Introduction
Topological quantum error correcting codes store information in nonlocal degrees of freedom that are protected against local perturbations [1 , 2 , 3 ] . Local weak measurements provide partial and noisy information about the quantum state [4 , 5 , 6 ] . This raises a natural inference problem: “given such a measurement record, which nonlocal observables can be inferred, and what constraints relate their learnability?”
For a binary observable in a stabilizer code, a natural estimate is its average conditioned on the measurement record, called the posterior expectation value. Its square, averaged over all records, diagnoses how reliably the observable can be inferred. The behavior of this quantity at large separations therefore indicates whether the corresponding nonlocal information remains learnable over long distances.
Statistical mechanical mappings provide a powerful framework for addressing this problem. They describe both Pauli errors in stabilizer codes [2 , 7 , 8 , 9 ] and the inference problem associated with weak measurements [10 , 11 , 12 , 13 , 14 , 15 , 16 , 17 ] . In the latter case, the spin configuration is unknown, and Bayes’ theorem uses the noisy measurement record to update the probability assigned to each configuration. Recent studies of the Ising learning problem and its dual description in terms of a deformed toric code wavefunction [18 ] identified a special relation between the model parameters, called the higher Nishimori line [19 , 20 ] . For Gaussian weak measurements, this line is obtained by setting the measurement strength equal to the strength of the wavefunction deformation. This condition leads to an enlarged permutation symmetry in a formulation using several copies, or replicas, of the classical model, allowing us to obtain exact relations between various record averaged correlators. Related higher Nishimori structures have recently been identified in Potts models and deformed Z q Z_q Z q toric codes [21 ] .
These results motivate a further question. Previous studies have mainly related averages of different powers of the posterior expectation of a single observable. In quantum codes with generalized symmetries, distinct nonlocal observables can also obey nontrivial multiplication relations. We ask whether the higher Nishimori symmetry can turn these relations into exact constraints on which observables can be learned simultaneously.
In this work, we address this question for a broad class of Calderbank–Shor–Steane (CSS) stabilizer codes and extend the analysis to finite group quantum double models, including non-Abelian ones. For CSS codes, the commutation relations between stabilizer generators and locally measured Pauli operators directly determine the interactions of an associated classical spin model. Each measured Pauli operator produces a local multispin interaction, while relations among the stabilizer generators appear as exact symmetries of the classical model. This construction accommodates ordinary global, spatially modulated, and subsystem symmetries within a common framework.
For these codes, matching each wavefunction deformation strength to the corresponding Gaussian measurement strength produces an enlarged replica permutation symmetry. We refer to the resulting parameter manifold as the higher Nishimori manifold . Using this symmetry, we derive exact identities relating the posterior expectations of distinct observables and their operator products. In particular, the record averaged second moments that quantify learnability are equal to correlations in the corresponding clean models. Here, the clean model describes the deformed state before conditioning on a measurement record. Combining this relation with correlation inequalities for clean models with nonnegative ferromagnetic couplings, we establish hierarchies of learnability: if two observables remain learnable at long distances, their product must also remain learnable. These inequalities constrain the possible patterns of learned and unlearned information without requiring a complete solution of the measurement dependent phase diagram.
As a concrete application, we consider the rank-2 toric code [22 ] . The associated classical model possesses spatially modulated symmetries inherited from relations among the rank-2 toric code stabilizers [23 , 24 , 25 ] . Correspondingly, the quantum code supports spatially modulated dipole loop operators. Multiplying a dipole loop operator by a lattice translation gives a homogeneous monopole loop operator. Applying the general hierarchy to the corresponding open string observables, we show that long range dipole learnability implies long range monopole learnability. Hence, the multiplication relations among these operators directly restrict the allowed patterns of learnability.
We obtain analogous exact relations in the X-cube model [26 , 27 , 28 ] . For local Pauli-Z Z Z measurements, relations among cube stabilizers generate planar subsystem symmetries in the associated classical plaquette model. Multiplication relations among subsystem invariant observables then yield exact inequalities among learning correlators that probe different pairs of planes.
Beyond CSS stabilizer codes, we extend the analysis to finite group quantum double models [1 , 29 , 30 , 31 , 32 ] , including the ones with non-Abelian anyons. We establish an exact higher Nishimori identity for an isotropic Gaussian measurement protocol. This identity relates the learning correlators of pure electric observables to correlations in the corresponding classical model. When the classical interactions admit a nonnegative expansion in terms of normal subgroups, we derive further inequalities. We illustrate these constraints in the S 3 S_3 S 3 and D 4 D_4 D 4 quantum doubles. Together, these examples demonstrate how operator multiplication and fusion constrain which types of nonlocal information can be learned simultaneously from noisy local measurements, providing a guide to learning phase diagrams in quantum codes.
The remainder of this paper is organized as follows. In § 2 , we review the classical spin representation of CSS stabilizer codes, including the symmetries arising from relations among stabilizer generators, and formulate the weak measurement problem as Bayesian inference. In § 3 , we derive the generalized higher Nishimori identities and the resulting hierarchy of learnability. We apply these results to the rank-2 toric code in § 4 , with the X-cube analysis presented in Appendix. In § 5 , we formulate the learning problem for finite group quantum doubles and derive constraints on the learnability of their electric observables. Finally, § 6 summarizes our results and discusses future directions. Additional technical details are provided in the appendices.
From stabilizer codes to classical spin models
Before formulating the learning problem, we review how a CSS stabilizer code gives rise to a classical spin model. We focus on one Pauli sector, since the two sectors of a CSS code can be treated independently in the same manner. Previous works have established the correspondence between quantum information observables of stabilizer codes and observables of the associated classical spin models [2 , 7 ] . Our purpose here is to formulate the measurement induced inference problem within this correspondence and to identify the role played by stabilizer relations.
Stabilizer relations and spin symmetries
Let { S k } k = 1 N s \{S_k\}_{k=1}^{N_s} { S k } k = 1 N s be a set of mutually commuting stabilizer generators, all of either X X X type or Z Z Z type. An element of the subgroup generated by these operators can be written as
g ( q ) = ∏ k = 1 N s S k q k , q k ∈ F 2 . (1) g(\boldsymbol{q}) = \prod_{k=1}^{N_s} S_k^{q_k}, \qquad q_k \in\mathbb{F}_2. \tag*{(1)} g ( q ) = k = 1 ∏ N s S k q k , q k ∈ F 2 . (1)
Here F 2 = { 0 , 1 } \mathbb{F}_2 = \{0,1\} F 2 = { 0 , 1 } is the binary field, with addition and multiplication defined modulo two. We associate an Ising variable with each binary variable q k q_k q k according to
σ k = ( − 1 ) q k , q k = 1 − σ k 2 . (2) \sigma_k = (-1)^{q_k}, \qquad q_k = \frac{1-\sigma_k}{2}. \tag*{(2)} σ k = ( − 1 ) q k , q k = 2 1 − σ k . (2)
Thus, q k = 1 q_k = 1 q k = 1 , corresponding to the inclusion of S k S_k S k in g ( q ) g(\boldsymbol{q}) g ( q ) , is represented by σ k = − 1 \sigma_k = -1 σ k = − 1 . We next introduce a family of local Pauli operators { P j } \{P_j\} { P j } of the complementary Pauli type. These operators represent local errors in § 2.2 and the observables subjected to weak measurements in § 2.3 . Their commutation relations with the stabilizer generators define the binary commutation matrix A A A :
P j S k = ( − 1 ) A j k S k P j , A j k = { 0 , [ P j , S k ] = 0 , 1 , { P j , S k } = 0. (3) P_jS_k = (-1)^{A_{jk}}S_kP_j,\qquad A_{jk} =
\begin{cases}
0, & [P_j,S_k]=0,\\
1, & \{P_j,S_k\}=0.
\end{cases} \tag*{(3)} P j S k = ( − 1 ) A j k S k P j , A j k = { 0 , 1 , [ P j , S k ] = 0 , { P j , S k } = 0. (3)
Commuting P j P_j P j through g ( q ) g(\boldsymbol{q}) g ( q ) gives
P j g ( q ) = ( − 1 ) ∑ k A j k q k g ( q ) P j = B j ( σ ) g ( q ) P j , (4) P_jg(\boldsymbol{q}) = (-1)^{\sum_k A_{jk}q_k}g(\boldsymbol{q})P_j = B_j(\boldsymbol{\sigma})g(\boldsymbol{q})P_j, \tag*{(4)} P j g ( q ) = ( − 1 ) ∑ k A j k q k g ( q ) P j = B j ( σ ) g ( q ) P j , (4)
where the sum in the exponent is evaluated modulo two and
B j ( σ ) = ∏ k = 1 N s σ k A j k . (5) B_j(\boldsymbol{\sigma}) = \prod_{k=1}^{N_s}\sigma_k^{A_{jk}}. \tag*{(5)} B j ( σ ) = k = 1 ∏ N s σ k A j k . (5)
Thus, each physical Pauli operator P j P_j P j determines an interaction B j B_j B j in the classical spin model. Its spatial support consists precisely of the stabilizer generators that anticommute with P j P_j P j .
The group multiplication also has a simple classical interpretation. For two stabilizer configurations σ \boldsymbol{\sigma} σ and s \boldsymbol{s} s , multiplication of the corresponding group elements is represented by the componentwise product
( σ s ) k : = σ k s k . (6) (\boldsymbol{\sigma}\boldsymbol{s})_k := \sigma_k s_k. \tag*{(6)} ( σ s ) k := σ k s k . (6)
Equation (5) then implies
B j ( σ s ) = B j ( σ ) B j ( s ) . (7) B_j(\boldsymbol{\sigma}\boldsymbol{s}) = B_j(\boldsymbol{\sigma})B_j(\boldsymbol{s}). \tag*{(7)} B j ( σ s ) = B j ( σ ) B j ( s ) . (7)
This property will be the central algebraic ingredient in the higher Nishimori identities derived in § 3 .
The local stabilizer generators need not be independent. Their relations form the binary vector space
R = { r ∈ F 2 N s | ∏ k = 1 N s S k r k = I } . (8) \mathcal{R} = \left\{r \in\mathbb{F}_2^{N_s}\,\middle|\,\prod_{k=1}^{N_s}S_k^{r_k}=\mathbb{I}\right\}. \tag*{(8)} R = { r ∈ F 2 N s k = 1 ∏ N s S k r k = I } . (8)
If r ∈ R r \in\mathcal{R} r ∈ R , the two binary configurations q \boldsymbol{q} q and q + r \boldsymbol{q}+r q + r describe the same stabilizer element. In terms of the Ising spins, this relation acts as
σ k ⟼ ( − 1 ) r k σ k . (9) \sigma_k \longmapsto(-1)^{r_k}\sigma_k. \tag*{(9)} σ k ⟼ ( − 1 ) r k σ k . (9)
Since the operator in (8) is the identity, it commutes with every P j P_j P j . It follows that
∑ k A j k r k = 0 ( m o d 2 ) \sum_k A_{jk}r_k = 0 \pmod{2} k ∑ A j k r k = 0 ( mod 2 )
and hence
B j ( { ( − 1 ) r k σ k } ) = B j ( σ ) . (10) B_j\left(\left\{(-1)^{r_k}\sigma_k\right\}\right)=B_j(\sigma). \tag*{(10)} B j ( { ( − 1 ) r k σ k } ) = B j ( σ ) . (10)
Stabilizer relations therefore appear as exact symmetries of the classical spin Hamiltonian. It is convenient to retain all local stabilizer generators rather than choosing a generally nonlocal independent subset. The resulting spin description contains a constant redundancy:
∑ g ∈ G F ( g ) = 1 ∣ R ∣ ∑ σ ∈ { + 1 } N s F ( g ( σ ) ) . (11) \sum_{g\in\mathcal{G}} F(g)=\frac{1}{|\mathcal{R}|}\sum_{\sigma\in\{+1\}^{N_s}}F(g(\sigma)). \tag*{(11)} g ∈ G ∑ F ( g ) = ∣ R ∣ 1 σ ∈ { + 1 } N s ∑ F ( g ( σ )) . (11)
for any function F F F of the stabilizer element. The factor ∣ R ∣ |\mathcal{R}| ∣ R ∣ only changes the overall normalization. Physical observables obtained from the stabilizer code must be invariant under (9) .
The relation between stabilizer identities and classical symmetries will be important in the examples in later sections. The global relation among toric code star operators produces the global Ising symmetry. The relations of the rank-2 toric code produce modulated symmetries (§ 4.2 ), while the subextensive relations of the X-cube code produce planar subsystem symmetries (Appendix § 10 ).
Classical Boltzmann weights from local Pauli noise
We next consider local Pauli noise. The classical Boltzmann weight can be seen directly from a local Pauli channel. For a state ρ 0 \rho_0 ρ 0 , we consider a channel induced by local Pauli noise with probability p j p_j p j :
E j [ ρ 0 ] = ( 1 − p j ) ρ 0 + p j P j ρ 0 P j , 0 ≤ p j < 1 2 . (12) \mathcal{E}_j[\rho_0]=(1-p_j)\rho_0+p_jP_j\rho_0P_j,\qquad0\leq p_j<\frac{1}{2}. \tag*{(12)} E j [ ρ 0 ] = ( 1 − p j ) ρ 0 + p j P j ρ 0 P j , 0 ≤ p j < 2 1 . (12)
Its action on a stabilizer element is given by
E j [ g ( q ) ] = [ ( 1 − p j ) + p j B j ( σ ) ] g ( q ) = ( 1 − 2 p j ) [ 1 − B j ( σ ) ] / 2 g ( q ) . (13) \mathcal{E}_j[g(q)]=\left[(1-p_j)+p_jB_j(\sigma)\right]g(q)=(1-2p_j)^{[1-B_j(\sigma)]/2}g(q). \tag*{(13)} E j [ g ( q )] = [ ( 1 − p j ) + p j B j ( σ ) ] g ( q ) = ( 1 − 2 p j ) [ 1 − B j ( σ )] /2 g ( q ) . (13)
For independent noise channels, we introduce
E : = ∏ j E j . (14) \mathcal{E}:=\prod_j\mathcal{E}_j. \tag*{(14)} E := j ∏ E j . (14)
Defining
μ j : = − log ( 1 − 2 p j ) , K j : = μ j 2 , (15) \mu_j:=-\log(1-2p_j),\qquad K_j:=\frac{\mu_j}{2}, \tag*{(15)} μ j := − log ( 1 − 2 p j ) , K j := 2 μ j , (15)
the coefficient produced by independent local channels becomes
w ( σ ) = ∏ j exp [ − μ j 2 ( 1 − B j ( σ ) ) ] ∝ exp [ ∑ j K j B j ( σ ) ] . (16) w(\sigma)=\prod_j\exp\left[-\frac{\mu_j}{2}\left(1-B_j(\sigma)\right)\right]\propto\exp\left[\sum_jK_jB_j(\sigma)\right]. \tag*{(16)} w ( σ ) = j ∏ exp [ − 2 μ j ( 1 − B j ( σ ) ) ] ∝ exp [ j ∑ K j B j ( σ ) ] . (16)
The associated classical Hamiltonian is therefore
H c l [ σ ] = − ∑ j K j B j ( σ ) . (17) H_{\mathrm{cl}}[\sigma]=-\sum_jK_jB_j(\sigma). \tag*{(17)} H cl [ σ ] = − j ∑ K j B j ( σ ) . (17)
This construction depends only on the commutation algebra between the local Pauli operators and the stabilizer generators. In the following subsection, we show that the same interaction functions B j ( σ ) B_j(\sigma) B j ( σ ) arise from the Born probabilities of a deformed code state. Weak measurements then promote the clean Boltzmann distribution to a posterior distribution conditioned on the measurement record.
Wavefunction deformation and weak measurement
We now show how the classical model introduced above arises from a deformed stabilizer code wave function subjected to weak measurements. We focus on Gaussian weak measurements, for which the replicated statistical mechanical representation derived below is exact at all length scales [19 ] .
Since the operators { P j } \{P_j\} { P j } mutually commute, we choose a common reference state ∣ Ω ⟩ \lvert\Omega\rangle ∣ Ω ⟩ satisfying
P j ∣ Ω ⟩ = ∣ Ω ⟩ for all j . (18) P_j\lvert\Omega\rangle=\lvert\Omega\rangle\qquad\text{for all }j. \tag*{(18)} P j ∣ Ω ⟩ = ∣ Ω ⟩ for all j . (18)
Acting with stabilizer group elements generates the orbit states
∣ σ ⟩ : = ∏ k S k ( 1 − σ k ) / 2 ∣ Ω ⟩ , σ k = ± 1. (19) \lvert\sigma\rangle:=\prod_k S_k^{(1-\sigma_k)/2}\lvert\Omega\rangle,\qquad\sigma_k=\pm1. \tag*{(19)} ∣ σ ⟩ := k ∏ S k ( 1 − σ k ) /2 ∣ Ω ⟩ , σ k = ± 1. (19)
Configurations related by elements of R \mathcal{R} R represent the same orbit state and are therefore identified. We assume that the resulting inequivalent orbit states form an orthonormal set. In this basis, we have
P j ∣ σ ⟩ = B j ( σ ) ∣ σ ⟩ , B j ( σ ) = ∏ k σ k A j k . (20) P_j\lvert\sigma\rangle=B_j(\sigma)\lvert\sigma\rangle,\qquad B_j(\sigma)=\prod_k \sigma_k^{A_{jk}}. \tag*{(20)} P j ∣ σ ⟩ = B j ( σ ) ∣ σ ⟩ , B j ( σ ) = k ∏ σ k A j k . (20)
Let C \mathcal{C} C denote the set of inequivalent orbit configurations. We first define the uniform orbit state
∣ Ψ 0 ⟩ = 1 ∣ C ∣ ∑ σ ∈ C ∣ σ ⟩ . (21) \lvert\Psi_0\rangle=\frac{1}{\sqrt{\lvert\mathcal{C}\rvert}}\sum_{\sigma\in\mathcal{C}}\lvert\sigma\rangle. \tag*{(21)} ∣ Ψ 0 ⟩ = ∣ C ∣ 1 σ ∈ C ∑ ∣ σ ⟩ . (21)
We also introduce the classical partition function
Z K : = ∑ σ ∈ C exp [ ∑ j K j B j ( σ ) ] . (22) Z_K:=\sum_{\sigma\in\mathcal{C}}\exp\left[\sum_j K_jB_j(\sigma)\right]. \tag*{(22)} Z K := σ ∈ C ∑ exp [ j ∑ K j B j ( σ ) ] . (22)
Here, K j K_j K j is a dimensionless coupling inherited from the wave function deformation and plays the role of an inverse temperature weighted coupling in the associated clean classical model. The normalized deformed code state is then
∣ Ψ K ⟩ : = ∣ C ∣ Z K exp [ 1 2 ∑ j K j P j ] ∣ Ψ 0 ⟩ = 1 Z K ∑ σ ∈ C exp [ 1 2 ∑ j K j B j ( σ ) ] ∣ σ ⟩ . (23) \lvert\Psi_K\rangle:=\sqrt{\frac{\lvert\mathcal{C}\rvert}{Z_K}}\exp\left[\frac{1}{2}\sum_jK_jP_j\right]\lvert\Psi_0\rangle
=\frac{1}{\sqrt{Z_K}}\sum_{\sigma\in\mathcal{C}}\exp\left[\frac{1}{2}\sum_jK_jB_j(\sigma)\right]\lvert\sigma\rangle. \tag*{(23)} ∣ Ψ K ⟩ := Z K ∣ C ∣ exp [ 2 1 j ∑ K j P j ] ∣ Ψ 0 ⟩ = Z K 1 σ ∈ C ∑ exp [ 2 1 j ∑ K j B j ( σ ) ] ∣ σ ⟩ . (23)
A projective measurement in the orbit basis therefore gives
P K ( σ ) = ∣ ⟨ σ ∣ Ψ K ⟩ ∣ 2 = 1 Z K exp [ ∑ j K j B j ( σ ) ] . (24) P_K(\sigma)=\lvert\langle\sigma\lvert\Psi_K\rangle\rvert^2
=\frac{1}{Z_K}\exp\left[\sum_jK_jB_j(\sigma)\right]. \tag*{(24)} P K ( σ ) = ∣⟨ σ ∣ Ψ K ⟩ ∣ 2 = Z K 1 exp [ j ∑ K j B j ( σ ) ] . (24)
Thus, the Born probabilities of the deformed code state coincide with the Boltzmann distribution of the associated clean classical model.
We now subject the operators P j P_j P j to independent Gaussian weak measurements. For a measurement record m = { m j } \boldsymbol{m}=\{m_j\} m = { m j } , the joint Kraus operator is described by
M m = ∏ j ( Δ j 2 π ) 1 / 4 exp [ − Δ j 4 ( m j − P j ) 2 ] , m j ∈ R . (25) M_{\boldsymbol{m}}=\prod_j\left(\frac{\Delta_j}{2\pi}\right)^{1/4}
\exp\left[-\frac{\Delta_j}{4}(m_j-P_j)^2\right],\qquad m_j\in\mathbb{R}. \tag*{(25)} M m = j ∏ ( 2 π Δ j ) 1/4 exp [ − 4 Δ j ( m j − P j ) 2 ] , m j ∈ R . (25)
The Kraus operators satisfy
∫ d m M m † M m = I , d m : = ∏ j d m j . (26) \int d\boldsymbol{m}\,M_{\boldsymbol{m}}^\dagger M_{\boldsymbol{m}}=\mathbb{I},\qquad d\boldsymbol{m}:=\prod_jdm_j. \tag*{(26)} ∫ d m M m † M m = I , d m := j ∏ d m j . (26)
Conditioned on an orbit configuration σ \boldsymbol{\sigma} σ , the measurement outcomes are independently distributed according to
P Δ ( m ∣ σ ) = ∏ j ( Δ j 2 π ) 1 / 2 exp [ − Δ j 2 ( m j − B j ( σ ) ) 2 ] . (27) P_{\Delta}(\boldsymbol{m} \mid\boldsymbol{\sigma})=\prod_j\left(\frac{\Delta_j}{2\pi}\right)^{1/2}\exp\left[-\frac{\Delta_j}{2}\left(m_j-B_j(\boldsymbol{\sigma})\right)^2\right]. \tag*{(27)} P Δ ( m ∣ σ ) = j ∏ ( 2 π Δ j ) 1/2 exp [ − 2 Δ j ( m j − B j ( σ ) ) 2 ] . (27)
Bayes’ theorem gives the posterior distribution
P ( σ ∣ m ) = 1 Z [ m ] exp [ ∑ j ( K j + Δ j m j ) B j ( σ ) ] , (28) P(\boldsymbol{\sigma} \mid\boldsymbol{m})=\frac{1}{Z[\boldsymbol{m}]}\exp\left[\sum_j\left(K_j+\Delta_jm_j\right)B_j(\boldsymbol{\sigma})\right], \tag*{(28)} P ( σ ∣ m ) = Z [ m ] 1 exp [ j ∑ ( K j + Δ j m j ) B j ( σ ) ] , (28)
where the record dependent partition function is
Z [ m ] : = ∑ σ ∈ C exp [ ∑ j ( K j + Δ j m j ) B j ( σ ) ] . (29) Z[\boldsymbol{m}]:=\sum_{\boldsymbol{\sigma}\in\mathcal{C}}\exp\left[\sum_j\left(K_j+\Delta_jm_j\right)B_j(\boldsymbol{\sigma})\right]. \tag*{(29)} Z [ m ] := σ ∈ C ∑ exp [ j ∑ ( K j + Δ j m j ) B j ( σ ) ] . (29)
In the chosen stabilizer orbit basis, the normalized post measurement state is therefore
∣ Ψ m ⟩ = ∑ σ ∈ C P ( σ ∣ m ) ∣ σ ⟩ . (30) \lvert\Psi_{\boldsymbol{m}}\rangle=\sum_{\boldsymbol{\sigma}\in\mathcal{C}}\sqrt{P(\boldsymbol{\sigma}\mid\boldsymbol{m})}\lvert\boldsymbol{\sigma}\rangle. \tag*{(30)} ∣ Ψ m ⟩ = σ ∈ C ∑ P ( σ ∣ m ) ∣ σ ⟩ . (30)
Consequently, posterior averages of B j ( σ ) B_j(\boldsymbol{\sigma}) B j ( σ ) coincide with quantum expectation values of the corresponding diagonal operators P j P_j P j in ∣ Ψ m ⟩ \lvert\Psi_{\boldsymbol{m}}\rangle ∣ Ψ m ⟩ .
The Born probability density of the measurement record is
P ( m ) = g Δ ( m ) Z K Z [ m ] , (31) P(\boldsymbol{m})=\frac{g_{\Delta}(\boldsymbol{m})}{Z_K}Z[\boldsymbol{m}], \tag*{(31)} P ( m ) = Z K g Δ ( m ) Z [ m ] , (31)
where
g Δ ( m ) : = ∏ j ( Δ j 2 π ) 1 / 2 exp [ − Δ j 2 ( m j 2 + 1 ) ] . (32) g_{\Delta}(\boldsymbol{m}):=\prod_j\left(\frac{\Delta_j}{2\pi}\right)^{1/2}\exp\left[-\frac{\Delta_j}{2}\left(m_j^2+1\right)\right]. \tag*{(32)} g Δ ( m ) := j ∏ ( 2 π Δ j ) 1/2 exp [ − 2 Δ j ( m j 2 + 1 ) ] . (32)
Note that record averaged products of posterior expectation values contain inverse powers of the normalization factor Z [ m ] Z[\boldsymbol{m}] Z [ m ] . For an observable O ( σ ) O(\boldsymbol{\sigma}) O ( σ ) , we denote its posterior expectation at a fixed measurement record by
⟨ O ⟩ m : = ∑ σ ∈ C P ( σ ∣ m ) O ( σ ) . \langle O\rangle_{\boldsymbol{m}}:=\sum_{\boldsymbol{\sigma}\in\mathcal{C}}P(\boldsymbol{\sigma}\mid\boldsymbol{m})O(\boldsymbol{\sigma}). ⟨ O ⟩ m := σ ∈ C ∑ P ( σ ∣ m ) O ( σ ) .
The average over measurement records with their Born probabilities is denoted by an overline, namely,
F ( m ) ‾ : = ∫ d m P ( m ) F ( m ) . (33) \overline{F(\boldsymbol{m})}:=\int d\boldsymbol{m}\,P(\boldsymbol{m})F(\boldsymbol{m}). \tag*{(33)} F ( m ) := ∫ d m P ( m ) F ( m ) . (33)
Introducing R > 0 R>0 R > 0 replicas of the posterior statistical model, we define the replicated generating function as
Z R : = 1 Z K ∫ d m g Δ ( m ) Z [ m ] R , d m : = ∏ j d m j . (34) \mathcal{Z}_R:=\frac{1}{Z_K}\int d\boldsymbol{m}\,g_{\Delta}(\boldsymbol{m})Z[\boldsymbol{m}]^R,\qquad d\boldsymbol{m}:=\prod_jdm_j. \tag*{(34)} Z R := Z K 1 ∫ d m g Δ ( m ) Z [ m ] R , d m := j ∏ d m j . (34)
The physical Born rule average is obtained by analytically continuing R R R to one. This differs from the conventional quenched-disorder replica limit R → 0 R\to0 R → 0 because the Born probability P ( m ) P(\boldsymbol{m}) P ( m ) itself contains one power of Z [ m ] Z[\boldsymbol{m}] Z [ m ] , as shown in (31) . In particular,
Z 1 = ∫ d m P ( m ) = 1. (35) \mathcal{Z}_1=\int d\boldsymbol{m}\,P(\boldsymbol{m})=1. \tag*{(35)} Z 1 = ∫ d m P ( m ) = 1. (35)
For integer R R R , the replicated power can be written as
Z [ m ] R = ∑ { σ ( a ) } a = 1 R exp [ ∑ j ∑ a = 1 R ( K j + Δ j m j ) B j ( σ ( a ) ) ] . (36) Z[\mathbf{m}]^{R}=\sum_{\{\boldsymbol{\sigma}^{(a)}\}_{a=1}^{R}}\exp\left[\sum_j\sum_{a=1}^{R}(K_j+\Delta_jm_j)B_j\left(\boldsymbol{\sigma}^{(a)}\right)\right]. \tag*{(36)} Z [ m ] R = { σ ( a ) } a = 1 R ∑ exp [ j ∑ a = 1 ∑ R ( K j + Δ j m j ) B j ( σ ( a ) ) ] . (36)
Here a = 1 , … , R a=1,\ldots,R a = 1 , … , R is the replica index. For each measurement outcome m j m_j m j , the Gaussian integral is
∫ − ∞ ∞ d m j ( Δ j 2 π ) 1 / 2 exp [ − Δ j 2 ( m j 2 + 1 ) + Δ j m j ∑ a = 1 R B j ( a ) ] = exp [ − Δ j 2 + Δ j 2 ( ∑ a = 1 R B j ( a ) ) 2 ] , (37) \int_{-\infty}^{\infty} \mathrm{d}m_j\left(\frac{\Delta_j}{2\pi}\right)^{1/2}\exp\left[-\frac{\Delta_j}{2}\left(m_j^2+1\right)+\Delta_jm_j\sum_{a=1}^{R}B_j^{(a)}\right]=\exp\left[-\frac{\Delta_j}{2}+\frac{\Delta_j}{2}\left(\sum_{a=1}^{R}B_j^{(a)}\right)^2\right], \tag*{(37)} ∫ − ∞ ∞ d m j ( 2 π Δ j ) 1/2 exp [ − 2 Δ j ( m j 2 + 1 ) + Δ j m j a = 1 ∑ R B j ( a ) ] = exp − 2 Δ j + 2 Δ j ( a = 1 ∑ R B j ( a ) ) 2 , (37)
where B j ( a ) : = B j ( σ ( a ) ) B_j^{(a)}:=B_j\left(\boldsymbol{\sigma}^{(a)}\right) B j ( a ) := B j ( σ ( a ) ) . Substituting (37) into (34) , we obtain
Z R = e − 1 2 ∑ j Δ j Z K ∑ { σ ( a ) } a = 1 R exp [ ∑ j ∑ a = 1 R K j B j ( a ) + 1 2 ∑ j Δ j ( ∑ a = 1 R B j ( a ) ) 2 ] . (38) \mathcal{Z}_{R}=\frac{e^{-\frac{1}{2}\sum_j\Delta_j}}{Z_K}\sum_{\{\boldsymbol{\sigma}^{(a)}\}_{a=1}^{R}}\exp\left[\sum_j\sum_{a=1}^{R}K_jB_j^{(a)}+\frac{1}{2}\sum_j\Delta_j\left(\sum_{a=1}^{R}B_j^{(a)}\right)^2\right]. \tag*{(38)} Z R = Z K e − 2 1 ∑ j Δ j { σ ( a ) } a = 1 R ∑ exp j ∑ a = 1 ∑ R K j B j ( a ) + 2 1 j ∑ Δ j ( a = 1 ∑ R B j ( a ) ) 2 . (38)
Since ( B j ( a ) ) 2 = 1 \left(B_j^{(a)}\right)^2=1 ( B j ( a ) ) 2 = 1 , we have
( ∑ a = 1 R B j ( a ) ) 2 = R + ∑ a , b = 1 a ≠ b R B j ( a ) B j ( b ) . (39) \left(\sum_{a=1}^{R}B_j^{(a)}\right)^2=R+\sum_{\substack{a,b=1\\a\ne b}}^{R}B_j^{(a)}B_j^{(b)}. \tag*{(39)} ( a = 1 ∑ R B j ( a ) ) 2 = R + a , b = 1 a = b ∑ R B j ( a ) B j ( b ) . (39)
Therefore, after dropping terms independent of the spin configurations, the replicated Hamiltonian is given by
− H R = ∑ j ∑ a = 1 R K j B j ( a ) + 1 2 ∑ j Δ j ∑ a , b = 1 a ≠ b R B j ( a ) B j ( b ) . (40) -H_R=\sum_j\sum_{a=1}^{R}K_jB_j^{(a)}+\frac{1}{2}\sum_j\Delta_j\sum_{\substack{a,b=1\\a\ne b}}^{R}B_j^{(a)}B_j^{(b)}. \tag*{(40)} − H R = j ∑ a = 1 ∑ R K j B j ( a ) + 2 1 j ∑ Δ j a , b = 1 a = b ∑ R B j ( a ) B j ( b ) . (40)
We now focus on the parameter manifold
Δ j = K j ∀ j (41) \Delta_j=K_j\ \forall j \tag*{(41)} Δ j = K j ∀ j (41)
which we call the higher Nishimori manifold . This condition matches the coupling K j K_j K j inherited from the wavefunction deformation with the coupling Δ j \Delta_j Δ j between replicas generated by the measurement. It is the analog, in the present learning problem, of the matching between the prior and likelihood on the conventional Nishimori line.
The special role of (41) can already be seen from the replicated Hamiltonian. On this manifold, it can be written, up to a spin-independent constant, as
− H R = 1 2 ∑ j K j [ 1 + ∑ a = 1 R B j ( σ ( a ) ) ] 2 + c o n s t . (42) -H_R=\frac{1}{2}\sum_jK_j\left[1+\sum_{a=1}^{R}B_j\left(\boldsymbol{\sigma}^{(a)}\right)\right]^2+\mathrm{const.} \tag*{(42)} − H R = 2 1 j ∑ K j [ 1 + a = 1 ∑ R B j ( σ ( a ) ) ] 2 + const. (42)
Away from the higher Nishimori manifold, the single replica and inter replica terms have independent coefficients and cannot be combined into this form. There is an additional replica symmetry in (42) . To see this, we introduce an auxiliary Ising configuration s = { s k } \boldsymbol{s}=\{s_k\} s = { s k } , average uniformly over s \boldsymbol{s} s , and perform the change of variables
σ ( a ) ⟼ σ ( a ) s , a = 1 , … , R . (43) \boldsymbol{\sigma}^{(a)}\longmapsto\boldsymbol{\sigma}^{(a)}\boldsymbol{s},\qquad a=1,\ldots,R. \tag*{(43)} σ ( a ) ⟼ σ ( a ) s , a = 1 , … , R . (43)
Using the multiplicative property
B j ( σ s ) = B j ( σ ) B j ( s ) , (44) B_j(\boldsymbol{\sigma}\boldsymbol{s})=B_j(\boldsymbol{\sigma})B_j(\boldsymbol{s}), \tag*{(44)} B j ( σ s ) = B j ( σ ) B j ( s ) , (44)
we find
[ 1 + ∑ a = 1 R B j ( σ ( a ) s ) ] 2 = [ B j ( s ) + ∑ a = 1 R B j ( σ ( a ) ) ] 2 . (45) \left[1+\sum_{a=1}^{R}B_j\left(\boldsymbol{\sigma}^{(a)}\boldsymbol{s}\right)\right]^2=\left[B_j(\boldsymbol{s})+\sum_{a=1}^{R}B_j\left(\boldsymbol{\sigma}^{(a)}\right)\right]^2. \tag*{(45)} [ 1 + a = 1 ∑ R B j ( σ ( a ) s ) ] 2 = [ B j ( s ) + a = 1 ∑ R B j ( σ ( a ) ) ] 2 . (45)
Identifying σ ( R + 1 ) : = s \boldsymbol{\sigma}^{(R+1)} := s σ ( R + 1 ) := s , the replicated partition function can therefore be rewritten, up to an overall normalization, as
Z R ∝ ∑ { σ ( a ) } a = 1 R + 1 exp [ 1 2 ∑ j K j ( ∑ a = 1 R + 1 B j ( σ ( a ) ) ) 2 ] . (46) \mathcal{Z}_R \propto\sum_{\{\boldsymbol{\sigma}^{(a)}\}_{a=1}^{R+1}} \exp\left[\frac{1}{2}\sum_j K_j\left(\sum_{a=1}^{R+1} B_j(\boldsymbol{\sigma}^{(a)})\right)^2\right]. \tag*{(46)} Z R ∝ { σ ( a ) } a = 1 R + 1 ∑ exp 2 1 j ∑ K j ( a = 1 ∑ R + 1 B j ( σ ( a ) ) ) 2 . (46)
This representation has a manifest enlarged replica permutation symmetry S R ⟶ S R + 1 S_R \longrightarrow S_{R+1} S R ⟶ S R + 1 , which treats the auxiliary configuration, interpreted as the truth replica, on the same footing as the R R R posterior replicas. In the next section, we make use of these properties to derive several useful identities.
Higher Nishimori identities
We now use the symmetry derived in § 2 to relate the learnability of different observables. In particular, we ask how learning two observables constrains the learnability of their product. We first review the ordinary Nishimori condition [10 , 12 ] . This will clarify the role of the additional replica before we turn to the weak measurement problem. We then derive an identity involving several observables and their product. Finally, we combine this identity with the Griffiths–Kelly–Sherman (GKS) correlation inequalities [33 , 34 ] in the clean model to obtain a hierarchy of learnability.
Ordinary Nishimori condition
We first review the ordinary Nishimori condition for the classical spin models introduced in § 2 .1 We consider quenched binary disorder τ j = ± 1 \tau_j = \pm1 τ j = ± 1 coupled to the local interactions B j ( σ ) B_j(\boldsymbol{\sigma}) B j ( σ ) ,
H τ ( σ ) = − ∑ j K j τ j B j ( σ ) , (47) H_{\tau}(\boldsymbol{\sigma}) = -\sum_j K_j \tau_j B_j(\boldsymbol{\sigma}), \tag*{(47)} H τ ( σ ) = − j ∑ K j τ j B j ( σ ) , (47)
with independent disorder distribution
P κ ( τ ) = ∏ j e κ j τ j 2 cosh κ j . (48) P_{\boldsymbol{\kappa}}(\boldsymbol{\tau}) = \prod_j \frac{e^{\kappa_j \tau_j}}{2\cosh\kappa_j}. \tag*{(48)} P κ ( τ ) = j ∏ 2 cosh κ j e κ j τ j . (48)
The signs are held fixed when taking a thermal expectation value and averaged over afterwards. We denote these two averages by ⟨ ⋅ ⟩ τ \langle\cdot\rangle_{\tau} ⟨ ⋅ ⟩ τ and [ ⋅ ] τ [\cdot]_{\tau} [ ⋅ ] τ , respectively. If p j = P κ ( τ j = − 1 ) p_j = P_{\boldsymbol{\kappa}}(\tau_j=-1) p j = P κ ( τ j = − 1 ) , then (48) gives κ j = 1 2 log [ ( 1 − p j ) / p j ] \kappa_j = \frac{1}{2}\log[(1-p_j)/p_j] κ j = 2 1 log [( 1 − p j ) / p j ] . The ordinary Nishimori condition sets this disorder parameter equal to the thermal coupling:
κ j = K j ∀ j . (49) \kappa_j = K_j \qquad\forall j. \tag*{(49)} κ j = K j ∀ j . (49)
Averaging the Boltzmann weight of n n n replicas over τ j \tau_j τ j gives the factor cosh [ κ j + K j ∑ a = 1 n B j ( σ ( a ) ) ] / cosh κ j \cosh[\kappa_j+K_j\sum_{a=1}^{n}B_j(\boldsymbol{\sigma}^{(a)})]/\cosh\kappa_j cosh [ κ j + K j ∑ a = 1 n B j ( σ ( a ) )] / cosh κ j . Under the condition (49) , the disorder bias and each replica contribute with the same coefficient. The replicated partition function therefore takes the following form:
[ Z τ n ] τ ∝ ∑ { σ ( a ) } ∏ j cosh [ K j ( 1 + ∑ a = 1 n B j ( σ ( a ) ) ) ] . (50) [Z_{\tau}^{n}]_{\tau} \propto\sum_{\{\boldsymbol{\sigma}^{(a)}\}} \prod_j \cosh\left[K_j\left(1+\sum_{a=1}^{n}B_j(\boldsymbol{\sigma}^{(a)})\right)\right]. \tag*{(50)} [ Z τ n ] τ ∝ { σ ( a ) } ∑ j ∏ cosh [ K j ( 1 + a = 1 ∑ n B j ( σ ( a ) ) ) ] . (50)
Introducing an arbitrary spin configuration s s s and using the evenness of the hyperbolic cosine yields
cosh [ K j ( 1 + ∑ a = 1 n B j ( σ s ( a ) ) ) ] = cosh [ K j ( B j ( s ) + ∑ a = 1 n B j ( σ ( a ) ) ) ] . (51) \cosh\left[K_j\left(1+\sum_{a=1}^{n}B_j(\boldsymbol{\sigma}^{(a)}_s)\right)\right] = \cosh\left[K_j\left(B_j(s)+\sum_{a=1}^{n}B_j(\boldsymbol{\sigma}^{(a)})\right)\right]. \tag*{(51)} cosh [ K j ( 1 + a = 1 ∑ n B j ( σ s ( a ) ) ) ] = cosh [ K j ( B j ( s ) + a = 1 ∑ n B j ( σ ( a ) ) ) ] . (51)
Averaging uniformly over s s s and identifying it with σ ( n + 1 ) \sigma^{(n+1)} σ ( n + 1 ) gives the following replicated theory up to an overall normalization:
[ Z τ n ] ∝ ∑ { σ ( a ) } a = 1 n + 1 ∏ j cosh [ K j ∑ a = 1 n + 1 B j ( σ ( a ) ) ] (52) [Z_{\tau}^{n}] \propto\sum_{\{\sigma^{(a)}\}_{a=1}^{n+1}} \prod_j \cosh\left[K_j\sum_{a=1}^{n+1} B_j(\sigma^{(a)})\right] \tag*{(52)} [ Z τ n ] ∝ { σ ( a ) } a = 1 n + 1 ∑ j ∏ cosh [ K j a = 1 ∑ n + 1 B j ( σ ( a ) ) ] (52)
which exhibits the familiar enlarged permutation symmetry S n → S n + 1 S_n \to S_{n+1} S n → S n + 1 together with the local gauge redundancy
σ k ( a ) ⟶ η k σ k ( a ) , a = 1 , … , n + 1 (53) \sigma_k^{(a)} \longrightarrow\eta_k \sigma_k^{(a)}, \qquad a=1,\ldots,n+1 \tag*{(53)} σ k ( a ) ⟶ η k σ k ( a ) , a = 1 , … , n + 1 (53)
with η k = ± 1 \eta_k=\pm1 η k = ± 1 . For a binary vector v ∈ F 2 N s v\in\mathbb{F}_2^{N_s} v ∈ F 2 N s , we define the spin product
O v ( σ ) : = ∏ k σ k v k . (54) O_v(\sigma):=\prod_k \sigma_k^{v_k}. \tag*{(54)} O v ( σ ) := k ∏ σ k v k . (54)
These observables satisfy O v ( σ s ) = O v ( σ ) O v ( s ) O_v(\sigma s)=O_v(\sigma)O_v(s) O v ( σ s ) = O v ( σ ) O v ( s ) and O v O w = O v ⊕ w O_vO_w=O_{v\oplus w} O v O w = O v ⊕ w , where ⊕ \oplus ⊕ denotes componentwise addition modulo two. Combining these properties with replica permutations and local gauge transformations gives the following identity:
[ ∏ λ = 1 M ⟨ O v λ ⟩ τ ] τ = [ ( ∏ λ = 1 M ⟨ O v λ ⟩ τ ) ⟨ O q ⟩ τ ] τ , q = ⨁ λ = 1 M v λ . (55) \left[\prod_{\lambda=1}^{M}\langle O_{v_\lambda}\rangle_{\tau}\right]_{\tau}
=
\left[\left(\prod_{\lambda=1}^{M}\langle O_{v_\lambda}\rangle_{\tau}\right)\langle O_q\rangle_{\tau}\right]_{\tau},
\qquad q=\bigoplus_{\lambda=1}^{M}v_\lambda. \tag*{(55)} [ λ = 1 ∏ M ⟨ O v λ ⟩ τ ] τ = [ ( λ = 1 ∏ M ⟨ O v λ ⟩ τ ) ⟨ O q ⟩ τ ] τ , q = λ = 1 ⨁ M v λ . (55)
For an odd number of identical insertions, v 1 = ⋯ = v 2 m − 1 = v v_1=\cdots=v_{2m-1}=v v 1 = ⋯ = v 2 m − 1 = v , this reduces to the identity
[ ⟨ O v ⟩ τ 2 m − 1 ] τ = [ ⟨ O v ⟩ τ 2 m ] τ . (56) \left[\langle O_v\rangle_{\tau}^{2m-1}\right]_{\tau}
=
\left[\langle O_v\rangle_{\tau}^{2m}\right]_{\tau}. \tag*{(56)} [ ⟨ O v ⟩ τ 2 m − 1 ] τ = [ ⟨ O v ⟩ τ 2 m ] τ . (56)
The learning problem has a similar replica symmetry, but a different physical interpretation. In the ordinary problem, the additional replica arises from the disorder distribution, and averages over quenched disorder are obtained in the limit n → 0 n\to0 n → 0 . In the learning problem, the additional replica represents the unknown spin configuration, and averages over measurement records are obtained in the limit R → 1 R\to1 R → 1 . We now derive the corresponding identity for learning. Its connection to correlations in the clean model will be established in § 3.3 .
Relation among record averaged correlators
We now return to the learning problem on the higher Nishimori manifold, Δ j = K j \Delta_j=K_j Δ j = K j . We restrict to observables that are invariant under the stabilizer relations. Let O v 1 , … , O v M O_{v_1},\ldots,O_{v_M} O v 1 , … , O v M be M M M such observables, which need not be distinct. We place them in different replicas with indices a 1 , … , a M a_1,\ldots,a_M a 1 , … , a M , and define the corresponding replica correlator:
C R ( v 1 , … , v M ) : = ⟨ ∏ λ = 1 M O v λ ( σ ( a λ ) ) ⟩ R \mathcal{C}_{R}(v_1,\ldots,v_M):=\left\langle\prod_{\lambda=1}^{M}O_{v_\lambda}\left(\sigma^{(a_\lambda)}\right)\right\rangle_R C R ( v 1 , … , v M ) := ⟨ λ = 1 ∏ M O v λ ( σ ( a λ ) ) ⟩ R
Here ⟨ ⋅ ⟩ R \langle\cdot\rangle_R ⟨ ⋅ ⟩ R denotes the normalized expectation value in the replicated model. Explicitly, $2
C R = 1 Z R ∑ { σ ( a ) } ∏ λ = 1 M O v λ ( σ ( a λ ) ) × exp [ 1 2 ∑ j Δ j ( 1 + ∑ a = 1 R B j ( σ ( a ) ) ) 2 ] . (57) \mathcal{C}_{R}=\frac{1}{Z_R}\sum_{\{\sigma^{(a)}\}}\prod_{\lambda=1}^{M}O_{v_\lambda}\left(\sigma^{(a_\lambda)}\right)\times\exp\left[\frac{1}{2}\sum_j\Delta_j\left(1+\sum_{a=1}^{R}B_j\left(\sigma^{(a)}\right)\right)^2\right]. \tag*{(57)} C R = Z R 1 { σ ( a ) } ∑ λ = 1 ∏ M O v λ ( σ ( a λ ) ) × exp 2 1 j ∑ Δ j ( 1 + a = 1 ∑ R B j ( σ ( a ) ) ) 2 . (57)
To proceed, we introduce an arbitrary Ising configuration s = { s k } \mathbf{s}=\{s_k\} s = { s k } and average over all its configurations. Since the original expression is independent of s \mathbf{s} s , we may write
C R = 1 2 N s Z R ∑ s ∑ { σ ( a ) } ∏ λ = 1 M O v λ ( σ ( a λ ) ) × exp [ 1 2 ∑ j Δ j ( 1 + ∑ a = 1 R B j ( σ ( a ) ) ) 2 ] . (58) \mathcal{C}_{R}=\frac{1}{2^{N_s}Z_R}\sum_{\mathbf{s}}\sum_{\{\boldsymbol{\sigma}^{(a)}\}}\prod_{\lambda=1}^{M}O_{\mathbf{v}_{\lambda}}\left(\boldsymbol{\sigma}^{(a_\lambda)}\right)\times\exp\left[\frac{1}{2}\sum_j\Delta_j\left(1+\sum_{a=1}^{R}B_j\left(\boldsymbol{\sigma}^{(a)}\right)\right)^2\right]. \tag*{(58)} C R = 2 N s Z R 1 s ∑ { σ ( a ) } ∑ λ = 1 ∏ M O v λ ( σ ( a λ ) ) × exp 2 1 j ∑ Δ j ( 1 + a = 1 ∑ R B j ( σ ( a ) ) ) 2 . (58)
For each fixed s \mathbf{s} s , make the invertible change of variables σ ~ k ( a ) = s k σ k ( a ) \widetilde{\sigma}^{(a)}_k=s_k\sigma^{(a)}_k σ k ( a ) = s k σ k ( a ) . The multiplication rules for the interactions and observables then give
B j ( σ ~ ( a ) ) = B j ( s ) B j ( σ ( a ) ) , O v λ ( σ ~ ( a λ ) ) = O v λ ( s ) O v λ ( σ ( a λ ) ) . (59) B_j\left(\widetilde{\boldsymbol{\sigma}}^{(a)}\right)=B_j(\mathbf{s})B_j\left(\boldsymbol{\sigma}^{(a)}\right),\qquad O_{\mathbf{v}_{\lambda}}\left(\widetilde{\boldsymbol{\sigma}}^{(a_\lambda)}\right)=O_{\mathbf{v}_{\lambda}}(\mathbf{s})O_{\mathbf{v}_{\lambda}}\left(\boldsymbol{\sigma}^{(a_\lambda)}\right). \tag*{(59)} B j ( σ ( a ) ) = B j ( s ) B j ( σ ( a ) ) , O v λ ( σ ( a λ ) ) = O v λ ( s ) O v λ ( σ ( a λ ) ) . (59)
The product of the observable factors generated by the change of variables is
∏ λ = 1 M O v λ ( s ) = O q ( s ) , (60) \prod_{\lambda=1}^{M}O_{\mathbf{v}_{\lambda}}(\mathbf{s})=O_{\mathbf{q}}(\mathbf{s}), \tag*{(60)} λ = 1 ∏ M O v λ ( s ) = O q ( s ) , (60)
where q : = v 1 ⊕ v 2 ⊕ ⋯ ⊕ v M \mathbf{q}:=\mathbf{v}_1\oplus\mathbf{v}_2\oplus\cdots\oplus\mathbf{v}_M q := v 1 ⊕ v 2 ⊕ ⋯ ⊕ v M . Moreover, since B j ( s ) 2 = 1 B_j(\mathbf{s})^2=1 B j ( s ) 2 = 1 , the Boltzmann weight transforms according to
[ 1 + ∑ a = 1 R B j ( σ ~ ( a ) ) ] 2 = [ 1 + B j ( s ) ∑ a = 1 R B j ( σ ( a ) ) ] 2 = [ B j ( s ) + ∑ a = 1 R B j ( σ ( a ) ) ] 2 . (61) \left[1+\sum_{a=1}^{R}B_j\left(\widetilde{\boldsymbol{\sigma}}^{(a)}\right)\right]^2=\left[1+B_j(\mathbf{s})\sum_{a=1}^{R}B_j\left(\boldsymbol{\sigma}^{(a)}\right)\right]^2=\left[B_j(\mathbf{s})+\sum_{a=1}^{R}B_j\left(\boldsymbol{\sigma}^{(a)}\right)\right]^2. \tag*{(61)} [ 1 + a = 1 ∑ R B j ( σ ( a ) ) ] 2 = [ 1 + B j ( s ) a = 1 ∑ R B j ( σ ( a ) ) ] 2 = [ B j ( s ) + a = 1 ∑ R B j ( σ ( a ) ) ] 2 . (61)
Dropping the tilde on the dummy summation variables, we therefore find
C R = 1 2 N s Z R ∑ s ∑ { σ ( a ) } [ ∏ λ = 1 M O v λ ( σ ( a λ ) ) ] O q ( s ) × exp [ 1 2 ∑ j Δ j ( B j ( s ) + ∑ a = 1 R B j ( σ ( a ) ) ) 2 ] . (62) \mathcal{C}_{R}=\frac{1}{2^{N_s}Z_R}\sum_{\mathbf{s}}\sum_{\{\boldsymbol{\sigma}^{(a)}\}}\left[\prod_{\lambda=1}^{M}O_{\mathbf{v}_{\lambda}}\left(\boldsymbol{\sigma}^{(a_\lambda)}\right)\right]O_{\mathbf{q}}(\mathbf{s})\times\exp\left[\frac{1}{2}\sum_j\Delta_j\left(B_j(\mathbf{s})+\sum_{a=1}^{R}B_j\left(\boldsymbol{\sigma}^{(a)}\right)\right)^2\right]. \tag*{(62)} C R = 2 N s Z R 1 s ∑ { σ ( a ) } ∑ [ λ = 1 ∏ M O v λ ( σ ( a λ ) ) ] O q ( s ) × exp 2 1 j ∑ Δ j ( B j ( s ) + a = 1 ∑ R B j ( σ ( a ) ) ) 2 . (62)
We now identify the auxiliary spin configuration with an additional replica, σ ( R + 1 ) : = s \boldsymbol{\sigma}^{(R+1)}:=\mathbf{s} σ ( R + 1 ) := s . The correlator (62) becomes
C R = 1 2 N s Z R ∑ { σ ( a ) } a = 1 R + 1 [ ∏ λ = 1 M O v λ ( σ ( a λ ) ) ] O q ( σ ( R + 1 ) ) × exp [ 1 2 ∑ j Δ j ( ∑ a = 1 R + 1 B j ( σ ( a ) ) ) 2 ] . (63) \mathcal{C}_{R}=\frac{1}{2^{N_s}Z_R}\sum_{\{\boldsymbol{\sigma}^{(a)}\}_{a=1}^{R+1}}\left[\prod_{\lambda=1}^{M}O_{\mathbf{v}_{\lambda}}\left(\boldsymbol{\sigma}^{(a_\lambda)}\right)\right]O_{\mathbf{q}}\left(\boldsymbol{\sigma}^{(R+1)}\right)\times\exp\left[\frac{1}{2}\sum_j\Delta_j\left(\sum_{a=1}^{R+1}B_j\left(\boldsymbol{\sigma}^{(a)}\right)\right)^2\right]. \tag*{(63)} C R = 2 N s Z R 1 { σ ( a ) } a = 1 R + 1 ∑ [ λ = 1 ∏ M O v λ ( σ ( a λ ) ) ] O q ( σ ( R + 1 ) ) × exp 2 1 j ∑ Δ j ( a = 1 ∑ R + 1 B j ( σ ( a ) ) ) 2 . (63)
The Boltzmann weight in (63) is manifestly invariant under the enlarged replica permutation group S R + 1 S_{R+1} S R + 1 . It is also invariant under the local transformation
σ k ( a ) ⟶ η k σ k ( a ) , a = 1 , … , R + 1 , η k = ± 1. (64) \sigma_k^{(a)}\longrightarrow\eta_k\sigma_k^{(a)},\qquad a=1,\ldots,R+1,\qquad\eta_k=\pm1. \tag*{(64)} σ k ( a ) ⟶ η k σ k ( a ) , a = 1 , … , R + 1 , η k = ± 1. (64)
Indeed, under such a transformation, we have
B j ( σ ( a ) ) ⟶ χ j ( η ) B j ( σ ( a ) ) , (65) B_j\left(\boldsymbol{\sigma}^{(a)}\right)\longrightarrow\chi_j(\boldsymbol{\eta})B_j\left(\boldsymbol{\sigma}^{(a)}\right), \tag*{(65)} B j ( σ ( a ) ) ⟶ χ j ( η ) B j ( σ ( a ) ) , (65)
where
χ j ( η ) : = ∏ k η k A j k = ± 1 , (66) \chi_j(\boldsymbol{\eta}):=\prod_k\eta_k^{A_{jk}}=\pm1, \tag*{(66)} χ j ( η ) := k ∏ η k A j k = ± 1 , (66)
so that [ ∑ a = 1 R + 1 B j ( σ ( a ) ) ] 2 \left[\sum_{a=1}^{R+1}B_j\left(\boldsymbol{\sigma}^{(a)}\right)\right]^2 [ ∑ a = 1 R + 1 B j ( σ ( a ) ) ] 2 is unchanged. Notice also that the complete operator insertion in (63) is gauge invariant. Under the change (64) , it acquires the factor
∏ λ = 1 M O v λ ( η ) O q ( η ) = O q ( η ) 2 = 1. (67) \prod_{\lambda=1}^{M}O_{\mathbf{v}_{\lambda}}(\boldsymbol{\eta})O_{\mathbf{q}}(\boldsymbol{\eta})=O_{\mathbf{q}}(\boldsymbol{\eta})^2=1. \tag*{(67)} λ = 1 ∏ M O v λ ( η ) O q ( η ) = O q ( η ) 2 = 1. (67)
We choose a replica index b b b satisfying b ≠ a 1 , … , a M b \ne a_1,\ldots,a_M b = a 1 , … , a M , 1 ≤ b ≤ R 1 \le b \le R 1 ≤ b ≤ R . For integer R R R sufficiently large such a replica always exists. Using the S R + 1 S_{R+1} S R + 1 symmetry, we exchange replicas b b b and R + 1 R+1 R + 1 . Since replica b b b initially carries no operator insertion, (63) becomes
C R = 1 2 N s Z R ∑ { σ ( a ) } a = 1 R + 1 [ ∏ λ = 1 M O v λ ( σ ( a λ ) ) ] O q ( σ ( b ) ) × exp [ 1 2 ∑ j Δ j ( ∑ a = 1 R + 1 B j ( σ ( a ) ) ) 2 ] . (68) \mathcal{C}_R = \frac{1}{2^{N_s} Z_R} \sum_{\{\sigma^{(a)}\}_{a=1}^{R+1}} \left[\prod_{\lambda=1}^{M} O_{v_\lambda}\left(\sigma^{(a_\lambda)}\right)\right] O_q\left(\sigma^{(b)}\right) \times\exp\left[\frac{1}{2}\sum_j \Delta_j\left(\sum_{a=1}^{R+1} B_j\left(\sigma^{(a)}\right)\right)^2\right]. \tag*{(68)} C R = 2 N s Z R 1 { σ ( a ) } a = 1 R + 1 ∑ [ λ = 1 ∏ M O v λ ( σ ( a λ ) ) ] O q ( σ ( b ) ) × exp 2 1 j ∑ Δ j ( a = 1 ∑ R + 1 B j ( σ ( a ) ) ) 2 . (68)
We finally fix the local gauge by imposing σ k ( R + 1 ) = + 1 ∀ k \sigma_k^{(R+1)}=+1\ \forall k σ k ( R + 1 ) = + 1 ∀ k . Since B j ( 1 ) = 1 B_j(\mathbf{1})=1 B j ( 1 ) = 1 , the Boltzmann weight reduces precisely to the original R R R -replica weight. The sum over the gauge orbit gives a factor 2 N s 2^{N_s} 2 N s , which cancels the explicit factor 2 − N s 2^{-N_s} 2 − N s in (68) . We therefore obtain
⟨ ∏ λ = 1 M O v λ ( a λ ) ⟩ R = ⟨ [ ∏ λ = 1 M O v λ ( a λ ) ] O q ( b ) ⟩ R , (69) \left\langle\prod_{\lambda=1}^{M} O_{v_\lambda}^{(a_\lambda)} \right\rangle_R = \left\langle\left[\prod_{\lambda=1}^{M} O_{v_\lambda}^{(a_\lambda)}\right] O_q^{(b)} \right\rangle_R, \tag*{(69)} ⟨ λ = 1 ∏ M O v λ ( a λ ) ⟩ R = ⟨ [ λ = 1 ∏ M O v λ ( a λ ) ] O q ( b ) ⟩ R , (69)
where q = ⨁ λ = 1 M v λ \mathbf{q}=\bigoplus_{\lambda=1}^{M}\mathbf{v}_\lambda q = ⨁ λ = 1 M v λ and O v ( a ) O_v^{(a)} O v ( a ) is shorthand for O v ( σ ( a ) ) O_v(\sigma^{(a)}) O v ( σ ( a ) ) . The identity applies to distinct observables as well as repeated copies of the same observable. Products of posterior expectation values are represented by insertions into pairwise distinct replicas. In particular,
∏ λ = 1 M ⟨ O v λ ⟩ m = lim R → 1 ⟨ ∏ λ = 1 M O v λ ( a λ ) ⟩ R . (70) \prod_{\lambda=1}^{M}\langle O_{v_\lambda}\rangle_m = \lim_{R\to1}\left\langle\prod_{\lambda=1}^{M} O_{v_\lambda}^{(a_\lambda)} \right\rangle_R. \tag*{(70)} λ = 1 ∏ M ⟨ O v λ ⟩ m = R → 1 lim ⟨ λ = 1 ∏ M O v λ ( a λ ) ⟩ R . (70)
Analytically continuing the relation (69) to the learning limit therefore yields
∏ λ = 1 M ⟨ O v λ ⟩ m ‾ = [ ∏ λ = 1 M ⟨ O v λ ⟩ m ] ⟨ O q ⟩ m ‾ , q = ⨁ λ = 1 M v λ . (71) \overline{\prod_{\lambda=1}^{M}\langle O_{v_\lambda}\rangle_m}=\overline{\left[\prod_{\lambda=1}^{M}\langle O_{v_\lambda}\rangle_m\right]\langle O_q\rangle_m},\qquad\mathbf{q}=\bigoplus_{\lambda=1}^{M}\mathbf{v}_\lambda. \tag*{(71)} λ = 1 ∏ M ⟨ O v λ ⟩ m = [ λ = 1 ∏ M ⟨ O v λ ⟩ m ] ⟨ O q ⟩ m , q = λ = 1 ⨁ M v λ . (71)
For the Gaussian measurements considered here, this is an exact identity. It can also be obtained directly from the joint distribution of the spin configuration and measurement record, without continuing in the replica number. For binary measurements, the corresponding relation is expected at long distances if the model approaches the same higher Nishimori fixed point.
As a consistency check, consider an odd number M = 2 k − 1 M=2k-1 M = 2 k − 1 of identical observables, v 1 = ⋯ = v 2 k − 1 = v v_1=\cdots=v_{2k-1}=v v 1 = ⋯ = v 2 k − 1 = v .
v 1 = ⋯ = v 2 k − 1 = v . (72) v_1=\cdots=v_{2k-1}=v. \tag*{(72)} v 1 = ⋯ = v 2 k − 1 = v . (72)
Since addition is modulo two,
q = v ⊕ ⋯ ⊕ v ⏟ 2 k − 1 t i m e s = v . (73) q=\underbrace{v\oplus\cdots\oplus v}_{2k-1\ \mathrm{times}}=v. \tag*{(73)} q = 2 k − 1 times v ⊕ ⋯ ⊕ v = v . (73)
The relation (71) then reduces to ⟨ O v ⟩ m 2 k − 1 ‾ = ⟨ O v ⟩ m 2 k ‾ \overline{\langle O_v\rangle_m^{2k-1}}=\overline{\langle O_v\rangle_m^{2k}} ⟨ O v ⟩ m 2 k − 1 = ⟨ O v ⟩ m 2 k , which reproduces the higher Nishimori identity [19 ] .
Exact hierarchy of record averaged correlators
We now use (71) to relate learnability to correlations in the clean model. For an observable O v O_v O v , define the average of its squared posterior expectation by
C v ( 2 ) : = ⟨ O v ⟩ m 2 ‾ . (74) \mathcal{C}_v^{(2)}:=\overline{\langle O_v\rangle_m^2}. \tag*{(74)} C v ( 2 ) := ⟨ O v ⟩ m 2 . (74)
When O v O_v O v is a spatial correlation operator, C v ( 2 ) \mathcal{C}_v^{(2)} C v ( 2 ) is its Edwards–Anderson (EA)-type correlator [35 , 36 ] and measures how reliably the corresponding nonlocal information can be inferred from the measurement record.
Taking M = 1 M=1 M = 1 in (71) gives
⟨ O v ⟩ ‾ m = C v ( 2 ) . (75) \overline{\langle O_{\boldsymbol{v}}\rangle}_{m}=\mathcal{C}_{\boldsymbol{v}}^{(2)}. \tag*{(75)} ⟨ O v ⟩ m = C v ( 2 ) . (75)
Averaging a posterior expectation over all possible records recovers the expectation in the original distribution:
⟨ O v ⟩ ‾ m = ∫ d m P ( m ) ∑ σ P ( σ ∣ m ) O v ( σ ) = ∑ σ P K ( σ ) O v ( σ ) = : ⟨ O v ⟩ c l e a n . (76) \overline{\langle O_{\boldsymbol{v}}\rangle}_{m}=\int\mathrm{d}m\,P(m)\sum_{\boldsymbol{\sigma}}P(\boldsymbol{\sigma}\mid m)O_{\boldsymbol{v}}(\boldsymbol{\sigma})=\sum_{\boldsymbol{\sigma}}P_{K}(\boldsymbol{\sigma})O_{\boldsymbol{v}}(\boldsymbol{\sigma})=:\langle O_{\boldsymbol{v}}\rangle_{\mathrm{clean}}. \tag*{(76)} ⟨ O v ⟩ m = ∫ d m P ( m ) σ ∑ P ( σ ∣ m ) O v ( σ ) = σ ∑ P K ( σ ) O v ( σ ) =: ⟨ O v ⟩ clean . (76)
Combining (75) and (76) , we obtain the general equality
C v ( 2 ) = ⟨ O v ⟩ c l e a n . (77) \mathcal{C}_{\boldsymbol{v}}^{(2)}=\langle O_{\boldsymbol{v}}\rangle_{\mathrm{clean}}. \tag*{(77)} C v ( 2 ) = ⟨ O v ⟩ clean . (77)
Therefore, on the higher Nishimori manifold, the average squared posterior expectation is equal to the corresponding correlation in the clean model. Here the clean model is the distribution P K P_K P K associated with the deformed state before conditioning on a measurement record.
We next assume that all couplings K j K_j K j multiplying the spin products B j B_j B j are nonnegative. In this sense, the clean model is ferromagnetic even when its interactions involve more than two spins. The GKS inequalities then give
⟨ O v O w ⟩ c l e a n ≥ ⟨ O v ⟩ c l e a n ⟨ O w ⟩ c l e a n (78) \langle O_{\boldsymbol{v}}O_{\boldsymbol{w}}\rangle_{\mathrm{clean}}\geq\langle O_{\boldsymbol{v}}\rangle_{\mathrm{clean}}\langle O_{\boldsymbol{w}}\rangle_{\mathrm{clean}} \tag*{(78)} ⟨ O v O w ⟩ clean ≥ ⟨ O v ⟩ clean ⟨ O w ⟩ clean (78)
for arbitrary product of spins O v O_{\boldsymbol{v}} O v and O w O_{\boldsymbol{w}} O w [33 , 34 ] (See Appendix § 8 for a review.). In words, the expectation of their product is at least the product of their expectations. Since O v O w = O v ⊕ w O_{\boldsymbol{v}}O_{\boldsymbol{w}}=O_{\boldsymbol{v}\oplus\boldsymbol{w}} O v O w = O v ⊕ w , (77) converts the inequality in the clean model directly into a hierarchy among learning correlators, leading to the following:
Theorem 1 (Exact correlator hierarchy). On the higher Nishimori manifold of a CSS stabilizer code learning problem whose associated clean spin model has nonnegative ferromagnetic couplings,
C v ⊕ w ( 2 ) ≥ C v ( 2 ) C w ( 2 ) . (79) \mathcal{C}_{\boldsymbol{v}\oplus\boldsymbol{w}}^{(2)}\geq\mathcal{C}_{\boldsymbol{v}}^{(2)}\mathcal{C}_{\boldsymbol{w}}^{(2)}. \tag*{(79)} C v ⊕ w ( 2 ) ≥ C v ( 2 ) C w ( 2 ) . (79) More generally, for any collection v 1 , … , v M \boldsymbol{v}_1,\ldots,\boldsymbol{v}_M v 1 , … , v M ,
C ⨁ λ = 1 M v λ ( 2 ) ≥ ∏ λ = 1 M C v λ ( 2 ) . (80) \mathcal{C}_{\bigoplus_{\lambda=1}^{M}\boldsymbol{v}_{\lambda}}^{(2)}\geq\prod_{\lambda=1}^{M}\mathcal{C}_{\boldsymbol{v}_{\lambda}}^{(2)}. \tag*{(80)} C ⨁ λ = 1 M v λ ( 2 ) ≥ λ = 1 ∏ M C v λ ( 2 ) . (80)
The inequality for two observables follows directly from (77) and (78) . Repeating the same argument gives (80) , since each intermediate product is again of the form (54) . These inequalities relate the learnability of different observables to that of their product. The hierarchy immediately constrains long range order. Suppose that O v ( r ) O_{\boldsymbol{v}}(r) O v ( r ) and O w ( r ) O_{\boldsymbol{w}}(r) O w ( r ) are families of correlation operators at separation r r r , and define their product channel by O q ( r ) = O v ( r ) O w ( r ) O_{\boldsymbol{q}}(r)=O_{\boldsymbol{v}}(r)O_{\boldsymbol{w}}(r) O q ( r ) = O v ( r ) O w ( r ) , with q = v ⊕ w \boldsymbol{q}=\boldsymbol{v}\oplus\boldsymbol{w} q = v ⊕ w . If
q v : = lim r → ∞ C v ( 2 ) ( r ) > 0 , q w : = lim r → ∞ C w ( 2 ) ( r ) > 0 , (81) q_{\boldsymbol{v}}:=\lim_{r\to\infty}\mathcal{C}_{\boldsymbol{v}}^{(2)}(r)>0,\qquad q_{\boldsymbol{w}}:=\lim_{r\to\infty}\mathcal{C}_{\boldsymbol{w}}^{(2)}(r)>0, \tag*{(81)} q v := r → ∞ lim C v ( 2 ) ( r ) > 0 , q w := r → ∞ lim C w ( 2 ) ( r ) > 0 , (81)
then
q q ≥ q v q w > 0. (82) q_{\boldsymbol{q}}\geq q_{\boldsymbol{v}}q_{\boldsymbol{w}}>0. \tag*{(82)} q q ≥ q v q w > 0. (82)
Therefore, if both observables remain learnable at long distances, their product must also remain learnable. The inequality does not imply the converse: the product may remain learnable even if one or both of the original observables do not. The same result gives a universal constraint at a continuous transition. Assume that the three second moment correlators have the asymptotic power law behaviors
C a ( 2 ) ( r ) ∼ A a r − 2 x a ( 2 ) , a = v , w , q , A a > 0. (83) \mathcal{C}_a^{(2)}(r) \sim A_a r^{-2x_a^{(2)}}, \qquad a = v,w,q, \qquad A_a > 0. \tag*{(83)} C a ( 2 ) ( r ) ∼ A a r − 2 x a ( 2 ) , a = v , w , q , A a > 0. (83)
The relation (79) can hold at arbitrarily large r r r only if
x v ⊕ w ( 2 ) ≤ x v ( 2 ) + x w ( 2 ) . (84) x_{v\oplus w}^{(2)} \le x_v^{(2)} + x_w^{(2)}. \tag*{(84)} x v ⊕ w ( 2 ) ≤ x v ( 2 ) + x w ( 2 ) . (84)
This bound does not assume a particular universality class and remains valid even when the individual scaling dimensions are not known exactly.
In Table 1 , we summarize the comparison with the ordinary Nishimori problem. In both settings, replica symmetry and the multiplication rules of observables lead to identities among averaged correlators. In the learning problem, (77) further relates these quantities to correlations in the clean model. For nonnegative couplings, the GKS inequalities then give the hierarchy (80) , which constrains the simultaneous learnability of distinct observables.
Table 1. Comparison between the ordinary and higher Nishimori structures.
Examples of CSS codes
In this section, we apply the higher Nishimori identities to concrete stabilizer code models. After reviewing learning problem of the toric code, we apply the general framework to the rank-2 toric code, whose associated classical model possesses modulated symmetries.
Warm-up: Toric code (review)
We first consider the two dimensional toric code with qubits on the links of a square lattice,
H T C = − ∑ v A v − ∑ p B p , (85) H_{\mathrm{TC}} = -\sum_v A_v - \sum_p B_p, \tag*{(85)} H TC = − v ∑ A v − p ∑ B p , (85)
where
A v = ∏ e ∋ v X e , B p = ∏ e ∈ ∂ p Z e . A_v = \prod_{e\ni v} X_e,\qquad B_p = \prod_{e\in\partial p} Z_e. A v = e ∋ v ∏ X e , B p = e ∈ ∂ p ∏ Z e .
For definiteness, let us consider single-link Z Z Z errors. A Z Z Z operator on a link e = ⟨ v v ′ ⟩ e = \langle vv'\rangle e = ⟨ v v ′ ⟩ anticommutes with the two star stabilizers located at the endpoints v v v and v ′ v' v ′ . Following the construction in § 2.1 , we assign an Ising spin σ v = ± 1 \sigma_v = \pm1 σ v = ± 1 to each X-type stabilizer. The classical interaction associated with the physical link e e e is then
b e ( σ ) = σ v σ v ′ . (86) b_e(\boldsymbol{\sigma}) = \sigma_v \sigma_{v'}. \tag*{(86)} b e ( σ ) = σ v σ v ′ . (86)
Thus, the classical model associated with this error sector is the ordinary nearest-neighbor Ising model,
H I s i n g [ σ ] = − J ∑ ⟨ v v ′ ⟩ σ v σ v ′ . (87) H_{\mathrm{Ising}}[\boldsymbol{\sigma}] = -J \sum_{\langle vv'\rangle} \sigma_v \sigma_{v'}. \tag*{(87)} H Ising [ σ ] = − J ⟨ v v ′ ⟩ ∑ σ v σ v ′ . (87)
Possible global relations among the star stabilizers only produce an overall redundancy and will not affect the following local argument.
We now regard the Boltzmann distribution of the clean model as a prior distribution,
P β ( σ ) = 1 Z β exp [ β J ∑ e b e ( σ ) ] . (88) P_{\beta}(\boldsymbol{\sigma}) = \frac{1}{Z_{\beta}} \exp\left[\beta J \sum_e b_e(\boldsymbol{\sigma})\right]. \tag*{(88)} P β ( σ ) = Z β 1 exp [ β J e ∑ b e ( σ ) ] . (88)
In the quantum problem, the parameter β \beta β can arise from a wavefunction deformation of the corresponding toric code state. We perform a weak Gaussian measurement of every bond variable b e = ± 1 b_e = \pm1 b e = ± 1 . For a fixed spin configuration σ \boldsymbol{\sigma} σ , the conditional probability for obtaining a continuous measurement outcome m e m_e m e is
P Δ ( m e ∣ σ ) ∝ exp [ − Δ 2 ( m e − b e ( σ ) ) 2 ] . (89) P_{\Delta}(m_e \mid\boldsymbol{\sigma}) \propto\exp\left[-\frac{\Delta}{2}\left(m_e-b_e(\boldsymbol{\sigma})\right)^2\right]. \tag*{(89)} P Δ ( m e ∣ σ ) ∝ exp [ − 2 Δ ( m e − b e ( σ ) ) 2 ] . (89)
The measurement outcomes therefore define a classical Bayesian inference problem. By Bayes’ theorem, the posterior distribution is
P ( σ ∣ m ) = 1 Z [ m ] exp [ β J ∑ e b e ( σ ) + Δ ∑ e m e b e ( σ ) ] , (90) P(\boldsymbol{\sigma}\mid\boldsymbol{m}) = \frac{1}{Z[\boldsymbol{m}]} \exp\left[\beta J \sum_e b_e(\boldsymbol{\sigma})+\Delta\sum_e m_e b_e(\boldsymbol{\sigma})\right], \tag*{(90)} P ( σ ∣ m ) = Z [ m ] 1 exp [ β J e ∑ b e ( σ ) + Δ e ∑ m e b e ( σ ) ] , (90)
where terms depending only on m \boldsymbol{m} m have been absorbed into the normalization.
To average products of posterior expectation values over measurement records, we introduce R R R replicas. After integrating over the Gaussian measurement outcomes, the replicated model is
− H R = β J ∑ e ∑ a = 1 R b e ( σ ( a ) ) + Δ 2 ∑ e ∑ a , b = 1 a ≠ b R b e ( σ ( a ) ) b e ( σ ( b ) ) , (91) -H_R = \beta J \sum_e \sum_{a=1}^{R} b_e\left(\boldsymbol{\sigma}^{(a)}\right) + \frac{\Delta}{2}\sum_e \sum_{\substack{a,b=1\\a\ne b}}^{R} b_e\left(\boldsymbol{\sigma}^{(a)}\right)b_e\left(\boldsymbol{\sigma}^{(b)}\right), \tag*{(91)} − H R = β J e ∑ a = 1 ∑ R b e ( σ ( a ) ) + 2 Δ e ∑ a , b = 1 a = b ∑ R b e ( σ ( a ) ) b e ( σ ( b ) ) , (91)
up to an additive constant. The physical learning problem is obtained in the replica limit R → 1 R \to1 R → 1 . A special simplification occurs on the line, Δ = β J \Delta= \beta J Δ = β J . On this line, (91) can be written as
− H R = Δ 2 ∑ e [ 1 + ∑ a = 1 R b e ( σ ( a ) ) ] 2 + c o n s t . (92) -H_R = \frac{\Delta}{2}\sum_e \left[1+\sum_{a=1}^{R} b_e\left(\boldsymbol{\sigma}^{(a)}\right)\right]^2+\mathrm{const.} \tag*{(92)} − H R = 2 Δ e ∑ [ 1 + a = 1 ∑ R b e ( σ ( a ) ) ] 2 + const. (92)
To expose the enhanced symmetry, introduce arbitrary Ising variables s v = ± 1 s_v = \pm1 s v = ± 1 and perform the change of variables
σ ~ v ( a ) = s v σ v ( a ) . (93) \widetilde{\sigma}_v^{(a)} = s_v \sigma_v^{(a)}. \tag*{(93)} σ v ( a ) = s v σ v ( a ) . (93)
Since
b e ( σ ~ ( a ) ) = s v s v ′ b e ( σ ( a ) ) , b_e\left(\widetilde{\boldsymbol{\sigma}}^{(a)}\right) = s_v s_{v'} b_e\left(\boldsymbol{\sigma}^{(a)}\right), b e ( σ ( a ) ) = s v s v ′ b e ( σ ( a ) ) ,
we have
[ 1 + ∑ a = 1 R b e ( σ ~ ( a ) ) ] 2 = [ s v s v ′ + ∑ a = 1 R b e ( σ ( a ) ) ] 2 . (94) \left[1+\sum_{a=1}^{R} b_e\left(\widetilde{\boldsymbol{\sigma}}^{(a)}\right)\right]^2 = \left[s_v s_{v'}+\sum_{a=1}^{R} b_e\left(\boldsymbol{\sigma}^{(a)}\right)\right]^2. \tag*{(94)} [ 1 + a = 1 ∑ R b e ( σ ( a ) ) ] 2 = [ s v s v ′ + a = 1 ∑ R b e ( σ ( a ) ) ] 2 . (94)
We may therefore promote the auxiliary variables to an additional replica, i.e., σ v ( R + 1 ) : = s v \sigma_v^{(R+1)} := s_v σ v ( R + 1 ) := s v . After summing over { s v } \{s_v\} { s v } , the partition function becomes
Z R = 1 2 N v ∑ { σ ( a ) } a = 1 R + 1 exp [ Δ 2 ∑ e ( ∑ a = 1 R + 1 b e ( σ ( a ) ) ) 2 ] . (95) Z_R = \frac{1}{2^{N_v}} \sum_{\{\sigma^{(a)}\}_{a=1}^{R+1}} \exp\left[\frac{\Delta}{2}\sum_e\left(\sum_{a=1}^{R+1} b_e\left(\sigma^{(a)}\right)\right)^2\right]. \tag*{(95)} Z R = 2 N v 1 { σ ( a ) } a = 1 R + 1 ∑ exp 2 Δ e ∑ ( a = 1 ∑ R + 1 b e ( σ ( a ) ) ) 2 . (95)
The formulation in (95) has an enlarged replica permutation symmetry S R + 1 S_{R+1} S R + 1 . In addition, it has a local gauge invariance
σ v ( a ) ⟶ η v σ v ( a ) , a = 1 , … , R + 1 , η v = ± 1. (96) \sigma_v^{(a)} \longrightarrow\eta_v\sigma_v^{(a)}, \qquad a=1,\ldots,R+1,\qquad\eta_v=\pm1. \tag*{(96)} σ v ( a ) ⟶ η v σ v ( a ) , a = 1 , … , R + 1 , η v = ± 1. (96)
Indeed, every b e ( σ ( a ) ) b_e\left(\sigma^{(a)}\right) b e ( σ ( a ) ) acquires the same factor η v η v ′ \eta_v\eta_{v'} η v η v ′ for all replicas, which disappears upon squaring the replica sum. The line Δ = β J \Delta=\beta J Δ = β J is therefore a higher Nishimori line: the R → 1 R\to1 R → 1 learning problem admits an equivalent gauge invariant formulation with R + 1 → 2 R+1\to2 R + 1 → 2 replicas and enlarged S R + 1 S_{R+1} S R + 1 symmetry [19 ] .
As a first consequence, consider the spin correlation
O v v ′ ( σ ) = σ v σ v ′ . (97) O_{vv'}(\sigma)=\sigma_v\sigma_{v'}. \tag*{(97)} O v v ′ ( σ ) = σ v σ v ′ . (97)
On the higher Nishimori line, the standard Nishimori argument gives
⟨ O v v ′ ⟩ m 2 q − 1 ‾ = ⟨ O v v ′ ⟩ m 2 q ‾ , q = 1 , 2 , … (98) \overline{\langle O_{vv'}\rangle_m^{2q-1}}=\overline{\langle O_{vv'}\rangle_m^{2q}},\qquad q=1,2,\ldots \tag*{(98)} ⟨ O v v ′ ⟩ m 2 q − 1 = ⟨ O v v ′ ⟩ m 2 q , q = 1 , 2 , … (98)
In particular,
⟨ O v v ′ ⟩ m 2 ‾ = ⟨ O v v ′ ⟩ m ‾ . (99) \overline{\langle O_{vv'}\rangle_m^2}=\overline{\langle O_{vv'}\rangle_m}. \tag*{(99)} ⟨ O v v ′ ⟩ m 2 = ⟨ O v v ′ ⟩ m . (99)
Furthermore, the law of total expectation gives
⟨ O v v ′ ⟩ m ‾ = ∫ D m P ( m ) ∑ σ P ( σ ∣ m ) O v v ′ ( σ ) = ∑ σ P β ( σ ) O v v ′ ( σ ) = ⟨ σ v σ v ′ ⟩ c l e a n . (100) \overline{\langle O_{vv'}\rangle_m}=\int DmP(m)\sum_{\sigma}P(\sigma\mid m)O_{vv'}(\sigma)=\sum_{\sigma}P_\beta(\sigma)O_{vv'}(\sigma)=\langle\sigma_v\sigma_{v'}\rangle_{\mathrm{clean}}. \tag*{(100)} ⟨ O v v ′ ⟩ m = ∫ D m P ( m ) σ ∑ P ( σ ∣ m ) O v v ′ ( σ ) = σ ∑ P β ( σ ) O v v ′ ( σ ) = ⟨ σ v σ v ′ ⟩ clean . (100)
Consequently,
⟨ σ v σ v ′ ⟩ m 2 ‾ = ⟨ σ v σ v ′ ⟩ c l e a n . (101) \overline{\langle\sigma_v\sigma_{v'}\rangle_m^2}=\langle\sigma_v\sigma_{v'}\rangle_{\mathrm{clean}}. \tag*{(101)} ⟨ σ v σ v ′ ⟩ m 2 = ⟨ σ v σ v ′ ⟩ clean . (101)
For the ordinary square-lattice Ising model, this reproduces the higher Nishimori identity discussed in [19 ] .
Rank-2 toric code and dipole Ising model
We next consider the Z 2 \mathbb{Z}_2 Z 2 rank-2 toric code [22 ] .3 We first introduce the model and its loop operators, and then apply the identities derived above.
Model and loop operators
The model contains three qubits, labeled by a = 0 , 1 , 2 a=0,1,2 a = 0 , 1 , 2 per unit cell. The type-0 qubit is located at the center of each plaquette, whereas the type-1 and type-2 qubits are located at the vertices. There are three types of mutually commuting stabilizers. The X-type stabilizer is given by
G r = X 1 , r − x ^ X 1 , r + x ^ X 2 , r − y ^ X 2 , r + y ^ ∏ ϵ x , ϵ y = ± 1 X 0 , r + ( ϵ x x ^ + ϵ y y ^ ) / 2 . (102) G_r=X_{1,r-\hat{x}}X_{1,r+\hat{x}}X_{2,r-\hat{y}}X_{2,r+\hat{y}}\prod_{\epsilon_x,\epsilon_y=\pm1}X_{0,r+(\epsilon_x\hat{x}+\epsilon_y\hat{y})/2}. \tag*{(102)} G r = X 1 , r − x ^ X 1 , r + x ^ X 2 , r − y ^ X 2 , r + y ^ ϵ x , ϵ y = ± 1 ∏ X 0 , r + ( ϵ x x ^ + ϵ y y ^ ) /2 . (102)
The two Z-type stabilizers can be written as
B l x = Z 2 , r Z 2 , r + x ^ Z 0 , r + x ^ / 2 + y ^ / 2 Z 0 , r + x ^ / 2 − y ^ / 2 , B l y = Z 1 , r Z 1 , r + y ^ Z 0 , r + x ^ / 2 + y ^ / 2 Z 0 , r − x ^ / 2 + y ^ / 2 . (103) B_{l_x}=Z_{2,r}Z_{2,r+\hat{x}}Z_{0,r+\hat{x}/2+\hat{y}/2}Z_{0,r+\hat{x}/2-\hat{y}/2},
\qquad
B_{l_y}=Z_{1,r}Z_{1,r+\hat{y}}Z_{0,r+\hat{x}/2+\hat{y}/2}Z_{0,r-\hat{x}/2+\hat{y}/2}. \tag*{(103)} B l x = Z 2 , r Z 2 , r + x ^ Z 0 , r + x ^ /2 + y ^ /2 Z 0 , r + x ^ /2 − y ^ /2 , B l y = Z 1 , r Z 1 , r + y ^ Z 0 , r + x ^ /2 + y ^ /2 Z 0 , r − x ^ /2 + y ^ /2 . (103)
where l x = r + x ^ / 2 l_x=r+\hat{x}/2 l x = r + x ^ /2 , l y = r + y ^ / 2 l_y=r+\hat{y}/2 l y = r + y ^ /2 . The Hamiltonian is defined by
H = − ∑ r G r − ∑ l x B l x − ∑ l y B l y . (104) H=-\sum_r G_r-\sum_{l_x}B_{l_x}-\sum_{l_y}B_{l_y}. \tag*{(104)} H = − r ∑ G r − l x ∑ B l x − l y ∑ B l y . (104)
See also Figure 1 .
Figure 1. Stabilizers of the rank-2 toric code (104) . The yellow star, blue triangle, and red square represent Pauli operators acting on qubits of types 1, 2, and 0, respectively.
This model admits three types of loop operators consisting of Pauli Z Z Z operators (Figure 2 ). Let C C C be the rectangular contour with corners ( x L , y B ) (x_L,y_B) ( x L , y B ) , ( x R , y B ) (x_R,y_B) ( x R , y B ) , ( x R , y T ) (x_R,y_T) ( x R , y T ) , and ( x L , y T ) (x_L,y_T) ( x L , y T ) , where
Figure 2. Three types of loop operators defined in (105) . The three loops are related via lattice translation as indicated in (106) . The yellow star, blue triangle, and red square represent Pauli Z1 _1 1 , Z2 _2 2 , and Z0 _0 0 , respectively.
x L = x 0 , x R = x 0 + 2 r x , y B = y 0 , y T = y 0 + 2 r y . x_L=x_0,\qquad x_R=x_0+2r_x,\qquad y_B=y_0,\qquad y_T=y_0+2r_y. x L = x 0 , x R = x 0 + 2 r x , y B = y 0 , y T = y 0 + 2 r y .
and x 0 , y 0 ∈ 2 Z x_0,y_0\in2\mathbb{Z} x 0 , y 0 ∈ 2 Z . The three loop operators supported on and immediately adjacent to C C C are defined as
W D ( C ) = ∏ x ^ = x L x R ( Z 1 , ( x ^ , y B ) ) x ^ × ∏ y ^ = y B y T ( Z 2 , ( x R , y ^ ) ) y ^ × ∏ x ^ = x L x R ( Z 1 , ( x ^ , y T ) ) x ^ × ∏ y ^ = y B y T ( Z 2 , ( x L , y ^ ) ) y ^ , W_D(C)=\prod_{\hat{x}=x_L}^{x_R}\left(Z_{1,(\hat{x},y_B)}\right)^{\hat{x}}\times\prod_{\hat{y}=y_B}^{y_T}\left(Z_{2,(x_R,\hat{y})}\right)^{\hat{y}}\times\prod_{\hat{x}=x_L}^{x_R}\left(Z_{1,(\hat{x},y_T)}\right)^{\hat{x}}\times\prod_{\hat{y}=y_B}^{y_T}\left(Z_{2,(x_L,\hat{y})}\right)^{\hat{y}}, W D ( C ) = x ^ = x L ∏ x R ( Z 1 , ( x ^ , y B ) ) x ^ × y ^ = y B ∏ y T ( Z 2 , ( x R , y ^ ) ) y ^ × x ^ = x L ∏ x R ( Z 1 , ( x ^ , y T ) ) x ^ × y ^ = y B ∏ y T ( Z 2 , ( x L , y ^ ) ) y ^ ,
W 1 ( C ) = ∏ x ^ = x L x R [ Z 1 , ( x ^ , y B ) Z 1 , ( x ^ , y T ) ] × ∏ y ^ = y B y T − 1 [ Z 0 , ( x L − 1 2 , y ^ + 1 2 ) Z 0 , ( x R + 1 2 , y ^ + 1 2 ) ] , W_1(C)=\prod_{\hat{x}=x_L}^{x_R}\left[Z_{1,(\hat{x},y_B)}Z_{1,(\hat{x},y_T)}\right]\times\prod_{\hat{y}=y_B}^{y_T-1}\left[Z_{0,(x_L-\frac{1}{2},\hat{y}+\frac{1}{2})}Z_{0,(x_R+\frac{1}{2},\hat{y}+\frac{1}{2})}\right], W 1 ( C ) = x ^ = x L ∏ x R [ Z 1 , ( x ^ , y B ) Z 1 , ( x ^ , y T ) ] × y ^ = y B ∏ y T − 1 [ Z 0 , ( x L − 2 1 , y ^ + 2 1 ) Z 0 , ( x R + 2 1 , y ^ + 2 1 ) ] ,
W 2 ( C ) = ∏ y ^ = y B y T [ Z 2 , ( x L , y ^ ) Z 2 , ( x R , y ^ ) ] × ∏ x ^ = x L x R − 1 [ Z 0 , ( x ^ + 1 2 , y B − 1 2 ) Z 0 , ( x ^ + 1 2 , y T + 1 2 ) ] . (105) W_2(C)=\prod_{\hat{y}=y_B}^{y_T}\left[Z_{2,(x_L,\hat{y})}Z_{2,(x_R,\hat{y})}\right]\times\prod_{\hat{x}=x_L}^{x_R-1}\left[Z_{0,(\hat{x}+\frac{1}{2},y_B-\frac{1}{2})}Z_{0,(\hat{x}+\frac{1}{2},y_T+\frac{1}{2})}\right]. \tag*{(105)} W 2 ( C ) = y ^ = y B ∏ y T [ Z 2 , ( x L , y ^ ) Z 2 , ( x R , y ^ ) ] × x ^ = x L ∏ x R − 1 [ Z 0 , ( x ^ + 2 1 , y B − 2 1 ) Z 0 , ( x ^ + 2 1 , y T + 2 1 ) ] . (105)
Here, the exponents in W D W_D W D are understood modulo two.
These loops are related via lattice translation. Indeed, within the fluxless subspace, B l x = B l y = + 1 B_{l_x}=B_{l_y}=+1 B l x = B l y = + 1 , after some algebra, one finds [39 ]
T x W D ( C ) T x − 1 = W D ( C ) W 1 ( C ) , T y W D ( C ) T y − 1 = W D ( C ) W 2 ( C ) . (106) T_xW_D(C)T_x^{-1}=W_D(C)W_1(C),
\qquad
T_yW_D(C)T_y^{-1}=W_D(C)W_2(C). \tag*{(106)} T x W D ( C ) T x − 1 = W D ( C ) W 1 ( C ) , T y W D ( C ) T y − 1 = W D ( C ) W 2 ( C ) . (106)
Equivalently, using W D 2 = 1 W_D^2=1 W D 2 = 1 ,
W 1 ( C ) = W D ( C ) T x W D ( C ) T x − 1 , W_1(C)=W_D(C)T_xW_D(C)T_x^{-1}, W 1 ( C ) = W D ( C ) T x W D ( C ) T x − 1 ,
W 2 ( C ) = W D ( C ) T y W D ( C ) T y − 1 . W_2(C)=W_D(C)T_yW_D(C)T_y^{-1}. W 2 ( C ) = W D ( C ) T y W D ( C ) T y − 1 .
Note that equality holds after projection onto the common + 1 +1 + 1 eigenspace of all Z-type stabilizers. Hence, the product of two translated modulated dipole loops gives a homogeneous charge loop. Relations in (106) are the loop-operator analogue of the transformation of a dipole moment under translations. For
D i = ∫ d 2 r r i ρ ( r ) , Q = ∫ d 2 r ρ ( r ) , D_i=\int d^2r\,r_i\rho(\boldsymbol{r}),\qquad Q=\int d^2r\,\rho(\boldsymbol{r}), D i = ∫ d 2 r r i ρ ( r ) , Q = ∫ d 2 r ρ ( r ) ,
a translation by one lattice spacing changes D i D_i D i by the total charge Q Q Q . Correspondingly, a lattice translation transforms the dipole loop into itself multiplied by a homogeneous charge loop. We therefore refer to W D W_D W D as the dipole loop and to W 1 W_1 W 1 and W 2 W_2 W 2 as the associated homogeneous charge, or monopole, loops [23 , 24 , 25 ] .
Learning problem
We now formulate the learning problem for this code. Based on the argument in § 2.1 , a single Z 1 , r Z_{1,r} Z 1 , r error anticommutes with G r − x ^ G_{r-\hat{x}} G r − x ^ and G r + x ^ G_{r+\hat{x}} G r + x ^ , and therefore gives the classical interaction
B 1 , r = σ r − x ^ σ r + x ^ . (107) B_{1,r}=\sigma_{r-\hat{x}}\sigma_{r+\hat{x}}. \tag*{(107)} B 1 , r = σ r − x ^ σ r + x ^ . (107)
Likewise, a single Z 2 , r Z_{2,r} Z 2 , r error anticommutes with G r − y ^ G_{r-\hat{y}} G r − y ^ and G r + y ^ G_{r+\hat{y}} G r + y ^ , giving the classical interaction
B 2 , r = σ r − y ^ σ r + y ^ . (108) B_{2,r}=\sigma_{r-\hat{y}}\sigma_{r+\hat{y}}. \tag*{(108)} B 2 , r = σ r − y ^ σ r + y ^ . (108)
Finally, for a type-0 qubit on the plaquette Z 0 , r + ( x ^ + y ^ ) / 2 Z_{0,r+(\hat{x}+\hat{y})/2} Z 0 , r + ( x ^ + y ^ ) /2 anticommutes with G r G_r G r , G r + x ^ G_{r+\hat{x}} G r + x ^ , G r + y ^ G_{r+\hat{y}} G r + y ^ , and G r + x ^ + y ^ G_{r+\hat{x}+\hat{y}} G r + x ^ + y ^ which gives the following classical interaction:
B 0 , r = σ r σ r + x ^ σ r + y ^ σ r + x ^ + y ^ . (109) B_{0,r}=\sigma_r\sigma_{r+\hat{x}}\sigma_{r+\hat{y}}\sigma_{r+\hat{x}+\hat{y}}. \tag*{(109)} B 0 , r = σ r σ r + x ^ σ r + y ^ σ r + x ^ + y ^ . (109)
Thus, up to a shift of the lattice coordinates, the statistical model associated with the Z-error sector of the rank-2 toric code is
H d i p = − ∑ r [ J x σ r σ r + 2 x ^ + J y σ r σ r + 2 y ^ + J □ σ r σ r + x ^ σ r + y ^ σ r + x ^ + y ^ ] . (110) H_{\mathrm{dip}}=-\sum_r\left[J_x\sigma_r\sigma_{r+2\hat{x}}+J_y\sigma_r\sigma_{r+2\hat{y}}+J_{\square}\sigma_r\sigma_{r+\hat{x}}\sigma_{r+\hat{y}}\sigma_{r+\hat{x}+\hat{y}}\right]. \tag*{(110)} H dip = − r ∑ [ J x σ r σ r + 2 x ^ + J y σ r σ r + 2 y ^ + J □ σ r σ r + x ^ σ r + y ^ σ r + x ^ + y ^ ] . (110)
The stabilizer relations of the rank-2 toric code become spatially modulated symmetries of the associated classical model. Explicitly, assuming the infinite lattice system, the Hamiltonian (110) respects the following modulated symmetry:
σ r → η ξ x r x ξ y r y σ r , η , ξ x , ξ y = ± 1. (111) \sigma_r\to\eta\,\xi_x^{r_x}\xi_y^{r_y}\sigma_r,\qquad\eta,\xi_x,\xi_y=\pm1. \tag*{(111)} σ r → η ξ x r x ξ y r y σ r , η , ξ x , ξ y = ± 1. (111)
Setting ζ x = − 1 \zeta_x=-1 ζ x = − 1 or ζ y = − 1 \zeta_y=-1 ζ y = − 1 , with the other parameters equal to + 1 +1 + 1 , generates the x x x - or y y y -modulated Z 2 \mathbb{Z}_2 Z 2 transformation, respectively. Setting η = − 1 \eta=-1 η = − 1 , with ζ x = ζ y = + 1 \zeta_x=\zeta_y=+1 ζ x = ζ y = + 1 , gives the uniform spin flip.
We weakly measure the three types of qubits of the rank-2 toric code with Gaussian measurement strengths Δ x , Δ y , Δ □ \Delta_x,\Delta_y,\Delta_{\square} Δ x , Δ y , Δ □ . After averaging over the measurement outcomes, the R R R -replica theory takes the form
− H R = ∑ r ∑ μ = x , y , □ [ K μ ∑ a = 1 R b r μ , ( a ) + Δ μ 2 ∑ a , b = 1 a ≠ b R b r μ , ( a ) b r μ , ( b ) ] . (112) -H_R=\sum_r\sum_{\mu=x,y,\square}\left[K_\mu\sum_{a=1}^{R}b_r^{\mu,(a)}+\frac{\Delta_\mu}{2}\sum_{\substack{a,b=1\\a\ne b}}^{R}b_r^{\mu,(a)}b_r^{\mu,(b)}\right]. \tag*{(112)} − H R = r ∑ μ = x , y , □ ∑ K μ a = 1 ∑ R b r μ , ( a ) + 2 Δ μ a , b = 1 a = b ∑ R b r μ , ( a ) b r μ , ( b ) . (112)
where K μ : = β J μ K_\mu:=\beta J_\mu K μ := β J μ .
The higher Nishimori condition is now a manifold in the coupling space,
Δ μ = K μ , μ = x , y , □ . (113) \Delta_\mu=K_\mu,\qquad\mu=x,y,\square. \tag*{(113)} Δ μ = K μ , μ = x , y , □ . (113)
On this manifold, we have
− H R = 1 2 ∑ r , μ K μ [ 1 + ∑ a = 1 R b r μ , ( a ) ] 2 + c o n s t . (114) -H_R=\frac{1}{2}\sum_{r,\mu}K_\mu\left[1+\sum_{a=1}^{R}b_r^{\mu,(a)}\right]^2+\mathrm{const.} \tag*{(114)} − H R = 2 1 r , μ ∑ K μ [ 1 + a = 1 ∑ R b r μ , ( a ) ] 2 + const. (114)
The key observation is that all three interactions satisfy the same multiplicative property. For an arbitrary Ising configuration s \boldsymbol{s} s ,
b r x ( σ s ) = b r x ( σ ) b r x ( s ) , (115) b_r^x(\boldsymbol{\sigma}\boldsymbol{s})=b_r^x(\boldsymbol{\sigma})b_r^x(\boldsymbol{s}), \tag*{(115)} b r x ( σ s ) = b r x ( σ ) b r x ( s ) , (115)
where ( σ s ) r : = σ r s r (\boldsymbol{\sigma}\boldsymbol{s})_r:=\sigma_rs_r ( σ s ) r := σ r s r . The same property holds for b r y ( σ s ) b_r^y(\boldsymbol{\sigma}\boldsymbol{s}) b r y ( σ s ) and b r □ ( σ s ) b_r^{\square}(\boldsymbol{\sigma}\boldsymbol{s}) b r □ ( σ s ) . Consequently, precisely the same auxiliary-spin construction as in the ordinary Ising model promotes s \boldsymbol{s} s to an ( R + 1 ) (R+1) ( R + 1 ) th replica. We obtain
Z R = 1 2 N s ∑ { σ ( a ) } a = 1 R + 1 exp [ 1 2 ∑ r , μ K μ ( ∑ a = 1 R + 1 b r μ ( σ ( a ) ) ) 2 ] . (116) Z_R=\frac{1}{2^{N_s}}\sum_{\{\boldsymbol{\sigma}^{(a)}\}_{a=1}^{R+1}}\exp\left[\frac{1}{2}\sum_{r,\mu}K_\mu\left(\sum_{a=1}^{R+1}b_r^\mu\left(\boldsymbol{\sigma}^{(a)}\right)\right)^2\right]. \tag*{(116)} Z R = 2 N s 1 { σ ( a ) } a = 1 R + 1 ∑ exp 2 1 r , μ ∑ K μ ( a = 1 ∑ R + 1 b r μ ( σ ( a ) ) ) 2 . (116)
This formulation again has S R + 1 S_{R+1} S R + 1 replica permutation symmetry as well as position dependent local gauge invariance
σ r ( a ) ⟶ η r σ r ( a ) , a = 1 , … , R + 1. (117) \sigma_r^{(a)}\longrightarrow\eta_r\sigma_r^{(a)},\qquad a=1,\ldots,R+1. \tag*{(117)} σ r ( a ) ⟶ η r σ r ( a ) , a = 1 , … , R + 1. (117)
with η r = ± 1 \eta_r=\pm1 η r = ± 1 . On the higher Nishimori manifold, the replicated theory possesses an enlarged position-dependent gauge redundancy.
We now introduce the three learning correlators associated with the three loop operators of the rank-2 toric code. To do so, we think of an open string of the dipole loop W D W_D W D which has the following form
∏ x ^ = x ^ L x R ( Z 1 , ( x ^ , y B ) ) x ^ × ∏ y ^ = y B y T ( Z 2 , ( x R , y ^ ) ) y ^ . (118) \prod_{\hat{x}=\hat{x}_L}^{x_R}\left(Z_{1,(\hat{x},y_B)}\right)^{\hat{x}}\times\prod_{\hat{y}=y_B}^{y_T}\left(Z_{2,(x_R,\hat{y})}\right)^{\hat{y}}. \tag*{(118)} x ^ = x ^ L ∏ x R ( Z 1 , ( x ^ , y B ) ) x ^ × y ^ = y B ∏ y T ( Z 2 , ( x R , y ^ ) ) y ^ . (118)
A single Z 1 , r Z_{1,r} Z 1 , r operator maps to the interaction B 1 , r B_{1,r} B 1 , r in (107) , while Z 2 , r Z_{2,r} Z 2 , r maps to B 2 , r B_{2,r} B 2 , r in (108) . Consequently, the classical image of the open dipole loop (118) is described by
∏ x ^ = x ^ L x R ( B 1 , ( x ^ , y B ) ) x ^ × ∏ y ^ = y B y T ( B 2 , ( x R , y ^ ) ) y ^ = σ ( x 0 , y B ) σ ( x R , y T ) . (119) \prod_{\hat{x}=\hat{x}_L}^{x_R}\left(B_{1,(\hat{x},y_B)}\right)^{\hat{x}}\times\prod_{\hat{y}=y_B}^{y_T}\left(B_{2,(x_R,\hat{y})}\right)^{\hat{y}}=\sigma_{(x_0,y_B)}\sigma_{(x_R,y_T)}. \tag*{(119)} x ^ = x ^ L ∏ x R ( B 1 , ( x ^ , y B ) ) x ^ × y ^ = y B ∏ y T ( B 2 , ( x R , y ^ ) ) y ^ = σ ( x 0 , y B ) σ ( x R , y T ) . (119)
More generally, we write such an interaction as
O D ( r , r ′ ) : = σ r σ r ′ . (120) O_D(\boldsymbol{r},\boldsymbol{r}'):=\sigma_{\boldsymbol{r}}\sigma_{\boldsymbol{r}'}. \tag*{(120)} O D ( r , r ′ ) := σ r σ r ′ . (120)
with r − r ′ ∈ ( 2 Z ) 2 \mathbf{r}-\mathbf{r}' \in(2\mathbb{Z})^2 r − r ′ ∈ ( 2 Z ) 2 (∣ r − r ′ ∣ ≫ 1 |\mathbf{r}-\mathbf{r}'| \gg1 ∣ r − r ′ ∣ ≫ 1 ), so that r \mathbf{r} r and r ′ \mathbf{r}' r ′ belong to the same parity sublattice. Likewise, we define classical spin interactions corresponding to open loops W 1 W_1 W 1 and W 2 W_2 W 2 as
O 1 ( r , r ′ ) : = σ r σ r + x ^ σ r ′ σ r ′ + x ^ , (121) O_1(\mathbf{r},\mathbf{r}') := \sigma_{\mathbf{r}}\sigma_{\mathbf{r}+\hat{x}}\sigma_{\mathbf{r}'}\sigma_{\mathbf{r}'+\hat{x}}, \tag*{(121)} O 1 ( r , r ′ ) := σ r σ r + x ^ σ r ′ σ r ′ + x ^ , (121)
O 2 ( r , r ′ ) : = σ r σ r + y ^ σ r ′ σ r ′ + y ^ . (122) O_2(\mathbf{r},\mathbf{r}') := \sigma_{\mathbf{r}}\sigma_{\mathbf{r}+\hat{y}}\sigma_{\mathbf{r}'}\sigma_{\mathbf{r}'+\hat{y}}. \tag*{(122)} O 2 ( r , r ′ ) := σ r σ r + y ^ σ r ′ σ r ′ + y ^ . (122)
Note that the following relation holds
O a ( r , r ′ ) = O D ( r , r ′ ) O D ( a ) ( r , r ′ ) ( a = 1 , 2 ) (123) O_a(\mathbf{r},\mathbf{r}')=O_D(\mathbf{r},\mathbf{r}')O_D^{(a)}(\mathbf{r},\mathbf{r}') \quad(a=1,2) \tag*{(123)} O a ( r , r ′ ) = O D ( r , r ′ ) O D ( a ) ( r , r ′ ) ( a = 1 , 2 ) (123)
where
O D ( a ) ( r , r ′ ) : = σ r + e a σ r ′ + e a , a = 1 , 2. (124) O_D^{(a)}(\mathbf{r},\mathbf{r}'):=\sigma_{\mathbf{r}+\mathbf{e}_a}\sigma_{\mathbf{r}'+\mathbf{e}_a},\qquad a=1,2. \tag*{(124)} O D ( a ) ( r , r ′ ) := σ r + e a σ r ′ + e a , a = 1 , 2. (124)
which is obtained by translation of O D ( r , r ′ ) O_D(\mathbf{r},\mathbf{r}') O D ( r , r ′ ) in the x x x - or y y y -direction. Using these interactions, we further define their EA-type learning correlators by
C D ( r , r ′ ) : = ⟨ O D ( r , r ′ ) ⟩ 2 ‾ m , C 1 ( r , r ′ ) : = ⟨ O 1 ( r , r ′ ) ⟩ 2 ‾ m , C 2 ( r , r ′ ) : = ⟨ O 2 ( r , r ′ ) ⟩ 2 ‾ m . (125) \mathcal{C}_D(\mathbf{r},\mathbf{r}'):=\overline{\langle O_D(\mathbf{r},\mathbf{r}')\rangle^2}_{m},\qquad
\mathcal{C}_1(\mathbf{r},\mathbf{r}'):=\overline{\langle O_1(\mathbf{r},\mathbf{r}')\rangle^2}_{m},\qquad
\mathcal{C}_2(\mathbf{r},\mathbf{r}'):=\overline{\langle O_2(\mathbf{r},\mathbf{r}')\rangle^2}_{m}. \tag*{(125)} C D ( r , r ′ ) := ⟨ O D ( r , r ′ ) ⟩ 2 m , C 1 ( r , r ′ ) := ⟨ O 1 ( r , r ′ ) ⟩ 2 m , C 2 ( r , r ′ ) := ⟨ O 2 ( r , r ′ ) ⟩ 2 m . (125)
On the higher Nishimori manifold, the generalized Nishimori identity gives
C α ( r , r ′ ) = ⟨ O α ( r , r ′ ) ⟩ c l e a n , α = D , 1 , 2. (126) \mathcal{C}_\alpha(\mathbf{r},\mathbf{r}')=\langle O_\alpha(\mathbf{r},\mathbf{r}')\rangle_{\mathrm{clean}},\qquad\alpha=D,1,2. \tag*{(126)} C α ( r , r ′ ) = ⟨ O α ( r , r ′ ) ⟩ clean , α = D , 1 , 2. (126)
Because the clean dipole Ising model has nonnegative ferromagnetic couplings, the GKS inequality can be applied to O D O_D O D and its translated copies. For a = 1 , 2 a=1,2 a = 1 , 2 , it gives
C a ( r , r ′ ) = ⟨ O D ( r , r ′ ) O D ( a ) ( r , r ′ ) ⟩ c l e a n ≥ ⟨ O D ( r , r ′ ) ⟩ c l e a n ⟨ O D ( a ) ( r , r ′ ) ⟩ c l e a n . (127) \mathcal{C}_a(\mathbf{r},\mathbf{r}')=\left\langle O_D(\mathbf{r},\mathbf{r}')O_D^{(a)}(\mathbf{r},\mathbf{r}')\right\rangle_{\mathrm{clean}}\geq\left\langle O_D(\mathbf{r},\mathbf{r}')\right\rangle_{\mathrm{clean}}\left\langle O_D^{(a)}(\mathbf{r},\mathbf{r}')\right\rangle_{\mathrm{clean}}. \tag*{(127)} C a ( r , r ′ ) = ⟨ O D ( r , r ′ ) O D ( a ) ( r , r ′ ) ⟩ clean ≥ ⟨ O D ( r , r ′ ) ⟩ clean ⟨ O D ( a ) ( r , r ′ ) ⟩ clean . (127)
Translation invariance implies
⟨ O D ( a ) ( r , r ′ ) ⟩ c l e a n = ⟨ O D ( r , r ′ ) ⟩ c l e a n . (128) \left\langle O_D^{(a)}(\mathbf{r},\mathbf{r}')\right\rangle_{\mathrm{clean}}=\left\langle O_D(\mathbf{r},\mathbf{r}')\right\rangle_{\mathrm{clean}}. \tag*{(128)} ⟨ O D ( a ) ( r , r ′ ) ⟩ clean = ⟨ O D ( r , r ′ ) ⟩ clean . (128)
Consequently,
C a ( r , r ′ ) ≥ [ C D ( r , r ′ ) ] 2 , a = 1 , 2. (129) \mathcal{C}_a(\mathbf{r},\mathbf{r}')\geq\left[\mathcal{C}_D(\mathbf{r},\mathbf{r}')\right]^2,\qquad a=1,2. \tag*{(129)} C a ( r , r ′ ) ≥ [ C D ( r , r ′ ) ] 2 , a = 1 , 2. (129)
This is the exact dipole–monopole hierarchy for the rank-2 toric code learning problem. In particular, long range learnability of the spatially modulated dipole loop W D W_D W D necessarily implies long range learnability of both homogeneous monopole loops W 1 W_1 W 1 and W 2 W_2 W 2 . The converse implication does not follow from the inequality: the homogeneous loops may remain learnable even when the dipole loop is not. If the correlators exhibit power law scaling,
C D ( r ) ∼ A D r − 2 h D , C a ( r ) ∼ A a r − 2 h a , a = 1 , 2 , (130) \mathcal{C}_D(r)\sim A_Dr^{-2h_D},\qquad\mathcal{C}_a(r)\sim A_ar^{-2h_a},\qquad a=1,2, \tag*{(130)} C D ( r ) ∼ A D r − 2 h D , C a ( r ) ∼ A a r − 2 h a , a = 1 , 2 , (130)
with positive amplitudes, the bound (129) implies h a ≤ 2 h D h_a\leq2h_D h a ≤ 2 h D . This explicit relation hints at the possibility of diagnosing and constraining the criticality of general learning problems in future studies. In previous works on learning problems [5 , 19 , 21 ] , conformal field theory (CFT) descriptions have been established on the higher Nishimori manifold. It would be interesting to investigate whether we have criticality of the present learning problem. We come back to this issue for a future work.
In Appendix § 9 , we discuss the corresponding learning problem with weak measurements of Pauli X X X operators. In this case, we have the complementary hierarchy: two distinct modulated parent loops share a common homogeneous loop. This reversal relative to the Pauli-Z Z Z measurement case reflects the opposite action of lattice translations on the electric and magnetic loop operators of the rank-2 toric code [39 ] . Also, in Appendix § 10 , we apply our results to the X-cube model [26 ] , whose classical representation possesses planar subsystem symmetries.
Learning problem in finite group quantum doubles
So far, we have studied the learning problem in Abelian topological stabilizer codes. In this section, we extend the discussion to quantum double D ( G ) D(G) D ( G ) for a finite group G G G [1 ] , which may admit non-Abelian anyons4 . The pure electric charges are labeled by irreducible representations of G G G and can be non-Abelian when G G G is non-Abelian. We consider Gaussian weak measurements of local operators that create pairs of these charges. We first establish an exact identity between the EA correlators of open electric strings and the corresponding clean correlators. Under an additional positivity condition on the classical interactions, this identity leads to inequalities relating the learnability of different electric charges.
Quantum double model and weak measurements
We consider an oriented square lattice with a finite, connected set of vertices V V V and edges E E E . Each oriented edge e e e carries a ∣ G ∣ |G| ∣ G ∣ -dimensional Hilbert space. The many-body Hilbert space is given by
H e = span { ∣ u ⟩ e : u ∈ G } , H = ⨂ e H e . (131) \mathcal{H}_e = \operatorname{span}\{ \lvert u\rangle_e : u \in G\}, \qquad\mathcal{H} = \bigotimes_e \mathcal{H}_e. \tag*{(131)} H e = span {∣ u ⟩ e : u ∈ G } , H = e ⨂ H e . (131)
For h ∈ G h \in G h ∈ G , we define L h ∣ u ⟩ e = ∣ h u ⟩ e L_h\lvert u\rangle_e = \lvert hu\rangle_e L h ∣ u ⟩ e = ∣ h u ⟩ e and R h ∣ u ⟩ e = ∣ u h ⟩ e R_h\lvert u\rangle_e = \lvert uh\rangle_e R h ∣ u ⟩ e = ∣ u h ⟩ e . The operator A v ( h ) A_v(h) A v ( h ) acts with L h L_h L h on edges oriented away from v v v and with R h − 1 R_{h^{-1}} R h − 1 on edges oriented toward v v v . The vertex projector is A v = ∣ G ∣ − 1 ∑ h ∈ G A v ( h ) A_v = |G|^{-1}\sum_{h\in G} A_v(h) A v = ∣ G ∣ − 1 ∑ h ∈ G A v ( h ) . The plaquette projector B p B_p B p imposes trivial holonomy around each plaquette. Since G G G need not be Abelian, we multiply the edge variables in the order along ∂ p \partial p ∂ p :
B p ∣ { u e } ⟩ = δ G ( ∏ e ∈ ∂ p → u e ε ( p , e ) , 1 G ) ∣ { u e } ⟩ . (132) B_p\lvert\{u_e\}\rangle= \delta_G\left(\overrightarrow{\prod_{e\in\partial p}}u_e^{\varepsilon(p,e)},1_G\right)\lvert\{u_e\}\rangle. \tag*{(132)} B p ∣{ u e }⟩ = δ G e ∈ ∂ p ∏ u e ε ( p , e ) , 1 G ∣{ u e }⟩ . (132)
Here, 1 G 1_G 1 G denotes the identity element of G G G , and ε ( p , e ) = ± 1 \varepsilon(p,e)=\pm1 ε ( p , e ) = ± 1 specifies whether the orientation of e e e agrees with that of the plaquette boundary. Also, δ G ( a , b ) \delta_G(a,b) δ G ( a , b ) is one for a = b a=b a = b and zero otherwise. The Hamiltonian is given by
H D ( G ) = − ∑ v A v − ∑ p B p . (133) H_{D(G)}=-\sum_v A_v-\sum_p B_p. \tag*{(133)} H D ( G ) = − v ∑ A v − p ∑ B p . (133)
An anyon of the quantum double D ( G ) D(G) D ( G ) is labeled by a pair ( [ g ] , π ) ([g],\pi) ([ g ] , π ) , where [ g ] [g] [ g ] is a conjugacy class of G G G and π ∈ Irr ( C G ( g ) ) \pi\in\operatorname{Irr}(C_G(g)) π ∈ Irr ( C G ( g )) is an irreducible representation of the centralizer C G ( g ) = { h ∈ G : h g = g h } C_G(g)=\{h\in G: hg=gh\} C G ( g ) = { h ∈ G : h g = g h } [1 ] . The pure electric sector has trivial flux, [ g ] = [ 1 G ] [g]=[1_G] [ g ] = [ 1 G ] . Since C G ( 1 G ) = G C_G(1_G)=G C G ( 1 G ) = G , a pure electric charge is labeled by ( [ 1 G ] , ρ ) ([1_G],\rho) ([ 1 G ] , ρ ) with ρ ∈ Irr ( G ) \rho\in\operatorname{Irr}(G) ρ ∈ Irr ( G ) . In the following, we omit the trivial flux label and denote pure electric charges by ρ \rho ρ . Their fusion rules are given by
( [ 1 G ] , λ ) ⊗ ( [ 1 G ] , η ) = ⨁ ν ∈ Irr ( G ) N λ η ν ( [ 1 G ] , ν ) , N λ η ν = dim Hom G ( ν , λ ⊗ η ) . (134) ([1_G],\lambda)\otimes([1_G],\eta)=\bigoplus_{\nu\in\operatorname{Irr}(G)}N_{\lambda\eta}^{\nu}([1_G],\nu), \qquad N_{\lambda\eta}^{\nu}=\dim\operatorname{Hom}_G(\nu,\lambda\otimes\eta). \tag*{(134)} ([ 1 G ] , λ ) ⊗ ([ 1 G ] , η ) = ν ∈ Irr ( G ) ⨁ N λ η ν ([ 1 G ] , ν ) , N λ η ν = dim Hom G ( ν , λ ⊗ η ) . (134)
We work in the zero-flux sector with trivial global holonomies. Introducing a group element g i ∈ G g_i\in G g i ∈ G at each vertex, we write the physical edge configurations as
u i j = g i g j − 1 ( e = ( i , j ) oriented from i to j ) . (135) u_{ij}=g_i g_j^{-1}\qquad(e=(i,j)\text{ oriented from }i\text{ to }j). \tag*{(135)} u ij = g i g j − 1 ( e = ( i , j ) oriented from i to j ) . (135)
The simultaneous right multiplication g i ↦ g i k g_i\mapsto g_i k g i ↦ g i k leaves all edge variables unchanged. Let L \mathcal{L} L denote the set of distinct pure-gauge edge configurations, and write ∣ u ⟩ = ⨂ e ∣ u e ⟩ e \lvert u\rangle=\bigotimes_e\lvert u_e\rangle_e ∣ u ⟩ = ⨂ e ∣ u e ⟩ e for the corresponding many-body basis state. The normalized ground state in this sector is the equal-amplitude superposition
∣ Ψ 0 ⟩ = 1 ∣ C ∣ ∑ u ∈ C ∣ u ⟩ . (136) \lvert\Psi_0\rangle= \frac{1}{\sqrt{|\mathcal{C}|}}\sum_{u\in\mathcal{C}}\lvert u\rangle. \tag*{(136)} ∣ Ψ 0 ⟩ = ∣ C ∣ 1 u ∈ C ∑ ∣ u ⟩ . (136)
For later convenience, we define the normalized sum over the vertex variables as
∫ D g F ( g ) : = 1 ∣ G ∣ ∣ V ∣ ∑ g ∈ G V F ( g ) . (137) \int\mathcal{D}_g F(g) := \frac{1}{|G|^{|V|}}\sum_{g\in G^V} F(g). \tag*{(137)} ∫ D g F ( g ) := ∣ G ∣ ∣ V ∣ 1 g ∈ G V ∑ F ( g ) . (137)
On a connected graph, exactly ∣ G ∣ |G| ∣ G ∣ choices of the vertex variables represent each pure-gauge edge configuration. These choices are related by g i ↦ g i k g_i\mapsto g_i k g i ↦ g i k , and the resulting factor cancels in normalized averages of functions of the edge variables. In particular, ∣ C ∣ = ∣ G ∣ ∣ V ∣ − 1 |\mathcal{C}|=|G|^{|V|-1} ∣ C ∣ = ∣ G ∣ ∣ V ∣ − 1 .
Let ρ ∈ Irr ( G ) \rho\in\operatorname{Irr}(G) ρ ∈ Irr ( G ) be a unitary irreducible representation of dimension d ρ d_\rho d ρ , with representation matrix D ρ ( u ) D^\rho(u) D ρ ( u ) . We define the following diagonal operators on a single edge:
T ^ a b , e ρ ∣ u ⟩ e = D a b ρ ( u ) d ρ ∣ u ⟩ e , a , b = 1 , … , d ρ . (138) \widehat{T}^{\rho}_{ab,e}\lvert u\rangle_e = \frac{D^\rho_{ab}(u)}{\sqrt{d_\rho}}\lvert u\rangle_e,\qquad a,b=1,\ldots,d_\rho. \tag*{(138)} T ab , e ρ ∣ u ⟩ e = d ρ D ab ρ ( u ) ∣ u ⟩ e , a , b = 1 , … , d ρ . (138)
When acting on the ground state, these operators create electric charge–anticharge pairs at the endpoints of the edge. The indices a , b a,b a , b label the representation components at the two endpoints. Note that these operators are not projectors onto a definite charge sector.
We weakly measure their real and imaginary parts. To this end, we collect all matrix components into the real vector
f ρ ( u ) : = 1 d ρ ( Re D a b ρ ( u ) , Im D a b ρ ( u ) ) a , b = 1 d ρ ∈ R n ρ , n ρ = 2 d ρ 2 . (139) f_\rho(u):=\frac{1}{\sqrt{d_\rho}}\left(\operatorname{Re}D^\rho_{ab}(u),\operatorname{Im}D^\rho_{ab}(u)\right)_{a,b=1}^{d_\rho}\in\mathbb{R}^{n_\rho},\qquad n_\rho=2d_\rho^2. \tag*{(139)} f ρ ( u ) := d ρ 1 ( Re D ab ρ ( u ) , Im D ab ρ ( u ) ) a , b = 1 d ρ ∈ R n ρ , n ρ = 2 d ρ 2 . (139)
The index A = 1 , … , n ρ A=1,\ldots,n_\rho A = 1 , … , n ρ labels a matrix element ( a , b ) (a,b) ( a , b ) together with its real or imaginary component. We use the Euclidean inner product x ⋅ y = ∑ A x A y A x\cdot y=\sum_A x_Ay_A x ⋅ y = ∑ A x A y A and the squared norm ∥ x ∥ 2 = ∑ A x A 2 \lVert x\rVert^2=\sum_A x_A^2 ∥ x ∥ 2 = ∑ A x A 2 . From the unitarity of D ρ D^\rho D ρ , we obtain
f ρ ( u ) ⋅ f ρ ( v ) = 1 d ρ Re ∑ a , b D a b ρ ( u ) D a b ρ ( v ) ∗ = 1 d ρ Re Tr [ D ρ ( u ) D ρ ( v ) † ] = Re χ ρ ( u v − 1 ) d ρ . (140) f_\rho(u)\cdot f_\rho(v)=\frac{1}{d_\rho}\operatorname{Re}\sum_{a,b}D^\rho_{ab}(u)D^\rho_{ab}(v)^*=\frac{1}{d_\rho}\operatorname{Re}\operatorname{Tr}\left[D^\rho(u)D^\rho(v)^\dagger\right]=\frac{\operatorname{Re}\chi_\rho(uv^{-1})}{d_\rho}. \tag*{(140)} f ρ ( u ) ⋅ f ρ ( v ) = d ρ 1 Re a , b ∑ D ab ρ ( u ) D ab ρ ( v ) ∗ = d ρ 1 Re Tr [ D ρ ( u ) D ρ ( v ) † ] = d ρ Re χ ρ ( u v − 1 ) . (140)
In particular, we have
∥ f ρ ( u ) ∥ 2 = 1 , c ρ ( u ) : = f ρ ( 1 G ) ⋅ f ρ ( u ) = Re χ ρ ( u ) d ρ . (141) \lVert f_\rho(u)\rVert^2=1,\qquad c_\rho(u):=f_\rho(1_G)\cdot f_\rho(u)=\frac{\operatorname{Re}\chi_\rho(u)}{d_\rho}. \tag*{(141)} ∥ f ρ ( u ) ∥ 2 = 1 , c ρ ( u ) := f ρ ( 1 G ) ⋅ f ρ ( u ) = d ρ Re χ ρ ( u ) . (141)
We choose a set S ⊆ Irr ( G ) \mathcal{S}\subseteq\operatorname{Irr}(G) S ⊆ Irr ( G ) of nontrivial representations to be measured. For each pair ( e , ρ ) (e,\rho) ( e , ρ ) , the measurement record is a real vector m e , ρ ∈ R n ρ m_{e,\rho}\in\mathbb{R}^{n_\rho} m e , ρ ∈ R n ρ . All components are measured with the same Gaussian precision Δ e , ρ > 0 \Delta_{e,\rho}>0 Δ e , ρ > 0 , which may depend on the edge and the representation. The corresponding Hermitian operators are given by
f ^ ρ , A ( u ^ e ) : = ∑ u ∈ G f ρ , A ( u ) ∣ u ⟩ e ⟨ u ∣ . (142) \widehat{f}_{\rho,A}(\widehat{u}_e):=\sum_{u\in G}f_{\rho,A}(u)\lvert u\rangle_e\langle u\rvert. \tag*{(142)} f ρ , A ( u e ) := u ∈ G ∑ f ρ , A ( u ) ∣ u ⟩ e ⟨ u ∣ . (142)
These operators are diagonal in the group basis and therefore commute. The joint Kraus operator is the product of the Gaussian Kraus operators for the individual components:
M e , ρ ( m ) = ∏ A = 1 n ρ ( Δ e , ρ 2 π ) 1 / 4 exp [ − Δ e , ρ 4 ( m A − f ^ ρ , A ) 2 ] = ( Δ e , ρ 2 π ) n ρ / 4 exp [ − Δ e , ρ 4 ∑ A = 1 n ρ ( m A − f ^ ρ , A ) 2 ] . (143) M_{e,\rho}(m)=\prod_{A=1}^{n_\rho}\left(\frac{\Delta_{e,\rho}}{2\pi}\right)^{1/4}\exp\left[-\frac{\Delta_{e,\rho}}{4}(m_A-\widehat{f}_{\rho,A})^2\right]=\left(\frac{\Delta_{e,\rho}}{2\pi}\right)^{n_\rho/4}\exp\left[-\frac{\Delta_{e,\rho}}{4}\sum_{A=1}^{n_\rho}(m_A-\widehat{f}_{\rho,A})^2\right]. \tag*{(143)} M e , ρ ( m ) = A = 1 ∏ n ρ ( 2 π Δ e , ρ ) 1/4 exp [ − 4 Δ e , ρ ( m A − f ρ , A ) 2 ] = ( 2 π Δ e , ρ ) n ρ /4 exp [ − 4 Δ e , ρ A = 1 ∑ n ρ ( m A − f ρ , A ) 2 ] . (143)
The sum in the exponent is the squared Euclidean norm ∥ m − f ^ ρ ∥ 2 \lVert m-\hat{f}_{\rho}\rVert^2 ∥ m − f ^ ρ ∥ 2 , taken componentwise. For a fixed group basis state ∣ u ⟩ e \lvert u\rangle_e ∣ u ⟩ e , we have f ^ ρ , A ∣ u ⟩ e = f ρ , A ( u ) ∣ u ⟩ e \hat{f}_{\rho,A}\lvert u\rangle_e=f_{\rho,A}(u)\lvert u\rangle_e f ^ ρ , A ∣ u ⟩ e = f ρ , A ( u ) ∣ u ⟩ e . The Born rule then gives the conditional probability density for the measurement record:
p Δ ( m ∣ u ) = ⟨ u ∣ M e , ρ ( m ) † M e , ρ ( m ) ∣ u ⟩ e = ( Δ e , ρ 2 π ) n ρ / 2 exp [ − Δ e , ρ 2 ∥ m − f ρ ( u ) ∥ 2 ] . (144) p_{\Delta}(m\mid u)=\langle u\lvert M_{e,\rho}(m)^{\dagger}M_{e,\rho}(m)\rvert u\rangle_e=\left(\frac{\Delta_{e,\rho}}{2\pi}\right)^{n_{\rho}/2}\exp\left[-\frac{\Delta_{e,\rho}}{2}\lVert m-f_{\rho}(u)\rVert^2\right]. \tag*{(144)} p Δ ( m ∣ u ) = ⟨ u ∣ M e , ρ ( m ) † M e , ρ ( m )∣ u ⟩ e = ( 2 π Δ e , ρ ) n ρ /2 exp [ − 2 Δ e , ρ ∥ m − f ρ ( u ) ∥ 2 ] . (144)
For fixed u u u , the record components are independent Gaussian variables with means f ρ , A ( u ) f_{\rho,A}(u) f ρ , A ( u ) and variances 1 / Δ e , ρ 1/\Delta_{e,\rho} 1/ Δ e , ρ . Thus, ∫ d n ρ m p Δ ( m ∣ u ) = 1 \int d^{n_{\rho}}m\,p_{\Delta}(m\mid u)=1 ∫ d n ρ m p Δ ( m ∣ u ) = 1 for every u u u , and the Kraus operators satisfy the completeness relation ∫ d n ρ m M e , ρ † M e , ρ = I \int d^{n_{\rho}}m\,M_{e,\rho}^{\dagger}M_{e,\rho}=I ∫ d n ρ m M e , ρ † M e , ρ = I .
As in the CSS code examples, we deform the initial state before applying these measurements:
∣ Ψ K ⟩ = 1 Z K exp [ 1 2 ∑ e , ρ ∈ S K e , ρ c ρ ( u ^ e ) ] ∣ Ψ 0 ⟩ , Z K = ∫ D g exp [ ∑ e , ρ ∈ S K e , ρ c ρ ( u e ) ] . (145) \lvert\Psi_K\rangle=\frac{1}{\sqrt{Z_K}}\exp\left[\frac{1}{2}\sum_{e,\rho\in\mathcal{S}}K_{e,\rho}c_{\rho}(\hat{u}_e)\right]\lvert\Psi_0\rangle,\qquad Z_K=\int\mathcal{D}g\,\exp\left[\sum_{e,\rho\in\mathcal{S}}K_{e,\rho}c_{\rho}(u_e)\right]. \tag*{(145)} ∣ Ψ K ⟩ = Z K 1 exp 2 1 e , ρ ∈ S ∑ K e , ρ c ρ ( u ^ e ) ∣ Ψ 0 ⟩ , Z K = ∫ D g exp e , ρ ∈ S ∑ K e , ρ c ρ ( u e ) . (145)
The diagonal probabilities of this state define the clean prior
P K ( g ) = 1 Z K exp [ ∑ e , ρ ∈ S K e , ρ c ρ ( u e ) ] , ∫ D g P K ( g ) = 1. (146) P_K(g)=\frac{1}{Z_K}\exp\left[\sum_{e,\rho\in\mathcal{S}}K_{e,\rho}c_{\rho}(u_e)\right],\qquad\int\mathcal{D}g\,P_K(g)=1. \tag*{(146)} P K ( g ) = Z K 1 exp e , ρ ∈ S ∑ K e , ρ c ρ ( u e ) , ∫ D g P K ( g ) = 1. (146)
Here, clean means that the average is taken before conditioning on a measurement record.
Let m = { m e , ρ } e , ρ ∈ S m=\{m_{e,\rho}\}_{e,\rho\in\mathcal{S}} m = { m e , ρ } e , ρ ∈ S denote the complete measurement record. For a fixed edge configuration, the likelihood is a product of the Gaussian factors for the measured components. Using Bayes’ rule and ∥ f ρ ( u ) ∥ 2 = 1 \lVert f_{\rho}(u)\rVert^2=1 ∥ f ρ ( u ) ∥ 2 = 1 , we obtain
P K , Δ ( g ∣ m ) ∝ P K ( g ) ∏ e , ρ ∈ S p Δ ( m e , ρ ∣ u e ) ∝ exp [ ∑ e , ρ ∈ S { K e , ρ c ρ ( u e ) + Δ e , ρ m e , ρ ⋅ f ρ ( u e ) } ] . (147) P_{K,\Delta}(g\mid m)\propto P_K(g)\prod_{e,\rho\in\mathcal{S}}p_{\Delta}(m_{e,\rho}\mid u_e)\propto\exp\left[\sum_{e,\rho\in\mathcal{S}}\left\{K_{e,\rho}c_{\rho}(u_e)+\Delta_{e,\rho}m_{e,\rho}\cdot f_{\rho}(u_e)\right\}\right]. \tag*{(147)} P K , Δ ( g ∣ m ) ∝ P K ( g ) e , ρ ∈ S ∏ p Δ ( m e , ρ ∣ u e ) ∝ exp e , ρ ∈ S ∑ { K e , ρ c ρ ( u e ) + Δ e , ρ m e , ρ ⋅ f ρ ( u e ) } . (147)
The posterior is normalized with respect to D g \mathcal{D}_g D g . The Born probability density of the complete record is obtained by averaging the likelihood over the prior:
p K , Δ ( m ) = ∫ D g P K ( g ) ∏ e , ρ ∈ S p Δ ( m e , ρ ∣ u e ) . p_{K,\Delta}(m)=\int\mathcal{D}g\,P_K(g)\prod_{e,\rho\in\mathcal{S}}p_{\Delta}(m_{e,\rho}\mid u_e). p K , Δ ( m ) = ∫ D g P K ( g ) e , ρ ∈ S ∏ p Δ ( m e , ρ ∣ u e ) .
Thus, the record distribution is determined by the deformed quantum state, rather than by a Gaussian distribution chosen independently of the state.
Matrix EA correlators
We next define the matrix EA correlator associated with an open electric string. For an oriented path γ : x → y \gamma:x\to y γ : x → y , we multiply the edge variables in path order, using u e − 1 u_e^{-1} u e − 1 whenever the path runs opposite to the orientation of e e e . In the pure-gauge sector, the product reduces to U x y = g x g y − 1 U_{xy}=g_xg_y^{-1} U x y = g x g y − 1 . For an irreducible representation λ \lambda λ , we define the open string matrix conditioned on the measurement record as
A λ ( m ; x , y ) : = ∫ D g P K , Δ ( g ∣ m ) D λ ( g x g y − 1 ) . (148) A_{\lambda}(m;x,y):=\int\mathcal{D}g\,P_{K,\Delta}(g\mid m)D^{\lambda}(g_xg_y^{-1}). \tag*{(148)} A λ ( m ; x , y ) := ∫ D g P K , Δ ( g ∣ m ) D λ ( g x g y − 1 ) . (148)
The corresponding matrix EA correlator is defined by
Q λ ( K , Δ ; x , y ) : = ∫ d m p K , Δ ( m ) Tr [ A λ ( m ) A λ ( m ) † ] d λ . (149) Q_{\lambda}(K,\Delta;x,y):=\int dm\,p_{K,\Delta}(m)\frac{\operatorname{Tr}\left[A_{\lambda}(m)A_{\lambda}(m)^{\dagger}\right]}{d_{\lambda}}. \tag*{(149)} Q λ ( K , Δ ; x , y ) := ∫ d m p K , Δ ( m ) d λ Tr [ A λ ( m ) A λ ( m ) † ] . (149)
The matrix A λ A_\lambda A λ is averaged over group configurations at a fixed record m m m . We then average its normalized squared Frobenius norm over records with their Born probabilities. This generalizes the squared spin–spin correlator of the Z 2 \mathbb{Z}_2 Z 2 toric code.
For comparison, we define the clean correlator by averaging over the same deformed prior without conditioning on a measurement record:
C λ ( K ; x , y ) : = ∫ D g P K ( g ) χ λ ( g x g y − 1 ) d λ = 1 Z K ∫ D g exp [ ∑ e , ρ ∈ S K e , ρ c ρ ( u e ) ] χ λ ( g x g y − 1 ) d λ . (150) C_\lambda(K;x,y) := \int\mathcal{D}g\,P_K(g)\frac{\chi_\lambda(g_xg_y^{-1})}{d_\lambda} = \frac{1}{Z_K}\int\mathcal{D}g\,\exp\left[\sum_{e,\rho\in\mathcal{S}}K_{e,\rho}c_\rho(u_e)\right]\frac{\chi_\lambda(g_xg_y^{-1})}{d_\lambda}. \tag*{(150)} C λ ( K ; x , y ) := ∫ D g P K ( g ) d λ χ λ ( g x g y − 1 ) = Z K 1 ∫ D g exp e , ρ ∈ S ∑ K e , ρ c ρ ( u e ) d λ χ λ ( g x g y − 1 ) . (150)
As a consistency check, for G = Z 2 G=\mathbb{Z}_2 G = Z 2 , the nontrivial character gives C λ ( K ; x , y ) = ⟨ σ x σ y ⟩ K C_\lambda(K;x,y)=\langle\sigma_x\sigma_y\rangle_K C λ ( K ; x , y ) = ⟨ σ x σ y ⟩ K . This is the ordinary Ising spin–spin correlator appearing in the Z 2 \mathbb{Z}_2 Z 2 toric code result (101) . In general, Q λ ( K , Δ ; x , y ) Q_\lambda(K,\Delta;x,y) Q λ ( K , Δ ; x , y ) and C λ ( K ; x , y ) C_\lambda(K;x,y) C λ ( K ; x , y ) are different: the former involves the squared conditional matrix expectation averaged over measurement records, whereas the latter is a single average in the prior ensemble.
We now derive an exact identity for the measurement protocol defined above. We consider a finite connected graph in the pure-gauge sector and use the isotropic component measurements specified above. We impose the matching condition
K e , ρ = Δ e , ρ for each measured ( e , ρ ) . (151) K_{e,\rho}=\Delta_{e,\rho}\qquad\text{for each measured }(e,\rho). \tag*{(151)} K e , ρ = Δ e , ρ for each measured ( e , ρ ) . (151)
Under this condition, we have
Q λ ( K , K ; x , y ) = C λ ( K ; x , y ) ∀ λ . (152) Q_\lambda(K,K;x,y)=C_\lambda(K;x,y)\qquad\forall\lambda. \tag*{(152)} Q λ ( K , K ; x , y ) = C λ ( K ; x , y ) ∀ λ . (152)
The proof proceeds in two steps. A group average cancels the posterior normalization factor, after which the Gaussian integrals reduce the result to the clean correlator. We introduce the shifted Gaussian vectors
J e , ρ = K e , ρ f ρ ( 1 G ) + Δ e , ρ m e , ρ , S J ( g ) = ∑ e = ( i , j ) , ρ ∈ S J e , ρ ⋅ f ρ ( g i g j − 1 ) . (153) J_{e,\rho}=K_{e,\rho}f_\rho(1_G)+\Delta_{e,\rho}m_{e,\rho},\qquad S_J(g)=\sum_{\substack{e=(i,j),\rho\in\mathcal{S}}}J_{e,\rho}\cdot f_\rho(g_ig_j^{-1}). \tag*{(153)} J e , ρ = K e , ρ f ρ ( 1 G ) + Δ e , ρ m e , ρ , S J ( g ) = e = ( i , j ) , ρ ∈ S ∑ J e , ρ ⋅ f ρ ( g i g j − 1 ) . (153)
We also define the record-dependent partition function and the unnormalized matrix correlator as
Z ( J ) : = ∫ D g e S J ( g ) , Z(J):=\int\mathcal{D}g\,e^{S_J(g)}, Z ( J ) := ∫ D g e S J ( g ) ,
N λ ( J ) : = ∫ D g e S J ( g ) D λ ( g x g y − 1 ) . (154) \mathcal{N}_\lambda(J):=\int\mathcal{D}g\,e^{S_J(g)}D^\lambda(g_xg_y^{-1}). \tag*{(154)} N λ ( J ) := ∫ D g e S J ( g ) D λ ( g x g y − 1 ) . (154)
Then A λ ( J ) = N λ ( J ) / Z ( J ) A_\lambda(J)=\mathcal{N}_\lambda(J)/Z(J) A λ ( J ) = N λ ( J ) / Z ( J ) .
Let d ν Δ ( J ) d\nu_\Delta(J) d ν Δ ( J ) denote the normalized product of centered, isotropic Gaussian measures for the real vectors J e , ρ J_{e,\rho} J e , ρ , each with covariance Δ e , ρ I n ρ \Delta_{e,\rho}I_{n_\rho} Δ e , ρ I n ρ . Under the matching condition, completing the square in the Born probability density gives
p ( J ) d J = e − ∑ e , ρ Δ e , ρ Z K d ν Δ ( J ) e S J ( 1 G ) Z ( J ) . (155) p(J)dJ=\frac{e^{-\sum_{e,\rho}\Delta_{e,\rho}}}{Z_K}d\nu_\Delta(J)e^{S_J(1_G)}Z(J). \tag*{(155)} p ( J ) dJ = Z K e − ∑ e , ρ Δ e , ρ d ν Δ ( J ) e S J ( 1 G ) Z ( J ) . (155)
Here, 1 G 1_G 1 G denotes the vertex configuration with g i = 1 G g_i=1_G g i = 1 G at every site. The factor Z ( J ) Z(J) Z ( J ) comes from averaging over the underlying group configuration. Substituting this probability density into the matrix EA correlator leaves one inverse power of Z ( J ) Z(J) Z ( J ) :
Q λ = e − ∑ e , ρ Δ e , ρ d λ Z K ∫ d ν Δ ( J ) e S J ( 1 G ) Tr [ N λ ( J ) N λ ( J ) † ] Z ( J ) . (156) Q_\lambda=\frac{e^{-\sum_{e,\rho}\Delta_{e,\rho}}}{d_\lambda Z_K}\int d\nu_\Delta(J)e^{S_J(1_G)}\frac{\operatorname{Tr}\left[\mathcal{N}_\lambda(J)\mathcal{N}_\lambda(J)^\dagger\right]}{Z(J)}. \tag*{(156)} Q λ = d λ Z K e − ∑ e , ρ Δ e , ρ ∫ d ν Δ ( J ) e S J ( 1 G ) Z ( J ) Tr [ N λ ( J ) N λ ( J ) † ] . (156)
For each ρ \rho ρ , we combine the real and imaginary components of J i j , ρ J_{ij,\rho} J ij , ρ into a complex matrix J i j , ρ J_{ij,\rho} J ij , ρ . We then make the simultaneous change of variables g i ↦ h i g i g_i \mapsto h_i g_i g i ↦ h i g i and J i j , ρ ↦ D ρ ( h i ) J i j , ρ D ρ ( h j ) † J_{ij,\rho} \mapsto D^\rho(h_i)J_{ij,\rho}D^\rho(h_j)^\dagger J ij , ρ ↦ D ρ ( h i ) J ij , ρ D ρ ( h j ) † . Viewed in real coordinates, this matrix transformation is orthogonal. It leaves the centered isotropic Gaussian measure, Z ( J ) Z(J) Z ( J ) , and Tr [ N λ ( J ) N λ ( J ) † ] \operatorname{Tr}[\mathcal{N}_\lambda(J)\mathcal{N}_\lambda(J)^\dagger] Tr [ N λ ( J ) N λ ( J ) † ] unchanged. The factor e S J ( 1 G ) e^{S_J(1_G)} e S J ( 1 G ) is not invariant, but averaging it over all h i ∈ G h_i \in G h i ∈ G gives
∫ D h e S J h ( 1 G ) = ∫ D h e S J ( h − 1 ) = Z ( J ) . (157) \int\mathcal{D}h\,e^{S_{J^h}(1_G)}=\int\mathcal{D}h\,e^{S_J(h^{-1})}=Z(J). \tag*{(157)} ∫ D h e S J h ( 1 G ) = ∫ D h e S J ( h − 1 ) = Z ( J ) . (157)
The last equality follows from the invariance of the normalized sum over h h h under inversion. This factor cancels the remaining 1 / Z ( J ) 1/Z(J) 1/ Z ( J ) in (156) .
Expanding N λ ( J ) N λ ( J ) † \mathcal{N}_\lambda(J)\mathcal{N}_\lambda(J)^\dagger N λ ( J ) N λ ( J ) † introduces two independently summed vertex configurations, g ( 1 ) g^{(1)} g ( 1 ) and g ( 2 ) g^{(2)} g ( 2 ) . For each edge and representation, the Gaussian integral together with (140) gives
∫ d ν Δ e , ρ ( J ) exp { J ⋅ [ f ρ ( u ( 1 ) ) + f ρ ( u ( 2 ) ) ] } = exp [ Δ e , ρ { 1 + c ρ ( u ( 1 ) ( u ( 2 ) ) − 1 ) } ] . (158) \int\mathrm{d}\nu_{\Delta_{e,\rho}}(J)\exp\left\{J\cdot\left[f_\rho\left(u^{(1)}\right)+f_\rho\left(u^{(2)}\right)\right]\right\}=\exp\left[\Delta_{e,\rho}\left\{1+c_\rho\left(u^{(1)}\left(u^{(2)}\right)^{-1}\right)\right\}\right]. \tag*{(158)} ∫ d ν Δ e , ρ ( J ) exp { J ⋅ [ f ρ ( u ( 1 ) ) + f ρ ( u ( 2 ) ) ] } = exp [ Δ e , ρ { 1 + c ρ ( u ( 1 ) ( u ( 2 ) ) − 1 ) } ] . (158)
The factor e Δ e , ρ e^{\Delta_{e,\rho}} e Δ e , ρ cancels the corresponding constant in the Born probability density. We now introduce the relative vertex variables
k i = ( g i ( 2 ) ) − 1 g i ( 1 ) . (159) k_i=\left(g_i^{(2)}\right)^{-1}g_i^{(1)}. \tag*{(159)} k i = ( g i ( 2 ) ) − 1 g i ( 1 ) . (159)
For u i j ( a ) = g i ( a ) ( g j ( a ) ) − 1 u_{ij}^{(a)}=g_i^{(a)}\left(g_j^{(a)}\right)^{-1} u ij ( a ) = g i ( a ) ( g j ( a ) ) − 1 , the product of the two edge variables can be written as
u i j ( 1 ) ( u i j ( 2 ) ) − 1 = g i ( 2 ) ( k i k j − 1 ) ( g j ( 2 ) ) − 1 . u_{ij}^{(1)}\left(u_{ij}^{(2)}\right)^{-1}=g_i^{(2)}\left(k_i k_j^{-1}\right)\left(g_j^{(2)}\right)^{-1}. u ij ( 1 ) ( u ij ( 2 ) ) − 1 = g i ( 2 ) ( k i k j − 1 ) ( g j ( 2 ) ) − 1 .
The same relation holds at the endpoints, with ( i , j ) (i,j) ( i , j ) replaced by ( x , y ) (x,y) ( x , y ) . Since characters are invariant under conjugation, the interaction obtained from the Gaussian integral depends only on k i k j − 1 k_i k_j^{-1} k i k j − 1 . The matrix insertion similarly reduces to the normalized character χ λ ( k x k y − 1 ) / d λ \chi_\lambda(k_x k_y^{-1})/d_\lambda χ λ ( k x k y − 1 ) / d λ . The change of variables ( g ( 1 ) , g ( 2 ) ) ↦ ( k , g ( 2 ) ) (g^{(1)},g^{(2)})\mapsto(k,g^{(2)}) ( g ( 1 ) , g ( 2 ) ) ↦ ( k , g ( 2 ) ) is bijective, and the normalized sum over g ( 2 ) g^{(2)} g ( 2 ) gives one. The remaining sum over k k k is the clean correlator in (150) , completing the proof of (152) .
Hierarchy of EA correlators
In this subsection, we study the clean correlator C λ ( K ; x , y ) C_\lambda(K;x,y) C λ ( K ; x , y ) and derive a hierarchy among the EA correlators of different electric charges. To this end, we use the Fortuin–Kasteleyn–Ginibre (FKG) inequality [45 ] and expand the interactions in terms of the normal subgroups of G G G , denoted by N N N . Recalling from (146) that the clean probability distribution is determined by the edge interactions V e ( u ) : = ∑ ρ ∈ S K e , ρ c ρ ( u ) V_e(u):=\sum_{\rho\in\mathcal{S}}K_{e,\rho}c_\rho(u) V e ( u ) := ∑ ρ ∈ S K e , ρ c ρ ( u ) , where u = g i g j − 1 u=g_i g_j^{-1} u = g i g j − 1 for an oriented edge e = ( i , j ) e=(i,j) e = ( i , j ) , we impose an additional condition on these interactions: each V e ( u ) V_e(u) V e ( u ) is assumed to admit an expansion in the so-called indicator functions of normal subgroups of G G G , δ N ( u ) \delta_N(u) δ N ( u ) with nonnegative coefficients, up to an additive constant c e c_e c e . Namely,
V e ( u ) : = ∑ ρ ∈ S K e , ρ c ρ ( u ) = c e + ∑ N ⊲ G J e , N δ N ( u ) , J e , N ≥ 0 , (160) V_e(u):=\sum_{\rho\in\mathcal{S}}K_{e,\rho}c_\rho(u)=c_e+\sum_{N\lhd G}J_{e,N}\delta_N(u),\qquad J_{e,N}\geq0, \tag*{(160)} V e ( u ) := ρ ∈ S ∑ K e , ρ c ρ ( u ) = c e + N ⊲ G ∑ J e , N δ N ( u ) , J e , N ≥ 0 , (160)
where δ N ( u ) = 1 \delta_N(u)=1 δ N ( u ) = 1 for u ∈ N u\in N u ∈ N and 0 0 0 otherwise. We work on the higher Nishimori manifold K e , ρ = Δ e , ρ K_{e,\rho}=\Delta_{e,\rho} K e , ρ = Δ e , ρ and assume the nonnegative expansion (160) .
Let J \mathcal{J} J be the finite set of pairs α = ( e , N ) \alpha=(e,N) α = ( e , N ) for which J e , N J_{e,N} J e , N is strictly positive. Here, e = ( i , j ) e=(i,j) e = ( i , j ) is an oriented edge and N ⊲ G N\lhd G N ⊲ G is a normal subgroup. Terms with J e , N = 0 J_{e,N}=0 J e , N = 0 can be omitted. For each pair, we define J α : = J e , N J_\alpha:=J_{e,N} J α := J e , N and δ α ( g ) : = δ N ( g i g j − 1 ) \delta_\alpha(g):=\delta_N(g_i g_j^{-1}) δ α ( g ) := δ N ( g i g j − 1 ) . Thus, δ α ( g ) = 1 \delta_\alpha(g)=1 δ α ( g ) = 1 when the group variables at the endpoints of e e e represent the same element of the quotient group G / N G/N G / N .
Since δ α \delta_\alpha δ α takes only the values zero and one, we can expand its Boltzmann factor as
e J α δ α ( g ) = 1 + v α δ α ( g ) , v α : = e J α − 1 ≥ 0. (161) e^{J_\alpha\delta_\alpha(g)}=1+v_\alpha\delta_\alpha(g),\qquad v_\alpha:=e^{J_\alpha}-1\geq0. \tag*{(161)} e J α δ α ( g ) = 1 + v α δ α ( g ) , v α := e J α − 1 ≥ 0. (161)
Both sides are 1 for δ α = 0 \delta_\alpha=0 δ α = 0 and e J α e^{J_\alpha} e J α for δ α = 1 \delta_\alpha=1 δ α = 1 . The terms c e c_e c e in (160) are independent of the configuration and contribute only an overall factor. Omitting this factor, we write the Boltzmann weight as
W ( g ) : = ∏ α ∈ J [ 1 + v α δ α ( g ) ] . (162) W(g):=\prod_{\alpha\in\mathcal{J}}\left[1+v_\alpha\delta_\alpha(g)\right]. \tag*{(162)} W ( g ) := α ∈ J ∏ [ 1 + v α δ α ( g ) ] . (162)
For each factor, we choose either 1 or v α δ α v_\alpha\delta_\alpha v α δ α . Let B ⊆ J B\subseteq\mathcal{J} B ⊆ J be the set of labels for which the second term is chosen. The product in (162) then becomes
W ( g ) = ∑ B ⊆ J ( ∏ α ∈ B v α ) ( ∏ α ∈ B δ α ( g ) ) . W(g)=\sum_{B\subseteq\mathcal{J}}\left(\prod_{\alpha\in B}v_\alpha\right)\left(\prod_{\alpha\in B}\delta_\alpha(g)\right). W ( g ) = B ⊆ J ∑ ( α ∈ B ∏ v α ) ( α ∈ B ∏ δ α ( g ) ) .
We define
H B : = { g ∈ G V : δ α ( g ) = 1 for every α ∈ B } . (163) H_B:=\left\{g\in G^V:\delta_\alpha(g)=1\text{ for every }\alpha\in B\right\}. \tag*{(163)} H B := { g ∈ G V : δ α ( g ) = 1 for every α ∈ B } . (163)
Summing over the group spins, we obtain
Z : = ∑ g ∈ G V W ( g ) = ∑ B ⊆ J ∣ H B ∣ ∏ α ∈ B v α . \mathcal{Z}:=\sum_{g\in G^V}W(g)=\sum_{B\subseteq\mathcal{J}}|H_B|\prod_{\alpha\in B}v_\alpha. Z := g ∈ G V ∑ W ( g ) = B ⊆ J ∑ ∣ H B ∣ α ∈ B ∏ v α .
The clean partition function in (145) is therefore Z K = e ∑ e c e ∣ G ∣ − ∣ V ∣ Z Z_K=e^{\sum_e c_e}|G|^{-|V|}\mathcal{Z} Z K = e ∑ e c e ∣ G ∣ − ∣ V ∣ Z . The terms in this expansion have nonnegative weights, which we normalize to define the probability distribution
μ ( B ) : = ∣ H B ∣ ∏ α ∈ B v α ∑ D ⊆ J ∣ H D ∣ ∏ α ∈ D v α , ∑ B ⊆ J μ ( B ) = 1. (164) \mu(B):=\frac{|H_B|\prod_{\alpha\in B}v_\alpha}{\sum_{D\subseteq\mathcal{J}}|H_D|\prod_{\alpha\in D}v_\alpha},\qquad\sum_{B\subseteq\mathcal{J}}\mu(B)=1. \tag*{(164)} μ ( B ) := ∑ D ⊆ J ∣ H D ∣ ∏ α ∈ D v α ∣ H B ∣ ∏ α ∈ B v α , B ⊆ J ∑ μ ( B ) = 1. (164)
Applying the same expansion to the numerator of the clean correlator in (150) , we obtain
C λ ( K ; x , y ) = ∑ B ⊆ J μ ( B ) C λ ( B ) ( x , y ) (165) C_\lambda(K;x,y)=\sum_{B\subseteq\mathcal{J}}\mu(B)C_\lambda^{(B)}(x,y) \tag*{(165)} C λ ( K ; x , y ) = B ⊆ J ∑ μ ( B ) C λ ( B ) ( x , y ) (165)
with
C λ ( B ) ( x , y ) : = 1 ∣ H B ∣ ∑ g ∈ H B χ λ ( g x g y − 1 ) d λ . (166) C_\lambda^{(B)}(x,y):=\frac{1}{|H_B|}\sum_{g\in H_B}\frac{\chi_\lambda(g_xg_y^{-1})}{d_\lambda}. \tag*{(166)} C λ ( B ) ( x , y ) := ∣ H B ∣ 1 g ∈ H B ∑ d λ χ λ ( g x g y − 1 ) . (166)
To evaluate the correlator in (165) , we introduce the following set for fixed B B B and endpoints x , y x,y x , y :
N B ( x , y ) : = { g x g y − 1 : g ∈ H B } . (167) N_B(x,y):=\left\{g_xg_y^{-1}:g\in H_B\right\}. \tag*{(167)} N B ( x , y ) := { g x g y − 1 : g ∈ H B } . (167)
Here N B ( x , y ) N_B(x,y) N B ( x , y ) is the set of relative group elements allowed by the constraints in B B B . We omit the endpoints when they are clear. In Appendix § 11 , we show that N B ( x , y ) N_B(x,y) N B ( x , y ) is a normal subgroup of G G G and that a uniform sample from H B H_B H B induces the uniform distribution on N B ( x , y ) N_B(x,y) N B ( x , y ) . The constrained character average is therefore
C λ ( B ) ( x , y ) = 1 d λ ∣ N B ∣ ∑ n ∈ N B χ λ ( n ) = { 1 , N B ⊆ ker D λ , 0 , otherwise . (168) C_\lambda^{(B)}(x,y)=\frac{1}{d_\lambda|N_B|}\sum_{n\in N_B}\chi_\lambda(n)=
\begin{cases}
1, & N_B\subseteq\ker D^\lambda,\\
0, & \text{otherwise}.
\end{cases} \tag*{(168)} C λ ( B ) ( x , y ) = d λ ∣ N B ∣ 1 n ∈ N B ∑ χ λ ( n ) = { 1 , 0 , N B ⊆ ker D λ , otherwise . (168)
We further define the event
A λ ( x , y ) : = { B ⊆ J : N B ( x , y ) ⊆ ker D λ } . \mathcal{A}_{\lambda}(x,y) := \left\{B \subseteq\mathcal{J}: N_B(x,y) \subseteq\ker D^{\lambda}\right\}. A λ ( x , y ) := { B ⊆ J : N B ( x , y ) ⊆ ker D λ } .
Let 1 A λ ( B ) \mathbf{1}_{\mathcal{A}_{\lambda}}(B) 1 A λ ( B ) equal one when B ∈ A λ B \in\mathcal{A}_{\lambda} B ∈ A λ and zero otherwise. Combining the constrained average with Q λ = C λ Q_{\lambda}=C_{\lambda} Q λ = C λ , we obtain
Q λ ( x , y ) = C λ ( x , y ) = ∑ B ⊆ J μ ( B ) 1 A λ ( B ) = μ ( A λ ) . (169) Q_{\lambda}(x,y)=C_{\lambda}(x,y)=\sum_{B\subseteq\mathcal{J}}\mu(B)\mathbf{1}_{\mathcal{A}_{\lambda}}(B)=\mu(\mathcal{A}_{\lambda}). \tag*{(169)} Q λ ( x , y ) = C λ ( x , y ) = B ⊆ J ∑ μ ( B ) 1 A λ ( B ) = μ ( A λ ) . (169)
This expresses the EA correlator as the probability of an event in the constraint ensemble. We can now compare the correlators of two representations. If ker D λ ⊆ ker D ν \ker D^{\lambda}\subseteq\ker D^{\nu} ker D λ ⊆ ker D ν , then A λ ⊆ A ν \mathcal{A}_{\lambda}\subseteq\mathcal{A}_{\nu} A λ ⊆ A ν . The inclusion of these events gives
ker D λ ⊆ ker D ν ⟹ Q λ ( x , y ) ≤ Q ν ( x , y ) . (170) \ker D^{\lambda}\subseteq\ker D^{\nu}\quad\Longrightarrow\quad Q_{\lambda}(x,y)\leq Q_{\nu}(x,y). \tag*{(170)} ker D λ ⊆ ker D ν ⟹ Q λ ( x , y ) ≤ Q ν ( x , y ) . (170)
Next, let ν ≺ λ ⊗ η \nu\prec\lambda\otimes\eta ν ≺ λ ⊗ η , meaning that ν \nu ν appears in the tensor-product decomposition of λ \lambda λ and η \eta η . If every element of N B N_B N B acts trivially in both representations, it also acts trivially in each constituent of their tensor product. Hence A λ ∩ A η ⊆ A ν \mathcal{A}_{\lambda}\cap\mathcal{A}_{\eta}\subseteq\mathcal{A}_{\nu} A λ ∩ A η ⊆ A ν . Adding constraints can only shrink N B N_B N B , so an event A λ \mathcal{A}_{\lambda} A λ that already holds continues to hold. The events A λ \mathcal{A}_{\lambda} A λ and A η \mathcal{A}_{\eta} A η are therefore increasing under inclusion of constraint sets. As discussed in Appendix § 12 , one can make use of the FKG inequality, which gives
μ ( A λ ∩ A η ) ≥ μ ( A λ ) μ ( A η ) . (171) \mu(\mathcal{A}_{\lambda}\cap\mathcal{A}_{\eta})\geq\mu(\mathcal{A}_{\lambda})\mu(\mathcal{A}_{\eta}). \tag*{(171)} μ ( A λ ∩ A η ) ≥ μ ( A λ ) μ ( A η ) . (171)
Combining this inequality with the event inclusion and the probability representation of the correlators yields
ν ≺ λ ⊗ η ⟹ Q ν ( x , y ) ≥ Q λ ( x , y ) Q η ( x , y ) . (172) \nu\prec\lambda\otimes\eta\quad\Longrightarrow\quad Q_{\nu}(x,y)\geq Q_{\lambda}(x,y)Q_{\eta}(x,y). \tag*{(172)} ν ≺ λ ⊗ η ⟹ Q ν ( x , y ) ≥ Q λ ( x , y ) Q η ( x , y ) . (172)
The relations (170) and (172) give the hierarchy among the EA correlators of different electric charges. We illustrate these bounds in the quantum doubles of S 3 S_3 S 3 and D 4 D_4 D 4 below.
Examples
We now apply the above inequalities to two concrete examples, S 3 S_3 S 3 and D 4 D_4 D 4 quantum double.
G=S3
The group S 3 S_3 S 3 consists of the six permutations of three objects. Choosing r = ( 123 ) r=(123) r = ( 123 ) and s = ( 12 ) s=(12) s = ( 12 ) , we write
S 3 = ⟨ r , s : r 3 = s 2 = 1 G , s r s = r − 1 ⟩ = { 1 G , r , r 2 , s , r s , r 2 s } . S_3=\langle r,s:r^3=s^2=1_G,\ srs=r^{-1}\rangle=\{1_G,r,r^2,s,rs,r^2s\}. S 3 = ⟨ r , s : r 3 = s 2 = 1 G , sr s = r − 1 ⟩ = { 1 G , r , r 2 , s , r s , r 2 s } .
The even permutations form a subgroup
A 3 : = { 1 G , r , r 2 } = ⟨ r ⟩ ≅ Z 3 . A_3:=\{1_G,r,r^2\}=\langle r\rangle\cong\mathbb{Z}_3. A 3 := { 1 G , r , r 2 } = ⟨ r ⟩ ≅ Z 3 .
This subgroup is normal: conjugation by r r r preserves it, while conjugation by s s s exchanges r r r and r − 1 = r 2 r^{-1}=r^2 r − 1 = r 2 .
The three irreducible representations of S 3 S_3 S 3 label its three pure electric sectors, including the vacuum. We denote the trivial, sign, and standard two dimensional representations by A A A , B B B , and C C C , respectively. Their dimensions are d A = d B = 1 d_A=d_B=1 d A = d B = 1 and d C = 2 d_C=2 d C = 2 . The sign representation satisfies D B ( r ) = 1 D^B(r)=1 D B ( r ) = 1 and D B ( s ) = − 1 D^B(s)=-1 D B ( s ) = − 1 . The characters and conjugacy class sizes are listed in Table 2.
caption: Character table of S 3 S_3 S 3 , including sizes of the conjugacy classes. The representations A A A , B B B , and C C C are the trivial, sign, and two dimensional irreducible representations, respectively.
:::
Characters multiply under tensor products and add under direct sums. The electric fusion rules are therefore
A ⊗ ρ = ρ ( ρ ∈ { A , B , C } ) , B ⊗ B = A , B ⊗ C = C , A\otimes\rho=\rho\quad(\rho\in\{A,B,C\}),\qquad B\otimes B=A,\qquad B\otimes C=C, A ⊗ ρ = ρ ( ρ ∈ { A , B , C }) , B ⊗ B = A , B ⊗ C = C ,
C ⊗ C = A ⊕ B ⊕ C . (173) C\otimes C=A\oplus B\oplus C. \tag*{(173)} C ⊗ C = A ⊕ B ⊕ C . (173)
In this quantum double, we think of weak measurement of two electric charges, corresponding to two representations S = { B , C } \mathcal{S}=\{B,C\} S = { B , C } . Since their characters are real, the corresponding classical spin interaction in (141) gives
V ( u ) = K B χ B ( u ) + K C 2 χ C ( u ) . (174) V(u)=K_B\chi_B(u)+\frac{K_C}{2}\chi_C(u). \tag*{(174)} V ( u ) = K B χ B ( u ) + 2 K C χ C ( u ) . (174)
To apply the hierarchy, we express this interaction in terms of indicator functions of normal subgroups. The two nonconstant indicators needed here are δ A 3 \delta_{A_3} δ A 3 and δ { 1 G } \delta_{\{1_G\}} δ { 1 G } , whose values are listed in Table 3.
caption: Indicator functions for the proper normal subgroups A 3 A_3 A 3 and { 1 G } \{1_G\} { 1 G } of S 3 S_3 S 3 . Their values are constant on each conjugacy class.
:::
Comparing Tables 2 and 3, we obtain
χ B ( u ) = 2 δ A 3 ( u ) − 1 , χ C ( u ) = 3 δ { 1 G } ( u ) − δ A 3 ( u ) . \chi_B(u)=2\delta_{A_3}(u)-1,\qquad\chi_C(u)=3\delta_{\{1_G\}}(u)-\delta_{A_3}(u). χ B ( u ) = 2 δ A 3 ( u ) − 1 , χ C ( u ) = 3 δ { 1 G } ( u ) − δ A 3 ( u ) .
Substituting these identities into (174) , we obtain
V ( u ) = K B [ 2 δ A 3 ( u ) − 1 ] + K C 2 [ 3 δ { 1 G } ( u ) − δ A 3 ( u ) ] = − K B + ( 2 K B − K C 2 ) δ A 3 ( u ) + 3 K C 2 δ { 1 G } ( u ) . (175) V(u)=K_B\left[2\delta_{A_3}(u)-1\right]+\frac{K_C}{2}\left[3\delta_{\{1_G\}}(u)-\delta_{A_3}(u)\right]
=-K_B+\left(2K_B-\frac{K_C}{2}\right)\delta_{A_3}(u)+\frac{3K_C}{2}\delta_{\{1_G\}}(u). \tag*{(175)} V ( u ) = K B [ 2 δ A 3 ( u ) − 1 ] + 2 K C [ 3 δ { 1 G } ( u ) − δ A 3 ( u ) ] = − K B + ( 2 K B − 2 K C ) δ A 3 ( u ) + 2 3 K C δ { 1 G } ( u ) . (175)
To apply the hierarchy derived in the preceding subsection, we require the coefficients in this expansion to be nonnegative:
K C ≥ 0 , K B ≥ K C / 4. K_C\geq0,\qquad K_B\geq K_C/4. K C ≥ 0 , K B ≥ K C /4.
On the higher Nishimori manifold, these conditions become
K B = Δ B , K C = Δ C , Δ C ≥ 0 , Δ B ≥ Δ C / 4. K_B=\Delta_B,\qquad K_C=\Delta_C,\qquad\Delta_C\geq0,\qquad\Delta_B\geq\Delta_C/4. K B = Δ B , K C = Δ C , Δ C ≥ 0 , Δ B ≥ Δ C /4.
From (170) and ker D C = { 1 G } ⊆ A 3 = ker D B \ker D^C=\{1_G\}\subseteq A_3=\ker D^B ker D C = { 1 G } ⊆ A 3 = ker D B , we have
Q B ( x , y ) ≥ Q C ( x , y ) . (176) Q_B(x,y)\geq Q_C(x,y). \tag*{(176)} Q B ( x , y ) ≥ Q C ( x , y ) . (176)
Thus, in this parameter region, long range learnability of the non-Abelian electric charge C C C implies long range learnability of the Abelian electric charge B B B .
G=D4
We next consider D 4 D_4 D 4 , given by
D 4 = ⟨ r , s : r 4 = s 2 = 1 G , s r s = r − 1 ⟩ = { r m s n : 0 ≤ m ≤ 3 , n = 0 , 1 } . D_4=\langle r,s:r^4=s^2=1_G,\ srs=r^{-1}\rangle=\{r^m s^n:0\le m\le3,\ n=0,1\}. D 4 = ⟨ r , s : r 4 = s 2 = 1 G , sr s = r − 1 ⟩ = { r m s n : 0 ≤ m ≤ 3 , n = 0 , 1 } .
Here, r r r is a quarter-turn and s s s is a reflection of a square. Besides the trivial representation 1 1 1 , there are three one-dimensional representations a , b , c = a b a,b,c=ab a , b , c = ab , defined by
a ( r m s n ) = ( − 1 ) m , b ( r m s n ) = ( − 1 ) n , c ( r m s n ) = ( − 1 ) m + n . a(r^m s^n)=(-1)^m,\qquad b(r^m s^n)=(-1)^n,\qquad c(r^m s^n)=(-1)^{m+n}. a ( r m s n ) = ( − 1 ) m , b ( r m s n ) = ( − 1 ) n , c ( r m s n ) = ( − 1 ) m + n .
The two dimensional representation E E E can be chosen as
D E ( r ) = ( 0 − 1 1 0 ) , D E ( s ) = ( 1 0 0 − 1 ) . D^E(r)=\begin{pmatrix}0&-1\\1&0\end{pmatrix},\qquad D^E(s)=\begin{pmatrix}1&0\\0&-1\end{pmatrix}. D E ( r ) = ( 0 1 − 1 0 ) , D E ( s ) = ( 1 0 0 − 1 ) .
The five irreducible representations 1 , a , b , c , E 1,a,b,c,E 1 , a , b , c , E label the pure electric sectors of D ( D 4 ) D(D_4) D ( D 4 ) . The trivial representation is the vacuum, while a , b , c , E a,b,c,E a , b , c , E label the four nontrivial electric charges.
The characters are listed in Table 4 .
Table 4. Character table of D 4 D_4 D 4 , the dihedral group of order eight, including sizes of the conjugacy classes.
The representations 1 , a , b , c 1,a,b,c 1 , a , b , c are one-dimensional, while E E E is two dimensional.
Since the one-dimensional characters take values ± 1 \pm1 ± 1 , multiplying their rows gives
a ⊗ a = b ⊗ b = c ⊗ c = 1 , a ⊗ b = c , b ⊗ c = a , c ⊗ a = b . (177) a\otimes a=b\otimes b=c\otimes c=1,\qquad a\otimes b=c,\qquad b\otimes c=a,\qquad c\otimes a=b. \tag*{(177)} a ⊗ a = b ⊗ b = c ⊗ c = 1 , a ⊗ b = c , b ⊗ c = a , c ⊗ a = b . (177)
Each of these characters is + 1 +1 + 1 at 1 G 1_G 1 G and r 2 r^2 r 2 , the only classes on which χ E \chi_E χ E is nonzero. Hence,
α ⊗ E = E ( α ∈ { 1 , a , b , c } ) . (178) \alpha\otimes E=E\qquad(\alpha\in\{1,a,b,c\}). \tag*{(178)} α ⊗ E = E ( α ∈ { 1 , a , b , c }) . (178)
We also have
E ⊗ E = 1 ⊕ a ⊕ b ⊕ c . (179) E\otimes E=1\oplus a\oplus b\oplus c. \tag*{(179)} E ⊗ E = 1 ⊕ a ⊕ b ⊕ c . (179)
The kernels of these representations are
ker D a = { 1 G , r 2 , s , r 2 s } , ker D b = { 1 G , r , r 2 , r 3 } , \ker D^a=\{1_G,r^2,s,r^2s\},\qquad\ker D^b=\{1_G,r,r^2,r^3\}, ker D a = { 1 G , r 2 , s , r 2 s } , ker D b = { 1 G , r , r 2 , r 3 } ,
ker D c = { 1 G , r 2 , r s , r 3 s } , ker D E = { 1 G } . \ker D^c=\{1_G,r^2,rs,r^3s\},\qquad\ker D^E=\{1_G\}. ker D c = { 1 G , r 2 , r s , r 3 s } , ker D E = { 1 G } .
The three kernels of order four are normal, and none is contained in another. Their intersection is the center Z ( D 4 ) = { 1 G , r 2 } Z(D_4)=\{1_G,r^2\} Z ( D 4 ) = { 1 G , r 2 } .
In the learning problem of the D 4 D_4 D 4 quantum double, we weakly measure four electric charges, S = { a , b , c , E } \mathcal{S}=\{a,b,c,E\} S = { a , b , c , E } . The corresponding classical spin interaction (141) reads
V ( u ) = K a a ( u ) + K b b ( u ) + K c c ( u ) + K E 2 χ E ( u ) . (180) V(u)=K_a a(u)+K_b b(u)+K_c c(u)+\frac{K_E}{2}\chi_E(u). \tag*{(180)} V ( u ) = K a a ( u ) + K b b ( u ) + K c c ( u ) + 2 K E χ E ( u ) . (180)
We again express the interaction in terms of indicator functions of normal subgroups. The five nonconstant indicators are δ ker D α \delta_{\ker D^\alpha} δ k e r D α for α = a , b , c \alpha=a,b,c α = a , b , c , δ Z ( D 4 ) \delta_{Z(D_4)} δ Z ( D 4 ) , and δ { 1 G } \delta_{\{1_G\}} δ { 1 G } . For a one-dimensional character χ α \chi_\alpha χ α (α = a , b , c \alpha=a,b,c α = a , b , c ), it is straightforward to verify that
α ( u ) = 2 δ ker D α ( u ) − 1. \alpha(u)=2\delta_{\ker D^\alpha}(u)-1. α ( u ) = 2 δ k e r D α ( u ) − 1.
Also, one can verify that
χ E ( u ) 2 = 2 δ { 1 G } ( u ) − δ Z ( D 4 ) ( u ) , δ Z ( D 4 ) ( u ) = 1 + a ( u ) + b ( u ) + c ( u ) 4 . \frac{\chi_E(u)}{2}=2\delta_{\{1_G\}}(u)-\delta_{Z(D_4)}(u),\qquad\delta_{Z(D_4)}(u)=\frac{1+a(u)+b(u)+c(u)}{4}. 2 χ E ( u ) = 2 δ { 1 G } ( u ) − δ Z ( D 4 ) ( u ) , δ Z ( D 4 ) ( u ) = 4 1 + a ( u ) + b ( u ) + c ( u ) .
Substituting these into (180) , we have
V ( u ) = 2 K E δ { 1 G } ( u ) − K E 4 + ∑ α = a , b , c ( K α − K E 4 ) α ( u ) . V(u)=2K_E\delta_{\{1_G\}}(u)-\frac{K_E}{4}+\sum_{\alpha=a,b,c}\left(K_\alpha-\frac{K_E}{4}\right)\alpha(u). V ( u ) = 2 K E δ { 1 G } ( u ) − 4 K E + α = a , b , c ∑ ( K α − 4 K E ) α ( u ) .
which is further rewritten as
V ( u ) = c 0 + 2 K E δ { 1 G } ( u ) + ∑ α = a , b , c ( 2 K α − K E 2 ) δ ker D α ( u ) , (181) V(u)=c_0+2K_E\delta_{\{1_G\}}(u)+\sum_{\alpha=a,b,c}\left(2K_\alpha-\frac{K_E}{2}\right)\delta_{\ker D^\alpha}(u), \tag*{(181)} V ( u ) = c 0 + 2 K E δ { 1 G } ( u ) + α = a , b , c ∑ ( 2 K α − 2 K E ) δ k e r D α ( u ) , (181)
with c 0 = − ( K a + K b + K c ) + K E / 2 c_0=-(K_a+K_b+K_c)+K_E/2 c 0 = − ( K a + K b + K c ) + K E /2 .
To apply the hierarchy relation regarding the EA correlators of electric charges, we require
K E ≥ 0 , K a , K b , K c ≥ K E / 4. (182) K_E\ge0,\qquad K_a,K_b,K_c\ge K_E/4. \tag*{(182)} K E ≥ 0 , K a , K b , K c ≥ K E /4. (182)
On the higher Nishimori manifold, these conditions read Δ E ≥ 0 \Delta_E\ge0 Δ E ≥ 0 and Δ a , Δ b , Δ c ≥ Δ E / 4 \Delta_a,\Delta_b,\Delta_c\ge\Delta_E/4 Δ a , Δ b , Δ c ≥ Δ E /4 . In this parameter region, the relation (170) yields
Q a ( x , y ) ≥ Q E ( x , y ) , Q b ( x , y ) ≥ Q E ( x , y ) , Q c ( x , y ) ≥ Q E ( x , y ) . (183) Q_a(x,y)\ge Q_E(x,y),\qquad Q_b(x,y)\ge Q_E(x,y),\qquad Q_c(x,y)\ge Q_E(x,y). \tag*{(183)} Q a ( x , y ) ≥ Q E ( x , y ) , Q b ( x , y ) ≥ Q E ( x , y ) , Q c ( x , y ) ≥ Q E ( x , y ) . (183)
Also, from (172) , one finds
Q c ( x , y ) ≥ Q a ( x , y ) Q b ( x , y ) , Q a ( x , y ) ≥ Q b ( x , y ) Q c ( x , y ) , Q b ( x , y ) ≥ Q c ( x , y ) Q a ( x , y ) . (184) Q_c(x,y)\ge Q_a(x,y)Q_b(x,y),\qquad Q_a(x,y)\ge Q_b(x,y)Q_c(x,y),\qquad Q_b(x,y)\ge Q_c(x,y)Q_a(x,y). \tag*{(184)} Q c ( x , y ) ≥ Q a ( x , y ) Q b ( x , y ) , Q a ( x , y ) ≥ Q b ( x , y ) Q c ( x , y ) , Q b ( x , y ) ≥ Q c ( x , y ) Q a ( x , y ) . (184)
In this parameter region, long range learnability of the non-Abelian electric charge E E E implies that of all three nontrivial Abelian electric charges. The fusion bounds further imply that learnability of any two of a , b , c a,b,c a , b , c entails learnability of the third.
Outlook
In this work, we formulated Bayesian learning under local weak measurements for a broad class of CSS stabilizer codes and extended the analysis to finite group quantum double models. For CSS codes, matching the wavefunction deformation and Gaussian measurement strengths on the higher Nishimori manifold leads to an enlarged S R + 1 S_{R+1} S R + 1 replica permutation symmetry and a local gauge redundancy. These structures yield exact identities among record averaged correlators that quantify the learnability of nonlocal observables associated with fractionalized excitations. In particular, these identities relate the learning correlators to correlations in the corresponding clean models. For nonnegative ferromagnetic couplings, classical correlation inequalities then establish exact hierarchies of learnability: simultaneous long range learnability of two observables implies that of their product.
The rank-2 toric code illustrates this mechanism through spatially modulated symmetries. Acting a lattice translation on a dipole loop operator gives a homogeneous monopole loop operator. Applying the hierarchy to the corresponding open string observables shows that long range dipole learnability implies long range learnability of both associated monopole observables. When the relevant correlators exhibit critical power law scaling, the same inequalities constrain their scaling dimensions. In the X-cube model discussed in Appendix § 10 , the same approach applies to planar subsystem symmetries. Multiplication relations among observables invariant under the subsystem symmetry yield a hierarchy under spatial subdivision and cyclic inequalities among extended observables formed from products of open strings.
For quantum double models, we established an analogous higher Nishimori identity for pure electric observables. When the classical interactions can each be written as a constant plus a sum of indicator functions of normal subgroups with nonnegative coefficients, representation kernels and the fusion rules of anyons with electric charges yield further inequalities among learning correlators. These inequalities apply to both Abelian and non-Abelian electric charges. In the parameter regions considered for the S 3 S_3 S 3 and D 4 D_4 D 4 examples, learning the non-Abelian electric anyons implies learning the associated nontrivial Abelian anyons. Together with the rank-2 toric code results, these examples show how learnability hierarchies reflect both the spatial structure of observables and the electric charges they probe. Operator multiplication and fusion thus constrain which types of nonlocal information can be learned simultaneously, providing a guide to learning phase diagrams and critical behavior in quantum codes.
Several questions remain open. One direction is to determine which aspects of the hierarchy survive away from the exactly matched higher Nishimori manifold and under more general measurement protocols. In Appendix § 13 , we briefly argue that the decoupled critical point of the rank-2 toric code is unstable under plaquette deformation and measurement. The hierarchy remains exact when the higher Nishimori condition is preserved, but its persistence away from this condition remains an open question. In particular, weaker constraints on simultaneous long range learnability may survive even when the exact correlator inequalities are no longer guaranteed. A second direction is to extend the quantum double analysis beyond pure electric observables to observables associated with magnetic and dyonic excitations. More broadly, it would be interesting to investigate analogous learning problems in topological models with categorical or higher group symmetries, and to determine whether their operator and fusion structures give rise to corresponding identities and learnability hierarchies.
Acknowledgments
We thank W. Cao, J. Haruna, B. Han, K. Kato, R. A. Patil, S. Trebst, H. Yan, G. Y. Zhu for helpful discussions. We also thank Bo Han for comments on the draft. This work is in part supported by JST CREST (Grant No. JPMJCR24I3), JSPS KAKENHI Grant Number JP22H01222 and the Royal Society grants ICA/R2/242058 and IEC/R3/243103. We are also grateful for the valuable discussions and feedback from the participants of the UK-Japan Workshop on Quantum Gravity 2026 at Durham University.
Griffiths–Kelly–Sherman inequality
In this appendix, we briefly review the Griffiths–Kelly–Sherman (GKS) inequalities [33 , 34 ] , which are used in § 3.3 to derive the hierarchy among record averaged correlators.
Consider Ising spins σ i = ± 1 \sigma_i=\pm1 σ i = ± 1 on a finite lattice Λ \Lambda Λ , with Hamiltonian
H ( σ ) = − ∑ X ⊆ Λ K X σ X , σ X : = ∏ i ∈ X σ i , K X ≥ 0. (185) H(\sigma)=-\sum_{X\subseteq\Lambda}K_X\sigma_X,\qquad\sigma_X:=\prod_{i\in X}\sigma_i,\qquad K_X\geq0. \tag*{(185)} H ( σ ) = − X ⊆ Λ ∑ K X σ X , σ X := i ∈ X ∏ σ i , K X ≥ 0. (185)
Here X X X denotes an arbitrary subset of lattice sites. The thermal expectation value of an observable F ( σ ) F(\sigma) F ( σ ) is
⟨ F ⟩ : = 1 Z ∑ σ F ( σ ) exp ( ∑ X ⊆ Λ K X σ X ) , (186) \langle F\rangle:=\frac{1}{Z}\sum_{\sigma}F(\sigma)\exp\left(\sum_{X\subseteq\Lambda}K_X\sigma_X\right), \tag*{(186)} ⟨ F ⟩ := Z 1 σ ∑ F ( σ ) exp ( X ⊆ Λ ∑ K X σ X ) , (186)
where
Z : = ∑ σ exp ( ∑ X ⊆ Λ K X σ X ) . (187) Z:=\sum_{\sigma}\exp\left(\sum_{X\subseteq\Lambda}K_X\sigma_X\right). \tag*{(187)} Z := σ ∑ exp ( X ⊆ Λ ∑ K X σ X ) . (187)
The first GKS inequality states that
⟨ σ A ⟩ ≥ 0 (188) \langle\sigma_A\rangle\geq0 \tag*{(188)} ⟨ σ A ⟩ ≥ 0 (188)
for every A ⊆ Λ A\subseteq\Lambda A ⊆ Λ . To see this, expand the Boltzmann weight as
Z ⟨ σ A ⟩ = ∑ σ σ A ∏ X ⊆ Λ e K X σ X = ∑ { n X ≥ 0 } ( ∏ X ⊆ Λ K X n X n X ! ) ∑ σ σ A ∏ X ⊆ Λ σ X n X . (189) Z\langle\sigma_A\rangle=\sum_{\sigma}\sigma^A\prod_{X\subseteq\Lambda}e^{K_X\sigma_X}=\sum_{\{n_X\geq0\}}\left(\prod_{X\subseteq\Lambda}\frac{K_X^{n_X}}{n_X!}\right)\sum_{\sigma}\sigma^A\prod_{X\subseteq\Lambda}\sigma_X^{n_X}. \tag*{(189)} Z ⟨ σ A ⟩ = σ ∑ σ A X ⊆ Λ ∏ e K X σ X = { n X ≥ 0 } ∑ ( X ⊆ Λ ∏ n X ! K X n X ) σ ∑ σ A X ⊆ Λ ∏ σ X n X . (189)
For each choice of { n X } \{n_X\} { n X } , the spin sum takes the form
∑ σ ∏ i ∈ Λ σ i m i = { 2 ∣ Λ ∣ , m i even for every i , 0 , otherwise . (190) \sum_{\sigma}\prod_{i\in\Lambda}\sigma_i^{m_i}=
\begin{cases}
2^{|\Lambda|}, & m_i\ \text{even for every }i,\\
0, & \text{otherwise}.
\end{cases} \tag*{(190)} σ ∑ i ∈ Λ ∏ σ i m i = { 2 ∣Λ∣ , 0 , m i even for every i , otherwise . (190)
Since every coefficient K X n X / n X ! K_X^{n_X}/n_X! K X n X / n X ! is nonnegative, every term in the expansion is nonnegative, proving the first inequality.
The second GKS inequality states that, for arbitrary A , B ⊆ Λ A,B \subseteq\Lambda A , B ⊆ Λ ,
⟨ σ A σ B ⟩ − ⟨ σ A ⟩ ⟨ σ B ⟩ ≥ 0. (191) \langle\sigma_A \sigma_B\rangle-\langle\sigma_A\rangle\langle\sigma_B\rangle\ge0. \tag*{(191)} ⟨ σ A σ B ⟩ − ⟨ σ A ⟩ ⟨ σ B ⟩ ≥ 0. (191)
A convenient proof uses two independent copies of the spin configuration, denoted by σ \boldsymbol{\sigma} σ and τ \boldsymbol{\tau} τ . The unnormalized connected correlator can be written as
Z 2 ( ⟨ σ A σ B ⟩ − ⟨ σ A ⟩ ⟨ σ B ⟩ ) = 1 2 ∑ σ , τ ( σ A − τ A ) ( σ B − τ B ) exp [ ∑ X ⊆ Λ K X ( σ X + τ X ) ] . (192) Z^2\left(\langle\sigma_A\sigma_B\rangle-\langle\sigma_A\rangle\langle\sigma_B\rangle\right)=\frac{1}{2}\sum_{\boldsymbol{\sigma},\boldsymbol{\tau}}(\sigma_A-\tau_A)(\sigma_B-\tau_B)\exp\left[\sum_{X\subseteq\Lambda}K_X(\sigma_X+\tau_X)\right]. \tag*{(192)} Z 2 ( ⟨ σ A σ B ⟩ − ⟨ σ A ⟩ ⟨ σ B ⟩ ) = 2 1 σ , τ ∑ ( σ A − τ A ) ( σ B − τ B ) exp [ X ⊆ Λ ∑ K X ( σ X + τ X ) ] . (192)
We introduce variables as
u i : = σ i + τ i 2 , v i : = σ i − τ i 2 . (193) u_i:=\frac{\sigma_i+\tau_i}{2},\qquad v_i:=\frac{\sigma_i-\tau_i}{2}. \tag*{(193)} u i := 2 σ i + τ i , v i := 2 σ i − τ i . (193)
For any C ⊆ Λ C\subseteq\Lambda C ⊆ Λ , one has
σ C + τ C = 2 ∑ D ⊆ C ∣ D ∣ e v e n u C ∖ D v D , σ C − τ C = 2 ∑ D ⊆ C ∣ D ∣ o d d u C ∖ D v D , (194) \sigma_C+\tau_C=2\sum_{\substack{D\subseteq C\\ |D|\ \mathrm{even}}}u_{C\setminus D}v_D,\qquad\sigma_C-\tau_C=2\sum_{\substack{D\subseteq C\\ |D|\ \mathrm{odd}}}u_{C\setminus D}v_D, \tag*{(194)} σ C + τ C = 2 D ⊆ C ∣ D ∣ even ∑ u C ∖ D v D , σ C − τ C = 2 D ⊆ C ∣ D ∣ odd ∑ u C ∖ D v D , (194)
where
u D : = ∏ i ∈ D u i , v D : = ∏ i ∈ D v i . (195) u_D:=\prod_{i\in D}u_i,\qquad v_D:=\prod_{i\in D}v_i. \tag*{(195)} u D := i ∈ D ∏ u i , v D := i ∈ D ∏ v i . (195)
All coefficients in these expansions are nonnegative. Expanding the exponential in (192) in powers of K X K_X K X therefore expresses its right-hand side as a sum of monomials in u i u_i u i and v i v_i v i with nonnegative coefficients.
For each site,
( u i , v i ) ∈ { ( 1 , 0 ) , ( − 1 , 0 ) , ( 0 , 1 ) , ( 0 , − 1 ) } . (u_i,v_i)\in\{(1,0),(-1,0),(0,1),(0,-1)\}. ( u i , v i ) ∈ {( 1 , 0 ) , ( − 1 , 0 ) , ( 0 , 1 ) , ( 0 , − 1 )} .
The sum of any monomial u i p v i q u_i^p v_i^q u i p v i q over these four possibilities is nonnegative: it vanishes if an exponent is odd or if both p , q > 0 p,q>0 p , q > 0 , and is positive otherwise. Consequently, every term that survives the sum over σ \boldsymbol{\sigma} σ and τ \boldsymbol{\tau} τ is nonnegative, proving (191) .
In the notation of § 3 , taking σ A = O v \sigma_A=O_v σ A = O v and σ B = O w \sigma_B=O_w σ B = O w gives
⟨ O v O w ⟩ c l e a n ≥ ⟨ O v ⟩ c l e a n ⟨ O w ⟩ c l e a n . (196) \langle O_vO_w\rangle_{\mathrm{clean}}\ge\langle O_v\rangle_{\mathrm{clean}}\langle O_w\rangle_{\mathrm{clean}}. \tag*{(196)} ⟨ O v O w ⟩ clean ≥ ⟨ O v ⟩ clean ⟨ O w ⟩ clean . (196)
Since O v O w = O v ⊕ w O_vO_w=O_{v\oplus w} O v O w = O v ⊕ w , this is precisely the inequality used to obtain the exact correlator hierarchy in Theorem 1.
Weak measurements of Pauli X in the rank-2 toric code
In this appendix, we consider weak measurements of local Pauli-X operators in the rank-2 toric code, with the wavefunction deformation also taken in the X X X basis. In contrast to the Pauli-Z Z Z measurement problem discussed in § 4.2.2 , the classical Ising variables are now associated with the two species of Z Z Z -type stabilizers. We introduce an Ising spin s x , y s_{x,y} s x , y on the horizontal link centered at ( x + 1 2 , y ) (x+\frac{1}{2},y) ( x + 2 1 , y ) and an Ising spin t x , y t_{x,y} t x , y on the vertical link centered at ( x , y + 1 2 ) (x,y+\frac{1}{2}) ( x , y + 2 1 ) :
s x , y : = τ ( x + 1 2 , y ) , t x , y : = τ ( x , y + 1 2 ) . (197) s_{x,y}:=\tau_{(x+\frac{1}{2},y)},\qquad t_{x,y}:=\tau_{(x,y+\frac{1}{2})}. \tag*{(197)} s x , y := τ ( x + 2 1 , y ) , t x , y := τ ( x , y + 2 1 ) . (197)
A local Pauli-X X X operator gives a classical interaction consisting of the Ising spins associated with the Z Z Z -type stabilizers with which it anticommutes. The three species of physical qubits therefore give
b 1 , x , y ( s , t ) = t x , y − 1 t x , y , (198) b_{1,x,y}(\boldsymbol{s},t)=t_{x,y-1}t_{x,y}, \tag*{(198)} b 1 , x , y ( s , t ) = t x , y − 1 t x , y , (198)
b 2 , x , y ( s , t ) = s x − 1 , y s x , y , (199) b_{2,x,y}(s,\boldsymbol{t})=s_{x-1,y}s_{x,y}, \tag*{(199)} b 2 , x , y ( s , t ) = s x − 1 , y s x , y , (199)
b 0 , x , y ( s , t ) = s x , y s x , y + 1 t x , y t x + 1 , y . (200) b_{0,x,y}(s,t)=s_{x,y}s_{x,y+1}t_{x,y}t_{x+1,y}. \tag*{(200)} b 0 , x , y ( s , t ) = s x , y s x , y + 1 t x , y t x + 1 , y . (200)
Here ( x , y ) (x,y) ( x , y ) in (200) labels the lower-left corner of the plaquette hosting the type-0 qubit. Thus, X 1 X_1 X 1 and X 2 X_2 X 2 generate two-spin interactions between neighboring collinear edges, whereas X 0 X_0 X 0 generates a four-spin interaction surrounding a plaquette.
The associated clean classical model is
H X ( s , t ) = − ∑ x , y [ K 1 t x , y − 1 t x , y + K 2 s x − 1 , y s x , y + K 0 s x , y s x , y + 1 t x , y t x + 1 , y ] . (201) H_X(s,t)=-\sum_{x,y}\left[K_1t_{x,y-1}t_{x,y}+K_2s_{x-1,y}s_{x,y}+K_0s_{x,y}s_{x,y+1}t_{x,y}t_{x+1,y}\right]. \tag*{(201)} H X ( s , t ) = − x , y ∑ [ K 1 t x , y − 1 t x , y + K 2 s x − 1 , y s x , y + K 0 s x , y s x , y + 1 t x , y t x + 1 , y ] . (201)
where K a ≥ 0 K_a\geq0 K a ≥ 0 . The model is invariant under
s x , y ⟼ η s η D y s x , y , t x , y ⟼ η t η D x t x , y , η s , η t , η D ∈ { ± 1 } . (202) s_{x,y}\longmapsto\eta_s\eta_D^y s_{x,y},\qquad t_{x,y}\longmapsto\eta_t\eta_D^x t_{x,y},\qquad\eta_s,\eta_t,\eta_D\in\{\pm1\}. \tag*{(202)} s x , y ⟼ η s η D y s x , y , t x , y ⟼ η t η D x t x , y , η s , η t , η D ∈ { ± 1 } . (202)
The first two transformations are uniform spin flips, while the last one is a spatially modulated symmetry. We assume an infinite lattice or periodic boundary conditions compatible with (202) .
For Gaussian measurements with strengths Δ 0 , Δ 1 , Δ 2 \Delta_0,\Delta_1,\Delta_2 Δ 0 , Δ 1 , Δ 2 , the replica theory obtained by averaging over measurement records is
− H R ( X ) = ∑ a = 0 , 1 , 2 ∑ q [ K a ∑ α = 1 R b a , q ( α ) + Δ a 2 ∑ α , β = 1 α ≠ β R b a , q ( α ) b a , q ( β ) ] , (203) -H_R^{(X)}=\sum_{a=0,1,2}\sum_q\left[K_a\sum_{\alpha=1}^{R}b_{a,q}^{(\alpha)}+\frac{\Delta_a}{2}\sum_{\substack{\alpha,\beta=1\\\alpha\ne\beta}}^{R}b_{a,q}^{(\alpha)}b_{a,q}^{(\beta)}\right], \tag*{(203)} − H R ( X ) = a = 0 , 1 , 2 ∑ q ∑ K a α = 1 ∑ R b a , q ( α ) + 2 Δ a α , β = 1 α = β ∑ R b a , q ( α ) b a , q ( β ) , (203)
where q q q runs over the appropriate vertices or plaquettes. The higher Nishimori manifold is
Δ a = K a , a = 0 , 1 , 2. (204) \Delta_a=K_a,\qquad a=0,1,2. \tag*{(204)} Δ a = K a , a = 0 , 1 , 2. (204)
Since all the interactions in (198) –(200) are Ising monomials, the enlarged replica symmetry and the exact higher Nishimori identities derived in Sec. 3 apply without modification.
There are two spatially modulated loops consisting of Pauli X X X operators, which we denote by W D , x W_{D,x} W D , x and W D , y W_{D,y} W D , y . Defining C C C be the rectangular contour with corners ( x L , y B ) (x_L,y_B) ( x L , y B ) , ( x R , y B ) (x_R,y_B) ( x R , y B ) , ( x R , y T ) (x_R,y_T) ( x R , y T ) , and ( x L , y T ) (x_L,y_T) ( x L , y T ) where
x L = x 0 , x R = x 0 + 2 r x , y B = y 0 , y T = y 0 + 2 r y , x_L=x_0,\qquad x_R=x_0+2r_x,\qquad y_B=y_0,\qquad y_T=y_0+2r_y, x L = x 0 , x R = x 0 + 2 r x , y B = y 0 , y T = y 0 + 2 r y ,
and all the coordinates being even integers, the spatially modulated loops are given by
W D , x = ∏ i = x L x R [ X 2 , ( i , y B ) X 2 , ( i , y B + 1 ) ] i × ∏ i = x L x R − 1 X 0 , ( i + 1 / 2 , y B + 1 / 2 ) × ∏ j = y B + 1 y T − 1 X 1 , ( x R , j ) × ∏ i = x L x R [ X 2 , ( i , y T − 1 ) X 2 , ( i , y T ) ] i × ∏ i = x L x R − 1 X 0 , ( i + 1 / 2 , y T − 1 / 2 ) × ∏ j = y B + 1 y T − 1 X 1 , ( x L , j ) , W D , y = ∏ i = x L + 1 x R − 1 X 2 , ( i , y B ) × ∏ j = y B + 1 y T − 1 [ X 1 , ( x R − 1 , j ) X 1 , ( x R , j ) ] j × ∏ j = y B y T − 1 X 0 , ( x R − 1 / 2 , j + 1 / 2 ) × ∏ i = x L + 1 x R − 1 X 2 , ( i , y T ) × ∏ j = y B + 1 y T − 1 [ X 1 , ( x L , j ) X 1 , ( x L + 1 , j ) ] j × ∏ j = y B y T − 1 X 0 , ( x L + 1 / 2 , j + 1 / 2 ) . (205) \begin{aligned}
W_{D,x}={}&\prod_{i=x_L}^{x_R}\left[X_{2,(i,y_B)}X_{2,(i,y_B+1)}\right]^i\times\prod_{i=x_L}^{x_R-1}X_{0,(i+1/2,y_B+1/2)}\times\prod_{j=y_B+1}^{y_T-1}X_{1,(x_R,j)}\\
&\times\prod_{i=x_L}^{x_R}\left[X_{2,(i,y_T-1)}X_{2,(i,y_T)}\right]^i\times\prod_{i=x_L}^{x_R-1}X_{0,(i+1/2,y_T-1/2)}\times\prod_{j=y_B+1}^{y_T-1}X_{1,(x_L,j)},\\
W_{D,y}={}&\prod_{i=x_L+1}^{x_R-1}X_{2,(i,y_B)}\times\prod_{j=y_B+1}^{y_T-1}\left[X_{1,(x_R-1,j)}X_{1,(x_R,j)}\right]^j\times\prod_{j=y_B}^{y_T-1}X_{0,(x_R-1/2,j+1/2)}\\
&\times\prod_{i=x_L+1}^{x_R-1}X_{2,(i,y_T)}\times\prod_{j=y_B+1}^{y_T-1}\left[X_{1,(x_L,j)}X_{1,(x_L+1,j)}\right]^j\times\prod_{j=y_B}^{y_T-1}X_{0,(x_L+1/2,j+1/2)}. \tag*{(205)}
\end{aligned} W D , x = W D , y = i = x L ∏ x R [ X 2 , ( i , y B ) X 2 , ( i , y B + 1 ) ] i × i = x L ∏ x R − 1 X 0 , ( i + 1/2 , y B + 1/2 ) × j = y B + 1 ∏ y T − 1 X 1 , ( x R , j ) × i = x L ∏ x R [ X 2 , ( i , y T − 1 ) X 2 , ( i , y T ) ] i × i = x L ∏ x R − 1 X 0 , ( i + 1/2 , y T − 1/2 ) × j = y B + 1 ∏ y T − 1 X 1 , ( x L , j ) , i = x L + 1 ∏ x R − 1 X 2 , ( i , y B ) × j = y B + 1 ∏ y T − 1 [ X 1 , ( x R − 1 , j ) X 1 , ( x R , j ) ] j × j = y B ∏ y T − 1 X 0 , ( x R − 1/2 , j + 1/2 ) × i = x L + 1 ∏ x R − 1 X 2 , ( i , y T ) × j = y B + 1 ∏ y T − 1 [ X 1 , ( x L , j ) X 1 , ( x L + 1 , j ) ] j × j = y B ∏ y T − 1 X 0 , ( x L + 1/2 , j + 1/2 ) . (205)
Acting translation operators on the modulated loops generates a homogeneous loop. Indeed, on the projected sector which does not admit excess electric charges G r = + 1 ∀ r G_r=+1\ \forall r G r = + 1 ∀ r , we obtain
T x W D , x T x − 1 = W D , x W H , T y W D , y T y − 1 = W D , y W H ′ , (206) T_xW_{D,x}T_x^{-1}=W_{D,x}W_H,\qquad T_yW_{D,y}T_y^{-1}=W_{D,y}W_{H'}, \tag*{(206)} T x W D , x T x − 1 = W D , x W H , T y W D , y T y − 1 = W D , y W H ′ , (206)
where
W H = ∏ i = x L + 1 x R X 2 , ( i , y B ) X 2 , ( i , y B + 1 ) × ∏ j = y B + 1 y T − 1 X 1 , ( x R , j ) X 1 , ( x R + 1 , j ) × X 0 , ( x R + 1 / 2 , y B + 1 / 2 ) X 0 , ( x R + 1 / 2 , y T − 1 / 2 ) × ∏ i = x L + 1 x R X 2 , ( i , y T − 1 ) X 2 , ( i , y T ) × ∏ j = y B + 1 y T − 1 X 1 , ( x L , j ) X 1 , ( x L + 1 , j ) × X 0 , ( x L + 1 / 2 , y B + 1 / 2 ) X 0 , ( x L + 1 / 2 , y T − 1 / 2 ) , W H ′ = ∏ i = x L + 1 x R − 1 X 2 , ( i , y B ) X 2 , ( i , y B + 1 ) × ∏ j = y B + 1 y T X 1 , ( x R − 1 , j ) X 1 , ( x R , j ) × X 0 , ( x R − 1 / 2 , y B + 1 / 2 ) X 0 , ( x R − 1 / 2 , y T + 1 / 2 ) × ∏ i = x L + 1 x R − 1 X 2 , ( i , y T ) X 2 , ( i , y T + 1 ) × ∏ j = y B + 1 y T X 1 , ( x L , j ) X 1 , ( x L + 1 , j ) × X 0 , ( x L + 1 / 2 , y B + 1 / 2 ) X 0 , ( x L + 1 / 2 , y T + 1 / 2 ) . (207) \begin{aligned}
W_H={}&\prod_{i=x_L+1}^{x_R}X_{2,(i,y_B)}X_{2,(i,y_B+1)}\times\prod_{j=y_B+1}^{y_T-1}X_{1,(x_R,j)}X_{1,(x_R+1,j)}\times X_{0,(x_R+1/2,y_B+1/2)}X_{0,(x_R+1/2,y_T-1/2)}\\
&\times\prod_{i=x_L+1}^{x_R}X_{2,(i,y_T-1)}X_{2,(i,y_T)}\times\prod_{j=y_B+1}^{y_T-1}X_{1,(x_L,j)}X_{1,(x_L+1,j)}\times X_{0,(x_L+1/2,y_B+1/2)}X_{0,(x_L+1/2,y_T-1/2)},\\
W_{H'}={}&\prod_{i=x_L+1}^{x_R-1}X_{2,(i,y_B)}X_{2,(i,y_B+1)}\times\prod_{j=y_B+1}^{y_T}X_{1,(x_R-1,j)}X_{1,(x_R,j)}\times X_{0,(x_R-1/2,y_B+1/2)}X_{0,(x_R-1/2,y_T+1/2)}\\
&\times\prod_{i=x_L+1}^{x_R-1}X_{2,(i,y_T)}X_{2,(i,y_T+1)}\times\prod_{j=y_B+1}^{y_T}X_{1,(x_L,j)}X_{1,(x_L+1,j)}\times X_{0,(x_L+1/2,y_B+1/2)}X_{0,(x_L+1/2,y_T+1/2)}. \tag*{(207)}
\end{aligned} W H = W H ′ = i = x L + 1 ∏ x R X 2 , ( i , y B ) X 2 , ( i , y B + 1 ) × j = y B + 1 ∏ y T − 1 X 1 , ( x R , j ) X 1 , ( x R + 1 , j ) × X 0 , ( x R + 1/2 , y B + 1/2 ) X 0 , ( x R + 1/2 , y T − 1/2 ) × i = x L + 1 ∏ x R X 2 , ( i , y T − 1 ) X 2 , ( i , y T ) × j = y B + 1 ∏ y T − 1 X 1 , ( x L , j ) X 1 , ( x L + 1 , j ) × X 0 , ( x L + 1/2 , y B + 1/2 ) X 0 , ( x L + 1/2 , y T − 1/2 ) , i = x L + 1 ∏ x R − 1 X 2 , ( i , y B ) X 2 , ( i , y B + 1 ) × j = y B + 1 ∏ y T X 1 , ( x R − 1 , j ) X 1 , ( x R , j ) × X 0 , ( x R − 1/2 , y B + 1/2 ) X 0 , ( x R − 1/2 , y T + 1/2 ) × i = x L + 1 ∏ x R − 1 X 2 , ( i , y T ) X 2 , ( i , y T + 1 ) × j = y B + 1 ∏ y T X 1 , ( x L , j ) X 1 , ( x L + 1 , j ) × X 0 , ( x L + 1/2 , y B + 1/2 ) X 0 , ( x L + 1/2 , y T + 1/2 ) . (207)
The loops W H W_H W H and W H ′ W_{H'} W H ′ are associated with the homogeneous loop. We regard these two loops as representing the same charge type, corresponding to the fact that there is only a single type of homogeneous magnetic superselection sector described by noncontractible loops (see [39 ] ). In contrast to the Pauli-Z sector, where a single dipole loop has two distinct homogeneous descendants, the Pauli-X sector contains two modulated loops whose appropriate lattice translates generate the same homogeneous loop.
Similar to the previous argument (§ 4.2.2 ), the classical images of the open counterparts of the two modulated loops (205) are described by
O D , x ( r , r ′ ) : = t r t r ′ , O D , y ( r , r ′ ) : = s r s r ′ , (208) O_{D,x}(r,r'):=t_rt_{r'},\qquad O_{D,y}(r,r'):=s_rs_{r'}, \tag*{(208)} O D , x ( r , r ′ ) := t r t r ′ , O D , y ( r , r ′ ) := s r s r ′ , (208)
where r r r and r ′ r' r ′ are chosen on the same parity sublattice. Likewise, the classical image of the homogeneous loop reads
O H ( r , r ′ ) : = t r t r + x ^ t r ′ t r ′ + x ^ = O D , x ( r , r ′ ) O D , x ( r + x ^ , r ′ + x ^ ) (209) O_H(r,r'):=t_rt_{r+\hat{x}}t_{r'}t_{r'+\hat{x}}=O_{D,x}(r,r')O_{D,x}(r+\hat{x},r'+\hat{x}) \tag*{(209)} O H ( r , r ′ ) := t r t r + x ^ t r ′ t r ′ + x ^ = O D , x ( r , r ′ ) O D , x ( r + x ^ , r ′ + x ^ ) (209)
O H ′ ( r , r ′ ) : = s r s r + y ^ s r ′ s r ′ + y ^ = O D , y ( r , r ′ ) O D , y ( r + y ^ , r ′ + y ^ ) (210) O_{H'}(r,r'):=s_rs_{r+\hat{y}}s_{r'}s_{r'+\hat{y}}=O_{D,y}(r,r')O_{D,y}(r+\hat{y},r'+\hat{y}) \tag*{(210)} O H ′ ( r , r ′ ) := s r s r + y ^ s r ′ s r ′ + y ^ = O D , y ( r , r ′ ) O D , y ( r + y ^ , r ′ + y ^ ) (210)
with the relation
O H ( r , r ′ ) = b 0 , r b 0 , r ′ O H ′ ( r , r ′ ) . (211) O_H(r,r')=b_{0,r}b_{0,r'}O_{H'}(r,r'). \tag*{(211)} O H ( r , r ′ ) = b 0 , r b 0 , r ′ O H ′ ( r , r ′ ) . (211)
To discuss the learnability of the loops, we define the EA-type correlators as
C D , x ( r , r ′ ) : = ⟨ O D , x ( r , r ′ ) ⟩ m 1 / 2 , C D , y ( r , r ′ ) : = ⟨ O D , y ( r , r ′ ) ⟩ m 1 / 2 , (212) C_{D,x}(r,r'):=\langle O_{D,x}(r,r')\rangle_m^{1/2},\qquad C_{D,y}(r,r'):=\langle O_{D,y}(r,r')\rangle_m^{1/2}, \tag*{(212)} C D , x ( r , r ′ ) := ⟨ O D , x ( r , r ′ ) ⟩ m 1/2 , C D , y ( r , r ′ ) := ⟨ O D , y ( r , r ′ ) ⟩ m 1/2 , (212)
C H ( r , r ′ ) : = ⟨ O H ( r , r ′ ) ⟩ m 1 / 2 , C H ′ ( r , r ′ ) : = ⟨ O H ′ ( r , r ′ ) ⟩ m 1 / 2 . (213) C_H(r,r'):=\langle O_H(r,r')\rangle_m^{1/2},\qquad C_{H'}(r,r'):=\langle O_{H'}(r,r')\rangle_m^{1/2}. \tag*{(213)} C H ( r , r ′ ) := ⟨ O H ( r , r ′ ) ⟩ m 1/2 , C H ′ ( r , r ′ ) := ⟨ O H ′ ( r , r ′ ) ⟩ m 1/2 . (213)
On the higher Nishimori manifold, these correlators are equal to the corresponding clean correlation functions. Applying the GKS inequality (Appendix § 8 ) to (209) and (210) , and using translation invariance, gives
C H ( r , r ′ ) ≥ [ C D , x ( r , r ′ ) ] 2 , C H ′ ( r , r ′ ) ≥ [ C D , y ( r , r ′ ) ] 2 . (214) C_H(r,r')\geq\left[C_{D,x}(r,r')\right]^2,\qquad C_{H'}(r,r')\geq\left[C_{D,y}(r,r')\right]^2. \tag*{(214)} C H ( r , r ′ ) ≥ [ C D , x ( r , r ′ ) ] 2 , C H ′ ( r , r ′ ) ≥ [ C D , y ( r , r ′ ) ] 2 . (214)
For K 0 > 0 K_0>0 K 0 > 0 , one can further make use of the GKS inequality to find
C H ( r , r ′ ) ≥ C H ′ ( r , r ′ ) ⟨ b 0 , r b 0 , r ′ ⟩ c l e a n ≥ C H ′ ( r , r ′ ) tanh 2 K 0 . (215) C_H(r,r') \ge C_{H'}(r,r')\langle b_{0,r}b_{0,r'}\rangle_{\mathrm{clean}} \ge C_{H'}(r,r')\tanh^2 K_0. \tag*{(215)} C H ( r , r ′ ) ≥ C H ′ ( r , r ′ ) ⟨ b 0 , r b 0 , r ′ ⟩ clean ≥ C H ′ ( r , r ′ ) tanh 2 K 0 . (215)
Analogous line of thoughts leads to C H ′ ( r , r ′ ) ≥ C H ( r , r ′ ) tanh 2 K 0 C_{H'}(r,r') \ge C_H(r,r')\tanh^2 K_0 C H ′ ( r , r ′ ) ≥ C H ( r , r ′ ) tanh 2 K 0 . Overall, we have
C H ( r , r ′ ) ≥ [ C D , x ( r , r ′ ) ] 2 , C H ′ ( r , r ′ ) ≥ tanh 2 K 0 [ C D , x ( r , r ′ ) ] 2 , C H ′ ( r , r ′ ) ≥ [ C D , y ( r , r ′ ) ] 2 , C H ( r , r ′ ) ≥ tanh 2 K 0 [ C D , y ( r , r ′ ) ] 2 . (216) \begin{aligned}
C_H(r,r') &\ge[C_{D,x}(r,r')]^2,\qquad C_{H'}(r,r') \ge\tanh^2 K_0 [C_{D,x}(r,r')]^2,\\
C_{H'}(r,r') &\ge[C_{D,y}(r,r')]^2,\qquad C_H(r,r') \ge\tanh^2 K_0 [C_{D,y}(r,r')]^2. \tag*{(216)}
\end{aligned} C H ( r , r ′ ) C H ′ ( r , r ′ ) ≥ [ C D , x ( r , r ′ ) ] 2 , C H ′ ( r , r ′ ) ≥ tanh 2 K 0 [ C D , x ( r , r ′ ) ] 2 , ≥ [ C D , y ( r , r ′ ) ] 2 , C H ( r , r ′ ) ≥ tanh 2 K 0 [ C D , y ( r , r ′ ) ] 2 . (216)
Therefore, long range learnability of either modulated loop necessarily implies long range learnability of homogeneous descendant. The converse does not follow from these inequalities.
X-cube model
In this appendix, we consider the learning problem for the X-cube stabilizer code. Qubits reside on the links of a cubic lattice and the Hamiltonian is given by [26 ]
H X C = − ∑ c A c − ∑ v , μ B v μ (217) H_{\mathrm{XC}}=-\sum_c A_c-\sum_{v,\mu}B_v^\mu \tag*{(217)} H XC = − c ∑ A c − v , μ ∑ B v μ (217)
where A c = ∏ e ∈ ∂ c X e A_c=\prod_{e\in\partial c}X_e A c = ∏ e ∈ ∂ c X e is the product over the twelve edges of cube c c c , while B v μ B_v^\mu B v μ is the product of four Z Z Z operators in the plane normal to μ \mu μ and adjacent to v v v .
Weak measurements of Pauli Z
We consider local Pauli-Z Z Z measurements. For convenience, we use the dual cubic lattice, in which cube centers become vertices and edges correspond to plaquettes. The qubits therefore reside on plaquettes, and the Hamiltonian becomes
H ^ X C = − ∑ v A ^ v − ∑ c , a B ^ c , a ( a = x , y , z ) (218) \widehat{H}_{\mathrm{XC}}=-\sum_v\widehat{A}_v-\sum_{c,a}\widehat{B}_{c,a}\quad(a=x,y,z) \tag*{(218)} H XC = − v ∑ A v − c , a ∑ B c , a ( a = x , y , z ) (218)
where A ^ v = ∏ v ⊂ p X p \widehat{A}_v=\prod_{v\subset p}X_p A v = ∏ v ⊂ p X p denotes the product over the twelve plaquettes which share the corner v v v . The other operator is defined as
B ^ c , x : = Z c − y ^ / 2 Z c + y ^ / 2 Z c − z ^ / 2 Z c + z ^ / 2 . (219) \widehat{B}_{c,x}:=Z_{c-\hat{y}/2}Z_{c+\hat{y}/2}Z_{c-\hat{z}/2}Z_{c+\hat{z}/2}. \tag*{(219)} B c , x := Z c − y ^ /2 Z c + y ^ /2 Z c − z ^ /2 Z c + z ^ /2 . (219)
This operator is the product of Z Z Z operators on four plaquettes on the boundary of the cube. The operators B ^ c , y \widehat{B}_{c,y} B c , y and B ^ c , z \widehat{B}_{c,z} B c , z are defined by cyclically permuting x x x , y y y , and z z z .
The corresponding classical spin model is then described by three-dimensional plaquette Ising model:
P K ( σ ) = 1 Z K exp ( ∑ p K p B p ( σ ) ) , B p ( σ ) = ∏ r ∈ corners ( p ) σ r , K p ≥ 0. (220) P_K(\boldsymbol{\sigma})=\frac{1}{Z_K}\exp\left(\sum_p K_pB_p(\boldsymbol{\sigma})\right),\qquad B_p(\boldsymbol{\sigma})=\prod_{r\in\operatorname{corners}(p)}\sigma_r,\qquad K_p\ge0. \tag*{(220)} P K ( σ ) = Z K 1 exp ( p ∑ K p B p ( σ ) ) , B p ( σ ) = r ∈ corners ( p ) ∏ σ r , K p ≥ 0. (220)
Here r = ( x ^ , y ^ , z ^ ) r=(\hat{x},\hat{y},\hat{z}) r = ( x ^ , y ^ , z ^ ) labels a node of a lattice. The distribution is invariant under independent flips of coordinate planes,
σ r ⟼ η x ( x ^ ) η y ( y ^ ) η z ( z ^ ) σ r , η μ = ± 1 , (221) \sigma_r\longmapsto\eta_x(\hat{x})\eta_y(\hat{y})\eta_z(\hat{z})\sigma_r,\qquad\eta_\mu=\pm1, \tag*{(221)} σ r ⟼ η x ( x ^ ) η y ( y ^ ) η z ( z ^ ) σ r , η μ = ± 1 , (221)
where η x ( x ) = − 1 \eta_x(x)=-1 η x ( x ) = − 1 corresponds to a spin flip on a y z yz y z plane at fixed x x x .
For x ^ 0 < x ^ 1 \hat{x}_0 < \hat{x}_1 x ^ 0 < x ^ 1 , y ^ 0 < y ^ 1 \hat{y}_0 < \hat{y}_1 y ^ 0 < y ^ 1 , z ^ 0 < z ^ 1 \hat{z}_0 < \hat{z}_1 z ^ 0 < z ^ 1 , the Hamiltonian (218) admits the following loops operator:
W z = ∏ z ^ 0 ≤ z ^ ≤ z ^ 1 − 1 ∏ x ^ 0 ≤ x ^ ≤ x ^ 1 Z ( x ^ + 1 / 2 , y ^ 0 , z ^ + 1 / 2 ) Z ( x ^ + 1 / 2 , y ^ 1 , z ^ + 1 / 2 ) × ∏ z ^ 0 ≤ z ^ ≤ z ^ 1 − 1 ∏ y ^ 0 ≤ y ^ ≤ y ^ 1 Z ( x ^ 0 , y ^ + 1 / 2 , z ^ + 1 / 2 ) Z ( x ^ 1 , y ^ + 1 / 2 , z ^ + 1 / 2 ) W_z = \prod_{\hat{z}_0 \le\hat{z} \le\hat{z}_1-1} \prod_{\hat{x}_0 \le\hat{x} \le\hat{x}_1} Z(\hat{x}+1/2,\hat{y}_0,\hat{z}+1/2)Z(\hat{x}+1/2,\hat{y}_1,\hat{z}+1/2) \times\prod_{\hat{z}_0 \le\hat{z} \le\hat{z}_1-1} \prod_{\hat{y}_0 \le\hat{y} \le\hat{y}_1} Z(\hat{x}_0,\hat{y}+1/2,\hat{z}+1/2)Z(\hat{x}_1,\hat{y}+1/2,\hat{z}+1/2) W z = z ^ 0 ≤ z ^ ≤ z ^ 1 − 1 ∏ x ^ 0 ≤ x ^ ≤ x ^ 1 ∏ Z ( x ^ + 1/2 , y ^ 0 , z ^ + 1/2 ) Z ( x ^ + 1/2 , y ^ 1 , z ^ + 1/2 ) × z ^ 0 ≤ z ^ ≤ z ^ 1 − 1 ∏ y ^ 0 ≤ y ^ ≤ y ^ 1 ∏ Z ( x ^ 0 , y ^ + 1/2 , z ^ + 1/2 ) Z ( x ^ 1 , y ^ + 1/2 , z ^ + 1/2 )
The classical images of the open counterparts of this loop is given by
O z 0 , z 1 ( u , u ′ ) = D z 0 , z 1 ( u ) D z 0 , z 1 ( u ′ ) = σ u , z 0 σ u , z 1 σ u ′ , z 0 σ u ′ , z 1 . (222) O_{z_0,z_1}(\mathbf{u},\mathbf{u}') = D_{z_0,z_1}(\mathbf{u})D_{z_0,z_1}(\mathbf{u}') = \sigma_{\mathbf{u},z_0}\sigma_{\mathbf{u},z_1}\sigma_{\mathbf{u}',z_0}\sigma_{\mathbf{u}',z_1}. \tag*{(222)} O z 0 , z 1 ( u , u ′ ) = D z 0 , z 1 ( u ) D z 0 , z 1 ( u ′ ) = σ u , z 0 σ u , z 1 σ u ′ , z 0 σ u ′ , z 1 . (222)
where u = ( x ^ , y ^ ) \mathbf{u}=(\hat{x},\hat{y}) u = ( x ^ , y ^ ) denotes a coordinate of a node on a x y xy x y plane. Physical meaning of this operator is as follows: Define
D z 0 , z 1 ( u ) : = σ u , z 0 σ u , z 1 . (223) D_{z_0,z_1}(\mathbf{u}) := \sigma_{\mathbf{u},z_0}\sigma_{\mathbf{u},z_1}. \tag*{(223)} D z 0 , z 1 ( u ) := σ u , z 0 σ u , z 1 . (223)
Under (221) ,
D z 0 , z 1 ( u ) ↦ η z ( z 0 ) η z ( z 1 ) D z 0 , z 1 ( u ) . (224) D_{z_0,z_1}(\mathbf{u}) \mapsto\eta_z(z_0)\eta_z(z_1)D_{z_0,z_1}(\mathbf{u}). \tag*{(224)} D z 0 , z 1 ( u ) ↦ η z ( z 0 ) η z ( z 1 ) D z 0 , z 1 ( u ) . (224)
Thus this operator is odd under a flip of either endpoint plane, but invariant under their simultaneous flip and under every plane flip normal to x x x or y y y . It probes the relative Z 2 \mathbb{Z}_2 Z 2 symmetry of the two endpoint planes. The four spin operator (222) is invariant under all planar subsystem symmetries. Although ⟨ D z 0 , z 1 ⟩ \langle D_{z_0,z_1}\rangle ⟨ D z 0 , z 1 ⟩ vanishes in a finite volume symmetry-preserving ensemble, long range order in its two-point function can diagnose spontaneous breaking of the relative planar symmetry.
We turn to evaluation of the corresponding EA correlator
C z 0 , z 1 ( u , u ′ ) : = ⟨ O z 0 , z 1 ( u , u ′ ) ⟩ m 2 ‾ . (225) \mathcal{C}_{z_0,z_1}(\mathbf{u},\mathbf{u}') := \overline{\langle O_{z_0,z_1}(\mathbf{u},\mathbf{u}')\rangle_m^2}. \tag*{(225)} C z 0 , z 1 ( u , u ′ ) := ⟨ O z 0 , z 1 ( u , u ′ ) ⟩ m 2 . (225)
where the overline denotes the average over measurement records. On the higher Nishimori manifold, Δ p = K p \Delta_p=K_p Δ p = K p for every plaquette, the general identity derived in the main text gives
C z 0 , z 1 ( u , u ′ ) = ⟨ O z 0 , z 1 ( u , u ′ ) ⟩ c l e a n . (226) \mathcal{C}_{z_0,z_1}(\mathbf{u},\mathbf{u}') = \langle O_{z_0,z_1}(\mathbf{u},\mathbf{u}')\rangle_{\mathrm{clean}}. \tag*{(226)} C z 0 , z 1 ( u , u ′ ) = ⟨ O z 0 , z 1 ( u , u ′ ) ⟩ clean . (226)
Consequently, its long distance limit diagnoses learnability in the same relative planar channel as the clean correlation function.
The key observation is that a finite height operator factorizes exactly into operators supported between intermediate planes. For any sequence z 0 < z 1 < ⋯ < z M z_0 < z_1 < \cdots< z_M z 0 < z 1 < ⋯ < z M , we have
D z 0 , z M ( u ) = ∏ a = 0 M − 1 D z a , z a + 1 ( u ) , O z 0 , z M ( u , u ′ ) = ∏ a = 0 M − 1 O z a , z a + 1 ( u , u ′ ) . (227) D_{z_0,z_M}(\mathbf{u}) = \prod_{a=0}^{M-1}D_{z_a,z_{a+1}}(\mathbf{u}), \qquad O_{z_0,z_M}(\mathbf{u},\mathbf{u}') = \prod_{a=0}^{M-1}O_{z_a,z_{a+1}}(\mathbf{u},\mathbf{u}'). \tag*{(227)} D z 0 , z M ( u ) = a = 0 ∏ M − 1 D z a , z a + 1 ( u ) , O z 0 , z M ( u , u ′ ) = a = 0 ∏ M − 1 O z a , z a + 1 ( u , u ′ ) . (227)
Every spin on an intermediate plane occurs twice and cancels because σ r 2 = 1 \sigma_r^2=1 σ r 2 = 1 . Equivalently, the intermediate planar symmetry charges cancel, leaving only the charges at the two endpoint planes.
Applying the GKS inequalities to these Ising monomials, and then using (226) , yields
C z 0 , z M ( u , u ′ ) = ⟨ ∏ a = 0 M − 1 O z a , z a + 1 ( u , u ′ ) ⟩ c l e a n ≥ ∏ a = 0 M − 1 ⟨ O z a , z a + 1 ( u , u ′ ) ⟩ c l e a n = ∏ a = 0 M − 1 C z a , z a + 1 ( u , u ′ ) . (228) \mathcal{C}_{z_0,z_M}(\mathbf{u},\mathbf{u}') = \left\langle\prod_{a=0}^{M-1}O_{z_a,z_{a+1}}(\mathbf{u},\mathbf{u}')\right\rangle_{\mathrm{clean}} \ge\prod_{a=0}^{M-1}\langle O_{z_a,z_{a+1}}(\mathbf{u},\mathbf{u}')\rangle_{\mathrm{clean}} = \prod_{a=0}^{M-1}\mathcal{C}_{z_a,z_{a+1}}(\mathbf{u},\mathbf{u}'). \tag*{(228)} C z 0 , z M ( u , u ′ ) = ⟨ a = 0 ∏ M − 1 O z a , z a + 1 ( u , u ′ ) ⟩ clean ≥ a = 0 ∏ M − 1 ⟨ O z a , z a + 1 ( u , u ′ ) ⟩ clean = a = 0 ∏ M − 1 C z a , z a + 1 ( u , u ′ ) . (228)
This is the exact hierarchy associated with subdivision in the vertical direction [33 , 34 ] . It requires neither statistical independence of the slabs nor a continuous phase transition. In particular, taking every pair of intermediate planes to be adjacent gives
C z , z + h ( u , u ′ ) ≥ ∏ a = 0 h − 1 C z + a , z + a + 1 ( u , u ′ ) . (229) \mathcal{C}_{z,z+h}(\mathbf{u},\mathbf{u}') \ge\prod_{a=0}^{h-1}\mathcal{C}_{z+a,z+a+1}(\mathbf{u},\mathbf{u}'). \tag*{(229)} C z , z + h ( u , u ′ ) ≥ a = 0 ∏ h − 1 C z + a , z + a + 1 ( u , u ′ ) . (229)
The right-hand side is a product of separately record averaged second moments, rather than a single record average of their product.
For a translation-invariant system and ensemble, let R = u ′ − u \mathbf{R} = \mathbf{u}' - \mathbf{u} R = u ′ − u and C h ( R ) : = C z , z + h ( u , u ′ ) \mathcal{C}_h(\mathbf{R}) := \mathcal{C}_{z,z+h}(\mathbf{u},\mathbf{u}') C h ( R ) := C z , z + h ( u , u ′ ) . The relation (229) becomes
C h ( R ) ≥ [ C 1 ( R ) ] h (230) \mathcal{C}_h(\mathbf{R}) \ge\left[\mathcal{C}_1(\mathbf{R})\right]^h \tag*{(230)} C h ( R ) ≥ [ C 1 ( R ) ] h (230)
Assuming the long distance limits exist, define
q h : = lim ∣ R ∣ → ∞ C h ( R ) (231) q_h := \lim_{|\mathbf{R}|\to\infty} \mathcal{C}_h(\mathbf{R}) \tag*{(231)} q h := ∣ R ∣ → ∞ lim C h ( R ) (231)
for fixed h h h . It follows that
q h ≥ q 1 h (232) q_h \ge q_1^h \tag*{(232)} q h ≥ q 1 h (232)
Hence, if the observable involving two adjacent planes remains learnable at long distances, the same is true for planes separated by any fixed finite height. Nor does the bound guarantee a nonzero limit when the height also tends to infinity.
Weak measurements of Pauli X X X
We next consider weak measurements of single-edge X X X operators, with the wavefunction deformation also taken in the X X X basis in the model (217) . The corresponding classical model differs from the plaquette Ising model discussed above. Since at each vertex, B v x B v y B v z = 1 B_v^xB_v^yB_v^z=1 B v x B v y B v z = 1 , we use B v x B_v^x B v x and B v y B_v^y B v y as local generators and introduce Ising variables t v t_v t v and s v s_v s v , respectively. Global relations among these generators give a configuration-independent redundancy in this parametrization. An X X X operator on an x x x -directed edge anticommutes with B y B^y B y at its two endpoints, while an X X X operator on a y y y -directed edge anticommutes with B x B^x B x . On a z z z -directed edge, it anticommutes with both generators at each endpoint. The associated classical monomials are thus
B v , x = s v s v + x ^ , B v , y = t v t v + y ^ , B v , z = s v t v s v + z ^ t v + z ^ (233) \mathcal{B}_{v,x}=s_vs_{v+\hat{x}}, \qquad\mathcal{B}_{v,y}=t_vt_{v+\hat{y}}, \qquad\mathcal{B}_{v,z}=s_vt_vs_{v+\hat{z}}t_{v+\hat{z}} \tag*{(233)} B v , x = s v s v + x ^ , B v , y = t v t v + y ^ , B v , z = s v t v s v + z ^ t v + z ^ (233)
Writing u v = s v t v u_v=s_vt_v u v = s v t v , the clean distribution takes the anisotropic Ashkin–Teller form [9 ]
P K ( s , t ) = 1 Z K exp [ ∑ v ( K x s v s v + x ^ + K y t v t v + y ^ + K z u v u v + z ^ ) ] (234) P_K(s,t)=\frac{1}{Z_K}\exp\left[\sum_v\left(K_xs_vs_{v+\hat{x}}+K_yt_vt_{v+\hat{y}}+K_zu_vu_{v+\hat{z}}\right)\right] \tag*{(234)} P K ( s , t ) = Z K 1 exp [ v ∑ ( K x s v s v + x ^ + K y t v t v + y ^ + K z u v u v + z ^ ) ] (234)
Here K μ ≥ 0 K_\mu\ge0 K μ ≥ 0 are dimensionless deformation parameters. Gaussian measurements replace the coupling on each edge by K μ + Δ μ m v , μ K_\mu+\Delta_\mu m_{v,\mu} K μ + Δ μ m v , μ in the posterior distribution. Below, we work on the higher Nishimori manifold Δ μ = K μ \Delta_\mu=K_\mu Δ μ = K μ .
The classical model has planar subsystem symmetries
s r ⟼ η y ( y ^ ) η z ( z ^ ) s r , t r ⟼ η x ( x ^ ) η z ( z ^ ) t r , u r ⟼ η x ( x ^ ) η y ( y ^ ) u r (235) s_r\longmapsto\eta_y(\hat{y})\eta_z(\hat{z})s_r,\qquad t_r\longmapsto\eta_x(\hat{x})\eta_z(\hat{z})t_r,\qquad u_r\longmapsto\eta_x(\hat{x})\eta_y(\hat{y})u_r \tag*{(235)} s r ⟼ η y ( y ^ ) η z ( z ^ ) s r , t r ⟼ η x ( x ^ ) η z ( z ^ ) t r , u r ⟼ η x ( x ^ ) η y ( y ^ ) u r (235)
where each η μ \eta_\mu η μ is an arbitrary Z 2 \mathbb{Z}_2 Z 2 -valued function. Defining S x = s S^x=s S x = s , S y = t S^y=t S y = t , and S z = u S^z=u S z = u , the classical image of a straight open X X X string in direction μ \mu μ gives
∏ k = 0 ℓ − 1 B v + k μ ^ , μ = S v μ S v + ℓ μ ^ μ (236) \prod_{k=0}^{\ell-1}\mathcal{B}_{v+k\hat{\mu},\mu}=S_v^\mu S_{v+\ell\hat{\mu}}^\mu \tag*{(236)} k = 0 ∏ ℓ − 1 B v + k μ ^ , μ = S v μ S v + ℓ μ ^ μ (236)
Its endpoints carry the corresponding lineon excitations.
Consider a rectangular box with positive integer side lengths a , b , c a,b,c a , b , c and vertex set
V ( a , b , c ) = { 0 , a } × { 0 , b } × { 0 , c } . (237) V(a,b,c)=\{0,a\}\times\{0,b\}\times\{0,c\}. \tag*{(237)} V ( a , b , c ) = { 0 , a } × { 0 , b } × { 0 , c } . (237)
We define three eight-spin observables on the same vertices as
Q x = ∏ v ∈ V s v , Q y = ∏ v ∈ V t v , Q z = ∏ v ∈ V u v . (238) Q_x=\prod_{v\in V}s_v,\qquad Q_y=\prod_{v\in V}t_v,\qquad Q_z=\prod_{v\in V}u_v. \tag*{(238)} Q x = v ∈ V ∏ s v , Q y = v ∈ V ∏ t v , Q z = v ∈ V ∏ u v . (238)
Each is invariant under all the planar subsystem symmetries. Moreover, Q μ Q_\mu Q μ is the classical image of the product of the four open X strings along the μ \mu μ -directed edges of the box. For example,
Q x = ∏ y ∈ { 0 , b } ∏ z ∈ { 0 , c } ( s 0 , y , z s a , y , z ) . (239) Q_x=\prod_{y\in\{0,b\}}\prod_{z\in\{0,c\}}(s_{0,y,z}s_{a,y,z}). \tag*{(239)} Q x = y ∈ { 0 , b } ∏ z ∈ { 0 , c } ∏ ( s 0 , y , z s a , y , z ) . (239)
The local identity u v = s v t v u_v=s_vt_v u v = s v t v implies
Q x Q y = Q z , Q y Q z = Q x , Q z Q x = Q y . (240) Q_xQ_y=Q_z,\qquad Q_yQ_z=Q_x,\qquad Q_zQ_x=Q_y. \tag*{(240)} Q x Q y = Q z , Q y Q z = Q x , Q z Q x = Q y . (240)
To proceed, we define the EA type correlators as
C μ ( 2 ) ( a , b , c ) : = ⟨ Q μ ⟩ m 2 ⟨ Q μ 2 ⟩ m , μ = x , y , z . (241) C_\mu^{(2)}(a,b,c):=\frac{\langle Q_\mu\rangle_m^2}{\langle Q_\mu^2\rangle_m},\qquad\mu=x,y,z. \tag*{(241)} C μ ( 2 ) ( a , b , c ) := ⟨ Q μ 2 ⟩ m ⟨ Q μ ⟩ m 2 , μ = x , y , z . (241)
On the higher Nishimori manifold, the identity established in § 3 gives C μ ( 2 ) = ⟨ Q μ ⟩ c l e a n C_\mu^{(2)}=\langle Q_\mu\rangle_{\mathrm{clean}} C μ ( 2 ) = ⟨ Q μ ⟩ clean . Applying the exact correlator hierarchy to (240) , we obtain
C x ( 2 ) ≥ C y ( 2 ) C z ( 2 ) , C y ( 2 ) ≥ C z ( 2 ) C x ( 2 ) , C z ( 2 ) ≥ C x ( 2 ) C y ( 2 ) . (242) \begin{aligned}
C_x^{(2)}&\ge C_y^{(2)}C_z^{(2)},\\
C_y^{(2)}&\ge C_z^{(2)}C_x^{(2)},\\
C_z^{(2)}&\ge C_x^{(2)}C_y^{(2)}. \tag*{(242)}
\end{aligned} C x ( 2 ) C y ( 2 ) C z ( 2 ) ≥ C y ( 2 ) C z ( 2 ) , ≥ C z ( 2 ) C x ( 2 ) , ≥ C x ( 2 ) C y ( 2 ) . (242)
All correlators in each inequality are evaluated on the same box. These relations hold at finite separation and do not require rotational symmetry or equality of the couplings K x , K y , K z K_x,K_y,K_z K x , K y , K z .
To state their long distance consequence, take a sequence of boxes with a / ℓ → α x a/\ell\to\alpha_x a / ℓ → α x , b / ℓ → α y b/\ell\to\alpha_y b / ℓ → α y , and c / ℓ → α z c/\ell\to\alpha_z c / ℓ → α z as ℓ → ∞ \ell\to\infty ℓ → ∞ , where all α μ > 0 \alpha_\mu>0 α μ > 0 . If the limits
q μ = lim ℓ → ∞ C μ ( 2 ) ( a , b , c ) (243) q_\mu=\lim_{\ell\to\infty}C_\mu^{(2)}(a,b,c) \tag*{(243)} q μ = ℓ → ∞ lim C μ ( 2 ) ( a , b , c ) (243)
exist, they satisfy q x ≥ q y q z q_x\ge q_yq_z q x ≥ q y q z and its cyclic permutations. Thus, long range learning of any two of these observables implies long range learning of the third. In particular, exactly two of the three limits cannot be positive while the remaining one vanishes. This is a constraint on the information carried by products of four strings; it does not imply that the individual strings are learnable. The result therefore provides an application of the general hierarchy to extended observables in a system with planar subsystem symmetry, without assuming a particular phase diagram.
Evaluation of the correlator
In this Appendix, we derive (168) . To this end, we first establish the group structure of the constraint set. For each selected constraint α = ( e , N ) \alpha=(e,N) α = ( e , N ) , with e = ( i , j ) e=(i,j) e = ( i , j ) , let π N : G → G / N \pi_N:G\to G/N π N : G → G / N be the quotient homomorphism. The constraint is equivalent to π N ( g i ) = π N ( g j ) \pi_N(g_i)=\pi_N(g_j) π N ( g i ) = π N ( g j ) . If g , h ∈ H B g,h\in H_B g , h ∈ H B , then
π N ( g i h i ) = π N ( g i ) π N ( h i ) = π N ( g j ) π N ( h j ) = π N ( g j h j ) . (244) \pi_N(g_i h_i)=\pi_N(g_i)\pi_N(h_i)=\pi_N(g_j)\pi_N(h_j)=\pi_N(g_j h_j). \tag*{(244)} π N ( g i h i ) = π N ( g i ) π N ( h i ) = π N ( g j ) π N ( h j ) = π N ( g j h j ) . (244)
The constraints are also preserved under componentwise inversion, and the identity configuration satisfies every constraint. Hence H B H_B H B is a subgroup of G V G^V G V .
We rewrite the correlator (166) as
C λ ( B ) ( x , y ) = 1 d λ ∣ H B ∣ ∑ n ∈ N B ∑ g ∈ H B g x g y − 1 = n χ λ ( n ) (245) C^{(B)}_{\lambda}(x,y)=\frac{1}{d_{\lambda}|H_B|}\sum_{n\in N_B}\sum_{\substack{g\in H_B\\g_xg_y^{-1}=n}}\chi_{\lambda}(n) \tag*{(245)} C λ ( B ) ( x , y ) = d λ ∣ H B ∣ 1 n ∈ N B ∑ g ∈ H B g x g y − 1 = n ∑ χ λ ( n ) (245)
with N B N_B N B being defined as
N B ( x , y ) : = { g x g y − 1 : g ∈ H B } . N_B(x,y):=\{g_xg_y^{-1}:g\in H_B\}. N B ( x , y ) := { g x g y − 1 : g ∈ H B } .
We have
∑ g ∈ H B g x g y − 1 = n χ λ ( n ) = Λ n χ λ ( n ) (246) \sum_{\substack{g\in H_B\\g_xg_y^{-1}=n}}\chi_{\lambda}(n)=\Lambda_n\chi_{\lambda}(n) \tag*{(246)} g ∈ H B g x g y − 1 = n ∑ χ λ ( n ) = Λ n χ λ ( n ) (246)
where Λ n \Lambda_n Λ n denotes the number of combinations of g ∈ H B g\in H_B g ∈ H B , satisfying g x g y − 1 = n g_xg_y^{-1}=n g x g y − 1 = n for given n n n . To evaluate Λ n \Lambda_n Λ n , we introduce the following projection:
p x y : H B ⟶ L B , p x y ( g ) = ( g x , g y ) , p_{xy}:H_B\longrightarrow L_B,\qquad p_{xy}(g)=(g_x,g_y), p x y : H B ⟶ L B , p x y ( g ) = ( g x , g y ) ,
L B : = { ( g x , g y ) : g ∈ H B } ⊆ G × G . (247) L_B:=\{(g_x,g_y):g\in H_B\}\subseteq G\times G. \tag*{(247)} L B := {( g x , g y ) : g ∈ H B } ⊆ G × G . (247)
This is a surjective homomorphism, with kernel
K B ; x , y : = ker p x y = { g ∈ H B : g x = 1 G , g y = 1 G } . (248) K_{B;x,y}:=\ker p_{xy}=\{g\in H_B:g_x=1_G,\ g_y=1_G\}. \tag*{(248)} K B ; x , y := ker p x y = { g ∈ H B : g x = 1 G , g y = 1 G } . (248)
Every constant configuration g i = b g_i=b g i = b satisfies all selected constraints, so ( b , b ) ∈ L B (b,b)\in L_B ( b , b ) ∈ L B for every b ∈ G b\in G b ∈ G . If ( a , b ) ∈ L B (a,b)\in L_B ( a , b ) ∈ L B , multiplication by ( b − 1 , b − 1 ) (b^{-1},b^{-1}) ( b − 1 , b − 1 ) on the right shows that ( a b − 1 , 1 G ) ∈ L B (ab^{-1},1_G)\in L_B ( a b − 1 , 1 G ) ∈ L B . Conversely, if ( n , 1 G ) ∈ L B (n,1_G)\in L_B ( n , 1 G ) ∈ L B , then ( n , 1 G ) ( b , b ) = ( n b , b ) ∈ L B (n,1_G)(b,b)=(nb,b)\in L_B ( n , 1 G ) ( b , b ) = ( nb , b ) ∈ L B . Hence, we have
N B = { n ∈ G : ( n , 1 G ) ∈ L B } , N_B=\{n\in G:(n,1_G)\in L_B\}, N B = { n ∈ G : ( n , 1 G ) ∈ L B } ,
L B = { ( n b , b ) : n ∈ N B , b ∈ G } . (249) L_B=\{(nb,b):n\in N_B,\ b\in G\}. \tag*{(249)} L B = {( nb , b ) : n ∈ N B , b ∈ G } . (249)
The first equality makes N B N_B N B a subgroup. Also, it is normal as
( h , h ) ( n , 1 G ) ( h − 1 , h − 1 ) = ( h n h − 1 , 1 G ) ∈ L B ( h ∈ G , n ∈ N B ) . (250) (h,h)(n,1_G)(h^{-1},h^{-1})=(hnh^{-1},1_G)\in L_B\quad(h\in G,\ n\in N_B). \tag*{(250)} ( h , h ) ( n , 1 G ) ( h − 1 , h − 1 ) = ( hn h − 1 , 1 G ) ∈ L B ( h ∈ G , n ∈ N B ) . (250)
Thus N B ⊴ G N_B\trianglelefteq G N B ⊴ G . Each endpoint pair in L B L_B L B has exactly ∣ K B ; x , y ∣ |K_{B;x,y}| ∣ K B ; x , y ∣ preimages under p x y p_{xy} p x y , because each nonempty preimage is a coset of the kernel. For a fixed n ∈ N B n\in N_B n ∈ N B , (249) lists precisely ∣ G ∣ |G| ∣ G ∣ endpoint pairs ( n b , b ) (nb,b) ( nb , b ) with relative value n n n . This leads to
Λ n = ∣ K B ; x , y ∣ ∣ G ∣ ( n ∈ N B ) , ∣ H B ∣ = ∣ K B ; x , y ∣ ∣ G ∣ ∣ N B ∣ . (251) \Lambda_n=|K_{B;x,y}||G|\quad(n\in N_B),\qquad|H_B|=|K_{B;x,y}||G||N_B|. \tag*{(251)} Λ n = ∣ K B ; x , y ∣∣ G ∣ ( n ∈ N B ) , ∣ H B ∣ = ∣ K B ; x , y ∣∣ G ∣∣ N B ∣. (251)
Substitution into (245) gives
C λ ( B ) ( x , y ) = ∣ K B ; x , y ∣ ∣ G ∣ d λ ∣ K B ; x , y ∣ ∣ G ∣ ∣ N B ∣ ∑ n ∈ N B χ λ ( n ) = 1 d λ ∣ N B ∣ ∑ n ∈ N B χ λ ( n ) . (252) C^{(B)}_{\lambda}(x,y)=\frac{|K_{B;x,y}||G|}{d_{\lambda}|K_{B;x,y}||G||N_B|}\sum_{n\in N_B}\chi_{\lambda}(n)=\frac{1}{d_{\lambda}|N_B|}\sum_{n\in N_B}\chi_{\lambda}(n). \tag*{(252)} C λ ( B ) ( x , y ) = d λ ∣ K B ; x , y ∣∣ G ∣∣ N B ∣ ∣ K B ; x , y ∣∣ G ∣ n ∈ N B ∑ χ λ ( n ) = d λ ∣ N B ∣ 1 n ∈ N B ∑ χ λ ( n ) . (252)
To proceed, we define the following matrix
P B λ : = 1 ∣ N B ∣ ∑ n ∈ N B D λ ( n ) . (253) P^\lambda_B:=\frac{1}{|N_B|}\sum_{n\in N_B}D^\lambda(n). \tag*{(253)} P B λ := ∣ N B ∣ 1 n ∈ N B ∑ D λ ( n ) . (253)
Since χ λ = Tr D λ \chi_\lambda=\operatorname{Tr}D^\lambda χ λ = Tr D λ , (252) is rewritten as C λ ( B ) = d λ − 1 Tr P B λ C_\lambda^{(B)}=d_\lambda^{-1}\operatorname{Tr}P_B^\lambda C λ ( B ) = d λ − 1 Tr P B λ . Using closure of the subgroup under multiplication and inversion, we obtain
( P B λ ) 2 = 1 ∣ N B ∣ 2 ∑ n , m ∈ N B D λ ( n m ) = 1 ∣ N B ∣ ∑ k ∈ N B D λ ( k ) = P B λ , ( P B λ ) † = 1 ∣ N B ∣ ∑ n ∈ N B D λ ( n − 1 ) = P B λ . (254) \left(P_B^\lambda\right)^2=\frac{1}{\lvert N_B\rvert^2}\sum_{n,m\in N_B}D^\lambda(nm)=\frac{1}{\lvert N_B\rvert}\sum_{k\in N_B}D^\lambda(k)=P_B^\lambda,
\\
\left(P_B^\lambda\right)^\dagger=\frac{1}{\lvert N_B\rvert}\sum_{n\in N_B}D^\lambda\left(n^{-1}\right)=P_B^\lambda. \tag*{(254)} ( P B λ ) 2 = ∣ N B ∣ 2 1 n , m ∈ N B ∑ D λ ( nm ) = ∣ N B ∣ 1 k ∈ N B ∑ D λ ( k ) = P B λ , ( P B λ ) † = ∣ N B ∣ 1 n ∈ N B ∑ D λ ( n − 1 ) = P B λ . (254)
For the first line, every k ∈ N B k\in N_B k ∈ N B has exactly ∣ N B ∣ \lvert N_B\rvert ∣ N B ∣ factorizations k = n m k=nm k = nm . Also D λ ( n ) P B λ = P B λ D^\lambda(n)P_B^\lambda=P_B^\lambda D λ ( n ) P B λ = P B λ for all n ∈ N B n\in N_B n ∈ N B . Thus P B λ P_B^\lambda P B λ is the orthogonal projector onto vectors fixed by the action of N B N_B N B . Further, since N B ⊴ G N_B\trianglelefteq G N B ⊴ G , one finds
D λ ( h ) P B λ D λ ( h ) − 1 = 1 ∣ N B ∣ ∑ n ∈ N B D λ ( h n h − 1 ) = P B λ ( h ∈ G ) . (255) D^\lambda(h)P_B^\lambda D^\lambda(h)^{-1}=\frac{1}{\lvert N_B\rvert}\sum_{n\in N_B}D^\lambda\left(hnh^{-1}\right)=P_B^\lambda\qquad(h\in G). \tag*{(255)} D λ ( h ) P B λ D λ ( h ) − 1 = ∣ N B ∣ 1 n ∈ N B ∑ D λ ( hn h − 1 ) = P B λ ( h ∈ G ) . (255)
By irreducibility and Schur’s lemma, P B λ P_B^\lambda P B λ is a scalar multiple of I d λ I_{d_\lambda} I d λ . Its idempotence restricts that scalar to zero or one. The identity case occurs exactly when every n ∈ N B n\in N_B n ∈ N B acts as I d λ I_{d_\lambda} I d λ , namely when N B ⊆ ker D λ N_B\subseteq\ker D^\lambda N B ⊆ ker D λ . Overall, we arrive at
C λ ( B ) ( x , y ) = Tr P B λ d λ = { 1 , N B ⊆ ker D λ , 0 , otherwise . (256) C_\lambda^{(B)}(x,y)=\frac{\operatorname{Tr}P_B^\lambda}{d_\lambda}=
\begin{cases}
1, & N_B\subseteq\ker D^\lambda,\\
0, & \text{otherwise}.
\end{cases} \tag*{(256)} C λ ( B ) ( x , y ) = d λ Tr P B λ = { 1 , 0 , N B ⊆ ker D λ , otherwise . (256)
This completes the derivation of (168) .
Derivation of (172)
In this appendix, we prove the relation (172) using positive correlation between increasing events; Let 2 J 2^{\mathcal{J}} 2 J be the collection of all subsets of the constraint label set J \mathcal{J} J . These subsets are ordered by inclusion: B ⊆ D B\subseteq D B ⊆ D means that every constraint selected in B B B is also selected in D D D . An event is increasing if it continues to hold whenever more constraints are selected. The FKG inequality [45 ] states that a strictly positive probability distribution on these subsets satisfying
μ ( B ∪ D ) μ ( B ∩ D ) ≥ μ ( B ) μ ( D ) for all B , D ⊆ J (257) \mu(B\cup D)\mu(B\cap D)\ge\mu(B)\mu(D)\qquad\text{for all }B,D\subseteq\mathcal{J} \tag*{(257)} μ ( B ∪ D ) μ ( B ∩ D ) ≥ μ ( B ) μ ( D ) for all B , D ⊆ J (257)
assigns nonnegative correlation to any two increasing events:
μ ( A ∩ B ) ≥ μ ( A ) μ ( B ) . (258) \mu(\mathcal{A}\cap\mathcal{B})\ge\mu(\mathcal{A})\mu(\mathcal{B}). \tag*{(258)} μ ( A ∩ B ) ≥ μ ( A ) μ ( B ) . (258)
We first verify the condition on μ \mu μ , and then apply the theorem to the events A λ \mathcal{A}_\lambda A λ and A η \mathcal{A}_\eta A η .
We consider B , D ⊆ J B,D\subseteq\mathcal{J} B , D ⊆ J . Satisfying all constraints in B ∪ D B\cup D B ∪ D is equivalent to satisfying both sets of constraints. Moreover, H B H_B H B and H D H_D H D are subgroups of H B ∩ D H_{B\cap D} H B ∩ D , so their product set is contained in H B ∩ D H_{B\cap D} H B ∩ D . It follows that
H B ∪ D = H B ∩ H D , H B H D ⊆ H B ∩ D . (259) H_{B\cup D}=H_B\cap H_D,\qquad H_BH_D\subseteq H_{B\cap D}. \tag*{(259)} H B ∪ D = H B ∩ H D , H B H D ⊆ H B ∩ D . (259)
For finite subgroups H , K H,K H , K of a group, ∣ H K ∣ = ∣ H ∣ ∣ K ∣ / ∣ H ∩ K ∣ \lvert HK\rvert=\lvert H\rvert\lvert K\rvert/\lvert H\cap K\rvert ∣ H K ∣ = ∣ H ∣ ∣ K ∣ / ∣ H ∩ K ∣ . Indeed, if h k = h ′ k ′ hk=h'k' hk = h ′ k ′ , then t = h − 1 h ′ = k k ′ − 1 ∈ H ∩ K t=h^{-1}h'=kk'^{-1}\in H\cap K t = h − 1 h ′ = k k ′ − 1 ∈ H ∩ K ; conversely, each such t t t gives the same product through ( h ′ , k ′ ) = ( h t , t − 1 k ) (h',k')=(ht,t^{-1}k) ( h ′ , k ′ ) = ( h t , t − 1 k ) . Every element of H K HK H K is consequently counted ∣ H ∩ K ∣ \lvert H\cap K\rvert ∣ H ∩ K ∣ times. Applying this identity to (259) gives
∣ H B ∣ ∣ H D ∣ = ∣ H B ∩ H D ∣ ∣ H B H D ∣ ≤ ∣ H B ∪ D ∣ ∣ H B ∩ D ∣ . (260) \lvert H_B\rvert\lvert H_D\rvert=\lvert H_B\cap H_D\rvert\lvert H_BH_D\rvert\le\lvert H_{B\cup D}\rvert\lvert H_{B\cap D}\rvert. \tag*{(260)} ∣ H B ∣ ∣ H D ∣ = ∣ H B ∩ H D ∣ ∣ H B H D ∣ ≤ ∣ H B ∪ D ∣ ∣ H B ∩ D ∣ . (260)
The other factors in μ \mu μ obey the exact identity
( ∏ α ∈ B v α ) ( ∏ α ∈ D v α ) = ( ∏ α ∈ B ∪ D v α ) ( ∏ α ∈ B ∩ D v α ) . \left(\prod_{\alpha\in B}v_\alpha\right)\left(\prod_{\alpha\in D}v_\alpha\right)=\left(\prod_{\alpha\in B\cup D}v_\alpha\right)\left(\prod_{\alpha\in B\cap D}v_\alpha\right). ( α ∈ B ∏ v α ) ( α ∈ D ∏ v α ) = ( α ∈ B ∪ D ∏ v α ) ( α ∈ B ∩ D ∏ v α ) .
Every label is counted the same number of times on both sides. Substituting this identity and (260) into (164) yields
μ ( B ∪ D ) μ ( B ∩ D ) ≥ μ ( B ) μ ( D ) . (261) \mu(B\cup D)\mu(B\cap D)\geq\mu(B)\mu(D). \tag*{(261)} μ ( B ∪ D ) μ ( B ∩ D ) ≥ μ ( B ) μ ( D ) . (261)
All labels in J \mathcal{J} J have v α > 0 v_\alpha>0 v α > 0 , and H B H_B H B is nonempty because it contains every constant configuration. Thus μ ( B ) > 0 \mu(B)>0 μ ( B ) > 0 for every subset B ⊆ J B\subseteq\mathcal{J} B ⊆ J , and the required condition on the probability distribution is satisfied. The events A λ \mathcal{A}_\lambda A λ are increasing under the inclusion of constraint sets. If B ⊆ D B\subseteq D B ⊆ D , then
H D ⊆ H B , N D ( x , y ) ⊆ N B ( x , y ) . (262) H_D\subseteq H_B,\qquad N_D(x,y)\subseteq N_B(x,y). \tag*{(262)} H D ⊆ H B , N D ( x , y ) ⊆ N B ( x , y ) . (262)
Hence N B ( x , y ) ⊆ ker D λ N_B(x,y)\subseteq\ker D^\lambda N B ( x , y ) ⊆ ker D λ implies N D ( x , y ) ⊆ ker D λ N_D(x,y)\subseteq\ker D^\lambda N D ( x , y ) ⊆ ker D λ . Thus B ∈ A λ B\in\mathcal{A}_\lambda B ∈ A λ implies D ∈ A λ D\in\mathcal{A}_\lambda D ∈ A λ , as required for the application of the FKG inequality.
The FKG inequality [45 ] applied to the increasing event indicators gives
μ ( A λ ∩ A η ) ≥ μ ( A λ ) μ ( A η ) . (263) \mu(\mathcal{A}_\lambda\cap\mathcal{A}_\eta)\geq\mu(\mathcal{A}_\lambda)\mu(\mathcal{A}_\eta). \tag*{(263)} μ ( A λ ∩ A η ) ≥ μ ( A λ ) μ ( A η ) . (263)
Suppose ν ≺ λ ⊗ η \nu\prec\lambda\otimes\eta ν ≺ λ ⊗ η . For any B ∈ A λ ∩ A η B\in\mathcal{A}_\lambda\cap\mathcal{A}_\eta B ∈ A λ ∩ A η and n ∈ N B n\in N_B n ∈ N B , both D λ ( n ) D^\lambda(n) D λ ( n ) and D η ( n ) D^\eta(n) D η ( n ) are identity matrices. Their tensor product is therefore the identity on the whole tensor product space, and in particular on every invariant subspace carrying ν \nu ν . Hence B ∈ A ν B\in\mathcal{A}_\nu B ∈ A ν , proving A λ ∩ A η ⊆ A ν \mathcal{A}_\lambda\cap\mathcal{A}_\eta\subseteq\mathcal{A}_\nu A λ ∩ A η ⊆ A ν . Therefore,
Q ν ( x , y ) = μ ( A ν ) ≥ μ ( A λ ∩ A η ) ≥ μ ( A λ ) μ ( A η ) = Q λ ( x , y ) Q η ( x , y ) . (264) Q_\nu(x,y)=\mu(\mathcal{A}_\nu)\geq\mu(\mathcal{A}_\lambda\cap\mathcal{A}_\eta)\geq\mu(\mathcal{A}_\lambda)\mu(\mathcal{A}_\eta)=Q_\lambda(x,y)Q_\eta(x,y). \tag*{(264)} Q ν ( x , y ) = μ ( A ν ) ≥ μ ( A λ ∩ A η ) ≥ μ ( A λ ) μ ( A η ) = Q λ ( x , y ) Q η ( x , y ) . (264)
The first inequality uses the inclusion implied by the fusion rules, and the second uses (263) , completing the proof of (172) .
Perturbations around the decoupled higher Nishimori lines
We examine the stability of the decoupled critical point in the rank-2 toric code learning problem with Pauli-Z Z Z measurements. When K □ = Δ □ = 0 K_\square=\Delta_\square=0 K □ = Δ □ = 0 , the interactions connect sites separated by two lattice spacings, and the model decomposes into four independent parity sublattices, labeled by α = 1 , … , 4 \alpha=1,\ldots,4 α = 1 , … , 4 . Setting K x = K y = Δ x = Δ y = K c K_x=K_y=\Delta_x=\Delta_y=K_c K x = K y = Δ x = Δ y = K c , where K c K_c K c is the critical coupling of the clean square lattice Ising model, places each sector at its higher Nishimori critical point [19 ] .
In each sector, the higher Nishimori identity gives
⟨ σ α ( r ) σ α ( 0 ) ⟩ m 2 = ⟨ σ α ( r ) σ α ( 0 ) ⟩ c l e a n ∼ ∣ r ∣ − 1 / 4 . (265) \left\langle\sigma_\alpha(\boldsymbol{r})\sigma_\alpha(0)\right\rangle_m^2=\left\langle\sigma_\alpha(\boldsymbol{r})\sigma_\alpha(0)\right\rangle_{\mathrm{clean}}\sim\lvert\boldsymbol{r}\rvert^{-1/4}. \tag*{(265)} ⟨ σ α ( r ) σ α ( 0 ) ⟩ m 2 = ⟨ σ α ( r ) σ α ( 0 ) ⟩ clean ∼ ∣ r ∣ − 1/4 . (265)
Thus the replica overlap field Q α a b = σ α ( a ) σ α ( b ) Q_\alpha^{ab}=\sigma_\alpha^{(a)}\sigma_\alpha^{(b)} Q α ab = σ α ( a ) σ α ( b ) , with a ≠ b a\neq b a = b , has scaling dimension x Q = 1 / 8 x_Q=1/8 x Q = 1/8 in the learning limit R → 1 R\to1 R → 1 . The single-replica spin field also has x σ = 1 / 8 x_\sigma=1/8 x σ = 1/8 , since its two-point function in this limit is the record averaged first moment, which equals the clean correlator by the law of total expectation.
Each elementary plaquette contains one site from each parity sublattice. Its leading continuum operator in replica a a a is therefore given by
Φ ( a ) ( r ) = ∏ α = 1 4 σ α ( a ) ( r ) , X a b ( r ) = Φ ( a ) ( r ) Φ ( b ) ( r ) = ∏ α = 1 4 Q α a b ( r ) . (266) \Phi^{(a)}(\boldsymbol{r})=\prod_{\alpha=1}^{4}\sigma_\alpha^{(a)}(\boldsymbol{r}),\qquad X^{ab}(\boldsymbol{r})=\Phi^{(a)}(\boldsymbol{r})\Phi^{(b)}(\boldsymbol{r})=\prod_{\alpha=1}^{4}Q_\alpha^{ab}(\boldsymbol{r}). \tag*{(266)} Φ ( a ) ( r ) = α = 1 ∏ 4 σ α ( a ) ( r ) , X ab ( r ) = Φ ( a ) ( r ) Φ ( b ) ( r ) = α = 1 ∏ 4 Q α ab ( r ) . (266)
The microscopic displacements within a plaquette contribute derivative corrections. Independence of the four sectors at the decoupled fixed point implies
x Φ = 4 x σ = 1 2 , x X = 4 x Q = 1 2 . (267) x_{\Phi}=4x_{\sigma}=\frac{1}{2}, \qquad x_{X}=4x_{Q}=\frac{1}{2}. \tag*{(267)} x Φ = 4 x σ = 2 1 , x X = 4 x Q = 2 1 . (267)
Here x Q x_{Q} x Q is the dimension of the overlap within an interacting learning theory; it is not obtained by adding two independent spin dimensions. Dimension additivity is used only between the four decoupled parity sectors.
To leading order, plaquette deformation and Gaussian measurement produce the perturbation
δ S = − ∫ d 2 r [ g K ∑ a = 1 R Φ ( a ) ( r ) + g Δ ∑ 1 ≤ a < b ≤ R X a b ( r ) ] . (268) \delta S=-\int d^{2}r\left[g_{K}\sum_{a=1}^{R}\Phi^{(a)}(r)+g_{\Delta}\sum_{1\leq a<b\leq R}X^{ab}(r)\right]. \tag*{(268)} δ S = − ∫ d 2 r [ g K a = 1 ∑ R Φ ( a ) ( r ) + g Δ 1 ≤ a < b ≤ R ∑ X ab ( r ) ] . (268)
where g K g_{K} g K and g Δ g_{\Delta} g Δ are continuum couplings proportional to K □ K_{\square} K □ and Δ □ \Delta_{\square} Δ □ , respectively, with nonuniversal normalization factors absorbed. Independent replica spin flips in each decoupled sector distinguish the components appearing in these sums, so summing over replica labels does not change their linear scaling dimensions. Both perturbations consequently have the positive RG eigenvalue
y □ = 2 − 1 2 = 3 2 . (269) y_{\square}=2-\frac{1}{2}=\frac{3}{2}. \tag*{(269)} y □ = 2 − 2 1 = 2 3 . (269)
For a dimensionless bare coupling g □ g_{\square} g □ along a fixed nonzero plaquette perturbation direction, the resulting crossover length is given by
ξ □ a 0 ∼ ∣ g □ ∣ − 1 / y □ = ∣ g □ ∣ − 2 / 3 . (270) \frac{\xi_{\square}}{a_{0}}\sim|g_{\square}|^{-1/y_{\square}}=|g_{\square}|^{-2/3}. \tag*{(270)} a 0 ξ □ ∼ ∣ g □ ∣ − 1/ y □ = ∣ g □ ∣ − 2/3 . (270)
Here a 0 a_{0} a 0 is a microscopic length. The intra sublattice couplings are kept at K x = K y = Δ x = Δ y = K c K_{x}=K_{y}=\Delta_{x}=\Delta_{y}=K_{c} K x = K y = Δ x = Δ y = K c , while only K □ K_{\square} K □ and Δ □ \Delta_{\square} Δ □ are varied.
The product of four critical learning theories is therefore unstable to plaquette deformation and measurement. Along K □ = Δ □ K_{\square}=\Delta_{\square} K □ = Δ □ , the higher Nishimori identities and the correlator hierarchy remain exact despite this instability. For unequal couplings, the same power counting establishes departure from the decoupled fixed point. We leave the subsequent RG flow and the resulting phase diagram for future investigation [46 ] .