]≤1/(d−12)\mathrm{Tr}[\rho^2]\le1/(d-\tfrac12) belongs to STABn\mathrm{STAB}_n, and the constant 12\tfrac12 cannot be improved [1]. The cases n≤4n\le4 are proved analytically; the case n=5n=5 is computer-assisted.

Proof: For n=1n=1 we have Tr[A2]=(1+∣r∣2)/2≤2\mathrm{Tr}[A^2]=(1+|r|^2)/2\le2. Let now 2≤n≤52\le n\le5, assume that Mn−1=2M_{n-1}=2, and suppose that some A∈ΛnA\in\Lambda_n has purity q=2+θq=2+\theta with θ>0\theta>0. We show that this is impossible. With Mn−1=2M_{n-1}=2 the Liu–Liu inequality reads

t ≥ 52+d+22 θ.t\ \ge\ \frac52+\frac{d+2}{2}\,\theta .

We combine it with the first two bounds on the third moment.

(a) From t≤ℓqt\le\ell q we get θ (d+2−2ℓ)≤4ℓ−5\theta\,(d+2-2\ell)\le4\ell-5. Hence ℓ>54\ell>\tfrac54, that is δ<v−54<0.1161\delta<v-\tfrac54<0.1161, and θ≤(4v−5)/(d+2−2v)\theta\le(4v-5)/(d+2-2v). In particular ℓ>1=Tr[A]\ell>1=\mathrm{Tr}[A], so AA has a negative eigenvalue, and both the lemma of Section 3.1 and the star inequality near A⋆A^\star apply.

(b) From t≤ℓ3−κ3+s23/2t\le\ell^3-\kappa^3+s_2^{3/2} and v3−c3=52v^3-c^3=\tfrac52 we get

d+22 θ ≤ −(v3−ℓ3)+(c3−κ3)+s23/2,s2=θ+(v2−ℓ2)+(c2−κ2).\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=2n=2 and n=3n=3. Since k≥0k\ge0 and θ>0\theta>0, the star inequality near A⋆A^\star requires δ [c D−d(v+3)]+δ2 [D+4d(v−1)+4]>0\delta\,[c\,D-d(v+3)]+\delta^2\,[D+4d(v-1)+4]>0. For d=4d=4 the two brackets are −14.14-14.14 and 18.9318.93, so δ>0.74\delta>0.74. For d=8d=8 they are −15.11-15.11 and 69.8669.86, so δ>0.21\delta>0.21. Both contradict (a). Hence M2=2M_2=2 (from M1=2M_1=2), and then M3=2M_3=2.

The case n=4n=4. Here D=240.29D=240.29, c D−d(v+3)=18.09c\,D-d(v+3)=18.09 and D+4d(v−1)+4=267.71D+4d(v-1)+4=267.71. By (a), 267.71 δ2<31.1 δ267.71\,\delta^2<31.1\,\delta, so the star inequality near A⋆A^\star (dropping the term −kDδ≤0-kD\delta\le0) gives 29.86 θ+21.86 k<49.2 δ29.86\,\theta+21.86\,k<49.2\,\delta, that is,

δ > 0.606 θ+0.444 k.\delta\ >\ 0.606\,\theta+0.444\,k .

In (b) we use v3−ℓ3≥3ℓ2δ≥4.6875 δv^3-\ell^3\ge3\ell^2\delta\ge4.6875\,\delta, c3−κ3≤3c2k≤0.402 kc^3-\kappa^3\le3c^2k\le0.402\,k and s2≤θ+2vδ+2ck≤0.616s_2\le\theta+2v\delta+2ck\le0.616 (by (a), θ≤0.0304\theta\le0.0304 and δ<0.1161\delta<0.1161), so that s23/2≤0.785 (θ+2.733 δ+0.733 k)s_2^{3/2}\le0.785\,(\theta+2.733\,\delta+0.733\,k). With d+22=9\tfrac{d+2}{2}=9 this gives 2.54 δ≤0.978 k−8.21 θ2.54\,\delta\le0.978\,k-8.21\,\theta, that is,

δ ≤ 0.386 k−3.2 θ.\delta\ \le\ 0.386\,k-3.2\,\theta .

The two bounds on δ\delta are incompatible for θ>0\theta>0 and k≥0k\ge0. Hence M4=2M_4=2.

The case n=5n=5 [computed]. For d=32d=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 (ℓ,κ)(\ell,\kappa) we have three conditions on θ\theta. The star inequality near A⋆A^\star reads θ≤θT(ℓ,κ)\theta\le\theta_T(\ell,\kappa), where

θT(ℓ,κ):=δ [(c−k) D−d (v+3)]+δ2 [D+4d(v−1)+4]−d v kd v2,δ=v−ℓ,k=c−κ.\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 θ≤θlin(ℓ):=(4ℓ−5)/(d+2−2ℓ)\theta\le\theta_{\rm lin}(\ell):=(4\ell-5)/(d+2-2\ell). The Liu–Liu inequality with the bound t≤ℓ3−κ3+Wt\le\ell^3-\kappa^3+W reads h(θ)≥0h(\theta)\ge0, where

h(θ):=2 (ℓ3−κ3+W)−5−(d+2) θh(\theta):=2\,\big(\ell^3-\kappa^3+W\big)-5-(d+2)\,\theta

is a convex function of θ\theta (because SS is linear in θ\theta). Therefore θ>0\theta>0 is possible only at pairs (ℓ,κ)(\ell,\kappa) with θT>0\theta_T>0, θlin>0\theta_{\rm lin}>0 and

max⁡{h(0), h(min⁡(θT,θlin))} ≥ 0.\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 3001×30013001\times3001 points of the box 54≤ℓ≤v\tfrac54\le\ell\le v, 0≤κ≤c0\le\kappa\le c, refined towards A⋆A^\star. Wherever θT>0\theta_T>0 and θlin>0\theta_{\rm lin}>0, it is at most −0.97 (δ+k)-0.97\,(\delta+k) for d=32d=32 (and at most −3.1 (δ+k)-3.1\,(\delta+k) for d=16d=16, in agreement with the analytical proof). Hence no witness in Λ5\Lambda_5 has purity above 22, and M5=2M_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 nn

For n≥6n\ge6 the argument of Section 3.2 stops working: the conditions used there are no longer contradictory, so we cannot conclude that Mn=2M_n=2, and we cannot use Mn−1=2M_{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 Mn−1M_{n-1}: if Mn−1≤mM_{n-1}\le m, then, because d−q≥0d-q\ge0,

dq+1−2t ≤ m (d−q)for every A∈Λn.dq+1-2t\ \le\ m\,(d-q)\qquad\text{for every }A\in\Lambda_n .

So from a number mn−1≥Mn−1m_{n-1}\ge M_{n-1} we can compute a number mn≥Mnm_n\ge M_n, from it a number mn+1≥Mn+1m_{n+1}\ge M_{n+1}, and so on. Each step can increase the bound only by an amount of order 1/d1/d, and ∑n2−n\sum_n2^{-n} converges, so the bounds stay close to 22 for all nn. We first make this precise in an elementary form, and then in the sharper form that gives the constant 0.49960.4996.

Lemma (one level of the recursion, elementary form). Let n≥2n\ge2 and let m≥2m\ge2 be any number with Mn−1≤mM_{n-1}\le m. Then

Mn ≤ gd(m):=m d−1d+m−2λ+,M_n\ \le\ g_d(m):=\frac{m\,d-1}{d+m-2\lambda_+},

and

gd(m)−m=m (2λ+−m)−1d+m−2λ+ ≤ 0.4642d−0.74.g_d(m)-m=\frac{m\,(2\lambda_+-m)-1}{d+m-2\lambda_+}\ \le\ \frac{0.4642}{d-0.74}.

Proof: Let A∈ΛnA\in\Lambda_n. By the first bound on the third moment and the theorem on the spectrum, t≤ℓ q≤λ+qt\le\ell\,q\le\lambda_+q. Inserting this in the Liu–Liu inequality with Mn−1≤mM_{n-1}\le m gives dq+1−2λ+q≤m (d−q)dq+1-2\lambda_+q\le m\,(d-q), that is, q (d+m−2λ+)≤m d−1q\,(d+m-2\lambda_+)\le m\,d-1. Maximising over AA proves the first claim. For the second, the numerator m(2λ+−m)−1m(2\lambda_+-m)-1 is decreasing in mm for m≥λ+m\ge\lambda_+ and equals 23−3<0.46422\sqrt3-3<0.4642 at m=2m=2, while the denominator is at least d+2−2λ+>d−0.74d+2-2\lambda_+>d-0.74. □\square

Corollary (a bound for all nn without computer assistance). For every n≥1n\ge1 we have Mn<2.03M_n<2.03. Hence the main theorem holds with a∗=0.4926a^*=0.4926 by an argument that uses no numerical computation.

Proof: By the cases n≤4n\le4 of the theorem in Section 3.2, which are proved analytically, Mn=2M_n=2 for n≤4n\le4. For n≥5n\ge5 we apply the lemma repeatedly, starting from M4=2M_4=2:

Mn ≤ 2+0.4642∑j=5n12j−0.74 < 2+0.4642⋅3231.26⋅∑j≥52−j < 2.03,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>0.49261/2.03>0.4926. □\square

If one also uses M5=2M_5=2, the same estimate starting from j=6j=6 gives Mn<2.0147M_n<2.0147 and a∗=0.4963a^*=0.4963. To reach 0.49960.4996 we replace the elementary step by a sharper one, which uses the star inequality and the Wilkins bound.

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

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

(i) the star inequality near A⋆A^\star, with θT\theta_T as in Section 3.2:

q ≤ qT(ℓ,κ):=2+θT(ℓ,κ);q\ \le\ q_T(\ell,\kappa):=2+\theta_T(\ell,\kappa);

(ii) the Liu–Liu inequality with t≤ℓ qt\le\ell\,q:

q ≤ qL(ℓ):=m d−1d+m−2ℓ;q\ \le\ q_L(\ell):=\frac{m\,d-1}{d+m-2\ell};

(iii) the Liu–Liu inequality with t≤ℓ3−κ3+Wt\le\ell^3-\kappa^3+W:

h(q):=2 (ℓ3−κ3+W)−(dq+1−m (d−q)) ≥ 0,h(q):=2\,\big(\ell^3-\kappa^3+W\big)-\big(dq+1-m\,(d-q)\big)\ \ge\ 0 ,

where WW depends on qq through S=q−ℓ2−κ2−(1−ℓ+κ)2/(d−2)S=q-\ell^2-\kappa^2-(1-\ell+\kappa)^2/(d-2).

Moreover, (ii) and q>mq>m force ℓ>(m2+1)/(2m)≥54\ell>(m^2+1)/(2m)\ge\tfrac54. So AA has a negative eigenvalue, and by the theorem on the spectrum 54<ℓ≤v\tfrac54<\ell\le v and 0<κ≤c0<\kappa\le c.

Step 2 (the function FdF_d). For m≥2m\ge2 we say that a number q>mq>m is admissible if there is a pair (ℓ,κ)(\ell,\kappa) with 1≤ℓ≤v1\le\ell\le v and 0≤κ≤c0\le\kappa\le c such that (i), (ii) and (iii) hold. We define

Fd(m):=sup⁡{ q>m: q is admissible },F_d(m):=\sup\{\,q>m:\ q\ \text{is admissible}\,\},

and Fd(m):=mF_d(m):=m if there is no admissible qq. By Step 1, the purity of every witness in Λn\Lambda_n is either at most mm or admissible, so

Mn−1≤m⟹Mn≤Fd(m),d=2n.M_{n-1}\le m\qquad\Longrightarrow\qquad M_n\le F_d(m),\qquad d=2^n .

Step 3 (how Fd(m)F_d(m) is evaluated). Fix a pair (ℓ,κ)(\ell,\kappa). Conditions (i) and (ii) say that q≤qU:=min⁡{qT(ℓ,κ),qL(ℓ)}q\le q_U:=\min\{q_T(\ell,\kappa),q_L(\ell)\}, and hh is a convex function of qq (on this box and for q≥2q\ge2 one has S≥0S\ge0, so WW is well defined, and SS is linear in qq). Hence the largest admissible qq for this pair is qUq_U if h(qU)≥0h(q_U)\ge0; it is the unique zero of hh in [m,qU)[m,q_U) if h(qU)<0≤h(m)h(q_U)<0\le h(m); and there is no admissible qq for this pair if h(m)<0h(m)<0 and h(qU)<0h(q_U)<0. The number Fd(m)F_d(m) is the maximum of this quantity over the box 1≤ℓ≤v1\le\ell\le v, 0≤κ≤c0\le\kappa\le c. We computed it numerically: a grid of 601×601601\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 10−810^{-8} under changes of the grid.

Step 4 (the recursion). We put m5:=2m_5:=2, which is the value of M5M_5, and

mn:=F2n(mn−1)for n≥6.m_n:=F_{2^n}(m_{n-1})\qquad\text{for }n\ge6 .

By Step 2, Mn≤mnM_n\le m_n for every n≥5n\ge5. For d≤32d\le32 no purity above 22 is admissible, in agreement with Section 3.2; the first level at which an admissible purity above the input appears is d=64d=64, where F64(2)=2.00009F_{64}(2)=2.00009. The results are collected in the following table. The bounds on MnM_n are rounded up, and the resulting constants a∗=1/Mna^*=1/M_n are rounded down.

nnd=2nd=2^nbound on MnM_nincrease from level n−1n-1resulting constant a∗a^*how the bound is obtained
122–0.5direct computation
24200.5analytical proof
38200.5analytical proof
416200.5analytical proof
532200.5computed: F32(2)=2F_{32}(2)=2
6642.000098.98⋅10−58.98\cdot10^{-5}0.49997computed: m6=F64(2)m_6=F_{64}(2)
71282.000524.29⋅10−44.29\cdot10^{-4}0.49987computed: m7=F128(m6)m_7=F_{128}(m_6)
82562.000913.83⋅10−43.83\cdot10^{-4}0.49977computed: m8=F256(m7)m_8=F_{256}(m_7)
95122.001162.52⋅10−42.52\cdot10^{-4}0.49971computed
1010242.001311.47⋅10−41.47\cdot10^{-4}0.49967computed
1120482.001398.05⋅10−58.05\cdot10^{-5}0.49965computed
1240962.001434.26⋅10−54.26\cdot10^{-5}0.49964computed
13 to 20up to 10485762.00147less than 2.3⋅10−52.3\cdot10^{-5} per level0.49963computed
all n≥21n\ge21–2.0015less than 5⋅10−75\cdot10^{-7} in total0.4996tail estimate (Step 5)

Step 5 (the tail). The computed value at level 2020 is m20=2.00146925…<2.0014693m_{20}=2.00146925\ldots<2.0014693. For n>20n>20 we use the elementary lemma, which bounds the increase at level jj by 0.4642/(2j−0.74)0.4642/(2^j-0.74):

Mn ≤ m20+0.4642∑j≥2112j−0.74 < 2.0014693+5⋅10−7 < 2.0015.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 Mn≤2.0015M_n\le2.0015 for every nn, and 1/2.0015>0.49961/2.0015>0.4996. □\square

What the numbers mean. The bound 2.000092.00009 at n=6n=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 Λ6\Lambda_6 with λmax⁡≈1.361\lambda_{\max}\approx1.361, λmin⁡≈−0.311\lambda_{\min}\approx-0.311 and purity 2.000092.00009. At the next level the Liu–Liu inequality is used with m6m_6 instead of 22, 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 2.001472.00147.

Why the exact value 22 is obtained only up to n=5n=5. Near A⋆A^\star the star inequality needs δ≥d v k/(c D−d(v+3))\delta\ge d\,v\,k/(c\,D-d(v+3)), while item (b) needs v2δ<c2kv^2\delta<c^2k. These two requirements are incompatible exactly as long as c D−d(v+3)<v3d/c2c\,D-d(v+3)<v^3d/c^2, that is, for d<64.6d<64.6. So the induction step closes for d≤32d\le32 and fails, by a very small amount, from d=64d=64 on. Closing this gap for all nn would prove the conjecture.

Permanent link:open-context.jan-bulling.com/contribution/5.12

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)
  2. Michael Zurel and Jack Davis, Basis-independent stabilizerness and maximally noisy magic states, arXiv:2602.22336 (2026)
  3. Zhenhuan Liu and Zi-Wen Liu, On the geometry, typicality, and detectability of quantum magic, arXiv:2609.03944 (2026)
  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)
  5. Ben W. Reichardt, Quantum universality by state distillation, arXiv:quant-ph/0608085 (2006)
  6. Hilbert–Schmidt inradius of the multiqubit stabilizer polytope, Quantum Information and Quantum Computation Open Problem Zoo (QIQCOP Zoo), (2026)
  7. John Watrous, The Theory of Quantum Information, Cambridge University Press (2018)
  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) doi: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)
  10. Michael Zurel, Cihan Okay, and Robert Raussendorf, Simulating Quantum Computation: How Many “Bits” for “It”?, PRX Quantum 5, 030343 (2024)

Discussion

Sign in to join the discussion.

No comments yet. Questions about this result, checks of its steps and pointers are welcome here.