betaFeedback

WorksPhysics and Astronomy

Stabilizer-Restricted Antidistinguishability in Quantum Computation with Magic States

Abstract

An antidistinguishing measurement for a family of preparations excludes at least one preparation for every possible outcome, constraining 'classical' representations of these processes, known as ontological models. We develop this connection within quantum computing with magic states (QCM), a universal model of quantum computation. Our constructions use adaptive Pauli measurements on tensor products of single-qubit non stabilizer states, including families associated with the Clifford hierarchy. These constructions yield stabilizer-restricted versions of the Pusey-Barrett-Rudolph theorem: any ontological model reproducing these measurement statistics and satisfying preparation independence must assign disjoint ontic supports to the preparations considered. We demonstrate this obstruction to preparation independence in a concrete ontological model based on convex polytopes used for the classical simulation of QCM.

Introduction

A central theme in the foundations of quantum theory has been to identify a precise notion of what is meant by ‘classical’. Throughout the development of the subject, a web of such notions has emerged, based on locality [1], noncontextuality [2, 3], its generalized variant [4], or nonnegativity in quasiprobability representations [5, 6]. A common framework for studying these notions describes an underlying physical reality in terms of an ontic state [7]. Quantum theory can be described in operational terms [8], i.e. by preparations, transformations, and measurements, while ontological models represent these operational primitives using probability distributions, transition kernels, and response functions defined on the ontic state space. Additional constraints on the mathematical form of the ontological model encode the relevant notion of classicality, with attendant ‘no-go’ results demonstrating incompatibility with quantum mechanics.

Pusey, Barrett, and Rudolph (PBR) [9] proved a fundamental ‘no-go’ theorem on the nature of the quantum state. Their argument is based on the principle of preparation independence, which reflects the classical intuition that operationally independent preparations should be represented in the ontological model by statistically independent product distributions. The PBR theorem places particular emphasis on the ontological representation of preparations: under preparation independence, distinct pure quantum states must be represented by distributions with disjoint supports. Another key notion is that of antidistinguishability [10, 11]. A quantum measurement is said to be weakly antidistinguishing for a set of states if every measurement outcome has zero probability for at least one preparation in the set [12]. The existence of such a measurement forces the ontological representations of the preparations to have vanishing common overlap. This is precisely the operational ingredient needed to prove the PBR theorem.

Complementary to these foundational developments has been an increased interest in the role of such nonclassical features in quantum computational advantage. Within the setting of measurement-based quantum computation [13], Bell nonlocality and contextuality have been identified as resources for increased computational power relative to classical counterparts [14, 15, 16, 17]. Similarly, negativity in quasiprobability representations quantifies the cost of particular classical simulation schemes [18, 19, 20, 21]. In several settings, contextuality and negativity have been shown to be equivalent notions [22, 23, 24, 25, 26]. By contrast, the computational implications of the PBR theorem and its underlying principle of preparation independence have received less attention; a related connection to classical simulation is developed in [27].

Ontological models also provide a natural framework for studying classical simulation of quantum circuits [27, 28, 29]. A prominent example is the discrete Wigner representation of Gross [30]. Restricted to states with nonnegative Wigner functions and stabilizer operations, it provides a noncontextual ontological model for odd-dimensional qudit systems that includes the stabilizer fragment [23] and facilitates a sampling-based classical simulation scheme [18]. In particular, the Wigner representation respects tensor products, so that independently prepared states with nonnegative Wigner functions are represented by product distributions. This gives a preparation-independent model of the corresponding operational fragment. For states exhibiting negativity, the same tensor-product property enables a multiplicative measure controlling the sampling cost [19].

For qubits, many of these appealing properties of the odd-qudit case do not persist, and understanding these distinctions has been the subject of continued investigation [21, 25, 31, 32, 33, 34]. An important development along these lines was made in [35] with the introduction of a family of geometric structures called Pauli polytopes, also known as Λ\Lambda polytopes, that are related by duality to the well-known stabilizer polytopes [36]. These objects furnish a concrete ψ\psi-epistemic ontological model that can be used for the classical simulation of quantum computing with magic states (QCM) [37], a universal model of quantum computation. As such, these ontological representations can be rigorously studied using polytope-theoretic methods, potentially yielding insights into the complexity of classical simulation and the boundaries of quantum advantage.

Results and outline

Summary of main results

In this paper we construct families of antidistinguishing measurements using adaptive and nonadaptive multiqubit Pauli measurements applied to preparation families consisting of completely separable states of computational interest, including those arising from the Clifford hierarchy [38]. These preparations and measurements lie within the basic computational primitives of QCM. Our results therefore present a direct operational obstruction for ontological models of QCM that assign overlapping distributions to the state pairs considered: such models cannot both satisfy preparation independence and reproduce the relevant Born-rule statistics. Although stabilizer implementations of particular PBR measurements are already known [9, 27], the general PBR construction does not impose this measurement restriction. Our work extends these stabilizer-restricted arguments to computationally relevant families of non-stabilizer states.

Let us summarize our main results. For the stabilizer fragment, we generalize existing constructions [9, 27] to demonstrate that any pair of distinct nonorthogonal pure stabilizer states yields a PBR proof (Proposition 5.2). A complementary argument based on contextuality appears in Appendix § 11. Additionally, we construct stabilizer-restricted weakly antidistinguishing measurements for magic-state preparations consisting of tensor products drawn from the single-qubit pairs Rk∣+⟩R_k|+\rangle and SRk∣+⟩SR_k|+\rangle, where

Rk=diag⁡(1,eiθk),θk:=2π2k,R_k = \operatorname{diag}(1,e^{i\theta_k}), \qquad\theta_k := \frac{2\pi}{2^k},

and k≥3k \ge3 labels the non-Clifford levels of the hierarchy [38] (Theorem 6.4). The number of qubits used by our construction scales exponentially with the level of the hierarchy according to Dk=2k−2D_k = 2^{k-2}. The construction uses an algebraic property of these phase-rotated states: a two-qubit parity measurement maps pairs of such states either to a maximally entangled stabilizer state on the odd-parity branch or to an encoded single-qubit state with phase level reduced by one on the even-parity branch.

It is also possible to construct a stabilizer-restricted PBR proof for every number n≥2n \ge2 of qubits (Theorem 6.5). For this, we use the preparation pair

∣ψa⟩=cos⁡θn∣0⟩+(−1)asin⁡θn∣1⟩,a∈{0,1},θn=π4n,|\psi_a\rangle= \cos\theta_n|0\rangle+ (-1)^a\sin\theta_n|1\rangle,\qquad a \in\{0,1\},\qquad\theta_n = \frac{\pi}{4n},

whose angle depends on the number of qubits. The nonadaptive Pauli measurements are specified by the canonical stabilizer generators of the graph state corresponding to KnK_n, i.e. the complete graph on nn vertices. Each outcome excludes exactly one preparation in the resulting product family.

The above results hold for otherwise arbitrary ontological models reproducing the specified preparation and Pauli-measurement statistics. We then specialize to the Pauli ontological model, based on the Pauli polytopes introduced in [35]. Within this model, the conceptual issue of preparation independence can be related to a structural question about closure under the operator tensor product. Our construction of antidistinguishing measurements for the TT-state pair (k=3k = 3), together with overlapping ontological representations, demonstrates an obstruction to preparation independence. In fact, we establish a stronger structural result: the tensor product of any vertex of an nn-qubit Pauli polytope with any vertex of an mm-qubit Pauli polytope falls outside the composite (n+m)(n+m)-qubit polytope. These results obstruct a direct strategy for constructing composite ontological representations by independently sampling and tensoring local vertices. They therefore preclude a straightforward tensor-product approach to analyzing this simulation scheme, of the kind used in other classical simulation methods [19, 20, 39], without themselves establishing a lower bound on simulation time.

Outline

The remainder of the paper is organized as follows. § 3 reviews stabilizer theory and quantum computing with magic states. In § 4 we introduce the operational and ontological frameworks, antidistinguishability, and the preparation-independence argument. § 5 and § 6 develop our antidistinguishing measurements for stabilizer and magic-state preparations, respectively. In § 7 we apply these constructions to the Pauli ontological model and establishes the tensor-product obstruction for Pauli-polytope vertices. We conclude with a discussion of our results in § 8.

Quantum computing with magic states

Stabilizer theory

We begin with some preliminaries on stabilizer theory. A system of nn qubits is modeled by a Hilbert space H=(C2)⊗n\mathcal{H} = (\mathbb{C}^{2})^{\otimes n}. An nn-qubit Pauli operator is any operator of the form

id σ1⊗⋯⊗σn,i^{d}\,\sigma_{1} \otimes\cdots\otimes\sigma_{n},

where d∈Z4d \in\mathbb{Z}_{4} and each σi∈{I,X,Y,Z}\sigma_{i} \in\{I, X, Y, Z\} is a single-qubit Pauli matrix or the identity

I=(1001),X=(0110),Y=(0−ii0),Z=(100−1).I = \begin{pmatrix} 1 & 0 \\ 0 & 1 \end{pmatrix}, \quad X = \begin{pmatrix} 0 & 1 \\ 1 & 0 \end{pmatrix}, \quad Y = \begin{pmatrix} 0 & -i \\ i & 0 \end{pmatrix}, \quad Z = \begin{pmatrix} 1 & 0 \\ 0 & -1 \end{pmatrix}.

For all nn, Pauli operators either commute or anticommute and are Hermitian if and only if d=0,2d = 0, 2. The nn-qubit Pauli group is the set of operators

Gn:={±1,±i}×{σ1⊗⋯⊗σn:σi∈{I,X,Y,Z}}\mathcal{G}_{n} := \{\pm1, \pm i\} \times\{\sigma_{1} \otimes\cdots\otimes\sigma_{n} : \sigma_{i} \in\{I, X, Y, Z\}\}

with matrix multiplication as the group operation. A stabilizer group S⊂Gn\mathcal{S} \subset\mathcal{G}_{n} is an abelian subgroup of the Pauli group not containing −I-I. Every element of S\mathcal{S} is therefore Hermitian and squares to the identity. Such a group has kk independent generators,

S=⟨s1,…,sk⟩,0≤k≤n,\mathcal{S} = \langle s_{1}, \ldots, s_{k} \rangle, \qquad0 \leq k \leq n,

and has order ∣S∣=2k|\mathcal{S}| = 2^{k}. Note that the generators may include an overall minus sign. The operator

ΠS:=1∣S∣∑s∈Ss\Pi_{\mathcal{S}} := \frac{1}{|\mathcal{S}|} \sum_{s \in\mathcal{S}} s

is the projector onto the joint +1+1 eigenspace of each s∈Ss \in\mathcal{S}. Measurement of a Pauli observable AA obtaining outcome σ∈{±1}\sigma\in\{\pm1\} is encoded by the projector ΠAσ:=ΠS\Pi_{A}^{\sigma} := \Pi_{\mathcal{S}}, where S=⟨σA⟩\mathcal{S} = \langle\sigma A \rangle. A stabilizer group S\mathcal{S} is maximal when it has nn independent generators. In this case, the projector ΠS=∣S⟩⟨S∣\Pi_{\mathcal{S}} = |S\rangle\langle S| has rank one and ∣S⟩|S\rangle is a stabilizer state.

The nn-qubit Clifford group Cn\mathcal{C}_{n} is a finite subgroup of the nn-qubit unitary group U(H)\mathcal{U}(\mathcal{H}) and, up to phases, is the normalizer of the Pauli group. Equivalently, the Clifford group is generated by the single-qubit Hadamard HH and phase SS gates together with the two-qubit CNOT gate. All stabilizer states can be written as

∣S⟩=US∣0⟩⊗n,|S\rangle= U_{S}|0\rangle^{\otimes n},

where ∣0⟩⊗n|0\rangle^{\otimes n} is a stabilizer state with stabilizer group S=⟨Zi⟩i=1n\mathcal{S} = \langle Z_{i}\rangle_{i=1}^{n} and USU_{S} is a Clifford unitary. It is known that any stabilizer state can be written in the normal form

∣S⟩=1∣A∣1/2∑x∈Aeiq(x)∣x⟩,(1)|S\rangle= \frac{1}{|\mathcal{A}|^{1/2}} \sum_{x \in\mathcal{A}} e^{i q(x)} |x\rangle, \tag*{(1)}

where A⊂Z2n\mathcal{A} \subset\mathbb{Z}_{2}^{n} is an affine subspace and q:A→Z4q : \mathcal{A} \to\mathbb{Z}_{4} is a quadratic form; see e.g., [40].

Any quantum computation initialized in a stabilizer state, then acted on by Clifford gates and measurements of Pauli observables, optionally with classical randomness and feedforward, can be efficiently simulated classically. This is the well-known Gottesman–Knill theorem [41].

Clifford hierarchy and magic state injection

One can promote the stabilizer fragment of quantum theory to a universal quantum computational model with access to non-stabilizer, or ‘magic’ resources. A common choice is T=diag⁡(1,eiπ/4)T = \operatorname{diag}(1, e^{i\pi/4}), which amounts geometrically to a π/4\pi/4 rotation about the ZZ-axis of the Bloch sphere. Particularly appealing about the TT gate is that it can be implemented in fault-tolerant architectures using state injection schemes [37].

In this setup, one can implement TT deterministically once supplied with a magic state

∣T⟩=T∣+⟩=12(∣0⟩+eiπ/4∣1⟩),(2)|T\rangle= T|+\rangle= \frac{1}{\sqrt{2}}\left(|0\rangle+ e^{i\pi/4}|1\rangle\right), \tag*{(2)}

together with adaptive stabilizer operations. Thus, the primitives for universal quantum computation can be taken to be adaptive Clifford gates and Pauli measurements together with a supply of ∣T⟩|T\rangle magic states. For drawing samples from a universal family of circuits, the primitives can be further refined to consist only of adaptive Pauli measurements and magic states [42]; see also [43, 44].

Other choices of universal gate sets admitting deterministic magic state injection using stabilizer operations are also possible, such as the controlled phase CSCS and controlled-controlled-ZZ CCZCCZ gate [20]. These gates all belong to the third level of the Clifford hierarchy [38]. More generally, the kkth-level of the hierarchy is defined recursively by

Cn(k):={U∈U(H):UAU†∈Cn(k−1) for every A∈Gn}.\mathcal{C}_{n}^{(k)} := \left\{U \in\mathcal{U}(\mathcal{H}) : U A U^\dagger\in\mathcal{C}_{n}^{(k-1)} \text{ for every } A \in\mathcal{G}_{n}\right\}.

with the first and second levels being the nn-qubit Pauli and Clifford groups, respectively. While the levels of the Clifford hierarchy do not generally form a group, the diagonal elements in each level do form a group. In particular, at each level kk we have the diagonal single-qubit unitary

Rk:=diag⁡(1,ζk),where ζk:=e2πi/2k,(3)R_k := \operatorname{diag}(1,\zeta_k), \qquad\text{where } \zeta_k := e^{2\pi i/2^k}, \tag*{(3)}

so that R1=ZR_1 = Z, R2=SR_2 = S, and R3=TR_3 = T. Here ζk\zeta_k is a root of unity. The diagonal Clifford hierarchy is naturally described by such roots of unity and phase polynomials [45].

The corresponding single-qubit resource states can be defined analogously to the TT state

∣Rk⟩=Rk∣+⟩=12(∣0⟩+ζk∣1⟩).|R_k\rangle= R_k|+\rangle= \frac{1}{\sqrt{2}}\left(|0\rangle+ \zeta_k|1\rangle\right).

For k≥4k \ge4, the standard deterministic injection scheme for RkR_k additionally requires a lower-level correction Rk−1R_{k-1} on one measurement branch. This correction can itself be implemented by state injection, with further lower-level corrections as required [38]. The nn-fold tensor product of ∣Rk⟩|R_k\rangle is given by

∣Rk⟩⊗n=2−n/2∑x∈Z2nζk∣x∣∣x⟩,(4)|R_k\rangle^{\otimes n} = 2^{-n/2}\sum_{x\in\mathbb{Z}_2^n}\zeta_k^{|x|}|x\rangle, \tag*{(4)}

where Z2\mathbb{Z}_2 is the binary ring and ∣x∣|x| is the Hamming weight of the bit string xx. For k≥3k \ge3, the coefficients of this normalized state lie in the algebraic field Q(ζk)\mathbb{Q}(\zeta_k). For every k≥1k \ge1, the coefficients lie in this field after removing the common normalization factor. (See [46] as well as Appendix § 10 for further details.) This fact will prove useful in the developments below.

Antidistinguishability

Operational quantum theory

The primitives of an operational theory are preparation, transformation, and measurement procedures, together with their empirical outcome probabilities. In quantum theory, preparations are represented by density operators, which are positive operators of unit trace. Here we consider projective measurements with post-measurement states determined by the Lüders rule. In particular, our setup will allow for any quantum state as a valid preparation, but we will focus on a subset of projective measurements M\mathcal{M} with the main example being sequences of Pauli measurements.

Consider a sequence of mutually commuting PVMs Mj={Πjk}k∈KjM_j = \{\Pi_j^{k}\}_{k\in K_j}, with j=1,…,tj = 1,\ldots,t, that we denote collectively as MM. For an outcome sequence k=(k1,…,kt)k = (k_1,\ldots,k_t), define the joint projector

ΠMk:=∏j=1tΠjkj.\Pi_M^k := \prod_{j=1}^{t}\Pi_j^{k_j}.

The probability of this sequence on a preparation ρ\rho is

p(k∣ρ,M)=Tr⁡(ρΠMk).p(k|\rho, M) = \operatorname{Tr}(\rho\Pi_M^k).

Whenever this probability is positive, the corresponding post-measurement state is

ρM,k=ΠMkρΠMkTr⁡(ρΠMk).\rho_{M,k} = \frac{\Pi_M^k \rho\Pi_M^k}{\operatorname{Tr}(\rho\Pi_M^k)}.

We also allow classical randomization over the permitted procedures. Pure states are the extremal quantum preparations, while rank-one PVMs are the finest projective measurements; higher-rank PVMs can be obtained from them by coarse-graining.

Given a restricted set of measurement procedures M\mathcal{M}, two preparations ρ,ρ′\rho,\rho' are single-shot distinguishable (SSD) if there exists a measurement M∈MM \in\mathcal{M} and a subset AA of its outcome set such that

p(A∣ρ,M)=1,p(A∣ρ′,M)=0.p(A|\rho, M) = 1,\qquad p(A|\rho', M) = 0.

Thus the outcome determines with certainty which preparation was performed. With access to all PVMs, orthogonal quantum states are always SSD since there exists a PVM that perfectly distinguishes them.

Following Ref. [27], consider now a finite set of pure state preparations P:={ρs:s∈S}P := \{\rho_s : s \in S\}. Performing a measurement M∈MM \in\mathcal{M} and obtaining outcome k∈Kk \in K, we denote the set of post-measurement states by PM,kP_{M,k}. The measurement MM is said to be a partitioning measurement for PP if the set PM,kP_{M,k} of post-measurement states contains at least one SSD pair for every outcome kk.

Ontological models framework

Ontological models aim to reproduce empirical statistics using ontic, or physical, microstates and classical probability theory. For simplicity, we take the set of ontic states Λ\Lambda to be finite.1 A preparation of ρ\rho is represented by a probability distribution μρ\mu_\rho on Λ\Lambda. For two preparation distributions μ0\mu_0 and μ1\mu_1, define their common subdistribution by

ν(λ):=min⁡{μ0(λ),μ1(λ)},\nu(\lambda) := \min\{\mu_0(\lambda),\mu_1(\lambda)\},

and their overlap by

ω(μ0,μ1):=∑λ∈Λν(λ).\omega(\mu_0,\mu_1) := \sum_{\lambda\in\Lambda} \nu(\lambda).

The support of a distribution μ\mu is defined as

Supp⁡(μ):={λ∈Λ:μ(λ)>0}.\operatorname{Supp}(\mu) := \{\lambda\in\Lambda: \mu(\lambda) > 0\}.

Drawing on standard terminology, we say that an ontological model is ψ\psi-ontic if ω(μρ,μρ′)=0\omega(\mu_\rho,\mu_{\rho'}) = 0 for every pair of distinct pure-state preparations ρ\rho and ρ′\rho'. Otherwise, it is ψ\psi-epistemic [7].

A projective measurement MM with outcome set KK is represented by a transition kernel

Γ(k,λ′∣M,λ)≥0,∑k∈K∑λ′∈ΛΓ(k,λ′∣M,λ)=1\Gamma(k,\lambda'|M,\lambda) \ge0,\qquad\sum_{k \in K}\sum_{\lambda' \in\Lambda}\Gamma(k,\lambda'|M,\lambda) = 1

for every λ∈Λ\lambda\in\Lambda. This kernel specifies the joint probability of obtaining outcome kk and transitioning to λ′\lambda', conditional on the initial ontic state λ\lambda. The measurement response functions are obtained by marginalizing over the final ontic state:

ξ(k∣M,λ):=∑λ′∈ΛΓ(k,λ′∣M,λ)\xi(k|M,\lambda) := \sum_{\lambda' \in\Lambda}\Gamma(k,\lambda'|M,\lambda)

and the model reproduces the operational probabilities when

p(k∣ρ,M)=∑λ∈Λμρ(λ) ξ(k∣M,λ).p(k|\rho,M) = \sum_{\lambda\in\Lambda}\mu_\rho(\lambda)\,\xi(k|M,\lambda).

Whenever this probability is positive, the conditional post-measurement distribution is

μρ;M,k(λ′)=∑λ∈ΛΓ(k,λ′∣M,λ)μρ(λ)p(k∣ρ,M).\mu_{\rho;M,k}(\lambda') = \frac{\sum_{\lambda\in\Lambda} \Gamma(k,\lambda'|M,\lambda)\mu_\rho(\lambda)}{p(k|\rho,M)}.

This distribution must reproduce the subsequent measurement statistics of the corresponding Lüders post-measurement state.

Antidistinguishing measurements

For preparations that are SSD, the measurement procedure MM that achieves this separation is called a distinguishing measurement. We will now consider antidistinguishing measurements (ADMs), which are measurements whose outcomes certify that some preparations were definitely not used in that run of the experiment, but typically do not confirm which preparation was actually used.

Consider a finite preparation family {ρs:s∈S}\{\rho_s : s \in S\} with measurement set M\mathcal{M}. Following [12], we say that a projective measurement M={ΠMk:k∈K}∈MM = \{\Pi_M^k : k \in K\} \in\mathcal{M} is a weakly antidistinguishing measurement (WADM) for this family if there exists a map f:K→Sf : K \to S such that

p(k∣ρf(k),M)=0for every k∈K.p(k|\rho_{f(k)},M) = 0 \qquad\text{for every } k \in K.

Equivalently, every measurement outcome is excluded by at least one member of the preparation family. We say MM is a strongly antidistinguishing measurement (SADM) if it is a WADM and, for every s∈Ss \in S, there exists an outcome ks∈Kk_s \in K such that

p(ks∣ρs,M)=0,p(ks∣ρs′,M)>0for all s′≠s.p(k_s|\rho_s,M) = 0,\qquad p(k_s|\rho_{s'},M) > 0 \qquad\text{for all } s' \ne s.

Since every SADM is also a WADM, we establish our results for the weak case unless stated otherwise. We will often use ADM simply as an abbreviation broadly referring to antidistinguishing measurement without invoking the specific qualifiers above.

Theorem 4.1. If a finite preparation family {ρs:s∈S}\{\rho_s : s \in S\} admits a WADM, then every ontological model reproducing its zero probabilities satisfies

⋂s∈SSupp⁡(μρs)=∅.\bigcap_{s\in S}\operatorname{Supp}(\mu_{\rho_s}) = \varnothing.

Proof. Suppose that λ\lambda belongs to the support of every μρs\mu_{\rho_s}. For each outcome kk, the WADM condition gives

p(k∣ρf(k),M)=∑λ′∈Λμρf(k)(λ′) ξ(k∣M,λ′)=0.p(k|\rho_{f(k)},M) = \sum_{\lambda'\in\Lambda} \mu_{\rho_{f(k)}}(\lambda')\,\xi(k|M,\lambda') = 0.

Every summand is nonnegative. Since μρf(k)(λ)>0\mu_{\rho_{f(k)}}(\lambda) > 0, it follows that ξ(k∣M,λ)=0\xi(k|M,\lambda) = 0. This holds for every kk, contradicting ∑k∈Kξ(k∣M,λ)=1\sum_{k\in K}\xi(k|M,\lambda) = 1.

Let ρ0,ρ1∈P\rho_0,\rho_1 \in\mathcal{P} be two preparations. For q≥1q \ge1, define their complete product family by

ρx:=ρx1⊗⋯⊗ρxq,x∈Z2q.\rho_x := \rho_{x_1} \otimes\cdots\otimes\rho_{x_q},\qquad x \in\mathbb{Z}_2^q.

The ontological model satisfies preparation independence for this family if its composite ontic space is Λq\Lambda^q and the preparation is represented by a product distribution

μρx(λ1,…,λq)=∏j=1qμρxj(λj).\mu_{\rho_x}(\lambda_1,\ldots,\lambda_q) = \prod_{j=1}^{q}\mu_{\rho_{x_j}}(\lambda_j).

Corollary 4.2 (PBR argument [9]). Suppose the complete qq-copy product family generated by ρ0,ρ1\rho_0,\rho_1 admits a WADM. If an ontological model reproduces its zero probabilities and satisfies preparation independence, then

Supp⁡(μρ0)∩Supp⁡(μρ1)=∅,\operatorname{Supp}(\mu_{\rho_0}) \cap\operatorname{Supp}(\mu_{\rho_1}) = \varnothing,

or equivalently, ω(μρ0,μρ1)=0\omega(\mu_{\rho_0},\mu_{\rho_1}) = 0.

Proof. Suppose that λ∈Supp⁡(μρ0)∩Supp⁡(μρ1)\lambda\in\operatorname{Supp}(\mu_{\rho_0}) \cap\operatorname{Supp}(\mu_{\rho_1}). Preparation independence implies that, for every x∈Z2qx \in\mathbb{Z}_2^q,

μρx(λ,…,λ)=∏j=1qμρxj(λ)>0.\mu_{\rho_x}(\lambda,\ldots,\lambda)=\prod_{j=1}^{q}\mu_{\rho_{x_j}}(\lambda)>0.

Thus (λ,…,λ)(\lambda,\ldots,\lambda) belongs to the common support of the entire product family, contradicting Theorem 4.1.

We remark that the PBR argument holds if and only if you regard preparation independence as fundamental. One could instead relinquish this principle by introducing additional ‘hidden’ correlations in the ontological representation of preparations, similar in spirit to how Bell’s theorem forces nonlocality in ontological models.

Proposition 4.3. If a preparation family {ρs:s∈S}\{\rho_s:s\in S\} admits a partitioning measurement whose conditional distinguishing measurements commute with its projectors, and the resulting adaptive measurement is allowed, then it admits a WADM.

Proof. Let MM be a partitioning measurement for {ρs:s∈S}\{\rho_s:s\in S\}. For each outcome k∈Kk\in K, choose two preparations ρs,ρs′\rho_s,\rho_{s'} whose conditional post-measurement states ρs;M,k,ρs′;M,k\rho_{s;M,k},\rho_{s';M,k} are SSD. Here the indices s,s′s,s' may depend on kk. Conditional on kk, perform an allowed measurement NkN_k that distinguishes these states. Writing RkR_k for its outcome set, there exists a subset Ak⊂RkA_k\subset R_k, with complement Ak⊥:=Rk\AkA_k^\perp:=R_k\backslash A_k, such that

p(Ak⊥∣ρs;M,k,Nk)=0,p(Ak∣ρs′;M,k,Nk)=0.p(A_k^\perp\mid\rho_{s;M,k},N_k)=0,\qquad p(A_k\mid\rho_{s';M,k},N_k)=0.

Let M~\widetilde{M} denote the resulting adaptive measurement, whose outcomes are pairs (k,r)(k,r) with r∈Rkr\in R_k. The joint probabilities satisfy

p(k,r∣ρs,M~)=p(k∣ρs,M) p(r∣ρs;M,k,Nk),p(k,r\mid\rho_s,\widetilde{M})=p(k\mid\rho_s,M)\,p(r\mid\rho_{s;M,k},N_k),

whenever the first factor is positive; otherwise the joint probability is zero. Consequently,

p(k,r∣ρs′,M~)=0if r∈Ak,p(k,r\mid\rho_{s'},\widetilde{M})=0\quad\text{if }r\in A_k,

while

p(k,r∣ρs,M~)=0if r∈Ak⊥.p(k,r\mid\rho_s,\widetilde{M})=0\quad\text{if }r\in A_k^\perp.

Writing PkP_k for the projectors of MM and Qr∣kQ_{r\mid k} for those of NkN_k, the joint effects are PkQr∣kPk=PkQr∣kP_kQ_{r\mid k}P_k=P_kQ_{r\mid k}. By the commutation assumption, these are mutually orthogonal projectors summing to the identity. Thus M~\widetilde{M} is a PVM, not necessarily of rank one. Every outcome of M~\widetilde{M} therefore excludes at least one original preparation, so M~\widetilde{M} is a WADM.

Antidistinguishing measurements for stabilizer states

Simple examples of ADMs restricted to Pauli measurements and stabilizer state preparations go back to the original work of PBR [9]; see also [27]. Consider the nonorthogonal single-qubit stabilizer states

∣ψx⟩:=12(∣0⟩+∣1⟩)and∣ψz⟩:=∣0⟩,\lvert\psi_x\rangle:=\frac{1}{\sqrt{2}}(\lvert0\rangle+\lvert1\rangle)\quad\text{and}\quad\lvert\psi_z\rangle:=\lvert0\rangle,

as well as the two-qubit state vectors

∣Φxx⟩=∣00⟩−∣11⟩2,(5)\lvert\Phi_{xx}\rangle=\frac{\lvert00\rangle-\lvert11\rangle}{\sqrt{2}}, \tag*{(5)}
∣Φxz⟩=∣00⟩+∣01⟩−∣10⟩+∣11⟩2,(6)\lvert\Phi_{xz}\rangle=\frac{\lvert00\rangle+\lvert01\rangle-\lvert10\rangle+\lvert11\rangle}{2}, \tag*{(6)}
∣Φzx⟩=∣00⟩−∣01⟩+∣10⟩+∣11⟩2,(7)\lvert\Phi_{zx}\rangle=\frac{\lvert00\rangle-\lvert01\rangle+\lvert10\rangle+\lvert11\rangle}{2}, \tag*{(7)}
∣Φzz⟩=∣01⟩+∣10⟩2.(8)\lvert\Phi_{zz}\rangle=\frac{\lvert01\rangle+\lvert10\rangle}{\sqrt{2}}. \tag*{(8)}

These rays form a PVM that can be implemented by the adaptive measurement sequence pictured in Figure 1. By direct calculation we have

Figure 1. The PVM in Eqns. (5a)-(5d) is implemented by an adaptive measurement tree: measure YYYY first, measure XXXX on the +1+1 branch, and measure XZXZ on the −1-1 branch. Since each terminal outcome is orthogonal to its correspondingly labelled product preparation it is a SADM for these preparations.

∣⟨Φab∣(∣ψa⟩⊗∣ψb⟩)∣2=0,and∣⟨Φab∣(∣ψc⟩⊗∣ψd⟩)∣2>0,ab≠cd.\left|\langle\Phi_{ab}| \left(|\psi_a\rangle\otimes|\psi_b\rangle\right)\right|^2=0,\qquad\text{and}\qquad\left|\langle\Phi_{ab}| \left(|\psi_c\rangle\otimes|\psi_d\rangle\right)\right|^2>0,\quad ab\ne cd.

Consequently, this is a SADM, which by Corollary 4.2 implies that a PBR proof exists in this case.

This example can be generalized to arbitrary nonorthogonal stabilizer states ∣S0⟩|S_0\rangle and ∣S1⟩|S_1\rangle. Our argument relies on the following standard result.

Lemma 5.1. Let ∣S0⟩|S_0\rangle and ∣S1⟩|S_1\rangle be distinct, non-orthogonal pure nn-qubit stabilizer states. There are a Clifford unitary CC, an integer r≥1r\ge1, and a relabelling of tensor factors such that

C∣S0⟩=∣0⟩⊗r⊗∣0⟩⊗(n−r),C∣S1⟩=∣+⟩⊗r⊗∣0⟩⊗(n−r).C|S_0\rangle=|0\rangle^{\otimes r}\otimes|0\rangle^{\otimes(n-r)},\qquad C|S_1\rangle=|+\rangle^{\otimes r}\otimes|0\rangle^{\otimes(n-r)}.

Proof. First send ∣S0⟩|S_0\rangle to ∣0⟩⊗n|0\rangle^{\otimes n} by a Clifford CC. From the stabilizer state normal form Eq. (1), we see that for the transformed C∣S1⟩C|S_1\rangle to be non-orthogonal that the corresponding affine subspace is, in fact, a linear subspace containing the origin. We can perform row reduction on a basis of this linear subspace using CNOT and SWAP gates, which fix ∣0⟩⊗n|0\rangle^{\otimes n}. Diagonal Clifford gates remove the remaining linear and quadratic phases, again while fixing ∣0⟩⊗n|0\rangle^{\otimes n} up to phase. The second state is then ∣+⟩⊗r∣0⟩⊗(n−r)|+\rangle^{\otimes r}|0\rangle^{\otimes(n-r)}. Requiring that the states be distinct gives r≥1r\ge1.

The generalization then proceeds by choosing an odd number mm of active qubits in each block: take m=rm=r for odd rr and m=r−1m=r-1 for even rr. The operators X‾=X⊗m\overline{X}=X^{\otimes m} and Z‾=Z⊗m\overline{Z}=Z^{\otimes m}, acting trivially on the remaining qubits, anticommute and define logical Pauli operators, with Y‾=iXZ\overline{Y}=iXZ. In the corresponding logical-qubit subsystem, ∣0⟩⊗r|0\rangle^{\otimes r} has logical state ∣0⟩|0\rangle and ∣+⟩⊗r|+\rangle^{\otimes r} has logical state ∣+⟩|+\rangle; their states on the remaining subsystem need not coincide. The common stabilizer tail ∣0⟩⊗(n−r)|0\rangle^{\otimes(n-r)} does not enter the measurement. We then perform the logical two-qubit adaptive Pauli PVM of Figure 1 on the two preparation blocks. Its observables commute along each branch, so its terminal effects are mutually orthogonal projectors, which need not have rank one.

Proposition 5.2. An ontological model reproducing the predictions of stabilizer theory and satisfying preparation independence must represent distinct stabilizer states ∣S0⟩|S_0\rangle and ∣S1⟩|S_1\rangle by probability distributions μ0\mu_0 and μ1\mu_1 with disjoint support, i.e. Supp⁡(μ0)∩Supp⁡(μ1)=∅\operatorname{Supp}(\mu_0)\cap\operatorname{Supp}(\mu_1)=\varnothing.

The result holds since every terminal outcome of the logical Pauli PVM excludes one of the four product preparations. Orthogonal pairs are already distinguishable by a Pauli measurement. For a contextuality-based proof of this result, see Appendix § 11; similar connections between preparation independence and the contextuality of the stabilizer fragment were also explored in [33].

One straightforward consequence of Proposition 5.2 is that an ontological model reproducing the predictions of qubit stabilizer theory, but that does not represent non-orthogonal stabilizer states by disjoint probability distributions, must then violate preparation independence. An example of such a model is the closed non-contextual (CNC) phase space introduced in [21] (see also [47]), which provides an operator-based ontological model for qubit stabilizer quantum mechanics. These operators, however, are not closed under tensor product [48] so that preparation independence fails in this model. We prove a generalization of this result in § 7.3 for an operator-based ontological model of universal QCM.

Antidistinguishing measurements for magic states

The input resource state for QCM is often taken to be a tensor product of magic states. In this section we construct WADMs for non-stabilizer input states. For convenience, we will often drop normalization factors, as they do not affect the zero-probability condition.

Partitioning measurement for third-level states

Here we establish WADMs for tensor products of TT magic states by constructing a partitioning measurement. Let ∣Ta⟩:=Za∣T⟩\lvert T_a\rangle:= Z^a\lvert T\rangle, where Z∣T⟩Z\lvert T\rangle is orthogonal to ∣T⟩\lvert T\rangle. Define now

∣ψab⟩:=∣Ta⟩⊗∣Tb⟩,a,b∈Z2,\lvert\psi_{ab}\rangle:= \lvert T_a\rangle\otimes\lvert T_b\rangle,\qquad a,b\in\mathbb{Z}_2,

or more explicitly

∣ψab⟩∝∣00⟩+(−1)a+bi∣11⟩+ζ3((−1)b∣01⟩+(−1)a∣10⟩).\lvert\psi_{ab}\rangle\propto\lvert00\rangle+ (-1)^{a+b}i\lvert11\rangle+ \zeta_3\left((-1)^b\lvert01\rangle+ (-1)^a\lvert10\rangle\right).

Consider now the parity measurement of the two-qubit Pauli ZZZZ yielding the post-measurement states represented, up to phase and normalization, by

∣uE(a+b)⟩:=∣00⟩+(−1)a+bi∣11⟩,(9)\lvert u_E(a+b)\rangle:= \lvert00\rangle+ (-1)^{a+b}i\lvert11\rangle, \tag*{(9)}
∣uO(a+b)⟩:=∣01⟩+(−1)a+b∣10⟩.(10)\lvert u_O(a+b)\rangle:= \lvert01\rangle+ (-1)^{a+b}\lvert10\rangle. \tag*{(10)}

where EE and OO correspond to the even and odd parity subspaces, respectively. These are, in fact, stabilizer states satisfying

XY∣uE(a+b)⟩=(−1)a+b∣uE(a+b)⟩,XX∣uO(a+b)⟩=(−1)a+b∣uO(a+b)⟩.XY\lvert u_E(a+b)\rangle= (-1)^{a+b}\lvert u_E(a+b)\rangle,\qquad XX\lvert u_O(a+b)\rangle= (-1)^{a+b}\lvert u_O(a+b)\rangle.

Thus, conditional on the outcome ZZ=+1ZZ=+1, measure XYXY; conditional on ZZ=−1ZZ=-1, measure XXXX. Each conditional measurement distinguishes the two possible post-measurement states. The Pauli ZZZZ is therefore a partitioning measurement for this set of states and the conditional measurements XYXY and XXXX extend this to a WADM. This is precisely the adaptive measurement sequence in Figure 1, up to single-qubit Clifford unitaries.

Proposition 6.1. For every n≥2n\ge2, the preparation family

ρx:=∣Tx⟩⟨Tx∣,∣Tx⟩:=⨂ℓ=1n∣Txℓ⟩,x∈Z2n,\rho_x := \lvert T_x\rangle\langle T_x\rvert,\qquad\lvert T_x\rangle:= \bigotimes_{\ell=1}^{n}\lvert T_{x_\ell}\rangle,\qquad x\in\mathbb{Z}_2^n,

admits a rank-one WADM implemented by adaptive Pauli measurements.

Proof. Partition the qubits into disjoint pairs. On each pair (i,j)(i,j), measure ZiZjZ_iZ_j, followed by XiYjX_iY_j on the +1+1 branch and XiXjX_iX_j on the −1-1 branch. If nn is odd, measure ZZ on the remaining qubit. The preceding two-qubit calculation shows that the second outcome on each pair is deterministically (−1)xi+xj(-1)^{x_i+x_j}. Thus, every complete outcome string excludes preparations with the opposite parity on any pair, establishing the WADM property. We see that along each branch, the protocol measures nn independent commuting Pauli observables, whose joint eigenspaces have rank one. Distinct branches give orthogonal projectors, since their first differing outcomes belong to orthogonal eigenspaces of the same observable. Thus these projectors form a rank-one PVM. □\square

Although ∣T⟩\lvert T\rangle and Z∣T⟩Z\lvert T\rangle are orthogonal, they are not SSD by a single Pauli measurement, and so their ontological representations could, in principle, overlap in this restricted setting. The WADM constructed above, together with Corollary 4.2, however, rules this out under preparation independence.

For later developments, the same protocol also gives a conclusion assuming independence only between two blocks. We write the block states explicitly as ∣T⟩⊗n\lvert T\rangle^{\otimes n} and (Z∣T⟩)⊗n(Z\lvert T\rangle)^{\otimes n}.

Corollary 6.2 (Block preparation independence). For n,n′≥1n,n' \ge1, let μn\mu_n, μn′\mu_{n'}, and μn′⊥\mu_{n'}^\perp represent ∣T⟩⊗n|T\rangle^{\otimes n}, ∣T⟩⊗n′|T\rangle^{\otimes n'}, and (Z∣T⟩)⊗n′(Z|T\rangle)^{\otimes n'}, respectively. Suppose an ontological model reproduces the adaptive Pauli protocol above and represents the composite preparations

∣T⟩⊗n⊗∣T⟩⊗n′,∣T⟩⊗n⊗(Z∣T⟩)⊗n′|T\rangle^{\otimes n} \otimes|T\rangle^{\otimes n'}, \qquad|T\rangle^{\otimes n} \otimes(Z|T\rangle)^{\otimes n'}

by μn⊗μn′\mu_n \otimes\mu_{n'} and μn⊗μn′⊥\mu_n \otimes\mu_{n'}^\perp, respectively. Then

Supp⁡(μn′)∩Supp⁡(μn′⊥)=∅,\operatorname{Supp}(\mu_{n'}) \cap\operatorname{Supp}(\mu_{n'}^\perp) = \varnothing,

or equivalently, ω(μn′,μn′⊥)=0\omega(\mu_{n'},\mu_{n'}^\perp)=0. Preparation independence is not assumed within either block.

Proof. Choose ii in the first block and jj in the second. For the two preparations, the selected qubits are in ∣ψ00⟩|\psi_{00}\rangle and ∣ψ01⟩|\psi_{01}\rangle, respectively. The two preparations can also differ on the unmeasured qubits, but these do not affect the selected-qubit probabilities. The two-qubit adaptive protocol described above therefore gives opposite deterministic second outcomes. Every terminal outcome excludes one of the two preparations, so the protocol is a WADM for this pair. Theorem 4.1 gives

Supp⁡(μn⊗μn′)∩Supp⁡(μn⊗μn′⊥)=∅.\operatorname{Supp}(\mu_n \otimes\mu_{n'}) \cap\operatorname{Supp}(\mu_n \otimes\mu_{n'}^\perp) = \varnothing.

The intersection equals

Supp⁡(μn)×(Supp⁡(μn′)∩Supp⁡(μn′⊥)).\operatorname{Supp}(\mu_n) \times\left(\operatorname{Supp}(\mu_{n'}) \cap\operatorname{Supp}(\mu_{n'}^\perp)\right).

Since μn\mu_n is normalized, its support is nonempty, and the result follows.

A hierarchy of antidistinguishing measurements

We now generalize the previous construction, based on the ZZZZ parity measurements, to construct WADMs for additional families of preparations arising from higher levels of the Clifford hierarchy. Define the single-qubit states

∣ψb(k)⟩:=SbRk∣+⟩=12(∣0⟩+ibζk∣1⟩),b∈Z2,(11)|\psi_b^{(k)}\rangle:= S^bR_k|+\rangle= \frac{1}{\sqrt{2}}\left(|0\rangle+ i^b\zeta_k|1\rangle\right), \qquad b \in\mathbb{Z}_2, \tag*{(11)}

where RkR_k and ζk\zeta_k are as in Section 3. Recall that ζk=e2πi/2k\zeta_k=e^{2\pi i/2^k}. The two states are nonorthogonal for every k≥2k \ge2. To each nn-bit string z∈Z2nz \in\mathbb{Z}_2^n, we associate the preparation

∣ψz(k)⟩:=⨂i=1n∣ψzi(k)⟩=2−n/2∑x∈Z2niz⋅xζk∣x∣∣x⟩,|\psi_z^{(k)}\rangle:= \bigotimes_{i=1}^{n}|\psi_{z_i}^{(k)}\rangle= 2^{-n/2}\sum_{x\in\mathbb{Z}_2^n}i^{z\cdot x}\zeta_k^{|x|}|x\rangle,

where z⋅x:=∑izixiz\cdot x := \sum_i z_i x_i is evaluated over the integers so that the exponent of ii is evaluated modulo four, not modulo two. Let ∣ϕ⟩|\phi\rangle be a stabilizer state. Using the normal form in Eq. (1), we have the overlap

⟨ϕ∣ψz(k)⟩=Pϕ,z(ζk)2n∣A∣,where Pϕ,z(t):=∑x∈Aiz⋅x−q(x)t∣x∣.(12)\langle\phi|\psi_z^{(k)}\rangle= \frac{P_{\phi,z}(\zeta_k)}{\sqrt{2^n|\mathcal{A}|}}, \qquad\text{where } P_{\phi,z}(t) := \sum_{x\in\mathcal{A}} i^{z\cdot x-q(x)}t^{|x|}. \tag*{(12)}

We will be interested in when these overlaps are precisely zero, since this corresponds to the outcome of a stabilizer PVM having zero probability.

To this end, it will be useful to interpret this overlap through the polynomial Pϕ,z(t)P_{\phi,z}(t) in a complex variable tt with coefficients in

Q(i):={a+ib:a,b∈Q},\mathbb{Q}(i):=\{a+ib:a,b\in\mathbb{Q}\},

which is a field [46]; see also Appendix § 10. This polynomial has degree at most nn. At the same time, the minimal polynomial of ζk\zeta_k over Q(i)\mathbb{Q}(i) is given by

mk(t)=tDk−i,(13)m_k(t)=t^{D_k}-i, \tag*{(13)}

which has degree Dk:=2k−2D_k:=2^{k-2}. These algebraic developments are important for making the following observation.

Observation 6.3. If n<Dkn < D_k, then a stabilizer overlap in (12) vanishes if and only if Pφ,z(t)P_{\varphi,z}(t) is the zero polynomial. Equivalently, every Hamming-weight coefficient vanishes separately. In particular, the first qubit number at which a zero can use the new algebraic relation ζkDk=i\zeta_k^{D_k} = i is n=Dkn = D_k.

The threshold establishes the minimal number of qubits at which a nonzero overlap polynomial can vanish at ζk\zeta_k. Below the threshold, zero overlaps can still occur through cancellation within each Hamming-weight coefficient. Indeed, at the threshold n=Dkn = D_k, we have the cat state

∣S1(k)⟩=12(∣0⟩⊗Dk−i∣1⟩⊗Dk),\lvert S_1^{(k)}\rangle= \frac{1}{\sqrt{2}}\left(\lvert0\rangle^{\otimes D_k} - i\lvert1\rangle^{\otimes D_k}\right),

with stabilizer group

S1(k)=⟨−Y1X2⋯XDk,ZiZi+1:i=1,…,Dk−1⟩.\mathcal{S}_1^{(k)} = \langle-Y_1X_2\cdots X_{D_k}, Z_iZ_{i+1}:i=1,\ldots,D_k-1\rangle.

This state gives an algebraic zero

⟨S1(k)∣(∣ψ0(k)⟩)⊗Dk⟩=2−(Dk+1)/2(1+iζkDk)=0.\langle S_1^{(k)}\mid(\lvert\psi_0^{(k)}\rangle)^{\otimes D_k}\rangle= 2^{-(D_k+1)/2}\left(1+i\zeta_k^{D_k}\right)=0.

The threshold alone does not establish the minimum number of qubits needed for a WADM. At the threshold, however, we have the following result for every k≥3k \ge3.

Theorem 6.4. For every k≥3k \ge3, the complete product family generated by the nonorthogonal single-qubit states in (11) admits a rank-one stabilizer WADM on

Dk=2k−2D_k=2^{k-2}

independently prepared copies.

The key observation is that a parity measurement acting on ∣ψa(k)⟩⊗∣ψb(k)⟩\lvert\psi_a^{(k)}\rangle\otimes\lvert\psi_b^{(k)}\rangle yields the even and odd parity post-measurement states represented, up to phase and normalization, by

∣vE(k)(a,b)⟩:=∣00⟩+ia+bζk2∣11⟩,(14)\lvert v_E^{(k)}(a,b)\rangle:= \lvert00\rangle+i^{a+b}\zeta_k^2\lvert11\rangle, \tag*{(14)}
∣vO(k)(a,b)⟩:=∣01⟩+ia−b∣10⟩.(15)\lvert v_O^{(k)}(a,b)\rangle:= \lvert01\rangle+i^{a-b}\lvert10\rangle. \tag*{(15)}

The odd case yields a stabilizer state. In the even case, since

ζk−1=ζk2,\zeta_{k-1}=\zeta_k^2,

we recognize the post-measurement state as an encoded single qubit with phase level k−1k-1. Here the encoding is ∣0⟩↦∣00⟩\lvert0\rangle\mapsto\lvert00\rangle and ∣1⟩↦∣11⟩\lvert1\rangle\mapsto\lvert11\rangle. Note that the exponent a+ba+b is an integer sum, so it may equal two. For the recursive construction, we use the inputs 00 and 01, which give the encoded states ∣ψ0(k−1)⟩\lvert\psi_0^{(k-1)}\rangle and ∣ψ1(k−1)⟩\lvert\psi_1^{(k-1)}\rangle, respectively. Our proof will use this recursive property to construct a WADM.

To see how this works, let us work out the cases k=3,4k=3,4, the first two nontrivial instances. For k=3k=3 we use D3=2D_3=2 qubits. Let z=abz=ab label the product preparation ∣ψab(3)⟩\lvert\psi_{ab}^{(3)}\rangle. Performing first the parity measurement ZZZZ we obtain

∣vE(3)(a,b)⟩:=∣00⟩+ia+b+1∣11⟩,(16)\lvert v_E^{(3)}(a,b)\rangle:= \lvert00\rangle+i^{a+b+1}\lvert11\rangle, \tag*{(16)}
∣vO(3)(a,b)⟩:=∣01⟩+ia−b∣10⟩.(17)\lvert v_O^{(3)}(a,b)\rangle:= \lvert01\rangle+i^{a-b}\lvert10\rangle. \tag*{(17)}

Notice that both post-measurement states are stabilizer states in this case. Measure next the Pauli XYXY, regardless of outcome. Equivalently, we perform the stabilizer PVM with rays

∣Ω00⟩=∣00⟩−i∣11⟩,∣Ω11⟩=∣00⟩+i∣11⟩,∣Ω01⟩=∣01⟩+i∣10⟩,∣Ω10⟩=∣01⟩−i∣10⟩.(18)\begin{aligned} \lvert\Omega_{00}\rangle&= \lvert00\rangle-i\lvert11\rangle, & \lvert\Omega_{11}\rangle&= \lvert00\rangle+i\lvert11\rangle,\\ \lvert\Omega_{01}\rangle&= \lvert01\rangle+i\lvert10\rangle, & \lvert\Omega_{10}\rangle&= \lvert01\rangle-i\lvert10\rangle. \tag*{(18)} \end{aligned}

It is straightforward to check that each vector is an eigenstate of both ZZZZ and XYXY. The labeling is chosen such that the preparation ∣ψab(3)⟩\lvert\psi_{ab}^{(3)}\rangle is excluded by the PVM element projecting onto ∣Ωab⟩\lvert\Omega_{ab}\rangle.

The higher-level case of k=4k=4 will build on the lower-level measurement scheme. Now we use D4=4D_4=4 qubits. Recall that ζ4=eiπ/8\zeta_4=e^{i\pi/8} and ζ42=ζ3=eiπ/4\zeta_4^2=\zeta_3=e^{i\pi/4}. The bit string z=z1z2z3z4z=z_1z_2z_3z_4 corresponds to the product preparation

∣ψz(4)⟩=∣ψz1(4)⟩⊗⋯⊗∣ψz4(4)⟩.|\psi_z^{(4)}\rangle=|\psi_{z_1}^{(4)}\rangle\otimes\cdots\otimes|\psi_{z_4}^{(4)}\rangle.

Pair qubits (1,2)(1,2) and (3,4)(3,4) and first measure Z1Z2Z_1Z_2 and Z3Z4Z_3Z_4. Introduce the even and odd code bases

∣0E⟩=∣00⟩,∣1E⟩=∣11⟩,∣0O⟩=∣01⟩,∣1O⟩=∣10⟩,|0_E\rangle=|00\rangle,\quad|1_E\rangle=|11\rangle,\quad|0_O\rangle=|01\rangle,\quad|1_O\rangle=|10\rangle,

together with the logical eigenstates, for σ∈{+1,−1}\sigma\in\{+1,-1\},

∣xℓσ⟩=∣0ℓ⟩+σ∣1ℓ⟩,∣yℓσ⟩=∣0ℓ⟩+σi∣1ℓ⟩,ℓ∈{E,O}.|x_\ell^\sigma\rangle=|0_\ell\rangle+\sigma|1_\ell\rangle,\quad|y_\ell^\sigma\rangle=|0_\ell\rangle+\sigma i|1_\ell\rangle,\quad\ell\in\{E,O\}.

After the parity measurement on either pair, the projected two-qubit states are

∣vE(b,c)⟩=∣0E⟩+ib+cζ3∣1E⟩,∣vO(b,c)⟩=∣0O⟩+ib−c∣1O⟩.(19)|v_E(b,c)\rangle=|0_E\rangle+i^{b+c}\zeta_3|1_E\rangle,\qquad|v_O(b,c)\rangle=|0_O\rangle+i^{b-c}|1_O\rangle. \tag*{(19)}

For inputs bc=00,01bc=00,01, the even pair forms an encoded version of ∣ψ0(3)⟩|\psi_0^{(3)}\rangle or ∣ψ1(3)⟩|\psi_1^{(3)}\rangle, respectively, one level lower in the hierarchy. On an odd pair we have that the string bc=01bc=01 corresponds to the stabilizer state ∣yO−⟩|y_O^-\rangle; similarly, 1010 to ∣yO+⟩|y_O^+\rangle, and both 0000 and 1111 correspond to ∣xO+⟩|x_O^+\rangle.

The initial two parity measurements produce the syndrome (ℓ,ℓ′)(\ell,\ell'), where ℓ,ℓ′∈{E,O}\ell,\ell'\in\{E,O\}. Whenever we obtain an odd parity syndrome outcome there is a qubit pair in a stabilizer state on which a logical YY measurement excludes input 0101 for one outcome and input 1010 for the other. The (E,E)(E,E) syndrome is the recursive branch since, as mentioned above, the selected inputs 00,0100,01 on each even pair give encoded single-qubit third-level states. Our stabilizer PVM will use the four stabilizer rays

∣Ξ00⟩=∣0000⟩−i∣1111⟩2,∣Ξ11⟩=∣0000⟩+i∣1111⟩2,|\Xi_{00}\rangle=\frac{|0000\rangle-i|1111\rangle}{\sqrt{2}},\qquad|\Xi_{11}\rangle=\frac{|0000\rangle+i|1111\rangle}{\sqrt{2}},
∣Ξ01⟩=∣0011⟩+i∣1100⟩2,∣Ξ10⟩=∣0011⟩−i∣1100⟩2.(20)|\Xi_{01}\rangle=\frac{|0011\rangle+i|1100\rangle}{\sqrt{2}},\qquad|\Xi_{10}\rangle=\frac{|0011\rangle-i|1100\rangle}{\sqrt{2}}. \tag*{(20)}

These are exactly the encoded versions of the measurement vectors ∣Ωab⟩|\Omega_{ab}\rangle in Eq. (18). All four syndrome subspaces are refined to rank-one elements by the following independent, conditional Pauli measurements:

syndromeconditional Pauli generators
EEEEZ1Z3, X1X2Y3X4Z_1Z_3,\ X_1X_2Y_3X_4
EOEOX1X2, Y3X4X_1X_2,\ Y_3X_4
OEOEY1X2, X3X4Y_1X_2,\ X_3X_4
OOOOY1X2, X3X4Y_1X_2,\ X_3X_4

Table 18.

The last three rows produce the remaining orthogonal vectors

BEO={∣xEt⟩12∣yOs⟩34:s,t=±1},\mathcal{B}_{EO}=\{|x_E^t\rangle_{12}|y_O^s\rangle_{34}:s,t=\pm1\},
BOE={∣yOs⟩12∣xEt⟩34:s,t=±1},\mathcal{B}_{OE}=\{|y_O^s\rangle_{12}|x_E^t\rangle_{34}:s,t=\pm1\},
BOO={∣yOs⟩12∣xOt⟩34:s,t=±1}.(21)\mathcal{B}_{OO}=\{|y_O^s\rangle_{12}|x_O^t\rangle_{34}:s,t=\pm1\}. \tag*{(21)}

Altogether, the four sets in (17) and (19) give the sixteen-ray stabilizer PVM. In Table 1 we identify a bijection between preparations and the corresponding impossible outcome. In the non-even rows, the first ket refers to qubits (1,2) and the second to qubits (3,4). Every one of the sixteen preparation words occurs exactly once.

SyndromePVM rayPreparation zz
EEEEΞ00\Xi_{00}0000
EEEEΞ01\Xi_{01}0001
EEEEΞ10\Xi_{10}0100
EEEEΞ11\Xi_{11}0011
EOEOxE+⊗yO+x_E^+ \otimes y_O^+1001
EOEOxE−⊗yO+x_E^- \otimes y_O^+1101
EOEOxE+⊗yO−x_E^+ \otimes y_O^-0010
EOEOxE−⊗yO−x_E^- \otimes y_O^-1110
OEOEyO+⊗xE+y_O^+ \otimes x_E^+0101
OEOEyO+⊗xE−y_O^+ \otimes x_E^-0110
OEOEyO−⊗xE+y_O^- \otimes x_E^+1000
OEOEyO−⊗xE−y_O^- \otimes x_E^-1010
OOOOyO+⊗xO+y_O^+ \otimes x_O^+0111
OOOOyO+⊗xO−y_O^+ \otimes x_O^-1100
OOOOyO−⊗xO+y_O^- \otimes x_O^+1011
OOOOyO−⊗xO−y_O^- \otimes x_O^-1111

Table 1. Each product preparation ∣ψz(4)⟩|\psi_z^{(4)}\rangle, labeled by z∈Z24z \in\mathbb{Z}_2^4, is orthogonal to its assigned stabilizer PVM ray.

Proof of Theorem 6.4. We recursively construct a measurement MkM_k. For k=3k = 3, use the two-qubit ADM defined above. For k>3k > 3, pair the DkD_k qubits and proceed as follows.

  1. Measure ZZZZ on every pair, obtaining a syndrome ℓ∈{E,O}Dk/2\ell\in\{E,O\}^{D_k/2}.

  2. If ℓ\ell contains an OO, choose the first odd pair. Measure logical YY on this pair and logical XX on every remaining pair.

  3. If ℓ\ell is all even, apply Mk−1M_{k-1} to the Dk/2=Dk−1D_k/2 = D_{k-1} encoded qubits, using ∣0E⟩=∣00⟩|0_E\rangle= |00\rangle and ∣1E⟩=∣11⟩|1_E\rangle= |11\rangle.

To verify the WADM property, consider a complete outcome. In step 2, the inputs 01 and 10 on the selected odd pair give opposite logical YY eigenstates. Either logical-YY outcome therefore excludes one of these inputs, independently of the inputs and subsequent outcomes on the remaining pairs.

In step 3, the physical inputs 00 and 01 give the encoded states ∣ψ0(k−1)⟩|\psi_0^{(k-1)}\rangle and ∣ψ1(k−1)⟩|\psi_1^{(k-1)}\rangle. By induction, every recursive outcome excludes a logical preparation word. Replacing each logical 0 by 00 and each logical 1 by 01 gives an original preparation word excluded by the complete outcome.

All logical measurements are physical Pauli measurements: on each parity subspace, logical XX, YY, and ZZ can be represented by XiXjX_iX_j, YiXjY_iX_j, and ZiZ_i, respectively. Each syndrome subspace is resolved into rank-one stabilizer projectors, directly in step 2 and recursively in step 3. Distinct syndrome subspaces are orthogonal, so the resulting measurement is a rank-one stabilizer PVM. Since every outcome excludes a preparation, it is a WADM.

Antidistinguishing measurements from complete graph states

We now introduce a somewhat different construction that gives a stabilizer SADM for every n≥2n \ge2, with the preparation angle depending on nn. This construction is closer to the original argument due to [9]. In particular, we define

θn:=π4n,∣ψa⟩:=cos⁡θn∣0⟩+(−1)asin⁡θn∣1⟩,a∈Z2.\theta_n := \frac{\pi}{4n}, \qquad|\psi_a\rangle:= \cos\theta_n|0\rangle+ (-1)^a \sin\theta_n|1\rangle,\qquad a \in\mathbb{Z}_2.

These states are nonorthogonal, since ⟨ψ0∣ψ1⟩=cos⁡(2θn)>0\langle\psi_0|\psi_1\rangle= \cos(2\theta_n) > 0. For x∈Z2nx \in\mathbb{Z}_2^n, write

∣ψx⟩:=⨂j=1n∣ψxj⟩.|\psi_x\rangle:= \bigotimes_{j=1}^{n}|\psi_{x_j}\rangle.

We recall some facts about graph states; see e.g., [49]. For a graph G=(V,E)G=(V,E) with vertex set V={1,⋯ ,n}V=\{1,\cdots,n\} and edge set EE, the corresponding graph state is defined by

∣G⟩=2−n/2∑x∈Z2n(−1)Γ(x)∣x⟩,where Γ(x)=∑(i,j)∈Exixj.|G\rangle=2^{-n/2}\sum_{x\in\mathbb{Z}_2^n}(-1)^{\Gamma(x)}|x\rangle,\qquad\text{where }\Gamma(x)=\sum_{(i,j)\in E}x_i x_j.

Every graph state is also a stabilizer state whose stabilizer group has the canonical generating set

Qi:=Xi⨂j∈N(i)Zj,i=1,⋯ ,n,Q_i:=X_i\bigotimes_{j\in N(i)}Z_j,\qquad i=1,\cdots,n,

where N(i)⊂VN(i)\subset V is the neighborhood of the vertex ii. The Pauli operator ZjZ_j commutes with QiQ_i for i≠ji\ne j and anticommutes with QjQ_j. Thus, we have the orthonormal stabilizer basis with rays

Zx∣G⟩,Zx:=⨂j=1nZjxj.Z^x|G\rangle,\qquad Z^x:=\bigotimes_{j=1}^{n}Z_j^{x_j}.

The complete graph KnK_n has an edge between every pair of distinct vertices. Its corresponding graph state is

∣Kn⟩=2−n/2∑z∈Z2n(−1)(∣z∣2)∣z⟩,|K_n\rangle=2^{-n/2}\sum_{z\in\mathbb{Z}_2^n}(-1)^{\binom{|z|}{2}}|z\rangle,

since Γ(z)=∑i<jzizj=(∣z∣2)\Gamma(z)=\sum_{i<j}z_i z_j=\binom{|z|}{2} and precisely ∣z∣|z| entries of zz are equal to one. We let the corresponding orthonormal basis rays be written as ∣ϕx⟩:=Zx∣Kn⟩|\phi_x\rangle:=Z^x|K_n\rangle.

Theorem 6.5. For every n≥2n\ge2, the product preparations

{∣ψx⟩⟨ψx∣:x∈Z2n}\{|\psi_x\rangle\langle\psi_x|:x\in\mathbb{Z}_2^n\}

admit a rank-one stabilizer ADM.

Proof. For preparation ∣ψy⟩|\psi_y\rangle, let d:=∣x+y∣d:=|x+y| be the Hamming distance between xx and yy, and write c=cos⁡θnc=\cos\theta_n, s=sin⁡θns=\sin\theta_n. Then

⟨ϕx∣ψy⟩=2−n/2∑z∈Z2n(−1)(∣z∣2)+(x+y)⋅zcn−∣z∣s∣z∣.\langle\phi_x|\psi_y\rangle=2^{-n/2}\sum_{z\in\mathbb{Z}_2^n}(-1)^{\binom{|z|}{2}+(x+y)\cdot z}c^{n-|z|}s^{|z|}.

The signs (−1)(w2)(-1)^{\binom{w}{2}} follow the repeating pattern 1,1,−1,−11,1,-1,-1, and hence

(−1)(w2)=Re⁡(iw)+Im⁡(iw).(-1)^{\binom{w}{2}}=\operatorname{Re}(i^w)+\operatorname{Im}(i^w).

Applying this identity and factoring the sum over bits gives

⟨ϕx∣ψy⟩=2−n/2(Re⁡[(c+is)n−d(c−is)d]+Im⁡[(c+is)n−d(c−is)d]).\langle\phi_x|\psi_y\rangle=2^{-n/2}\left(\operatorname{Re}\left[(c+is)^{n-d}(c-is)^d\right]+\operatorname{Im}\left[(c+is)^{n-d}(c-is)^d\right]\right).

Since c±is=e±iθnc\pm is=e^{\pm i\theta_n}, this becomes

⟨ϕx∣ψy⟩=2−n/2[cos⁡((n−2d)θn)+sin⁡((n−2d)θn)].\langle\phi_x|\psi_y\rangle=2^{-n/2}\left[\cos((n-2d)\theta_n)+\sin((n-2d)\theta_n)\right].

For θn=π/(4n)\theta_n=\pi/(4n), the argument lies in [−π/4,π/4][-\pi/4,\pi/4]. On this interval, cos⁡t+sin⁡t\cos t+\sin t vanishes only at t=−π/4t=-\pi/4. Consequently,

⟨ϕx∣ψy⟩=0⟺d=n⟺y=x+1.\langle\phi_x|\psi_y\rangle=0\quad\Longleftrightarrow\quad d=n\quad\Longleftrightarrow\quad y=x+1.

Thus each measurement outcome excludes exactly one preparation, and each preparation has exactly one excluding outcome. In particular, the measurement is a SADM.

Our construction yields a PBR proof for every number n≥2n\ge2 of qubits. Although less general than the SADM proposed by PBR, it uses a measurement scheme built from Pauli measurements and graph states. The required angles θn=π/(4n)\theta_n=\pi/(4n) enter through the preparations. Whether this simplification of the measurement protocol makes actual implementation simpler requires separate consideration.

Preparation independence in the Pauli ontological model

We now specialize our preceding results to the Pauli ontological model [35], which furnishes a classical simulation algorithm for universal QCM based on convex polytopes. (For preliminaries on convex polytopes, we refer the reader to [50].) Particularly relevant to any classical simulation algorithm is its scaling in nn, the number of qubits. In many approaches, including those based on quasiprobability distributions [19, 20] or stabilizer rank methods [39, 42, 51], tensor products provide a way to construct representations of larger systems from those of their constituent parts. Here we use our WADMs for magic states to show that such tensor-product constructions in the Pauli ontological model encounter obstructions arising from the PBR argument.

The Pauli polytopes

Let SnS_n denote the set of pure nn-qubit stabilizer states. The nn-qubit Pauli polytope, also called the Λ\Lambda polytope, is defined as the intersection of halfspace inequalities

Pn:={W∈Herm⁡(H):Tr⁡(W)=1, ⟨S∣W∣S⟩≥0 ∀ ∣S⟩∈Sn}.(22)P_n := \{W \in\operatorname{Herm}(\mathcal{H}) : \operatorname{Tr}(W) = 1,\ \langle S|W|S\rangle\ge0\ \forall\ |S\rangle\in S_n\}. \tag*{(22)}

This is a bounded polytope [35], and hence the convex hull of its finitely many vertices [50]. We label the extreme points ext⁡(Pn)\operatorname{ext}(P_n) of PnP_n by Λn\Lambda_n.

Every quantum state belongs to PnP_n and therefore admits a decomposition

ρ=∑λ∈Λnμρ(λ)Aλ,μρ(λ)≥0,∑λ∈Λnμρ(λ)=1.\rho= \sum_{\lambda\in\Lambda_n} \mu_\rho(\lambda) A_\lambda,\qquad\mu_\rho(\lambda) \ge0,\qquad\sum_{\lambda\in\Lambda_n} \mu_\rho(\lambda) = 1.

The labels λ\lambda are interpreted as ontic states, and μρ\mu_\rho as a preparation distribution. Such decompositions, however, are generally not unique. A measurement of a Pauli observable QQ is represented by a stochastic kernel Γ\Gamma satisfying

ΠQσAλΠQσ=∑λ′∈ΛnΓ(σ,λ′∣Q,λ)Aλ′.\Pi^\sigma_Q A_\lambda\Pi^\sigma_Q = \sum_{\lambda' \in\Lambda_n} \Gamma(\sigma,\lambda' \mid Q,\lambda) A_{\lambda'}.

where ΠQσ\Pi^\sigma_Q is the projector associated with the outcome σ∈{±1}\sigma\in\{\pm1\}, and

Γ(σ,λ′∣Q,λ)≥0,∑σ=±1∑λ′∈ΛnΓ(σ,λ′∣Q,λ)=1.\Gamma(\sigma,\lambda' \mid Q,\lambda) \ge0,\qquad\sum_{\sigma=\pm1}\sum_{\lambda' \in\Lambda_n} \Gamma(\sigma,\lambda' \mid Q,\lambda) = 1.

The existence of these kernels follows from the behavior of PnP_n under Pauli measurements; see [26, 35].

We are interested in whether independent preparations can be represented by independently sampled local ontic states. More technically, we study the corresponding operators combined by the tensor product. To this end, one compatibility result has already been established. For any Aλ∈ext⁡(Pn)A_\lambda\in\operatorname{ext}(P_n), any stabilizer state ∣S⟩∈Sm|S\rangle\in S_m, and any (n+m)(n+m)-qubit Clifford unitary UU, the operator

U(Aλ⊗∣S⟩⟨S∣)U†U(A_\lambda\otimes|S\rangle\langle S|)U^\dagger

is a vertex of Pn+mP_{n+m} [48]. This construction adds only stabilizer ancillas and Clifford processing, and therefore introduces no additional computational resource.

More relevant for quantum computation are tensor products of non-stabilizer resource states, such as those supplied by magic-state distillation protocols [37]. For general preparations, consider decompositions

ρ=∑λ∈Λnμρ(λ)Aλ,andρ′=∑λ′∈Λmμρ′(λ′)Aλ′.\rho= \sum_{\lambda\in\Lambda_n} \mu_\rho(\lambda) A_\lambda,\qquad\text{and}\qquad\rho' = \sum_{\lambda' \in\Lambda_m} \mu_{\rho'}(\lambda') A_{\lambda'}.

Their tensor product satisfies

ρ⊗ρ′=∑λ∈Λn∑λ′∈Λmμρ(λ)μρ′(λ′)(Aλ⊗Aλ′).\rho\otimes\rho' = \sum_{\lambda\in\Lambda_n}\sum_{\lambda' \in\Lambda_m} \mu_\rho(\lambda)\mu_{\rho'}(\lambda') (A_\lambda\otimes A_{\lambda'}).

The coefficients form a product probability distribution, as required by preparation independence. The remaining question is whether the tensored operators are valid elements of Pn+mP_{n+m}.

Were we to have Aλ⊗Aλ′′∈Pn+mA_{\lambda} \otimes A^{\prime}_{\lambda^{\prime}} \in P_{n+m}, then this construction yields an ontological model for universal QCM satisfying preparation independence. Indeed, if Aλ⊗Aλ′∈Pn+mA_{\lambda} \otimes A_{\lambda^{\prime}} \in P_{n+m}, then every stabilizer projective measurement MM has a normalized, non-negative response function

ξ(k∣M,λ,λ′)=⟨Sk∣Aλ⊗Aλ′∣Sk⟩,∣Sk⟩∈M.\xi(k|M,\lambda,\lambda^{\prime}) = \langle S^{k}|A_{\lambda} \otimes A_{\lambda^{\prime}}|S^{k}\rangle,\quad|S_{k}\rangle\in M.

Moreover, if Aλ⊗Aλ′A_{\lambda} \otimes A_{\lambda^{\prime}} belongs to Pn+mP_{n+m}, it is in fact a vertex of Pn+mP_{n+m}.2 Consequently, closure under tensor products of local vertices would identify each pair (λ,λ′)(\lambda,\lambda^{\prime}) with a global ontic state and provide an operator realization of preparation independence.

Our results up until this point suggest that preparation independence should fail for the magic states considered thus far, provided that their preparation distributions have overlapping support. In the next sections, we show that the geometry of the Pauli polytopes allows us to choose such distributions for distinct magic states.

Convex geometry of Pauli polytopes

The vertices that can appear in a decomposition of W∈PnW \in P_n are determined by its vanishing stabilizer probabilities. Indeed, if a stabilizer state ∣S⟩|S\rangle satisfies ⟨S∣W∣S⟩=0\langle S|W|S\rangle= 0, then

⟨S∣W∣S⟩=0=∑λ∈ΛnμW(λ)⟨S∣Aλ∣S⟩=0.\langle S|W|S\rangle= 0 = \sum_{\lambda\in\Lambda_n} \mu_W(\lambda)\langle S|A_{\lambda}|S\rangle= 0.

Since every term is nonnegative, this implies ⟨S∣Aλ∣S⟩=0\langle S|A_{\lambda}|S\rangle= 0 for all λ∈Supp⁡(μW)\lambda\in\operatorname{Supp}(\mu_W). These feasible vertices are elements of the carrier face of WW introduced below; see also [53].

Recall that F⊂PnF \subset P_n is a face of PnP_n if it is of the form

FV={W∈Pn:⟨S∣W∣S⟩=0,∀ ∣S⟩∈V},F_V = \{W \in P_n : \langle S|W|S\rangle= 0,\quad\forall\ |S\rangle\in V\},

where V⊂SnV \subset S_n is any subset of stabilizer states. Note that FF can be the empty set. Fix some W∈PnW \in P_n and let ZWZ_W be the set of stabilizer rays with zero expectation on WW. The carrier face of WW is then the inclusion minimal face of PnP_n containing WW. It is given by

FW:={A∈Pn:⟨S∣A∣S⟩=0 for every ∣S⟩∈ZW}.F_W := \{A \in P_n : \langle S|A|S\rangle= 0\text{ for every }|S\rangle\in Z_W\}.

In particular, we have that W∈relint⁡(FW)W \in\operatorname{relint}(F_W) is in the relative interior of FWF_W [50]. Geometrically, this means that there exists a ball of dimension dim⁡(FW)\operatorname{dim}(F_W) centered at WW with suitable radius ε\varepsilon contained in FWF_W.

The vertices of FWF_W are therefore precisely those vertices of PnP_n satisfying the above orthogonality conditions [53]. We define the vertex support of WW by

ΛW:={λ∈Λn:Aλ∈FW}.\Lambda_W := \{\lambda\in\Lambda_n : A_{\lambda} \in F_W\}.

Every vertex decomposition of WW satisfies

Supp⁡(μW)⊆ΛW.\operatorname{Supp}(\mu_W) \subseteq\Lambda_W.

Conversely, every label in ΛW\Lambda_W can occur with positive probability in some decomposition of WW.

Lemma 7.1. Let FF be a nonempty face of PnP_n and W∈relint⁡(F)W \in\operatorname{relint}(F). For every Aλ∈ext⁡(F)A_{\lambda} \in\operatorname{ext}(F), there exists a vertex decomposition of WW with μW(λ)>0\mu_W(\lambda) > 0.

Proof. Since W∈relint⁡(F)W \in\operatorname{relint}(F), for sufficiently small ε∈(0,1)\varepsilon\in(0,1) the point

R=W−εAλ1−εR = \frac{W-\varepsilon A_{\lambda}}{1-\varepsilon}

belongs to FF. Decomposing RR into vertices of FF and substituting into

W=εAλ+(1−ε)RW = \varepsilon A_{\lambda} + (1-\varepsilon)R

gives the required decomposition.

The following proposition identifies exactly when two operators admit decompositions with a common ontic state.

Proposition 7.2. For W,W′∈PnW,W' \in P_n, the following are equivalent:

  1. There exist vertex decompositions μW,μW′\mu_W,\mu_{W'} such that

Supp⁡(μW)∩Supp⁡(μW′)≠∅.\operatorname{Supp}(\mu_W) \cap\operatorname{Supp}(\mu_{W'}) \ne\varnothing.
  1. Their carrier faces satisfy

FW∩FW′≠∅.F_W \cap F_{W'} \ne\varnothing.

In particular, if FW=FW′F_W = F_{W'}, then any vertex of this common face can occur with positive coefficient in decompositions of both WW and W′W'.

Proof. If λ\lambda belongs to the support of both decompositions, then Aλ∈FW∩FW′A_{\lambda} \in F_W \cap F_{W'}. Conversely, a nonempty intersection FW∩FW′F_W \cap F_{W'} is a face of PnP_n and therefore contains a vertex AλA_{\lambda} of PnP_n. This is a vertex of both carrier faces. Applying Lemma 7.1 separately to WW and W′W' one can construct preparation distributions such that μW(λ)>0\mu_W(\lambda)>0 and μW′(λ)>0\mu_{W'}(\lambda)>0.

We can extend this result so that every vertex of FWF_W appears simultaneously with strictly positive weight in a single decomposition.

Corollary 7.3. Let FF be a face of a polytope. Every W∈relint⁡(F)W \in\operatorname{relint}(F) admits a decomposition

W=∑A∈ext⁡(F)pAA,pA>0,∑A∈ext⁡(F)pA=1.W = \sum_{A \in\operatorname{ext}(F)} p_A A,\qquad p_A > 0,\qquad\sum_{A \in\operatorname{ext}(F)} p_A = 1.

In particular, states with the same carrier face admit vertex decompositions with the same support.

Proof. Let Aˉ\bar{A} be the uniform average of the vertices of FF. Since W∈relint⁡(F)W \in\operatorname{relint}(F), for sufficiently small ε>0\varepsilon> 0 we have

B:=W−εAˉ1−ε∈F.B := \frac{W-\varepsilon\bar{A}}{1-\varepsilon} \in F.

Expressing BB as a convex combination of the vertices of FF and substituting into W=εAˉ+(1−ε)BW = \varepsilon\bar{A} + (1-\varepsilon)B gives a strictly positive coefficient for every vertex.

Thus, to establish the existence of overlapping preparation distributions, it suffices to show that the two states have the same carrier face. Notice, however, that this is a statement about which admissible decompositions exist. Not every pair of decompositions need overlap, even if the points share the same carrier face. An exception is when the carrier face is a simplex since the extremal points are then affinely independent.

Ontic supports for magic states

We now analyze the carrier faces of magic states, with a particular emphasis on TT-state preparations. A similar analysis for higher-level phases is straightforward. Introduce the notation

∣Tn⟩:=∣T⟩⊗n,Πn:=∣Tn⟩⟨Tn∣,Πn⊥:=Z⊗nΠnZ⊗n,Fn:=FΠn.\lvert T_n\rangle:= \lvert T\rangle^{\otimes n},\qquad\Pi_n := \lvert T_n\rangle\langle T_n\rvert,\qquad\Pi_n^\perp:= Z^{\otimes n}\Pi_n Z^{\otimes n},\qquad F_n := F_{\Pi_n}.

Here Πn⊥\Pi_n^\perp denotes the projector onto (Z∣T⟩)⊗n(Z\lvert T\rangle)^{\otimes n}, rather than the complementary projector I−ΠnI-\Pi_n. Our goal is to show that Πn\Pi_n and Πn⊥\Pi_n^\perp have the same carrier face. They will then admit vertex decompositions with common strictly positive support.

For a stabilizer ray ∣ϕ⟩\lvert\phi\rangle, the Born probability

Tr⁡(Πn∣ϕ⟩⟨ϕ∣)=∣⟨ϕ∣Tn⟩∣2\operatorname{Tr}(\Pi_n\lvert\phi\rangle\langle\phi\rvert)=\lvert\langle\phi\vert T_n\rangle\rvert^2

vanishes precisely when ⟨ϕ∣Tn⟩=0\langle\phi\vert T_n\rangle=0. Recall the overlap in Eq. (12) and specialize to k=3k=3 and z=0z=0. Since ζ32=i\zeta_3^2=i, separate the even and odd Hamming weights of the overlap polynomial according to

Pϕ,0(ζ3)=C0+ζ3C1,Cr∈Q(i),r=0,1.(23)P_{\phi,0}(\zeta_3)=C_0+\zeta_3C_1,\qquad C_r\in\mathbb{Q}(i),\quad r=0,1. \tag*{(23)}

Because 11 and ζ3\zeta_3 are linearly independent over Q(i)\mathbb{Q}(i), a zero overlap requires C0=C1=0C_0=C_1=0. Physically, the amplitudes cancel separately in the even and odd sectors. Observe that applying the unitary Z⊗nZ^{\otimes n} changes only the relative sign of these sectors, so it cannot undo either cancellation.

Proposition 7.4. For every n≥1n\ge1, the states Πn\Pi_n and Πn⊥\Pi_n^\perp have the same carrier face FnF_n, and hence the same vertex support.

Proof. The operator Z⊗nZ^{\otimes n} multiplies the amplitude at xx by (−1)∣x∣(-1)^{\lvert x\rvert}. The overlap with Z⊗n∣Tn⟩Z^{\otimes n}\lvert T_n\rangle is therefore proportional to C0−ζ3C1C_0-\zeta_3C_1. By linear independence, this vanishes if and only if C0=C1=0C_0=C_1=0. Thus the two states have exactly the same orthogonal stabilizer projectors and the same carrier face.

The same argument permits a continuous change of relative phase between the parity sectors. Let

Uα:=exp⁡(iα2Z⊗n)=cos⁡α2I+isin⁡α2Z⊗n.U_\alpha:=\exp\left(\frac{i\alpha}{2}Z^{\otimes n}\right)=\cos\frac{\alpha}{2}I+i\sin\frac{\alpha}{2}Z^{\otimes n}.

Proposition 7.5. For every n≥1n\ge1 and α∈R\alpha\in\mathbb{R}, UαΠnUα†∈FnU_\alpha\Pi_nU_\alpha^\dagger\in F_n. If e−iαζ3∉Q(i)e^{-i\alpha\zeta_3}\notin\mathbb{Q}(i), the transformed state has carrier face exactly FnF_n. In particular, this holds for α=π\alpha=\pi.

Proof. The transformed overlap is proportional to

eiα/2C0+e−iα/2ζ3C1=eiα/2(C0+e−iαζ3C1).(24)e^{i\alpha/2}C_0+e^{-i\alpha/2}\zeta_3C_1=e^{i\alpha/2}\left(C_0+e^{-i\alpha\zeta_3}C_1\right). \tag*{(24)}

Every original zero is preserved since C0=C1=0C_0=C_1=0. For e−iαζ3∉Q(i)e^{-i\alpha\zeta_3}\notin\mathbb{Q}(i), we have exact equality of the zero sets.

At exceptional angles, new stabilizer zeros can appear and the state can move to a proper subface. A second remark concerns symmetries. The Pauli polytopes are dual to the stabilizer polytopes and therefore have the extended Clifford group as their symmetry group [36]. The unitary above is therefore not a symmetry of PnP_n, except in special cases. Instead, it is a symmetry of a convex set within FnF_n.

Tensor obstruction for magic state supports

In this section we show that tensoring operator decompositions of TT-states need not give a valid decomposition in the composite Pauli polytope. The argument combines the carrier face results above with the two-block TT-state protocol of § 6.1. The underlying obstruction is that independent local sampling can produce a product operator that assigns a negative probability to a stabilizer measurement outcome.

Since Πm\Pi_m and Πm⊥\Pi_m^\perp have the same carrier face, they admit decompositions with strictly positive weights on overlapping vertices (Corollary 7.3). Tensoring these decompositions with a fixed decomposition of Πn\Pi_n preserves this overlap. However, the protocol of § 6.1, together with Corollary 6.2, perfectly distinguishes these two composite preparations using adaptive Pauli measurements. Consequently, they cannot share an ontic state with valid response functions. At least one of the product operators in the common support must therefore lie outside the composite Pauli polytope.

Proposition 7.6. For every n,m≥1n,m \ge1, there exist vertices A∈ext⁡(Fn)A \in\operatorname{ext}(F_n) and B∈ext⁡(Fm)B \in\operatorname{ext}(F_m) such that

A⊗B∉Pn+m.A \otimes B \notin P_{n+m}.

Proof. Choose decompositions of Πn\Pi_n positive on every vertex of FnF_n, and of both Πm\Pi_m and Πm⊥\Pi_m^\perp positive on every vertex of FmF_m. The composite preparations

ρ0:=Πn⊗Πm,ρ1:=Πn⊗Πm⊥\rho_0 := \Pi_n \otimes\Pi_m,\qquad\rho_1 := \Pi_n \otimes\Pi_m^\perp

then have strictly positive decompositions over the product ontic space ext⁡(Fn)×ext⁡(Fm)\operatorname{ext}(F_n) \times\operatorname{ext}(F_m). Suppose every A⊗BA \otimes B in this product set belongs to Pn+mP_{n+m}. Let {Φl}l\{\Phi_l\}_l be the stabilizer measurement from Corollary 6.2. For each outcome ll, at least one of Tr⁡(Φlρ0)\operatorname{Tr}(\Phi_l\rho_0) and Tr⁡(Φlρ1)\operatorname{Tr}(\Phi_l\rho_1) is zero. Expanding that preparation gives a strictly positive sum of nonnegative terms Tr⁡(Φl(A⊗B))\operatorname{Tr}(\Phi_l(A \otimes B)). Every such term must vanish. For any fixed pair of vertices, however,

∑lTr⁡(Φl(A⊗B))=Tr⁡(A⊗B)=1,\sum_l \operatorname{Tr}(\Phi_l(A \otimes B)) = \operatorname{Tr}(A \otimes B) = 1,

a contradiction.

This argument shows only that some product operator violates a stabilizer inequality. One could imagine, however, that a carefully chosen decomposition could avoid the problematic product operator(s), thereby salvaging the construction. In the next section we show that this is not the case.

A general tensor obstruction for qubits

We now establish a stronger structural result about Pauli polytopes: every tensor product of two qubit Pauli-polytope vertices lies outside the composite polytope.

For A∈PnA \in P_n and B∈PmB \in P_m, every product stabilizer state satisfies

⟨Sn∣⟨Sm∣A⊗B∣Sn⟩∣Sm⟩=⟨Sn∣A∣Sn⟩⟨Sm∣B∣Sm⟩≥0.\langle S_n|\langle S_m|A \otimes B|S_n\rangle|S_m\rangle= \langle S_n|A|S_n\rangle\langle S_m|B|S_m\rangle\ge0.

Thus, any violation must be witnessed by an entangled stabilizer state. For vertices, we construct such a witness by superposing two carefully chosen product stabilizer states. Each product state has a local factor with zero expectation on the corresponding vertex.

Lemma 7.7. Let A∈ext⁡(Pn)A \in\operatorname{ext}(P_n), with n≥1n \ge1. There exist two orthogonal stabilizer states ∣a0⟩|a_0\rangle, ∣a1⟩|a_1\rangle, agreeing on n−1n-1 independent stabilizer generators and differing in the sign of one remaining generator, such that

⟨a0∣A∣a0⟩=0,α:=⟨a0∣A∣a1⟩≠0.\langle a_0|A|a_0\rangle= 0,\qquad\alpha:= \langle a_0|A|a_1\rangle\ne0.

Their phases can be chosen so that they are the encoded computational basis of a rank-two stabilizer code.

Proof. Let ZAZ_A denote the set of tight stabilizer rays, namely those satisfying ⟨a∣A∣a⟩=0\langle a|A|a\rangle= 0. Since AA is a vertex, the identity together with the projectors formed from ZAZ_A span the real vector space of Hermitian operators. Otherwise, there would be a nonzero traceless Hermitian direction preserving all tight equalities, along which both sufficiently small positive and negative perturbations of AA remain in PnP_n, contradicting extremality; see e.g., [54].

First, we need to show that not every ∣a⟩∈ZA|a\rangle\in Z_A satisfies A∣a⟩=0A|a\rangle= 0. We proceed by contradiction. Suppose instead that every tight stabilizer ray belongs to K:=ker⁡AK := \ker A. Since Tr⁡A=1\operatorname{Tr} A = 1, AA is not the zero operator and ker⁡A\ker A is a proper subspace. This means we can choose unit vectors ∣u⟩∈K|u\rangle\in K and ∣v⟩∈K⊥|v\rangle\in K^\perp, and define the Hermitian operator

H:=∣u⟩⟨v∣+∣v⟩⟨u∣.H := |u\rangle\langle v| + |v\rangle\langle u|.

Observe that Tr⁡H=0\operatorname{Tr} H = 0 and ⟨a∣H∣a⟩=0\langle a|H|a\rangle= 0 for every tight stabilizer ray ∣a⟩|a\rangle, since ∣a⟩∈K|a\rangle\in K. In particular, we have ⟨a∣(A±εH)∣a⟩=0\langle a|(A \pm\varepsilon H)|a\rangle= 0 for all ∣a⟩∈ZA|a\rangle\in Z_A so that all tight equalities are preserved along HH. Since all remaining stabilizer inequalities are strict, there exists a sufficiently small ε>0\varepsilon> 0 such that A±εH∈PnA \pm\varepsilon H \in P_n. This contradicts extremality. Hence some tight stabilizer ray ∣a0⟩|a_0\rangle satisfies A∣a0⟩≠0A|a_0\rangle\ne0.

Recall that any stabilizer state is of the form ∣a0⟩=U∣0⟩⊗n|a_0\rangle= U|0\rangle^{\otimes n}, for some Clifford unitary UU. Complete ∣a0⟩|a_0\rangle to an orthonormal stabilizer basis B\mathcal{B} by applying UU to the computational basis. Since ∣a0⟩|a_0\rangle is not in the kernel of AA and because BB is a basis, there exists some ∣a1⟩∈B\lvert a_1\rangle\in B such that ⟨a0∣A∣a1⟩≠0\langle a_0\lvert A\lvert a_1\rangle\ne0. It is known that two stabilizer rays related by the same unitary to the computational basis can rearrange their stabilizer generators so that they differ in the sign of just a single generator. Removing that generator yields a rank-two stabilizer code.

Theorem 7.8. For all n,m≥1n,m \ge1 and all vertices A∈ext⁡(Pn)A \in\operatorname{ext}(P_n) and B∈ext⁡(Pm)B \in\operatorname{ext}(P_m),

A⊗B∉Pn+m.A \otimes B \notin P_{n+m}.

More precisely, there is an entangled rank-one stabilizer projector Π\Pi with Tr⁡(Π(A⊗B))<0\operatorname{Tr}(\Pi(A \otimes B)) < 0.

Proof. Apply Lemma 7.7 to AA and BB. Choose encoded stabilizer pairs satisfying

⟨a0∣A∣a0⟩=0,α:=⟨a0∣A∣a1⟩≠0,⟨b0∣B∣b0⟩=0,β:=⟨b0∣B∣b1⟩≠0.\langle a_0\lvert A\lvert a_0\rangle= 0,\qquad\alpha:= \langle a_0\lvert A\lvert a_1\rangle\ne0,\qquad\langle b_0\lvert B\lvert b_0\rangle= 0,\qquad\beta:= \langle b_0\lvert B\lvert b_1\rangle\ne0.

Put W:=A⊗BW := A \otimes B and consider the product states

∣u⟩:=∣a0⟩⊗∣b1⟩,∣v⟩:=∣a1⟩⊗∣b0⟩.\lvert u\rangle:= \lvert a_0\rangle\otimes\lvert b_1\rangle,\qquad\lvert v\rangle:= \lvert a_1\rangle\otimes\lvert b_0\rangle.

Their expectations vanish:

⟨u∣W∣u⟩=⟨v∣W∣v⟩=0.\langle u\lvert W\lvert u\rangle= \langle v\lvert W\lvert v\rangle= 0.

Their off-diagonal matrix element, however, is

c:=⟨u∣W∣v⟩=αβ‾≠0.c := \langle u\lvert W\lvert v\rangle= \alpha\overline{\beta} \ne0.

Let ZAZ_A and ZBZ_B denote the logical Pauli operators distinguishing the two states in each local pair. The vectors u,vu,v share the n+m−2n+m-2 local code generators and the additional generator −ZAZB-Z_AZ_B. They therefore span a rank-two stabilizer code. With the phases fixed by the local Clifford encodings, its complementary logical Pauli bases consist of

∣x±⟩:=∣u⟩±∣v⟩2,∣y±⟩:=∣u⟩±i∣v⟩2.\lvert x_\pm\rangle:= \frac{\lvert u\rangle\pm\lvert v\rangle}{\sqrt{2}},\qquad\lvert y_\pm\rangle:= \frac{\lvert u\rangle\pm i\lvert v\rangle}{\sqrt{2}}.

These are rank-one stabilizer states, each entangled across the n∣mn|m partition. Direct calculation gives

⟨x±∣W∣x±⟩=±Re⁡(c),⟨y±∣W∣y±⟩=∓Im⁡(c).\langle x_\pm\lvert W\lvert x_\pm\rangle= \pm\operatorname{Re}(c),\qquad\langle y_\pm\lvert W\lvert y_\pm\rangle= \mp\operatorname{Im}(c).

Since c≠0c \ne0, at least one of these expectations is strictly negative. In fact, the violating projector can be chosen so that

Tr⁡(ΠW)=−max⁡{∣Re⁡(αβ‾)∣,∣Im⁡(αβ‾)∣}<0.\operatorname{Tr}(\Pi W) = -\max\{|\operatorname{Re}(\alpha\overline{\beta})|,|\operatorname{Im}(\alpha\overline{\beta})|\} < 0.

Thus A⊗BA \otimes B violates a defining stabilizer inequality of Pn+mP_{n+m}.

Consequently, independently sampling local vertex decompositions and returning the operator A⊗BA \otimes B never produces a valid composite ontic operator. In short: preparation independence fails for the Pauli ontological model.

Conclusion

In this paper we initiated the development of PBR proofs for non-stabilizer preparation families using Pauli measurements. For the single-qubit preparation families considered at levels k≥3k \ge3 of the Clifford hierarchy, we showed that stabilizer-restricted PBR proofs exist using Dk=2k−2D_k = 2^{k-2} independently prepared systems. Beyond these Clifford-hierarchy families, we also constructed such measurement schemes for every n≥2n \ge2, with the preparation states depending on nn. Although these constructions do not necessarily provide new conceptual insights into the PBR theorem, they open an avenue for studying particularly simple and elegant proofs, similar in spirit to the study of Kochen–Specker proofs [55, 56]. Particularly interesting is the role of algebraic number theory in the geometry of stabilizer-restricted antidistinguishability, which our work has only begun to explore. Related connections between number theory and quantum-state geometry have been developed, for instance, in the study of SIC measurements [57].

We also demonstrated a general obstruction to preparation independence in ontological representations of qubit stabilizer theory. Our results imply that any such model satisfying preparation independence must be ψ\psi-ontic on pure stabilizer states, suggesting the stabilizer states themselves as a natural candidate ontic state space. Indeed, these states also serve as the underlying state space in quasiprobability representations of qubit stabilizer theory [20]. This contrasts with the odd-prime-dimensional qudit case, where stabilizer states have nonnegative Gross Wigner representations and the phase-point operators are closed under tensor products. Several distinctions between the even- and odd-dimensional cases have been related to topological obstructions [34]. Whether related obstructions also account for the difference in tensor-product structure appears plausible (see Appendix § 11), but remains an interesting question for future work.

Our results also connect the operational constraints imposed by antidistinguishability to the geometry of Pauli polytopes. In this representation, the obstruction to preparation independence appears concretely as a failure of closure under tensor products: the tensor product of any two local vertices lies outside the corresponding composite-system polytope. This restricts how these polytopes can represent independently prepared systems. From a categorical perspective, it appears to obstruct a monoidal representation in which parallel composition is implemented by the ordinary operator tensor product. While a full categorical treatment of polyhedral classical simulation schemes [58] remains open, some steps in this direction were taken in [59].

This question also bears on the computational complexity of polytope-based classical simulation. Zurel et al. [60] showed that the simulation requires only polynomially many bits to describe its instantaneous state, leaving the presumed computational hardness in the initial sampling or subsequent measurement updates. An efficiently implementable tensor-product rule would make initialization of fixed single-qubit product inputs straightforward: one could sample the local distributions independently and compose the resulting ontic states. Our result rules out this construction using operator tensor products of local vertices. It does not establish the hardness of initial sampling, but it does remove a natural reason to expect initialization to be easy. The difficulty of classical simulation may therefore arise before the first measurement, in representing completely unentangled inputs within an ontological state space not admitting a product structure.

Acknowledgments

The author thanks P. Mitosek, R. Raussendorf, and A. T. Yücel for useful discussions. The author acknowledges support from the Niedersächsisches Ministerium für Wissenschaft und Kultur and the Alexander von Humboldt Foundation. The author acknowledges the use of ChatGPT in literature review, manuscript revision, and the development of mathematical arguments. The author has verified the final results and takes full responsibility for the manuscript.

Algebra background

Here we recall the facts about field extensions and cyclotomic fields used in the main text. For a detailed treatment, see, e.g., [46].

Field extensions

Let F\mathbb{F} be a field and let α\alpha be algebraic over F\mathbb{F}, meaning that it is a root of a nonzero polynomial in F[x]\mathbb{F}[x]. The minimal polynomial mα(x)m_{\alpha}(x) is the unique monic polynomial of least degree satisfying mα(α)=0m_{\alpha}(\alpha)=0. It is irreducible over F\mathbb{F}, and

P(α)=0⟺mα(x)∣P(x)for P∈F[x].P(\alpha)=0 \Longleftrightarrow m_{\alpha}(x)\mid P(x) \quad\text{for } P\in\mathbb{F}[x].

In particular, no nonzero polynomial of degree less than deg⁡mα\deg m_{\alpha} can vanish at α\alpha.

The field F(α)\mathbb{F}(\alpha) is the smallest field containing F\mathbb{F} and α\alpha. Its degree over F\mathbb{F} is its dimension as an F\mathbb{F}-vector space, and

[F(α):F]=deg⁡mα.[\mathbb{F}(\alpha):\mathbb{F}]=\deg m_{\alpha}.

For a tower of finite extensions F⊆K⊆L\mathbb{F}\subseteq K\subseteq L, the degrees satisfy

[L:F]=[L:K][K:F].[L:\mathbb{F}]=[L:K][K:\mathbb{F}].

Cyclotomic fields

A primitive rrth root of unity is ξr=e2πi/r\xi_r=e^{2\pi i/r}. Its minimal polynomial over Q\mathbb{Q} is the cyclotomic polynomial

Φr(x)=∏1≤j≤rgcd⁡(j,r)=1(x−ξrj).\Phi_r(x)=\prod_{\substack{1\leq j\leq r\\ \gcd(j,r)=1}}(x-\xi_r^j).

This polynomial is irreducible over Q\mathbb{Q} and has degree φ(r)\varphi(r), where φ\varphi is Euler’s totient function. Consequently,

[Q(ξr):Q]=φ(r).[\mathbb{Q}(\xi_r):\mathbb{Q}]=\varphi(r).

In the main text we use ζk=ξ2k\zeta_k=\xi_{2^k}. For k≥2k\geq2, we have i=ζk2k−2i=\zeta_k^{2^{k-2}}, so Q(i)⊆Q(ζk)\mathbb{Q}(i)\subseteq\mathbb{Q}(\zeta_k). The tower law therefore gives

[Q(ζk):Q(i)]=φ(2k)[Q(i):Q]=2k−2=:Dk.[\mathbb{Q}(\zeta_k):\mathbb{Q}(i)]=\frac{\varphi(2^k)}{[\mathbb{Q}(i):\mathbb{Q}]}=2^{k-2}=:D_k.

Since ζkDk=i\zeta_k^{D_k}=i, the polynomial xDk−ix^{D_k}-i vanishes at ζk\zeta_k and has precisely this degree. It is therefore the minimal polynomial of ζk\zeta_k over Q(i)\mathbb{Q}(i). Thus, for P∈Q(i)[x]P \in\mathbb{Q}(i)[x],

P(ζk)=0⟺xDk−i∣P(x).P(\zeta_k)=0 \Longleftrightarrow x^{D_k}-i \mid P(x).

In particular, if deg⁡P<Dk\deg P < D_k, then P(ζk)=0P(\zeta_k)=0 only when PP is the zero polynomial. This is the algebraic constraint used in the stabilizer-overlap argument.

Contextuality-based PBR theorem for stabilizer states

We now give a contextuality-based proof using the notation and conventions of the main body together with the results of [27], which connect contextuality to partitioning measurements. For a preparation family P={ρs:s∈S}P=\{\rho_s:s\in S\}, let ΩP\Omega_P be the set of Hermitian Pauli observables whose measurement has a deterministic outcome on at least one ρs\rho_s. Denote its closure under products of commuting observables by Ω~P\widetilde{\Omega}_P. The relevant Kochen–Specker relations are

ν(QR)=ν(Q)ν(R)if [Q,R]=0,ν(±I)=±1.\nu(QR)=\nu(Q)\nu(R) \quad\text{if } [Q,R]=0,\qquad\nu(\pm I)=\pm1.

These relations express the algebraic constraints on compatible measurement outcomes.

In the ontological model, preparations ρs\rho_s are represented by probability distributions μs\mu_s. We assume that the response functions are noncontextual under fine-graining: if N={ΠNj}jN=\{\Pi_N^j\}_j refines M={ΠMk}kM=\{\Pi_M^k\}_k, with

ΠMk=∑j:f(j)=kΠNj,\Pi_M^k=\sum_{j:f(j)=k}\Pi_N^j,

then

ξ(k∣M,λ)=∑j:f(j)=kξ(j∣N,λ).\xi(k\mid M,\lambda)=\sum_{j:f(j)=k}\xi(j\mid N,\lambda).

In particular, the response to a coarse-grained outcome is independent of the chosen refinement.

Suppose now that there exists some λ\lambda in the joint support of the preparations, i.e. λ∈⋂s∈SSupp⁡(μs)\lambda\in\bigcap_{s\in S}\operatorname{Supp}(\mu_s). For every Q∈ΩPQ\in\Omega_P, some preparation has a deterministic outcome ν(Q)∈{±1}\nu(Q)\in\{\pm1\}. Nonnegativity of the ontological representation therefore gives

ξ(k∣MQ,λ)=δk,ν(Q).\xi(k\mid M_Q,\lambda)=\delta_{k,\nu(Q)}.

If two preparations impose different values on the same observable, this already excludes a common λ\lambda. Now let Q,RQ,R be commuting observables whose responses at λ\lambda are deterministic. Their joint measurement refines both MQM_Q and MRM_R, so its response is concentrated on (ν(Q),ν(R))(\nu(Q),\nu(R)). The same joint measurement refines MQRM_{QR}, with product outcome qrqr. Consequently,

ξ(k∣MQR,λ)=δk,ν(Q)ν(R).\xi(k\mid M_{QR},\lambda)=\delta_{k,\nu(Q)\nu(R)}.

Iterating this argument extends the deterministic assignment to Ω~P\widetilde{\Omega}_P, where it satisfies

ν(QR)=ν(Q)ν(R)whenever [Q,R]=0,ν(±I)=±1.\nu(QR)=\nu(Q)\nu(R) \quad\text{whenever } [Q,R]=0,\qquad\nu(\pm I)=\pm1.

If Ω~P\widetilde{\Omega}_P admits a proof of state-independent contextuality, no such assignment exists. Thus we conclude

⋂s∈SSupp⁡(μs)=∅.\bigcap_{s\in S}\operatorname{Supp}(\mu_s)=\varnothing.

Similar notions have been developed in [27].

Consider now two distinct, nonorthogonal pure stabilizer preparations ρ0,ρ1\rho_0,\rho_1, with stabilizer groups S0,S1\mathcal{S}_0,\mathcal{S}_1. There exist anticommuting g∈S0g\in\mathcal{S}_0 and h∈S1h\in\mathcal{S}_1. Otherwise, maximality would imply that the two groups have the same unsigned observables, making the states either identical or orthogonal. Define the preparations

ρab:=ρa⊗ρb,P:={ρab:a,b∈Z2}.\rho_{ab}:=\rho_a\otimes\rho_b,\qquad P:=\{\rho_{ab}:a,b\in\mathbb{Z}_2\}.

Since these are stabilizer states, the set of observables ΩP\Omega_P and its closure Ω~P\widetilde{\Omega}_P are nonempty.

Proposition B.1. The closure Ω~P\widetilde{\Omega}_{P} contains a Mermin square. Consequently, every ontological model reproducing the allowed measurements and state updates satisfies

⋂a,b∈Z2Supp⁡(μρab)=∅.\bigcap_{a,b\in\mathbb{Z}_{2}} \operatorname{Supp}(\mu_{\rho_{ab}})=\varnothing.

Proof. Put l:=ighl := \mathrm{i}gh, so that ll is a Hermitian Pauli observable. Consider the array

g⊗II⊗gg⊗gI⊗hh⊗Ih⊗hg⊗hh⊗gl⊗l\begin{array}{ccc} g\otimes I & I\otimes g & g\otimes g \\ I\otimes h & h\otimes I & h\otimes h \\ g\otimes h & h\otimes g & l\otimes l \end{array}

This is precisely the structure of the Peres-Mermin square [61]. The first eight observables belong to ΩP\Omega_{P}, each having deterministic value +1+1 on some preparation. The final observable L:=l⊗lL := l\otimes l belongs to Ω~P\widetilde{\Omega}_{P} since

(g⊗h)(h⊗g)=L,(g⊗g)(h⊗h)=−L.(g\otimes h)(h\otimes g)=L,\qquad(g\otimes g)(h\otimes h)=-L.

Each row and column consists of commuting observables. The row products are II, while the column products are II, II, −I-I, precluding a consistent value assignment. By the construction of Karanjai et al. [27 Theorem 2], MLM_{L} is a partitioning measurement, thus excluding common ontic support for the states in PP.

Corollary B.2. An ontological model reproducing qubit stabilizer theory and satisfying preparation independence must represent distinct pure stabilizer preparations ρ0,ρ1\rho_{0},\rho_{1} by distributions satisfying

Supp⁡(μρ0)∩Supp⁡(μρ1)=∅.\operatorname{Supp}(\mu_{\rho_{0}})\cap\operatorname{Supp}(\mu_{\rho_{1}})=\varnothing.

Proof. Orthogonal stabilizer preparations are SSD and therefore have disjoint supports. For a distinct nonorthogonal pair, the result follows from Proposition B.1.

Thus a model with overlapping distributions for distinct pure qubit stabilizer states must violate preparation independence. Note that a similar result for qudit stabilizer theory with odd local dimension does not hold. In that case the Wigner representation yields a noncontextual model that is also closed under tensor product, thus satisfying preparation independence. As a concluding remark, we note that the Mermin-Peres square paradox [61] can be understood topologically [62].

References

  1. [1]J. S. Bell, “On the Einstein Podolsky Rosen paradox,” Physics Physique Fizika, vol. 1, no. 3, p. 195, 1964. https://doi.org/10.1103/physicsphysiquefizika.1.195.
  2. [2]S. Kochen and E. P. Specker, “The problem of hidden variables in quantum mechanics,” in Ernst Specker Selecta, pp. 235–263, Springer, 2011. https://doi.org/10.1007/978-3-0348-9259-9_21.
  3. [3]J. S. Bell, “On the problem of hidden variables in quantum mechanics,” Reviews of Modern physics, vol. 38, no. 3, p. 447, 1966. https://doi.org/10.1103/revmodphys.38.447.
  4. [4]R. W. Spekkens, “Contextuality for preparations, transformations, and unsharp measurements,” Physical Review A—Atomic, Molecular, and Optical Physics, vol. 71, no. 5, p. 052108, 2005. https://doi.org/10.1103/physreva.71.052108.
  5. [5]E. Wigner, “On the quantum correction for thermodynamic equilibrium,” Physical review, vol. 40, no. 5, p. 749, 1932. https://doi.org/10.1103/physrev.40.749.
  6. [6]C. Ferrie and J. Emerson, “Frame representations of quantum mechanics and the necessity of negativity in quasi-probability representations,” Journal of Physics A: Mathematical and Theoretical, vol. 41, no. 35, p. 352001, 2008. https://doi.org/10.1088/1751-8113/41/35/352001.
  7. [7]N. Harrigan and R. W. Spekkens, “Einstein, incompleteness, and the epistemic view of quantum states,” Foundations of Physics, vol. 40, no. 2, pp. 125–157, 2010. https://doi.org/10.1007/s10701-009-9347-0.
  8. [8]L. Hardy, “Quantum theory from five reasonable axioms,” arXiv preprint quant-ph/0101012, 2001. https://arxiv.org/abs/quant-ph/0101012.
  9. [9]M. F. Pusey, J. Barrett, and T. Rudolph, “On the reality of the quantum state,” Nature Physics, vol. 8, pp. 475–478, 2012. https://doi.org/10.1038/nphys2309.
  10. [10]C. M. Caves, C. A. Fuchs, and R. Schack, “Conditions for compatibility of quantum-state assignments,” Physical Review A, vol. 66, no. 6, p. 062111, 2002. https://doi.org/10.1103/physreva.66.062111.
  11. [11]M. S. Leifer, “Is the quantum state real? an extended review of ψ-ontology theorems,” arXiv preprint arXiv:1409.1570, 2014. https://arxiv.org/abs/1409.1570.
  12. [12]M. Srikumar, S. D. Bartlett, and A. Karanjai, “How contextuality and antidistinguishability are related,” Physical Review Letters, vol. 136, p. 080203, 2026. https://doi.org/10.1103/tlh1-xwks.
  13. [13]R. Raussendorf and H. J. Briegel, “A one-way quantum computer,” Phys. Rev. Lett., vol. 86, pp. 5188–5191, May 2001. https://doi.org/10.1103/PhysRevLett.86.5188.
  14. [14]J. Anders and D. E. Browne, “Computational power of correlations,” Physical Review Letters, vol. 102, no. 5, p. 050502, 2009. https://doi.org/10.1103/physrevlett.102.050502.
  15. [15]M. J. Hoban, E. T. Campbell, K. Loukopoulos, and D. E. Browne, “Non-adaptive measurement-based quantum computation and multi-party bell inequalities,” New Journal of Physics, vol. 13, no. 2, p. 023014, 2011. https://doi.org/10.1088/1367-2630/13/2/023014.
  16. [16]R. Raussendorf, “Contextuality in measurement-based quantum computation,” Phys. Rev. A, vol. 88, p. 022322, Aug 2013. https://doi.org/10.1103/PhysRevA.88.022322.
  17. [17]M. Frembs, S. Roberts, E. T. Campbell, and S. D. Bartlett, “Hierarchies of resources for measurement-based quantum computation,” New Journal of Physics, vol. 25, no. 1, p. 013002, 2023. https://doi.org/10.1088/1367-2630/acaee2.
  18. [18]V. Veitch, C. Ferrie, D. Gross, and J. Emerson, “Negative quasi-probability as a resource for quantum computation,” New Journal of Physics, vol. 14, no. 11, p. 113011, 2012. https://doi.org/10.1088/1367-2630/14/11/113011.
  19. [19]H. Pashayan, J. J. Wallman, and S. D. Bartlett, “Estimating outcome probabilities of quantum circuits using quasiprobabilities,” Physical review letters, vol. 115, no. 7, p. 070501, 2015. https://doi.org/10.1103/physrevlett.115.070501.
  20. [20]M. Howard and E. Campbell, “Application of a resource theory for magic states to fault-tolerant quantum computing,” Physical review letters, vol. 118, no. 9, p. 090501, 2017. https://doi.org/10.1103/physrevlett.118.090501.
  21. [21]R. Raussendorf, J. Bermejo-Vega, E. Tyhurst, C. Okay, and M. Zurel, “Phase-space-simulation method for quantum computation with magic states on qubits,” Physical Review A, vol. 101, no. 1, p. 012350, 2020. https://doi.org/10.1103/physreva.101.012350.
  22. [22]M. Howard, J. Wallman, V. Veitch, and J. Emerson, “Contextuality supplies the ‘magic’ for quantum computation,” Nature, vol. 510, no. 7505, pp. 351–355, 2014. https://doi.org/10.1038/nature13460.
  23. [23]N. Delfosse, C. Okay, J. Bermejo-Vega, D. E. Browne, and R. Raussendorf, “Equivalence between contextuality and negativity of the Wigner function for qudits,” New Journal of Physics, vol. 19, no. 12, p. 123024, 2017. https://doi.org/10.1088/1367-2630/aa8fe3.
  24. [24]D. Schmid, H. Du, J. H. Selby, and M. F. Pusey, “Uniqueness of noncontextual models for stabilizer subtheories,” Physical Review Letters, vol. 129, no. 12, p. 120403, 2022. https://doi.org/10.1103/physrevlett.129.120403.
  25. [25]R. Raussendorf, D. E. Browne, N. Delfosse, C. Okay, and J. Bermejo-Vega, “Contextuality and Wigner-function negativity in qubit quantum computation,” Physical Review A, vol. 95, no. 5, p. 052334, 2017. https://doi.org/10.1103/physreva.95.052334.
  26. [26]C. Okay, A. T. Yucel, and S. Ipek, “Classical simulation of universal measurement-based quantum computation using multipartite Bell scenarios,” arXiv preprint arXiv:2410.23734, 2024. https://arxiv.org/abs/2410.23734.
  27. [27]A. Karanjai, J. J. Wallman, and S. D. Bartlett, “Contextuality bounds the efficiency of classical simulation of quantum processes,” arXiv preprint arXiv:1802.07744, 2018. https://arxiv.org/abs/1802.07744.
  28. [28]A. Montina, “Exponential complexity and ontological theories of quantum mechanics,” Physical Review A—Atomic, Molecular, and Optical Physics, vol. 77, no. 2, p. 022104, 2008. https://doi.org/10.1103/physreva.77.022104.
  29. [29]M. Kleinmann, O. Gühne, J. R. Portillo, J.-Å. Larsson, and A. Cabello, “Memory cost of quantum contextuality,” New Journal of Physics, vol. 13, no. 11, p. 113011, 2011. https://doi.org/10.1088/1367-2630/13/11/113011.
  30. [30]D. Gross, “Hudson’s theorem for finite-dimensional quantum systems,” Journal of mathematical physics, vol. 47, no. 12, 2006. https://doi.org/10.1063/1.2393152.
  31. [31]N. Delfosse, P. Allard Guerin, J. Bian, and R. Raussendorf, “Wigner function negativity and contextuality in quantum computation on rebits,” Physical Review X, vol. 5, no. 2, p. 021003, 2015. https://doi.org/10.1103/physrevx.5.021003.
  32. [32]J. Bermejo-Vega, N. Delfosse, D. E. Browne, C. Okay, and R. Raussendorf, “Contextuality as a resource for models of quantum computation with qubits,” Physical review letters, vol. 119, no. 12, p. 120505, 2017. https://doi.org/10.1103/physrevlett.119.120505.
  33. [33]P. Lillystone and J. Emerson, “A contextual ψ-epistemic model of the n-qubit stabilizer formalism,” arXiv preprint arXiv:1904.04268, 2019. https://arxiv.org/abs/1904.04268.
  34. [34]R. Raussendorf, C. Okay, M. Zurel, and P. Feldmann, “The role of cohomology in quantum computation with magic states,” Quantum, vol. 7, p. 979, 2023. https://doi.org/10.22331/q-2023-04-13-979.
  35. [35]M. Zurel, C. Okay, and R. Raussendorf, “Hidden variable model for universal quantum computation with magic states on qubits,” Physical Review Letters, vol. 125, no. 26, p. 260404, 2020. https://doi.org/10.1103/physrevlett.125.260404.
  36. [36]M. Heinrich and D. Gross, “Robustness of magic and symmetries of the stabiliser polytope,” Quantum, vol. 3, p. 132, 2019. https://doi.org/10.22331/q-2019-04-08-132.
  37. [37]S. Bravyi and A. Kitaev, “Universal quantum computation with ideal Clifford gates and noisy ancillas,” Physical Review A, vol. 71, no. 2, p. 022316, 2005. https://doi.org/10.1103/physreva.71.022316.
  38. [38]D. Gottesman and I. L. Chuang, “Demonstrating the viability of universal quantum computation using teleportation and single-qubit operations,” Nature, vol. 402, no. 6760, pp. 390–393, 1999. https://doi.org/10.1038/46503.
  39. [39]S. Bravyi and D. Gosset, “Improved classical simulation of quantum circuits dominated by Clifford gates,” Physical review letters, vol. 116, no. 25, p. 250501, 2016. https://doi.org/10.1103/physrevlett.116.250501.
  40. [40]M. Van den Nest, “Classical simulation of quantum computation, the Gottesman-Knill theorem, and slightly beyond,” Quantum Information and Computation, vol. 10, no. 3&4, pp. 258–271, 2010. https://doi.org/10.26421/QIC10.3–4–6.
  41. [41]D. Gottesman, “The Heisenberg representation of quantum computers,” in Group22: Proceedings of the XXIII International Colloquium on Group Theoretical Methods in Physics (S. P. Corney, R. Delbourgo, and P. D. Jarvis, eds.), (Cambridge, MA), pp. 32–43, International Press, 1999. https://arxiv.org/abs/quant-ph/9807006.
  42. [42]S. Bravyi, G. Smith, and J. A. Smolin, “Trading classical and quantum computational resources,” Physical Review X, vol. 6, no. 2, p. 021043, 2016. https://doi.org/10.1103/physrevx.6.021043.
  43. [43]M. Yoganathan, R. Jozsa, and S. Strelchuk, “Quantum advantage of unitary Clifford circuits with magic state inputs,” Proceedings of the Royal Society A, vol. 475, no. 2225, p. 20180427, 2019. https://doi.org/10.1098/rspa.2018.0427.
  44. [44]F. C. R. Peres, S. Ipek, D. Lovsted, J. Nevin, and A. Broadbent, “Normal-form characterization of adaptivity and entanglement in Pauli-based computation,” arXiv preprint arXiv:2610.30315, 2026. https://arxiv.org/abs/2610.30315.
  45. [45]S. X. Cui, D. Gottesman, and A. Krishna, “Diagonal gates in the Clifford hierarchy,” Physical Review A, vol. 95, no. 1, p. 012329, 2017. https://doi.org/10.1103/PhysRevA.95.012329.
  46. [46]I. N. Herstein, Topics in algebra. John Wiley & Sons, 1991.
  47. [47]S. Ipek, A. T. Yucel, F. Shahi, C. Ozdemir, and C. Okay, “Phase-space tableau simulation for quantum computation,” Phys. Rev. A, vol. 113, p. 032409, Mar 2026. https://link.aps.org/doi/10.1103/11f2-nqht.DOI
  48. [48]C. Okay, M. Zurel, and R. Raussendorf, “On the extremal points of the Λ-polytopes and classical simulation of quantum computation with magic states,” arXiv preprint arXiv:2104.05822, 2021. https://arxiv.org/abs/2104.05822.
  49. [49]M. Hein, W. Dür, J. Eisert, R. Raussendorf, M. Van den Nest, and H.-J. Briegel, “Entanglement in graph states and its applications,” arXiv preprint quant-ph/0602096, 2006. https://arxiv.org/abs/quant-ph/0602096.
  50. [50]G. M. Ziegler, Lectures on polytopes, vol. 152. Springer Science & Business Media, 2012. https://doi.org/10.1007/978-1-4613-8431-1.
  51. [51]S. Bravyi, D. Browne, P. Calpin, E. Campbell, D. Gosset, and M. Howard, “Simulation of quantum circuits by low-rank stabilizer decompositions,” Quantum, vol. 3, p. 181, 2019. https://doi.org/10.22331/q-2019-09-02-181.
  52. [52]M. Plávala, “General probabilistic theories: An introduction,” Physics Reports, 2023. https://doi.org/10.1016/j.physrep.2023.09.001.
  53. [53]R. T. Rockafellar, Convex analysis, vol. 28. Princeton university press, 1997. https://press.princeton.edu/books/paperback/9780691015866/convex-analysis.
  54. [54]V. Chvátal, Linear programming. Macmillan, 1983.DOI
  55. [55]A. Cabello, J. Estebaranz, and G. García-Alcaine, “Bell-Kochen-Specker theorem: A proof with 18 vectors,” Physics Letters A, vol. 212, no. 4, pp. 183–187, 1996. https://doi.org/10.1016/0375-9601(96)00134-X.
  56. [56]M. Waegell and P. Aravind, “Proofs of the Kochen–Specker theorem based on a system of three qubits,” Journal of Physics A: Mathematical and Theoretical, vol. 45, no. 40, p. 405301, 2012. https://doi.org/10.1088/1751-8113/45/40/405301.
  57. [57]M. Appleby, S. Flammia, G. McConnell, and J. Yard, “Sics and algebraic number theory,” Foundations of Physics, vol. 47, no. 8, pp. 1042–1059, 2017. https://doi.org/10.1007/s10701-017-0090-7.
  58. [58]C. Okay, “Polyhedral classical simulators for quantum computation,” Research in the Mathematical Sciences, vol. 13, no. 2, p. 55, 2026. https://doi.org/10.1007/s40687-026-00626-x.
  59. [59]C. Okay, W. Stern, R. Haderi, and S. Ipek, “Double categories for adaptive quantum computation,” arXiv preprint arXiv:2510.25915, 2025. https://arxiv.org/abs/2510.25915.
  60. [60]M. Zurel, C. Okay, and R. Raussendorf, “Simulating quantum computation: How many “bits” for “it”?,” PRX Quantum, vol. 5, no. 3, p. 030343, 2024. https://doi.org/10.1103/prxquantum.5.030343.
  61. [61]N. D. Mermin, “Hidden variables and the two theorems of John Bell,” Reviews of Modern Physics, vol. 65, no. 3, p. 803, 1993. https://doi.org/10.1103/RevModPhys.65.803.
  62. [62]C. Okay, S. Roberts, S. D. Bartlett, and R. Raussendorf, “Topological proofs of contextuality in quantum mechanics,” Quantum Information & Computation, vol. 17, no. 13-14, pp. 1135–1166, 2017. https://doi.org/10.26421/QIC17.13-14-5.

Paper details

vlnc:W_xzq7xk25
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

ChatGPT

View statement
Looking for related papers…
Contents