betaFeedback

WorksPhysics and Astronomy

Third-Order Quantum Information from Distributed Randomized Measurements

Abstract

Randomized measurements offer a way to probe nonlinear properties of quantum states without reconstructing their full density matrices. For states shared between distant laboratories, we study products of three outcome probabilities measured at the same local settings and then averaged over independent local random unitaries. We determine, for an arbitrary number of parties, which third-order local-unitary invariant combinations these averages retain and which they cannot separate. The accessible information is encoded in algebraic invariants of the quantum state and is reconstructed by analysing the permutation symmetries induced by the distributed measurement protocol. Exploiting these symmetries, the reconstruction problem collapses, regardless of the number of measurement outcomes, to the symbolic inversion of a 3 x 3 matrix. As a result, nonlinear information can be extracted without inverting the exponentially large matrices that typically arise in distributed multipartite quantum systems.

Introduction

Many properties relevant to quantum-state characterization are nonlinear functions of the density matrix, yet can be determined without full-state reconstruction [1, 2, 3, 4, 5, 6]. Randomized measurements provide access to such information through products of outcome probabilities averaged over random measurement bases [7, 8, 9]. For spatially separated subsystems, the central question is which nonlinear properties can be recovered from local measurements and the classical records shared between laboratories.

We study third-order correlators obtained by measuring three independent preparations at the same local unitary setting and averaging over Haar distribution. The protocol requires only local rotations, projective local measurements, and classical post-processing, with no global measurement or coherent coupling between copies. After averaging, each local outcome triple is specified by its equality pattern: all distinct, exactly one equal pair, or all equal. Replacing repeated averaged entries by these patterns loses no further information. The local Haar averages assign identical coefficients to some distinct invariants, so they appear in the measurement equations only through sums. We ask which invariant combinations these equations determine and which they cannot separate.

Classical-shadow protocols use recorded settings and outcomes to estimate selected observables [10]. Elben et al. [11] combine records from independently sampled settings to estimate partial-transpose moments individually. Other randomized-measurement schemes access additional invariant polynomials or optimize estimation costs [12, 13, 14, 15].

Our reconstruction takes only the averaged outcome correlators as input. Unlike estimators that use recorded settings, it requires neither calibration nor storage of each individual unitary matrix: one only needs to identify shots taken at the same setting. This removes the need to characterize every random rotation for post-processing, while retaining the assumptions of Haar sampling (or an exact 3-design) and the specified measurement model. For resource tests built from the retained invariant combinations, only the combinations required by the test need to be reconstructed.

We solve this identifiability and reconstruction problem for an arbitrary number of parties and local dimensions. Simultaneous relabelling of the three copies organises the retained coordinates into permutation orbits. A change of basis, independent of the local Hilbert-space dimensions, then separates the reconstruction map into independent sectors built from 3×33 \times3 matrices and scalars.

The construction recovers the known bipartite reconstruction structure [16, 17], including the fact that the global and partial-transpose third moments enter only through their sum, and extends that structure to arbitrary party number. Its contribution is an exact characterization of the nonlinear information available from distributed measurements.

Section II defines the protocol. Sections III and IV develop the one- and two-party constructions, and Sec. V establishes their multipartite extension. Section VI constructs the reconstruction sectors, followed by the main theorem in Sec. VII. Section VIII then examines partial-transpose invariance and the information limits of the averaged correlators.

Randomized measurements

Let ρ\rho be a density matrix on ⨂k=1nCdk\bigotimes_{k=1}^{n}\mathbb{C}^{d_k}. Party kk applies a unitary UkU_k and measures in the computational basis {∣ik⟩}ik=1dk\{\lvert i_k\rangle\}_{i_k=1}^{d_k}. Write U=(U1,…,Un)U=(U_1,\ldots,U_n) for the setting and i=(i1,…,in)\mathbf{i}=(i_1,\ldots,i_n) for a joint outcome tuple. The conditional probability is

pρ(i∣U)=Tr⁡[ρ⨂k=1nUk†∣ik⟩⟨ik∣Uk].(1)p_{\rho}(\mathbf{i}\mid U)=\operatorname{Tr}\left[\rho\bigotimes_{k=1}^{n}U_k^{\dagger}\lvert i_k\rangle\langle i_k\rvert U_k\right]. \tag*{(1)}

For three independently prepared copies measured at the same product setting UU, the joint conditional probability of the outcome tuples i1,i2,i3i_1,i_2,i_3 factorizes as

pρ(i1,i2,i3∣U)=∏h=13pρ(ih∣U).(2)p_{\rho}(i_1,i_2,i_3\mid U)=\prod_{h=1}^{3}p_{\rho}(i_h\mid U). \tag*{(2)}

Each probability is linear in ρ\rho, so their product is cubic. The three shots use the same setting, and then their product is averaged over independent local Haar distributions,

yout(i1,i2,i3)=∫∏k=1ndUk∏h=13pρ(ih∣U).=∑qh,phh=1,2,3(∏h=13ρqhph)∏k=1n∫dUk∏h=13(Uk)ikhqkh(Uk‾)ikhpkh.(3)\begin{aligned} y_{\mathrm{out}}(i_1,i_2,i_3)=\int\prod_{k=1}^{n}\mathrm{d}U_k\prod_{h=1}^{3}p_{\rho}(i_h\mid U). \\ &=\sum_{\substack{q_h,p_h\\h=1,2,3}}\left(\prod_{h=1}^{3}\rho_{q_hp_h}\right)\prod_{k=1}^{n}\int\mathrm{d}U_k\prod_{h=1}^{3}(U_k)_{i_k^{h}q_k^{h}}(\overline{U_k})_{i_k^{h}p_k^{h}}. \tag*{(3)} \end{aligned}

This probes third-order information without a coherent measurement coupling the copies. Collecting all outcome triples gives the data vector yout∈R∏kdk3\mathbf{y}_{\mathrm{out}}\in\mathbb{R}^{\prod_k d_k^3}. The integrand contains three matrix entries of each local unitary and three conjugate entries. An exact unitary 3-design, an ensemble matching these Haar moments, therefore gives the same mean correlators [18, 19]. For a qubit register at each party, the multiqubit Clifford group provides such an ensemble [20]. We first develop the construction for n=1n=1 to introduce the building blocks for the main theorem.

One party

Equality patterns and local cycle merging

For one party, the Haar average in Eq. (3) connects measured outcome triples to scalar invariants of ρ\rho. Weingarten calculus evaluates the required integral as a sum over permutations of the three copy positions [21, 22, 23]:

∫dU∏h=13UihqhU‾ihph=∑σ,τ∈S3Wg⁡(στ−1,d)∏h=13δih,iσ(h)δqh,pτ(h).(4)\int\mathrm{d}U\prod_{h=1}^{3}U_{i_hq_h}\overline{U}_{i_hp_h} =\sum_{\sigma,\tau\in S_3}\operatorname{Wg}(\sigma\tau^{-1},d)\prod_{h=1}^{3}\delta_{i_h,i_{\sigma(h)}}\delta_{q_h,p_{\tau(h)}}. \tag*{(4)}

and the formula (2.3) takes the form:

yout(i1,i2,i3)=∑σ,τ∈S3∏h=13δih,iσ(h)Wg⁡(στ−1,d)∑p1,p2,p3,q1,q2,q3ρq1p1ρq2p2ρq3p3∏h=13δqh,pτ(h).(5)y_{\mathrm{out}}(i_1,i_2,i_3)=\sum_{\sigma,\tau\in S_3}\prod_{h=1}^{3}\delta_{i_h,i_{\sigma(h)}}\operatorname{Wg}(\sigma\tau^{-1},d) \sum_{p_1,p_2,p_3,q_1,q_2,q_3}\rho_{q_1p_1}\rho_{q_2p_2}\rho_{q_3p_3}\prod_{h=1}^{3}\delta_{q_h,p_{\tau(h)}}. \tag*{(5)}

Here σ\sigma tests equalities among the measured outcomes, whereas τ\tau connects indices of the three density-matrix factors. These two roles define the data and invariant coordinates below.

Equality patterns

The factor ∏h=13δih,iσ(h)=1\prod_{h=1}^{3}\delta_{i_h,i_{\sigma(h)}}=1 iff all outcomes linked by σ\sigma are equal, e.g. σ=(12)\sigma=(12) requires i1=i2i_1=i_2, whereas σ=(123)\sigma=(123) requires i1=i2=i3i_1=i_2=i_3. Consequently, the averaged entry depends only on the equality pattern of the outcome triple. We use the ordered set

Π=(D,T12,T23,T13,E),(6)\Pi=(D,T_{12},T_{23},T_{13},E), \tag*{(6)}

where DD denotes three distinct outcomes, TabT_{ab} denotes equality only between positions a,ba,b, and EE denotes three equal outcomes.

For example, (1,1,2)(1,1,2) and (2,2,3)(2,2,3) both have pattern T12T_{12} and therefore the same averaged entry yT12y_{T_{12}}. The vector ypatt∈R5\mathbf{y}_{\mathrm{patt}}\in\mathbb{R}^{5} collects one such value per pattern; its entries are representative correlators, rather than total pattern-event probabilities.

Evaluating the outcome equality tests on each pattern defines the matrix SπσS_{\pi\sigma}. With permutation columns ordered as (1,(12),(23),(13),(123),(132))(1,(12),(23),(13),(123),(132)), it is

11(12)(12)(23)(23)(13)(13)(123)(123)(132)(132)
DD110000000000
T12T_{12}111100000000
T23T_{23}110011000000
T13T_{13}110000110000
EE111111111111

Table 3.4.

Thus SS encodes which permutation constraints each equality pattern satisfies.

Permutation contractions

The index connections specified by τ\tau can be expressed through the unitary permutation operator Pτ:(Cd)⊗3→(Cd)⊗3P_{\tau}:(\mathbb{C}^{d})^{\otimes3}\to(\mathbb{C}^{d})^{\otimes3}, defined by

Pτ∣i1,i2,i3⟩=∣iτ−1(1),iτ−1(2),iτ−1(3)⟩.(7)P_{\tau}|i_1,i_2,i_3\rangle=|i_{\tau^{-1}(1)},i_{\tau^{-1}(2)},i_{\tau^{-1}(3)}\rangle. \tag*{(7)}

This convention gives PσPτ=PστP_{\sigma}P_{\tau}=P_{\sigma\tau}. The scalar obtained by closing the index connections is

x(τ)=Tr⁡[Pτρ⊗3],xfull=(x(τ))τ∈S3∈C6.(8)x(\tau)=\operatorname{Tr}[P_{\tau}\rho^{\otimes3}],\qquad\mathbf{x}_{\mathrm{full}}=(x(\tau))_{\tau\in S_3}\in\mathbb{C}^{6}. \tag*{(8)}

These contractions are invariant under unitary changes of basis because PτP_{\tau} commutes with U⊗3U^{\otimes3} [24, 25, 26, 27, 28, 29].

For one party, the cycle structure of τ\tau determines which state factors are joined. The identity, a transposition, and a three-cycle give, respectively, q0=(Tr⁡ρ)3q_0=(\operatorname{Tr}\rho)^3, q1=(Tr⁡ρ)Tr⁡ρ2q_1=(\operatorname{Tr}\rho)\operatorname{Tr}\rho^2, q2=Tr⁡ρ3q_2=\operatorname{Tr}\rho^3. Accordingly,

xfull=(q0,q1,q1,q1,q2,q2)T.(9)\mathbf{x}_{\mathrm{full}}=(q_0,q_1,q_1,q_1,q_2,q_2)^{\mathsf{T}}. \tag*{(9)}

The six labelled coordinates therefore contain only three distinct moment values.

Local cycle merging

Let Wστ(d)=Wg⁡(στ−1,d)W_{\sigma\tau}(d)=\operatorname{Wg}(\sigma\tau^{-1},d). It assigns to a σ\sigma, permitted by the equality pattern by SS, a linear combination of unitary invariants encoded by τ\tau. It depends only on whether its permutation argument is the identity [1][1], a transposition [12][12], or a three-cycle [123][123]:

Wg⁡([1],d)=d2−2d(d2−1)(d2−4)=:e,(10)\operatorname{Wg}([1],d)=\frac{d^2-2}{d(d^2-1)(d^2-4)}=:e, \tag*{(10)}
Wg⁡([12],d)=−dd(d2−1)(d2−4)=:t,(11)\operatorname{Wg}([1^{2}],d)=\frac{-d}{d(d^{2}-1)(d^{2}-4)}=:t, \tag*{(11)}
Wg⁡([13],d)=2d(d2−1)(d2−4)=:c.(12)\operatorname{Wg}([1^{3}],d)=\frac{2}{d(d^{2}-1)(d^{2}-4)}=:c. \tag*{(12)}

giving (s=e+3t+2c)(s=e+3t+2c) [30, 31]

W(d)=(etttcctecctttcectttccettctttecctttce)∈R6×6.(13)W(d)= \begin{pmatrix} e&t&t&t&c&c\\ t&e&c&c&t&t\\ t&c&e&c&t&t\\ t&c&c&e&t&t\\ c&t&t&t&e&c\\ c&t&t&t&c&e \end{pmatrix}\in\mathbb{R}^{6\times6}. \tag*{(13)}

Therefore Eq. (5) can be reformulated as

ypatt=SW(d)xfull(14)\mathbf{y}_{\mathrm{patt}}=SW(d)\mathbf{x}_{\mathrm{full}} \tag*{(14)}

The two last columns of SW(d)SW(d) associated with (123)(123) and (132)(132) coincide, hence they can be rewritten as one column multiplying the sum of last two entries of xfull\mathbf{x}_{\mathrm{full}}. Define the local cycle-merging map C:C6→C5C:\mathbb{C}^{6}\to\mathbb{C}^{5} by

C(a1,a12,a23,a13,a123,a132)T=(a1,a12,a23,a13,a123+a132)T.(15)C(a_{1},a_{12},a_{23},a_{13},a_{123},a_{132})^{T} =(a_{1},a_{12},a_{23},a_{13},a_{123}+a_{132})^{T}. \tag*{(15)}

It retains the sum of the cycle coordinates and annihilates their difference. For one party, those coordinates already coincide, so this merging removes no information. Its information-losing consequence for distinct invariants will appear in the two-party case. Defining M(d)M(d) as SW(d)SW(d) with the last column removed and z=Cxfull\mathbf{z}=C\mathbf{x}_{\mathrm{full}} one can rewrite the equation (14) as

ypatt=M(d)z,(16)\mathbf{y}_{\mathrm{patt}}=M(d)\mathbf{z}, \tag*{(16)}

where

M(d)=(etttce+te+tc+tc+tc+te+tc+te+tc+tc+te+tc+tc+te+tc+tsssss).(17)M(d)= \begin{pmatrix} e&t&t&t&c\\ e+t&e+t&c+t&c+t&c+t\\ e+t&c+t&e+t&c+t&c+t\\ e+t&c+t&c+t&e+t&c+t\\ s&s&s&s&s \end{pmatrix}. \tag*{(17)}

We label the retained coordinates in z\mathbf{z} by Π\Pi, matching the identity to DD, each transposition (ab)(ab) to TabT_{ab}, and the cycle sum to EE. These labels match coordinates with different meanings: EE denotes a cycle sum in z\mathbf{z}, whereas three equal outcomes in ypatt\mathbf{y}_{\mathrm{patt}}.

The map M(d)M(d) is the local building block used at every party in the multipartite construction. The pole at d=2d=2 in Eqs. (10)–(12) occurs because one of the five outcome patterns is absent for qubits [32].

Orbit reduction

Relabelling the three identical copies does not change a contraction. Indeed, PηPτPη−1=Pητη−1P_{\eta}P_{\tau}P_{\eta}^{-1}=P_{\eta\tau\eta^{-1}} and ⊗3Pη−1=ρη⊗3\mathop{\otimes}\limits^{3}P_{\eta}^{-1}=\rho_{\eta}^{\otimes3}, so

x(ητη−1)=x(τ),η∈S3.(18)x(\eta\tau\eta^{-1})=x(\tau),\qquad\eta\in S_{3}. \tag*{(18)}

Thus contractions are equal on each conjugation orbit. The same relabelling leaves the probability product unchanged and acts on equality patterns as

η⋅D=D,η⋅E=E,η⋅Tab=Tη(a)η(b).(19)\eta\cdot D=D,\qquad\eta\cdot E=E,\qquad\eta\cdot T_{ab}=T_{\eta(a)\eta(b)}. \tag*{(19)}

There are therefore three pattern orbits: DD, the three TT labels, and EE. Their data coordinates are yD,yT,yEy_{D},y_{T},y_{E}, where yTy_{T} is the common value of the three labelled TT entries.

As contractions on an orbit are equal, we change the equality-pattern-labelled coordinates to orbit-labelled coordinates, summing in M(d)M(d) the columns related to permutations of the same orbit. The column-merging is realised by the right multiplication with the matrix R1R_{1}:

R1=(100010010010001),M(d)R1=(e3tce+tsc+te+tsc+te+tsc+ts3ss).(20)R_{1}= \begin{pmatrix} 1&0&0\\ 0&1&0\\ 0&1&0\\ 0&1&0\\ 0&0&1 \end{pmatrix}, \qquad M(d)R_{1}= \begin{pmatrix} e&3t&c\\ e+t&s&c+t\\ e+t&s&c+t\\ e+t&s&c+t\\ s&3s&s \end{pmatrix}. \tag*{(20)}

The same applies to renumbering of outcomes in ypatt\mathbf{y}_{\mathrm{patt}} — these belonging to the same permutation orbit are equal and hence are realised by identical rows in M(d)R1M(d)R_{1}. We remove the duplicates by averaging over orbits:

L1=(100000131313000001),L1R1=I3.(21)L_{1}= \begin{pmatrix} 1&0&0&0&0\\ 0&\frac{1}{3}&\frac{1}{3}&\frac{1}{3}&0\\ 0&0&0&0&1 \end{pmatrix}, \qquad L_{1}R_{1}=I_{3}. \tag*{(21)}

Substituting z=R1(q0,q1,2q2)T\mathbf{z}=R_{1}(q_{0},q_{1},2q_{2})^{T} and applying from left L1L_{1} to Eq. (17) gives the in the orbit space

\begin{pmatrix} y_{D}\\ y_{T}\\ y_{E} \end{pmatrix} =H(d) \begin{pmatrix} q_{0}\\ q_{1}\\ 2q_{2} \end{pmatrix}, \tag*{(22)}\end{pmatrix}
H(d)=(e3tce+tsc+ts3ss)=L1M(d)R1.(23)H(d)= \begin{pmatrix} e&3t&c\\ e+t&s&c+t\\ s&3s&s \end{pmatrix} =L_{1}M(d)R_{1}. \tag*{(23)}

The factor 33 collects the three identical transposition coordinates, whereas 2q22q_{2} retains the cycle-sum convention. To reconstruct the invariant coordinates q0,q1,2q2q_{0},q_{1},2q_{2}, one only needs to invert the 3×33\times3 one-party matrix H(d)H(d):

H(d)−1=(d(d−1)(d−2)3d(d−1)d−d(d−1)(d−2)d(d−1)(d−2)d22d(d−1)(d−2)−3d(d−1)2d3+d).(24)H(d)^{-1}= \begin{pmatrix} d(d-1)(d-2)&3d(d-1)&d\\ -d(d-1)(d-2)&d(d-1)(d-2)&d^{2}\\ 2d(d-1)(d-2)&-3d(d-1)^{2}&d^{3}+d \end{pmatrix}. \tag*{(24)}

The striking consequence, which we prove at the end of the paper, is that the many-party reconstruction problem collapses to this same 3×33\times3 inverse and scalar blocks. Thus the apparent growth of the full reconstruction matrix, studied in the proper basis, does not introduce any new nontrivial matrix inversion.

For a normalized state, q0=1q_{0}=1; retaining this coordinate keeps the cubic map homogeneous and makes its normalization explicit. The matrix H(d)H(d) will also form the blocks of the multipartite reconstruction.

Finally, representative correlators are converted to event probabilities by counting outcome realizations. The local multiplicities are

mD(d)=d(d−1)(d−2),mT(d)=d(d−1),mE(d)=d.(25)m_D(d)=d(d-1)(d-2),\quad m_T(d)=d(d-1),\quad m_E(d)=d. \tag*{(25)}

There are three labelled TT patterns, so normalization reads

q0=mD(d)yD+3mT(d)yT+mE(d)yE=1.(26)q_0=m_D(d)y_D+3m_T(d)y_T+m_E(d)y_E=1. \tag*{(26)}

These multiplicities convert measured pattern frequencies into the representative coordinates required by the inverse map. Appendix B extends this conversion to multipartite orbit events.

Two parties

Equality patterns and local cycle merging

For two parties, independent local Haar integration gives the tensor product of the one-party maps. The new ingredient is the relative arrangement of the copy labels at AA and BB.

Let ρ\rho act on CdA⊗CdB\mathbb{C}^{d_A}\otimes\mathbb{C}^{d_B}. Grouping the three copies of each party together, define

x(τA,τB)=Tr⁡[(PτAA⊗PτBB)ρ⊗3],(τA,τB)∈S32.(27)x(\tau_A,\tau_B)=\operatorname{Tr}\left[\left(P_{\tau_A}^{A}\otimes P_{\tau_B}^{B}\right)\rho^{\otimes3}\right],\quad(\tau_A,\tau_B)\in S_3^2. \tag*{(27)}

The vector xfull(2)∈C36x_{\mathrm{full}}^{(2)}\in\mathbb{C}^{36} collects these labelled contractions. Each party also supplies one equality-pattern label, giving ypatt(2)∈R25y_{\mathrm{patt}}^{(2)}\in\mathbb{R}^{25}. Tensoring the one-party relation SW(d)=M(d)CSW(d)=M(d)C yields

ypatt(2)=[SW(dA)⊗SW(dB)]xfull(2)=[M(dA)⊗M(dB)]z(2),z(2)=(C⊗C)xfull(2).(28)\begin{aligned} y_{\mathrm{patt}}^{(2)}=[SW(d_A)\otimes SW(d_B)]x_{\mathrm{full}}^{(2)} \\ &=[M(d_A)\otimes M(d_B)]z^{(2)},\qquad z^{(2)}=(C\otimes C)x_{\mathrm{full}}^{(2)}. \tag*{(28)} \end{aligned}

Thus the same local cycle-merging map acts independently on the contraction label at each party.

A coordinate without an EE label is unchanged by C⊗2C^{\otimes2}, for example, zT12T23(2)=x((12),(23))z_{T_{12}T_{23}}^{(2)}=x((12),(23)). One EE label sums two cycle orientations,

zET12(2)=x((123),(12))+x((132),(12)),(29)z_{ET_{12}}^{(2)}=x((123),(12))+x((132),(12)), \tag*{(29)}

whereas two EE labels sum all four orientation pairs:

zEE(2)=x((123),(123))+x((123),(132))+x((132),(123))+x((132),(132)).(30)\begin{aligned} z_{EE}^{(2)}={}&x((123),(123))+x((123),(132))\\ &+x((132),(123))+x((132),(132)). \tag*{(30)} \end{aligned}

The measurement map therefore retains this sum before any orbit reduction is performed. The orbit classification below identifies the physical invariants entering it.

Contraction coordinateOrbit size
x0=x(1,1)=(Tr⁡ρ)3x_0=x(1,1)=(\operatorname{Tr}\rho)^31
x1=x(1,τ)=(Tr⁡ρ)Tr⁡ρB2x_1=x(1,\tau)=(\operatorname{Tr}\rho)\operatorname{Tr}\rho_B^23
x2=x(1,κ)=Tr⁡ρB3x_2=x(1,\kappa)=\operatorname{Tr}\rho_B^32
x3=x(τ,1)=(Tr⁡ρ)Tr⁡ρA2x_3=x(\tau,1)=(\operatorname{Tr}\rho)\operatorname{Tr}\rho_A^23
x4=x(τ,τ′)=Tr⁡[ρ(ρA⊗ρB)]x_4=x(\tau,\tau')=\operatorname{Tr}[\rho(\rho_A\otimes\rho_B)]6
x5=x(τ,τ)=(Tr⁡ρ)Tr⁡ρ2x_5=x(\tau,\tau)=(\operatorname{Tr}\rho)\operatorname{Tr}\rho^23
x6=x(τ,κ)=Tr⁡[(Tr⁡Aρ2)ρB]x_6=x(\tau,\kappa)=\operatorname{Tr}[(\operatorname{Tr}_A\rho^2)\rho_B]6
x7=x(κ,1)=Tr⁡ρA3x_7=x(\kappa,1)=\operatorname{Tr}\rho_A^32
x8=x(κ,τ)=Tr⁡[(Tr⁡Bρ2)ρA]x_8=x(\kappa,\tau)=\operatorname{Tr}[(\operatorname{Tr}_B\rho^2)\rho_A]6
x9=x(κ,κ)=Tr⁡ρ3x_9=x(\kappa,\kappa)=\operatorname{Tr}\rho^32
x10=x(κ,κ−1)=Tr⁡[(ρΓB)3]x_{10}=x(\kappa,\kappa^{-1})=\operatorname{Tr}[(\rho^{\Gamma_B})^3]2

caption: TABLE I. The eleven diagonal-conjugation orbits of bipartite contractions [16, 17]. Here τ=(12)\tau=(12), τ′=(23)\tau'=(23), κ=(123)\kappa=(123), ρA=Tr⁡Bρ\rho_A=\operatorname{Tr}_B\rho, and ρB=Tr⁡Aρ\rho_B=\operatorname{Tr}_A\rho. The operation ΓB\Gamma_B is partial transposition on BB.

Orbit reduction

The three copy positions label complete bipartite preparations, so a relabelling must act simultaneously at both parties. The resulting diagonal conjugation action is

η⋅(τA,τB)=(ητAη−1,ητBη−1),η∈S3.(31)\eta\cdot(\tau_A,\tau_B)=(\eta\tau_A\eta^{-1},\eta\tau_B\eta^{-1}),\qquad\eta\in S_3. \tag*{(31)}

Invariance due to identical copies ρ⊗3\rho^{\otimes3}, as in the one-party case, gives x(η⋅(τA,τB))=x(τA,τB)x(\eta\cdot(\tau_A,\tau_B))=x(\tau_A,\tau_B). An orbit consists of the distinct pairs generated by these six relabellings. There are eleven contraction orbits. Table I lists their representative values containing the physical information—the local unitary invariants of ρ\rho.

In particular, equal cycle orientations give x9x_9, while opposite orientations give x10x_{10}. Each class occurs twice in Eq. (30), so

zEE(2)=2(x9+x10).(32)z_{EE}^{(2)}=2(x_9+x_{10}). \tag*{(32)}

Unlike the two cycle coordinates for one party, x9x_9 and x10x_{10} need not agree. The local cycle merging therefore retains their sum but cannot separate their difference.

On the data side, simultaneous relabelling acts on equality-pattern pairs through

η⋅(πA,πB)=(η⋅πA,η⋅πB),(πA,πB)∈Π2.(33)\eta\cdot(\pi_A,\pi_B)=(\eta\cdot\pi_A,\eta\cdot\pi_B),\qquad(\pi_A,\pi_B)\in\Pi^2. \tag*{(33)}

The eight local-type pairs other than TTTT each form one orbit. The TTTT pairs split into two orbits according to whether the equalities involve the same pair of shots, T=TT=T, or different pairs, T≠TT\ne T. Table II gives all ten orbits. In the table order, the representative data values and matching merged contractions are

yorb(2)=(yDD,yDT,yDE,yTD,yT=T,yT≠T,yTE,yED,yET,yEE)T,y_{\mathrm{orb}}^{(2)}=(y_{DD},y_{DT},y_{DE},y_{TD},y_{T=T},y_{T\ne T},y_{TE},y_{ED},y_{ET},y_{EE})^{T},
zorb(2)=(x0,x1,2x2,x3,x5,x4,2x6,2x7,2x8,2(x9+x10))T.(34)z_{\mathrm{orb}}^{(2)}=(x_0,x_1,2x_2,x_3,x_5,x_4,2x_6,2x_7,2x_8,2(x_9+x_{10}))^{T}. \tag*{(34)}

Table 1.

The factors of two follow from C⊗2C^{\otimes2}: each single-EE coordinate sums two equal contractions, while the EEEE coordinate contains two copies of each cycle orbit. The normalization convention for the data coordinates yorb(2)\mathbf{y}^{(2)}_{\mathrm{orb}} is the same as in the one-party case of Eq. (26).

To express the map that relates data and invariants in these ten coordinates for both yorb(2)\mathbf{y}^{(2)}_{\mathrm{orb}} and zorb(2)\mathbf{z}^{(2)}_{\mathrm{orb}}, let O1,…,O10\mathcal{O}_1,\ldots,\mathcal{O}_{10} be the pattern orbits in Table II. Generalize R1R_1 and L1L_1, the embedding and extraction maps of Eqs. (20) and (21), by

OrbitEquality pattern of Π2\Pi^2Size
DDDD{(D,D)}\{(D,D)\}1
DTDT{(D,T12),(D,T23),(D,T13)}\{(D,T_{12}),(D,T_{23}),(D,T_{13})\}3
DEDE{(D,E)}\{(D,E)\}1
TDTD{(T12,D),(T23,D),(T13,D)}\{(T_{12},D),(T_{23},D),(T_{13},D)\}3
T=TT=T{(Tab,Tab):ab∈{12,23,13}}\{(T_{ab},T_{ab}):ab\in\{12,23,13\}\}3
T≠TT\ne T{(Tab,Tcd):ab≠cd}\{(T_{ab},T_{cd}):ab\ne cd\}6
TETE{(T12,E),(T23,E),(T13,E)}\{(T_{12},E),(T_{23},E),(T_{13},E)\}3
EDED{(E,D)}\{(E,D)\}1
ETET{(E,T12),(E,T23),(E,T13)}\{(E,T_{12}),(E,T_{23}),(E,T_{13})\}3
EEEE{(E,E)}\{(E,E)\}1

Table II. The ten orbits of labelled equality-pattern pairs. The pair labels ab,cdab, cd range over {12,23,13}\{12,23,13\}. Common relabelling preserves whether the parties select the same or different pairs of shots.

(R2)qμ=1{q∈Oμ},(L2)μq=1{q∈Oμ}∣Oμ∣.(35)(R_2)_{q\mu}=\mathbf{1}_{\{q\in\mathcal{O}_\mu\}},\qquad(L_2)_{\mu q}=\frac{\mathbf{1}_{\{q\in\mathcal{O}_\mu\}}}{|\mathcal{O}_\mu|}. \tag*{(35)}

where q∈Π2q\in\Pi^2 and μ=1,…,10\mu=1,\ldots,10. Here R2:R10→R25R_2:\mathbb{R}^{10}\to\mathbb{R}^{25} embeds the ten orbit coordinates into the 25 labelled coordinates by assigning one if the pair qq belongs to the orbit Oμ\mathcal{O}_\mu. Conversely, L2:R25→R10L_2:\mathbb{R}^{25}\to\mathbb{R}^{10} extracts one coordinate per orbit by averaging the labelled entries in that orbit. Hence L2R2=I10L_2R_2=I_{10}.

Both labelled vectors ypatt(2)\mathbf{y}^{(2)}_{\mathrm{patt}} and z(2)\mathbf{z}^{(2)} are constant on the orbits, so

zpatt(2)=R2zorb(2),yorb(2)=L2ypatt(2),\mathbf{z}^{(2)}_{\mathrm{patt}}=R_2\mathbf{z}^{(2)}_{\mathrm{orb}},\qquad\mathbf{y}^{(2)}_{\mathrm{orb}}=L_2\mathbf{y}^{(2)}_{\mathrm{patt}},
zorb(2)=L2zpatt(2)=R2yorb(2).\mathbf{z}^{(2)}_{\mathrm{orb}}=L_2\mathbf{z}^{(2)}_{\mathrm{patt}}=\mathbf{R}_2\mathbf{y}^{(2)}_{\mathrm{orb}}.

Substituting the first relation into Eq. (28) and applying L2L_2 gives

yorb(2)=Morb(dA,dB)zorb(2),\mathbf{y}^{(2)}_{\mathrm{orb}}=M_{\mathrm{orb}}(d_A,d_B)\mathbf{z}^{(2)}_{\mathrm{orb}},
Morb(dA,dB)=L2[M(dA)⊗M(dB)]R2.(36)M_{\mathrm{orb}}(d_A,d_B)=L_2[M(d_A)\otimes M(d_B)]R_2. \tag*{(36)}

This orbit reduction removes repeated coordinates without further loss of the information retained by C⊗CC\otimes C. Figure 1 summarizes these steps in the multipartite notation introduced in the next section and the solution for n=2n=2 that expresses zorb(2)\mathbf{z}^{(2)}_{\mathrm{orb}} from yorb(2)\mathbf{y}^{(2)}_{\mathrm{orb}} will be obtained as an instance of the general theorem for arbitrary nn in Sec. VII A.

From full labelled coordinates to the reconstruction map

Figure 1. From full labelled coordinates to the reconstruction map MorbM_{\mathrm{orb}}. Horizontal arrows relate state invariants to measured averages through Haar integration. Equality-pattern reduction removes repeated outcome entries, whereas local cycle merging by C⊗nC^{\otimes n} adds the two cycle orientations (e.g. (32)) and removes access to their differences. Simultaneous copy relabelling then gives one coordinate per orbit in each column, without further information loss. The resulting map MorbM_{\mathrm{orb}} is reconstructed blockwise by the theorem.

Arbitrary parties

Equality patterns and local cycle merging

The two-party construction extends by applying the same local maps SW(di)SW(d_i) at every party. A tuple τ=(τ1,…,τn)∈S3n\tau=(\tau_1,\ldots,\tau_n)\in S_3^n specifies the contraction

xρ(τ)=Tr⁡[(⨂k=1nPτk(k))ρ⊗3],xfull(n)∈C6n.(37)x_\rho(\tau)=\operatorname{Tr}\left[\left(\bigotimes_{k=1}^{n}P_{\tau_k}^{(k)}\right)\rho^{\otimes3}\right],\qquad\mathbf{x}_{\mathrm{full}}^{(n)}\in\mathbb{C}^{6^n}. \tag*{(37)}

As before in Eq. (27), tensor factors are grouped by party. For arbitrary party number, the full contractions need not be real. Since ρ⊗3\rho^{\otimes3} is Hermitian and Pτ†=Pτ−1P_\tau^\dagger=P_{\tau^{-1}}, complex conjugation gives

xρ(τ1,…,τn)‾=xρ(τ1−1,…,τn−1).(38)\overline{x_\rho(\tau_1,\ldots,\tau_n)}=x_\rho(\tau_1^{-1},\ldots,\tau_n^{-1}). \tag*{(38)}

For one or two parties, the inverse tuple belongs to the same diagonal-conjugation orbit, so each contraction is real. For three or more parties, this need not hold.

The merged vector z(n)=C⊗nxfull(n)\mathbf{z}^{(n)}=C^{\otimes n}\mathbf{x}_{\mathrm{full}}^{(n)} nevertheless remains real. To see this, write the operators associated with its local labels as

OD=I,OTab=P(ab),OE=P(123)+P(132).(39)O_D=I,\qquad O_{T_{ab}}=P_{(ab)},\qquad O_E=P_{(123)}+P_{(132)}. \tag*{(39)}

Each component is then

zπ(n)=Tr⁡[(⨂k=1nOπk(k))ρ⊗3],π∈Πn.(40)z_\pi^{(n)}=\operatorname{Tr}\left[\left(\bigotimes_{k=1}^{n}O_{\pi_k}^{(k)}\right)\rho^{\otimes3}\right],\qquad\pi\in\Pi^n. \tag*{(40)}

Both operators inside the trace are Hermitian, so cyclicity of the trace makes this expression real.

Denote the one-party coordinate space by A=span⁡R{eπ:π∈Π}≅R5A = \operatorname{span}_{\mathbb{R}}\{e_{\pi} : \pi\in\Pi\} \cong\mathbb{R}^{5}, with its standard orthonormal basis. Tensoring the local relation SW(d)=M(d)CSW(d) = M(d)C gives

ypatt(n)=Mnz(n),Mn=⨂k=1nM(dk):A⊗n→A⊗n.(41)y_{\mathrm{patt}}^{(n)} = M_n z^{(n)}, \qquad M_n = \bigotimes_{k=1}^{n} M(d_k) : A^{\otimes n} \to A^{\otimes n}. \tag*{(41)}

Although the map MnM_n factorizes, the input and data vectors, respectively z(n)z^{(n)} and ypatt(n)y_{\mathrm{patt}}^{(n)}, generally do not: both retain correlations among the parties.

Orbit reduction

As for two parties, one common η∈S3\eta\in S_3 relabels the three preparations. The local pattern action in Eq. (19) defines the matrices

γ(η)eπ=eη⋅π,Γn(η)=γ(η)⊗n,(42)\gamma(\eta)e_{\pi} = e_{\eta\cdot\pi}, \qquad\Gamma_n(\eta) = \gamma(\eta)^{\otimes n}, \tag*{(42)}

acting on AA and A⊗nA^{\otimes n}, respectively. The retained operators in Eq. (39) obey PηOπPη−1=Oη⋅πP_{\eta}O_{\pi}P_{\eta}^{-1} = O_{\eta\cdot\pi}, defining the same action both for the pattern equality of experimental data and the invariant coordinates. Indeed, the action fixes II, sends P(ab)P_{(ab)} to P(η(a)η(b))P_{(\eta(a)\eta(b))}, and permutes the two three-cycles while preserving their sum.

Relabelling the three probability factors leaves their product unchanged. Likewise, relabelling identical global copies leaves ρ⊗3\rho^{\otimes3} unchanged. Thus

Γn(η)ypatt(n)=ypatt(n),Γn(η)z(n)=z(n).(43)\Gamma_n(\eta)y_{\mathrm{patt}}^{(n)} = y_{\mathrm{patt}}^{(n)}, \qquad\Gamma_n(\eta)z^{(n)} = z^{(n)}. \tag*{(43)}

Both labelled vectors therefore lie in the invariant subspace (A⊗n)S3(A^{\otimes n})^{S_3}: their entries are constant on each simultaneous-relabelling orbit in Πn\Pi^n. One coordinate per orbit suffices. Denote the number of orbits by NnN_n; its count is derived in Appendix A.

Extend the two-party embedding and extraction maps by setting, for each orbit o⊂Πno \subset\Pi^n,

(Rn)π,o=1{π∈o},(Ln)o,π=1{π∈o}∣o∣.(44)(R_n)_{\pi,o} = \mathbf{1}_{\{\pi\in o\}}, \qquad(L_n)_{o,\pi} = \frac{\mathbf{1}_{\{\pi\in o\}}}{|o|}. \tag*{(44)}

Here Rn:RNn→A⊗nR_n : \mathbb{R}^{N_n} \to A^{\otimes n} embeds the NnN_n orbit coordinates into the full labelled space by assigning one to all labelled pattern tuples in the same diagonal orbit. Conversely, Ln:A⊗n→RNnL_n : A^{\otimes n} \to\mathbb{R}^{N_n} extracts one coordinate per orbit by averaging over its labelled members. Consequently, LnRn=INnL_nR_n = I_{N_n}. Define

yorb(n)=Lnypatt(n),zorb(n)=Lnz(n).(45)y_{\mathrm{orb}}^{(n)} = L_n y_{\mathrm{patt}}^{(n)}, \qquad z_{\mathrm{orb}}^{(n)} = L_n z^{(n)}. \tag*{(45)}

Since all entries within an orbit agree, repeating each extracted value recovers the original labelled vectors:

ypatt(n)=Rnyorb(n),z(n)=Rnzorb(n).(46)y_{\mathrm{patt}}^{(n)} = R_n y_{\mathrm{orb}}^{(n)}, \qquad z^{(n)} = R_n z_{\mathrm{orb}}^{(n)}. \tag*{(46)}

Thus these coordinate reductions are reversible on the invariant subspace.

Substituting z(n)=Rnzorb(n)z^{(n)} = R_n z_{\mathrm{orb}}^{(n)} into Eq. (41) and applying LnL_n gives

yorb(n)=LnMnRnzorb(n)=Morb(d)zorb(n),(47)y_{\mathrm{orb}}^{(n)} = L_n M_n R_n z_{\mathrm{orb}}^{(n)} = M_{\mathrm{orb}}(d)z_{\mathrm{orb}}^{(n)}, \tag*{(47)}

where d=(d1,…,dn)d = (d_1,\ldots,d_n). The dimension sum in Eq. (A5) gives Nn∼5n/6N_n \sim5^n/6. Storing a dense Nn×NnN_n \times N_n matrix requires O(Nn2)O(N_n^2) memory, and conventional elimination costs O(Nn3)O(N_n^3) arithmetic operations; both grow exponentially with party number. A numerical inverse would recover the coordinates without revealing which invariant combinations are actually independent. We therefore decompose the map into symmetry-resolved sectors, where every nontrivial matrix inversion reduces to the same 3×33 \times3 one-party blocks, together with scalar weights.

Reconstruction sectors

The orbit reduction identifies the independent coordinates of the data and invariant vectors. In this section, we introduce a second coordinate system that makes the reconstruction map easier to analyze and invert.

Local invariant and zero-sum subspaces

Copy relabelling fixes DD and EE, but permutes the three TT labels. We separate their common component from two orthogonal zero-sum components by choosing

νT=eT12+eT23+eT13,ξ1=eT12−eT23,ξ2=eT12+eT23−2eT13.(48)\nu_T = e_{T_{12}} + e_{T_{23}} + e_{T_{13}}, \qquad\xi_1 = e_{T_{12}} - e_{T_{23}}, \qquad\xi_2 = e_{T_{12}} + e_{T_{23}} - 2e_{T_{13}}. \tag*{(48)}

In the ordered basis (eD,eT12,eT23,eT13,eE)(e_D,e_{T_{12}},e_{T_{23}},e_{T_{13}},e_E), the change-of-basis matrix is

F=(eD,νT,eE,ξ1,ξ2)=(1000001011010−110100−200100).(49)F = (e_D,\nu_T,e_E,\xi_1,\xi_2) = \begin{pmatrix} 1 & 0 & 0 & 0 & 0 \\ 0 & 1 & 0 & 1 & 1 \\ 0 & 1 & 0 & -1 & 1 \\ 0 & 1 & 0 & 0 & -2 \\ 0 & 0 & 1 & 0 & 0 \end{pmatrix}. \tag*{(49)}

Its columns form an orthogonal basis, without normalizing the TT vectors: FTF=diag⁡(1,3,1,2,6)F^{T}F = \operatorname{diag}(1,3,1,2,6). For a=(aD,a12,a23,a13,aE)Ta = (a_D,a_{12},a_{23},a_{13},a_E)^T, the new coordinates are

F−1a=(aD(a12+a23+a13)/3aE(a12−a23)/2(a12+a23−2a13)/6).F^{-1}a = \begin{pmatrix} a_D \\ (a_{12}+a_{23}+a_{13})/3 \\ a_E \\ (a_{12}-a_{23})/2 \\ (a_{12}+a_{23}-2a_{13})/6 \end{pmatrix}.

The first three columns of FF are R1R_1, and the first three rows of F−1F^{-1} are L1L_1.

In this basis, relabelling leaves the first three coordinates unchanged. The remaining two transform among themselves:

F−1γ(η)F=I3⊕γW(η).(50)F^{-1}\gamma(\eta)F = I_3 \oplus\gamma_W(\eta). \tag*{(50)}

where, for two generators of S3S_3,

γW((12))=12(131−1),γW((23))=12(1−3−1−1).\gamma_W((12)) = \frac{1}{2}\begin{pmatrix} 1 & 3 \\ 1 & -1 \end{pmatrix}, \qquad\gamma_W((23)) = \frac{1}{2}\begin{pmatrix} 1 & -3 \\ -1 & -1 \end{pmatrix}.

If a vector [α,β]T∈W[\alpha,\beta]^{T}\in W were unchanged by both generators, the two invariance equations would give α=β=0\alpha=\beta=0. The vectors invariant under every relabelling are therefore precisely those whose last two coordinates vanish: AS3=V=span⁡{eD,vT,eE}A^{S_3}=V=\operatorname{span}\{e_D,v_T,e_E\}, W=span⁡{ξ1,ξ2}W=\operatorname{span}\{\xi_1,\xi_2\}, so that A=V⊕WA=V\oplus W. Relabelling preserves WW as a subspace, although WW contains no nonzero vector invariant under every relabelling.

As [M(d),γ(η)]=0[M(d),\gamma(\eta)]=0, then M(d)M(d) does not mix VV and WW, hence it is block-diagonal with respect to this basis. Its 2×22\times2 block commuting with both displayed matrices γW((12))\gamma_W((12)) and γW((23))\gamma_W((23)) must be scalar; this is the instance of Schur’s lemma used here. The change of basis gives

F−1M(d)F=H(d)⊕λ(d)I2,(51)F^{-1}M(d)F=H(d)\oplus\lambda(d)I_2, \tag*{(51)}

with λ(d)=e(d)−c(d)=1d(d2−1)\lambda(d)=e(d)-c(d)=\frac{1}{d(d^2-1)}. The first block is the one-party map H(d)=L1M(d)R1H(d)=L_1M(d)R_1. The second says that M(d)M(d) multiplies every zero-sum vector by the same scalar. Although WW alone has no invariant vector, tensor products of WW factors can have invariant combinations, as the next subsection shows.

Invariant tensor sectors

The orbit map still couples many coordinates. We now use the local decomposition to identify subspaces on which the measurement map acts independently; their invariant parts will become the diagonal blocks in the theorem.

The nn-party orbit space A⊗nA^{\otimes n} splits into a direct sum of tensor sectors:

A⊗n=(V⊕W)⊗n=⨁X∈{V,W}nHX,(52)A^{\otimes n}=(V\oplus W)^{\otimes n}=\bigoplus_{X\in\{V,W\}^n}\mathcal{H}_X, \tag*{(52)}

where HX=⨂i=1nXi\mathcal{H}_X=\bigotimes_{i=1}^{n}X_i. To select the component in each sector, use the local orthogonal projectors. In the ordered basis (D,T12,T23,T13,E)(D,T_{12},T_{23},T_{13},E), they are

PV=eDeDT+13vTvTT+eEeET=1⊕13(111111111)⊕1,P_V=e_De_D^T+\frac{1}{3}v_Tv_T^T+e_Ee_E^T =1\oplus\frac{1}{3}\begin{pmatrix}1&1&1\\1&1&1\\1&1&1\end{pmatrix}\oplus1,
PW=I5−PV=01⊕13(2−1−1−12−1−1−12)⊕01.P_W=I_5-P_V =0_1\oplus\frac{1}{3}\begin{pmatrix}2&-1&-1\\-1&2&-1\\-1&-1&2\end{pmatrix}\oplus0_1.

The factor 1/31/3 accounts for ∥vT∥2=3\lVert v_T\rVert^2=3. Thus PV=R1L1P_V=R_1L_1 keeps the D,ED,E entries and replaces the three TT entries by their average; PWP_W keeps their zero-sum deviations. For a choice XX, the tensor projector PX=⨂i=1nPXiP_X=\bigotimes_{i=1}^{n}P_{X_i} projects onto HX\mathcal{H}_X. The local relations PVPW=0P_VP_W=0 and PV+PW=I5P_V+P_W=I_5 give

PXPX′=δXX′PX,∑X∈{V,W}nPX=I5n.P_XP_{X'}=\delta_{XX'}P_X,\qquad\sum_{X\in\{V,W\}^n}P_X=I_{5^n}.

Both local maps γ(η)\gamma(\eta) and M(d)M(d) preserve VV and WW, by the Eq. (51). They therefore commute with PV,PWP_V,P_W. The commutation property extends to each sector:

MnPX=⨂i=1n[M(di)PXi]=⨂i=1n[PXiM(di)]=PXMn,M_nP_X=\bigotimes_{i=1}^{n}[M(d_i)P_{X_i}] =\bigotimes_{i=1}^{n}[P_{X_i}M(d_i)] =P_XM_n,

and hence MnM_n factorises by sectors (has block-diagonal form with blocks supported on HX\mathcal{H}_X), as Γ(η)\Gamma(\eta) do.

Preserving a sector does not mean that every vector in it is unchanged by relabelling. On the other hand, the physical data and merged-contraction vectors yorby_{\rm orb} and zorbz_{\rm orb} are invariant by Eq. (5.7), and their sector components inherit this property. Therefore, their sector components live in HXS3\mathcal{H}_X^{S_3}—the subspace of vectors in the sector invariant with respect to common relabelling. The group-average Reynolds projector onto the full invariant space (A⊗n)S3=⨁XHXS3(A^{\otimes n})^{S_3}=\bigoplus_X\mathcal{H}_X^{S_3} is

Pinv(n)=16∑η∈S3Γn(η)=RnLn.(53)P_{\rm inv}^{(n)}=\frac{1}{6}\sum_{\eta\in S_3}\Gamma_n(\eta)=R_nL_n. \tag*{(53)}

Since PXP_X commutes with every Γn(η)\Gamma_n(\eta), the product Pinv(n)PXP_{\rm inv}^{(n)}P_X projects onto HXS3\mathcal{H}_X^{S_3}. Tr⁡Pinv(n)\operatorname{Tr}P_{\rm inv}^{(n)} tells us how many independent reconstruction coordinates that sector supplies. As M(d)M(d) commutes with γ(η)\gamma(\eta): the latter is the identity on VV, while M(d)M(d) is scalar on WW. Tensoring the commutation relation gives MnΓn(η)=Γn(η)MnM_n\Gamma_n(\eta)=\Gamma_n(\eta)M_n. Thus MnM_n preserves the invariant subspace. In particular, its action on the orbit embedding satisfies

MnRn=RnLnMnRn=RnMorb.(54)M_nR_n=R_nL_nM_nR_n=R_nM_{\rm orb}. \tag*{(54)}

This identifies the map Morb(d)M_{\rm orb}(d) on the space of orbits as the restriction of the map Mn(d)M_n(d) on the equality-pattern space to the invariant coordinates. To count explicitly the dimension of the invariant subspace HXS3\mathcal{H}_X^{S_3}, define

IV(X)={i:Xi=V},IW(X)={i:Xi=W},I_V(X)=\{i:X_i=V\},\qquad I_W(X)=\{i:X_i=W\},

and let m=∣IV(X)∣m=|I_V(X)|, k=∣IW(X)∣k=|I_W(X)|. After grouping the VV factors first, pointwise invariance of VV gives

HXS3≃V⊗m⊗(W⊗k)S3.(55)\mathcal{H}_X^{S_3}\simeq V^{\otimes m}\otimes(W^{\otimes k})^{S_3}. \tag*{(55)}

Hence only the invariant subspace of W⊗kW^{\otimes k} is left to be identified. The dimension of the invariant subspace, derived in Appendix A, is

wk=dim⁡(W⊗k)S3=2k+3δk0+2(−1)k6,(56)w_k=\dim(W^{\otimes k})^{S_3}=\frac{2^k+3\delta_{k0}+2(-1)^k}{6}, \tag*{(56)}

dim⁡HXS3=3mwk\dim\mathcal{H}_X^{S_3}=3^mw_k.

Here k=0k=0 denotes the one-dimensional empty tensor product. The first values are w0=1w_0=1, w1=0w_1=0, and w2=w3=1w_2=w_3=1. The vanishing value w1=0w_1=0 means that a sector with just one zero-sum factor carries no invariant component of the measured data. Therefore the sectors contributing to the invariant subspace are

Sn={X∈{V,W}n:∣IW(X)∣≠1}.(57)\mathcal{S}_n=\{X\in\{V,W\}^n:|I_W(X)|\ne1\}. \tag*{(57)}

For two parties, VVVV contributes nine coordinates and WWWW contributes one; VWVW and WVWV contribute none; constituting the 10-dimensional orbit space. For three parties, VVVVVV, the three placements of WVVWVV, and WWWWWW give 27+3×3+1=3727+3\times3+1=37 coordinates. More generally, the dimension sum

∑k=0n(nk)3n−kwk=Nn,(58)\sum_{k=0}^{n}\binom{n}{k}3^{n-k}w_k=N_n, \tag*{(58)}

agrees with the orbit count derived in Appendix A. Thus the invariant sectors supply a complete coordinate basis for the data already identified by orbit reduction.

Sector coordinates

The preceding subsection identifies the invariant subspace within each tensor sector. To write the restriction of Morb(d)M_{\mathrm{orb}}(d) to these subspaces, we need a basis for their coordinates and a way to convert back to the orbit coordinates defined by the measured data.

For each k≠1k\ne1, choose a matrix Jk:Rwk→A⊗kJ_k:\mathbb{R}^{w_k}\to A^{\otimes k} whose independent columns span (W⊗k)S3(W^{\otimes k})^{S_3}, and set J0=[1]J_0=[1]. One explicit construction is to select wkw_k independent columns of the projector

Pinv(k)PW⊗k,Pinv(k)=16∑η∈S3γ(η)⊗k.(59)P_{\mathrm{inv}}^{(k)}P_W^{\otimes k},\qquad P_{\mathrm{inv}}^{(k)}=\frac{1}{6}\sum_{\eta\in S_3}\gamma(\eta)^{\otimes k}. \tag*{(59)}

Its image is exactly im⁡(Pinv(k)PW⊗k)=(W⊗k)S3\operatorname{im}(P_{\mathrm{inv}}^{(k)}P_W^{\otimes k})=(W^{\otimes k})^{S_3}. The factor Pinv(k)P_{\mathrm{inv}}^{(k)} averages over the same relabelling at all kk positions, while PW⊗kP_W^{\otimes k} restricts each factor to WW. The two projectors commute, so their product is a projector onto an intersection of (A⊗n)S3(A^{\otimes n})^{S_3} and W⊗kW^{\otimes k} and hence has the rank wkw_k. As no non-zero invariant vector exists in WW, a single WW factor kills the invariant subspace of a tensor sector.

At each VV position we pick one of the three columns of R1R_1; for the kk WW positions jointly, we use one of the wkw_k columns of JkJ_k. Taking all tensor products of these basis vectors and restoring the original party order gives a basis of HXS3\mathcal{H}^{S_3}_X with 3mwk3^m w_k columns. The embedding of HXS3\mathcal{H}^{S_3}_X is:

RXsec=ΣX−1(R1⊗m⊗Jk)∈R5n×3mwk.(60)R_X^{\mathrm{sec}}=\Sigma_X^{-1}(R_1^{\otimes m}\otimes J_k)\in\mathbb{R}^{5^n\times3^m w_k}. \tag*{(60)}

Its columns form a basis of HXS3\mathcal{H}^{S_3}_X. Indeed, both R1⊗mR_1^{\otimes m} and JkJ_k have independent columns, and ΣX\Sigma_X only permutes tensor factors.

The inverse permutation restores the original party order; the sector coordinates retain the column order of R1⊗m⊗JkR_1^{\otimes m}\otimes J_k. This product convention determines the tensor form of the blocks in the theorem.

Each RXsecR_X^{\mathrm{sec}} embeds coordinates from one sector. To describe the embedding of all invariant spaces of subsequent sectors, one performs the horizontal concatenation of the sector embeddings: Rnsec=⨆X∈SnRXsec∈R5n×NnR_n^{\mathrm{sec}}=\bigsqcup_{X\in\mathcal{S}_n}R_X^{\mathrm{sec}}\in\mathbb{R}^{5^n\times N_n}.

Because orbits do not change under the action of the group, the orbit space is a subspace of (A⊗n)S3(A^{\otimes n})^{S_3}. While the dimension of the orbit space and the invariant space agree, it is the same subspace. The two embeddings RnR_n and RnsecR_n^{\mathrm{sec}} assign different coordinates to the same vector: RnR_n uses orbit basis, whereas RnsecR_n^{\mathrm{sec}} uses the sector bases. We need the matrix that converts the latter coordinates into the former. Applying the orbit extraction map LnL_n to each column of RnsecR_n^{\mathrm{sec}} gives

Qn=LnRnsec,Rnsec=RnQn.(61)Q_n=L_nR_n^{\mathrm{sec}},\qquad R_n^{\mathrm{sec}}=R_nQ_n. \tag*{(61)}

The second equality holds because RnLnR_nL_n acts as the identity on the invariant subspace, spanned by the columns of RnsecR_n^{\mathrm{sec}}. Thus QnQ_n changes coordinates of a fixed vector in (A⊗n)S3(A^{\otimes n})^{S_3} from sector coordinates to orbit coordinates:

yorb(n)=Qnysec(n),zorb(n)=Qnzsec(n).y_{\mathrm{orb}}^{(n)}=Q_ny_{\mathrm{sec}}^{(n)},\qquad z_{\mathrm{orb}}^{(n)}=Q_nz_{\mathrm{sec}}^{(n)}.

Vice versa, define

LXsec=[(RXsec)TRXsec]−1(RXsec)T.(62)L_X^{\mathrm{sec}}=[(R_X^{\mathrm{sec}})^{\mathsf{T}}R_X^{\mathrm{sec}}]^{-1}(R_X^{\mathrm{sec}})^{\mathsf{T}}. \tag*{(62)}

The concatenation of these left inverses vertically in the same sector order forms LnsecL_n^{\mathrm{sec}}. Orthogonality between sectors then gives Ln(n)Rnsec=INnL_n^{(n)}R_n^{\mathrm{sec}}=I_{N_n}, while RnsecLnsec=Pinv(n)R_n^{\mathrm{sec}}L_n^{\mathrm{sec}}=P_{\mathrm{inv}}^{(n)}. Consequently, Qn(n)=LnRnQ_n^{(n)}=L_nR_n, and Qn−1Qn=INnQ_n^{-1}Q_n=I_{N_n}. See (53). The inverse Qn−1Q_n^{-1} converts measured orbit data into the sector coordinates on the direct sum of H(di)H(d_i) and λ(di)\lambda(d_i). This is the change of basis used in the main theorem below as represented in Figure 2, which completes the path begun in Figure 1.

Diagram showing retained invariant orbits, measured pattern orbits, invariant sectors, and data sectors

Figure 2. From the orbit coordinates at the bottom of Fig. 1 to independent reconstruction sectors. The vertical arrows are reversible changes of coordinates, with the same matrix for the retained invariants and measured data.

Structural theorem

The theorem shows how to obtain the physical information in z(n)=C⊗nxfull(n)z^{(n)}=C^{\otimes n}x_{\mathrm{full}}^{(n)} from the averaged data youty_{\mathrm{out}} of (3) analytically.

Theorem 1. Let n≥1n \ge1, and use the orbit map Morb(d)M_{\mathrm{orb}}(\mathbf{d}) and sector coordinate change QnQ_n defined in Eqs. (47) and (61). For each contributing sector X∈SnX \in\mathcal{S}_n, put k=∣IW(X)∣k = |I_W(X)| and

ΛX={1∣IW(X)∣=0,∏j∈IW(X)λ(dj)∣IW(X)∣>1.\Lambda_X = \begin{cases} 1 & |I_W(X)|=0,\\ \prod_{j\in I_W(X)}\lambda(d_j) & |I_W(X)|>1. \end{cases}

In the product basis of Eq. (60),

Qn−1Morb(d)Qn=⨁X∈SnBX,(63)Q_n^{-1}M_{\mathrm{orb}}(\mathbf{d})Q_n=\bigoplus_{X\in\mathcal{S}_n}B_X, \tag*{(63)}

where

BX=ΛXΣX−1[(⨂i∈IV(X)H(di))⊗Iwk]ΣX.B_X=\Lambda_X\Sigma_X^{-1}\left[\left(\bigotimes_{i\in I_V(X)}H(d_i)\right)\otimes I_{w_k}\right]\Sigma_X.

The block BXB_X has dimension 3∣IV(X)∣wk3^{|I_V(X)|}w_k. Because each H(di)H(d_i) is invertible and each λ(dj)\lambda(d_j) is nonzero, Morb(d)M_{\mathrm{orb}}(\mathbf{d}) is invertible.

Proof. Fix a contributing sector X∈SnX\in\mathcal{S}_n, and retain its factor sets IV,IWI_V,I_W and their sizes m,km,k from the sector construction. The operator ΣX\Sigma_X in Eq. (60) groups its VV factors before its WW factors. Hence

ΣXMnΣX−1=(⨂i∈IVM(di))⊗(⨂j∈IWM(dj)).(64)\Sigma_XM_n\Sigma_X^{-1}=\left(\bigotimes_{i\in I_V}M(d_i)\right)\otimes\left(\bigotimes_{j\in I_W}M(d_j)\right). \tag*{(64)}

This reordering keeps each dimension did_i attached to its party; the dimensions need not be equal. From the two local actions in Eq. (51) then

(⨂i∈IVM(di))R1⊗m=R1⊗m(⨂i∈IVH(di)),\left(\bigotimes_{i\in I_V}M(d_i)\right)R_1^{\otimes m}=R_1^{\otimes m}\left(\bigotimes_{i\in I_V}H(d_i)\right),
(⨂j∈IWM(dj))Jk=JkΛXIwk.(65)\left(\bigotimes_{j\in I_W}M(d_j)\right)J_k=J_k\Lambda_XI_{w_k}. \tag*{(65)}

The operator before JkJ_k acts as the same scalar on all of W⊗kW^{\otimes k}, and therefore on the columns of JkJ_k. Substituting Eq. (65) into the sector embedding (6.13), and using Eq. (64), gives

MnRXsec=RXsecBX,(66)M_nR_X^{\mathrm{sec}}=R_X^{\mathrm{sec}}B_X, \tag*{(66)}

with BXB_X as stated in Eq. (63). Concatenating Eq. (66) over the contributing sectors yields

MnRnsec=Rnsec(⨁X∈SnBX).(67)M_nR_n^{\mathrm{sec}}=R_n^{\mathrm{sec}}\left(\bigoplus_{X\in\mathcal{S}_n}B_X\right). \tag*{(67)}

The relation (6.7) and the basis relation (6.14) also give

MnRnsec=RnMorbQn=Rnsec(Qn−1MorbQn).(68)M_nR_n^{\mathrm{sec}}=R_nM_{\mathrm{orb}}Q_n=R_n^{\mathrm{sec}}\left(Q_n^{-1}M_{\mathrm{orb}}Q_n\right). \tag*{(68)}

Applying the left inverse LnsecL_n^{\mathrm{sec}} from Eq. (62) to Eqs. (67) and (68) proves the block identity Eq. (63). Finally, every H(di)H(d_i) has the inverse in Eq. (24), and each λ(dj)\lambda(d_j) in Eq. (51) is nonzero. Thus every block is invertible. The basis change QnQ_n then makes MorbM_{\mathrm{orb}} invertible as well. □\square

Thus each sector satisfies its own reconstruction system,

y~sec,X(n)=BXzsec,X(n).\widetilde{\mathbf{y}}_{\mathrm{sec},X}^{(n)}=B_X\mathbf{z}_{\mathrm{sec},X}^{(n)}.

Its inverse requires only the known one-party inverse matrices H(di)−1H(d_i)^{-1} and a scalar rescaling λ(di)−1\lambda(d_i)^{-1}:

BX−1=ΛX−1[(⨂i∈IV(X)H(di)−1)⊗Iwk],B_X^{-1}=\Lambda_X^{-1}\left[\left(\bigotimes_{i\in I_V(X)}H(d_i)^{-1}\right)\otimes I_{w_k}\right],
zorb(n)=Qn(⨁X∈SnBX−1)Qn−1yorb(n).(69)\mathbf{z}_{\mathrm{orb}}^{(n)}=Q_n\left(\bigoplus_{X\in\mathcal{S}_n}B_X^{-1}\right)Q_n^{-1}\mathbf{y}_{\mathrm{orb}}^{(n)}. \tag*{(69)}

Equation (69) is the central result. As we announce below Eq. (24), even if MorbM_{\mathrm{orb}} explodes as Nn∝5n/6N_n\propto5^{n/6}, we discover that inverting never takes more than 3×33\times3. A python notebook, where the reconstruction is implemented symbolically in SymPy is available in Ref. [33].

For measured data, Appendix B explains how to obtain y~orb(n)\widetilde{\mathbf{y}}_{\mathrm{orb}}^{(n)} from event frequencies, including the outcome and orbit multiplicities. Inserting this vector into Eq. (69) gives estimates of the retained invariant combinations. If a resource test uses only selected combinations, its linear input map can be composed with Eq. (69) in advance, so the remaining invariant coordinates need not be estimated separately. Finite-size entanglement certification from second- and third-order invariants provides a concrete application [17]. This reduces the intermediate reconstruction steps; a comparison of the measurement time needed for a prescribed precision requires a separate statistical analysis.

The bipartite case

For two parties, the theorem gives a nine-dimensional VVVV block H(dA)⊗H(dB)H(d_A)\otimes H(d_B) and a one-dimensional WWWW block λ(dA)λ(dB)\lambda(d_A)\lambda(d_B). The transformation from the orbit basis to the sector basis goes as follows.

A vector u∈{eTA,eTB,eTC}⊗2u\in\{e_{T_A},e_{T_B},e_{T_C}\}^{\otimes2} is S3S_3-invariant iff it is of the form

α∑i∈{TA,TB,TC}ei⊗ei+β∑i≠ji,j∈{TA,TB,TC}ei⊗ej=α(T=T)+β(T≠T).(70)\alpha\sum_{i\in\{T_A,T_B,T_C\}}e_i\otimes e_i+\beta\sum_{\substack{i\ne j\\i,j\in\{T_A,T_B,T_C\}}}e_i\otimes e_j =\alpha(T=T)+\beta(T\ne T). \tag*{(70)}

in the orbit basis. Hence the only S3S_3-invariant vector u2u_2 in W⊗WW\otimes W has α=−2β\alpha=-2\beta and the only S3S_3-invariant vector u1u_1 in V⊗VV\otimes V has α=β\alpha=\beta. Summarising:

\begin{bmatrix}u_1\\u_2\end{bmatrix} = \begin{bmatrix}1&1\\2&-1\end{bmatrix} \begin{bmatrix}T=T\\T\ne T\end{bmatrix}. \tag*{(71)}\end{bmatrix}

and the coordinate transformation in this basis transition is described by the transposed matrix:

QTT=(121−1),QTT−1=13(121−1).(72)Q_{TT}=\begin{pmatrix}1&2\\1&-1\end{pmatrix},\qquad Q_{TT}^{-1}=\frac{1}{3}\begin{pmatrix}1&2\\1&-1\end{pmatrix}. \tag*{(72)}

All other orbit coordinates already match the VVVV product coordinates, up to their ordering. The two displayed directions have data coordinates (yT=T+2yT≠T)/3(y_{T=T}+2y_{T\ne T})/3 and (yT=T−yT≠T)/3(y_{T=T}-y_{T\ne T})/3, respectively. The same transformation applies to the merged input.

In particular, the scalar WWWW response gives

yT=T−yT≠T=λ(dA)λ(dB)(x5−x4).(73)y_{T=T}-y_{T\ne T}=\lambda(d_A)\lambda(d_B)(x_5-x_4). \tag*{(73)}

For a normalized state, this isolates the difference between the global purity Tr⁡ρ2\operatorname{Tr}\rho^2 and Tr⁡[ρ(ρA⊗ρB)]\operatorname{Tr}[\rho(\rho_A\otimes\rho_B)]. The known bipartite nine-plus-one factorization [16] is thus the first instance of the general sector reconstruction.

Partial transposition and information limits

Theorem 1 reconstructs all combinations retained by the cycle-merging map C⊗nC^{\otimes n} (3.13). To see why the discarded cycle orientations cannot be separated, recall that the three probability factors in Eq. (3) share each local setting. After Haar averaging, their local measurement coefficients obey

[SW(d)]π,(123)=[SW(d)]π,(132),π∈Π.(74)\left[SW(d)\right]_{\pi,(123)}=\left[SW(d)\right]_{\pi,(132)},\qquad\pi\in\Pi. \tag*{(74)}

The two orientations therefore enter these equations only through their sum. The map CC records this information loss; inverting the averaged equations cannot recover their difference.

Let Γj\Gamma_j denote transposition on party jj. A negative eigenvalue of ρΓj\rho^{\Gamma_j} certifies entanglement between that party and the rest [34]; its moments can also support entanglement tests [35]. Partial transposition makes the lost cycle distinction explicit. Moving it from the state to the permutation operator in Eq. (37), using PτjTj=Pτj−1P_{\tau_j}^{T_j}=P_{\tau_j^{-1}}, gives

xρΓj(τ1,…,τn)=xρ(τ1,…,τj−1,…,τn).(75)x_{\rho^{\Gamma_j}}(\tau_1,\ldots,\tau_n)=x_\rho(\tau_1,\ldots,\tau_j^{-1},\ldots,\tau_n). \tag*{(75)}

Thus it exchanges the two three-cycles at party jj alone; the identity and transpositions are unchanged. This differs from complex conjugation, which inverts every local label, and common copy relabelling, which conjugates every label by the same permutation.

Equation (8.2) exchanges the two cycle orientations at party jj, while CC maps both to eEe_E. Consequently,

z(n)(ρΓj)=z(n)(ρ).(76)z^{(n)}(\rho^{\Gamma_j})=z^{(n)}(\rho). \tag*{(76)}

The discarded local direction is ker⁡C=span⁡{e(123)−e(132)}\ker C=\operatorname{span}\{e_{(123)}-e_{(132)}\}; ker⁡(C⊗n)\ker(C^{\otimes n}) is spanned by tensors containing at least one such factor. These directions are removed before orbit reduction.

The same equality has a Born-rule explanation. Set U(j∗)=(U1,…,Uj‾,…,Un)U^{(j*)}=(U_1,\ldots,\overline{U_j},\ldots,U_n), where the bar denotes entrywise conjugation. Equation (2.1) gives

pρΓj(i∣U)=pρ(i∣U(j∗)).(77)p_{\rho^{\Gamma_j}}(i|U)=p_\rho(i|U^{(j*)}). \tag*{(77)}

This relates generally different settings; it does not equate the two expressions at the same UU. Applying Eq. (77) to each of the three factors in Eq. (3) gives, with dU=∏kdUk\mathrm{d}U=\prod_k\mathrm{d}U_k,

∫dU∏h=13pρΓj(ih∣U)=∫dU∏h=13pρ(ih∣U(j∗))=∫dU∏h=13pρ(ih∣U).(78)\begin{aligned} \int\mathrm{d}U\prod_{h=1}^{3}p_{\rho^{\Gamma_j}}(i_h|U)=\int\mathrm{d}U\prod_{h=1}^{3}p_\rho(i_h|U^{(j*)}) \\ &=\int\mathrm{d}U\prod_{h=1}^{3}p_\rho(i_h|U). \tag*{(78)} \end{aligned}

The change of variables Uj↦Uj‾U_j\mapsto\overline{U_j} preserves Haar measure and gives the last equality. Averaging therefore removes the distinction between the two settings, in agreement with Eq. (76). For two parties, partial transposition on BB interchanges x9=Tr⁡ρ3x_9=\operatorname{Tr}\rho^3 and x10=Tr⁡[(ρΓB)3]x_{10}=\operatorname{Tr}[(\rho^{\Gamma_B})^3]. Equation (4.6) retains only 2(x9+x10)2(x_9+x_{10}), so these correlators cannot determine x9−x10x_9-x_{10} without additional information.

To access independently only x10x_{10} one can define for all R=(i,U)R=(i,U) an unbiased estimator ρ^(R)\hat{\rho}(R), that does not need to be a density matrix [10, 15], such that Eq[ρ^(R)]=ρ\mathbb{E}_q[\hat{\rho}(R)]=\rho where qρ(dU,i)=μ(dU)pρ(i∣U)q_\rho(dU,i)=\mu(dU)p_\rho(i|U). In this approach, each shot RR retains both the random setting and the measured outcome [11]. Therefore to access only x10x_{10}, the price to pay is an experiment where each applied setting UU must be recorded. Instead, in our protocol, we have access only to x9+x10x_9+x_{10}, but we only require that the same unknown UU acting on each copy of ρ\rho is Haar-distributed. Therefore the lack of information on UU turns into an easier and more economical experimental procedure.

In summary, the information limit proved here is forced by the common-setting Haar averages and Theorem 1 completes the reconstruction of what survives those averages.

Conclusion

We have determined which third-order local-unitary invariant combinations can be reconstructed from products of three outcome probabilities measured at a common local random setting and then averaged. For any number of parties, the theorem recovers all combinations retained by local cycle merging by accessing to the physical information with a symbolically inversion of 3×33\times3 matrices. This avoids to handle a numerical inversion of a matrix that explodes as the number of subsystems increases. The protocol to access those physical quantities uses independent local rotations, measurements on separately prepared copies, and classical processing, making it compatible with distributed experiments without the need of calibrating or storing each unitary. The result distinguishes information present in these correlators from information their equations cannot recover. In particular, the local Haar averages force the bipartite third moments Tr⁡ρ3\operatorname{Tr}\rho^3 and Tr⁡[(ρΓB)3]\operatorname{Tr}[(\rho^{\Gamma_B})^3] to appear only through their sum. The number of retained coordinates still grows exponentially, and the theorem does not establish favorable statistical cost or conditioning. Its practical appeal is that reconstruction does not require recording the applied unitary matrices. Developing efficient estimators for the retained combinations remains an open task, as does extending the analysis to higher orders.

Acknowledgements

We thank D. Chruściński, C. Lupo and M. Ziman for the fruitful scientific discussion. GSc is supported by grant NCN SONATA 21 UMO2025/59/D/ST2/02708. GSc thanks the Institute of Physics at the Slovak Academy of Sciences for its hospitality through the SAIA Support Fund.

References

Characters and dimensions of the invariant sectors

This appendix derives the dimensions of the invariant tensor sectors and verifies that their sum equals the orbit count NnN_n. The required information is contained in the characters: the traces of the representation matrices.

For a representation on H\mathcal{H} with character χ\chi, averaging its matrices over the group projects onto HS3\mathcal{H}^{S_3}. The average maps every vector to an invariant vector and acts as the identity on vectors already invariant. Its trace therefore equals the dimension of that subspace, giving [36]

dim⁡HS3=16∑η∈S3χ(η)=χ(1)+3χ((12))+2χ((123))6.(79)\dim\mathcal{H}^{S_3} = \frac{1}{6}\sum_{\eta\in S_3}\chi(\eta) = \frac{\chi(1)+3\chi((12))+2\chi((123))}{6}. \tag*{(79)}

The weights count the group elements in each conjugacy class. Thus invariant-subspace dimensions can be calculated without constructing a basis of the subspace of invariant vectors.

The pattern action in Eq. (19) gives the local characters

11(12)(12)(123)(123)
class size113322
χγ\chi_\gamma553322
χW\chi_W2200−1-1

The last row follows by subtracting the three invariant directions in VV from χγ\chi_\gamma. Characters are constant on each conjugacy class, so these values specify all six group elements.

Since γ\gamma is a permutation representation, its trace counts local patterns invariant by the group element: the identity fixes all five, a transposition fixes DD, EE and its matching TT label, and a three-cycle fixes only DD, $E. A tuple of local π∈Πn\pi\in\Pi^n is invariant under relabelling π↦η⋅π\pi\mapsto\eta\cdot\pi, ∀η∈S3\forall\eta\in S_3. Equivalently, traces multiply under tensor products, so Γn=γ⊗n\Gamma_n = \gamma^{\otimes n} has class character (5n,3n,2n)(5^n, 3^n, 2^n). Averaging with the class sizes $1, 3, 2$ in Eq. (79) gives

Nn=∣Πn/S3∣=5n+3⋅3n+2⋅2n6=dim⁡(A⊗n)S3.(80)N_n = \lvert\Pi^n/S_3\rvert= \frac{5^n+3\cdot3^n+2\cdot2^n}{6} = \dim\left(A^{\otimes n}\right)_{S_3}. \tag*{(80)}

This is also Burnside’s orbit-counting formula. It gives N1=3N_1=3, N2=10N_2=10, and N3=37N_3=37, counting orbit coordinates rather than (∏i=1ndi)2\left(\prod_{i=1}^n d_i\right)^2 independent parameters of a normalized physical state.

For k≥1k \ge1, the character of W⊗kW^{\otimes k} instead has class values

χW⊗k=(2k,0,(−1)k).\chi_{W^{\otimes k}} = (2^k, 0, (-1)^k).

We calculate the dimension of its invariant subspace using the formula Eq. (79). At k=0k=0, the empty tensor product is one-dimensional and every group element acts as the identity, giving character values (1,1,1)(1,1,1). Combining these cases yields

wk=dim⁡(W⊗k)S3=2k+3δk0+2(−1)k6.(81)w_k = \dim\left(W^{\otimes k}\right)_{S_3} = \frac{2^k+3\delta_{k0}+2(-1)^k}{6}. \tag*{(81)}

From measured frequencies to reconstruction coordinates

(69) takes representative orbit correlators yorb(n)y_{\mathrm{orb}}^{(n)} as input. Measured event frequencies instead count all outcome triples belonging to a pattern orbit. This appendix converts those frequencies into the coordinates required by the inverse map in Eq. (25).

For an orbit o∈Πno \in\Pi^n, the corresponding event contains every outcome triple whose local equality-pattern tuple belongs to oo. Choose a representative π∈o\pi\in o. Using the local multiplicities mπk(dk)m_{\pi_k}(d_k) introduced with Eq. (26), this representative has ∏k=1nmπk(dk)\prod_{k=1}^n m_{\pi_k}(d_k) numerical outcome realizations. Common copy relabelling preserves every local type, so the same multiplicity applies to all ∣o∣\lvert o\rvert pattern tuples in the orbit.

All these realizations have the same averaged correlator yo=(yorb(n))oy_o = \left(y_{\mathrm{orb}}^{(n)}\right)_o. The probability mass of the complete orbit event is therefore

Po=∣o∣[∏k=1nmπk(dk)]yo.(82)P_o = \lvert o\rvert\left[\prod_{k=1}^n m_{\pi_k}(d_k)\right]y_o. \tag*{(82)}

Here ∣o∣\lvert o\rvert counts pattern tuples, while the product counts numerical outcome realizations of each tuple.

To estimate PoP_o, form triples from three distinct measurement shots at each product setting, as required by Eq. (2), and count the fraction belonging to the event oo.

For an orbit with nonzero outcome multiplicity, the representative-coordinate estimate is

y^o=Po∣o∣∏k=1nmπk(dk).(83)\widehat{y}_o = \frac{P_o}{\lvert o\rvert\prod_{k=1}^n m_{\pi_k}(d_k)}. \tag*{(83)}

Collecting these estimates in the prescribed orbit order gives y^orb(n)\widehat{y}_{\mathrm{orb}}^{(n)}, to which the reconstruction formula (7.7) applies.

References

  1. [1]S. J. van Enk and C. W. J. Beenakker, Physical Review Letters 108, 110503 (2012), arXiv:1112.1027.DOI
  2. [2]A. K. Ekert, C. Moura Alves, D. K. L. Oi, M. Horodecki, P. Horodecki, and L. C. Kwek, Physical Review Letters 88, 217901 (2002), arXiv:quant-ph/0203016.DOI
  3. [3]T. A. Brun, Measuring polynomial functions of states, Quantum Information and Computation 4, 401 (2004), arXiv:quant-ph/0401067.DOI
  4. [4]M. S. Leifer, N. Linden, and A. Winter, Physical Review A 69, 052304 (2004), arXiv:quant-ph/0308008.DOI
  5. [5]A. Elben, S. T. Flammia, H.-Y. Huang, R. Kueng, J. Preskill, B. Vermersch, and P. Zoller, Nature Reviews Physics 5, 9 (2023), arXiv:2203.11374.DOI
  6. [6]P. Cieśliński, S. Imai, J. Dziewior, O. Gühne, L. Knips, W. Laskowski, J. Meinecke, T. Paterek, and T. Vértesi, Physics Reports 1095, 1 (2024), arXiv:2307.01251.DOI
  7. [7]A. Elben, B. Vermersch, C. F. Roos, and P. Zoller, Phys. Rev. A 99, 052323 (2019), arXiv:1812.02624.DOI
  8. [8]A. Elben, B. Vermersch, M. Dalmonte, J. I. Cirac, and P. Zoller, Physical Review Letters 120, 050406 (2018), arXiv:1710.05060.
  9. [9]T. Brydges, A. Elben, P. Jurcevic, B. Vermersch, C. Maier, B. P. Lanyon, P. Zoller, R. Blatt, and C. F. Roos, Science 364, 260 (2019), arXiv:1806.05747.DOI
  10. [10]H.-Y. Huang, R. Kueng, and J. Preskill, Nature Physics 16, 1050 (2020), arXiv:2002.08953.DOI
  11. [11]A. Elben, R. Kueng, H.-Y. R. Huang, R. van Bijnen, C. Kokail, M. Dalmonte, P. Calabrese, B. Kraus, J. Preskill, P. Zoller, and B. Vermersch, Phys. Rev. Lett. 125, 200501 (2020), arXiv:2007.06305.DOI
  12. [12]Z. Liu, Y. Tang, H. Dai, P. Liu, S. Chen, and X. Ma, Physical Review Letters 129, 260501 (2022), arXiv:2203.08391.DOI
  13. [13]N. Wyderka and A. Ketterer, PRX Quantum 4, 020325 (2023), arXiv:2211.09610.DOI
  14. [14]A. Rath, R. van Bijnen, A. Elben, P. Zoller, and B. Vermersch, Physical Review Letters 127, 200503 (2021), arXiv:2102.13524.DOI
  15. [15]B. Vermersch, A. Rath, B. Sundar, C. Branciard, J. Preskill, and A. Elben, PRX Quantum 5, 010352 (2024), arXiv:2304.12292.DOI
  16. [16]G. Scala, A. Bera, and G. Sarbicki, Quantum Science and Technology 11, 015022 (2026), arXiv:2506.18303.arxiv.org/abs/2506.18303
  17. [17]G. Scala, C. C. Mattiacci, D. Pietryka, and G. Sarbicki, (2026), arXiv:2609.08635.arxiv.org/abs/2609.08635
  18. [18]D. Gross, K. Audenaert, and J. Eisert, Journal of Mathematical Physics 48, 052104 (2007), arXiv:quant-ph/0611002.DOI
  19. [19]A. A. Mele, Quantum 8, 1340 (2024), arXiv:2307.08956.DOI
  20. [20]H. Zhu, Phys. Rev. A 96, 062336 (2017), arXiv:1510.02619.DOI
  21. [21]B. Collins and P. Śniady, Communications in Mathematical Physics 264, 773 (2006), arXiv:math-ph/0402073.DOI
  22. [22]B. Collins, International Mathematics Research Notices 2003, 953 (2003), arXiv:math-ph/0205010.DOI
  23. [23]B. Collins, S. Matsumoto, and J. Novak, Notices of the American Mathematical Society 69, 734 (2022), arXiv:2109.14890.DOI
  24. [24]M. Grassl, M. Rötteler, and T. Beth, Physical Review A 58, 1833 (1998), arXiv:quant-ph/9712040.DOI
  25. [25]E. M. Rains, IEEE Transactions on Information Theory 46, 54 (2000), arXiv:quant-ph/9704042.
  26. [26]Y. Makhlin, Quantum Information Processing 1, 243 (2002), arXiv:quant-ph/0002045.
  27. [27]J. Kempe, Physical Review A 60, 910 (1999), arXiv:quant-ph/9902036.DOI
  28. [28]N. Linden, S. Popescu, and A. Sudbery, Physical Review Letters 83, 243 (1999), arXiv:quant-ph/9801706.
  29. [29]P. Vrana, Journal of Physics A: Mathematical and Theoretical 44, 115302 (2011), arXiv:1007.0163.DOI
  30. [30]P. W. Brouwer and C. W. J. Beenakker, Journal of Mathematical Physics 37, 4904 (1996), arXiv:cond-mat/9604059.
  31. [31]Z. Puchała and J. A. Miszczak, Bulletin of the Polish Academy of Sciences: Technical Sciences 65, 21 (2017), arXiv:1109.4244.
  32. [32]G. Scala, C. C. Mattiacci, and G. Sarbicki (2026), in preparation.
  33. [33]C. C. Mattiacci, G. Scala, and G. Sarbicki, 3-rd Order Symbolic Calculus, (2026), GitHub repository.
  34. [34]A. Peres, Physical Review Letters 77, 1413 (1996), arXiv:quant-ph/9604005.DOI
  35. [35]X.-D. Yu, S. Imai, and O. Gühne, Physical Review Letters 127, 060504 (2021), arXiv:2103.06897.DOI
  36. [36]J.-P. Serre, Linear Representations of Finite Groups, Graduate Texts in Mathematics, Vol. 42 (Springer-Verlag, New York, 1977).DOI

Paper details

vlnc:W_d5efuvxg
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 2

downloads · 0

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

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents