Introduction
Stabilizer measurements provide a direct way to prepare quantum error-correcting codes by imposing commuting constraints on a many-body state [1, 2]. Preparing topological order from a product state by bounded local unitary dynamics requires a time that grows with system size [3]. Measurements and classical feed-forward can instead prepare such states with constant quantum depth [5 ] [4], with constructions extending to a broad range of topological orders [6 , 7 ] . A toric-code implementation using mid-circuit measurements and feed-forward has also been demonstrated experimentally [8 ] . In the ideal code space, the measured constraints both protect information against local errors and make different physical representatives of a logical operator act identically.
In practice, imperfect control of the system–ancilla interaction can produce deviations from the intended projective measurement [9 , 10 ] . We describe such deviations through finite-strength measurement operators that retain coherent amplitudes across different check sectors and thereby deform the encoding map itself [11 , 12 ] . In many-body states, such measurements can drive phase transitions [10 , 13 , 14 , 15 , 16 ] yet preserve long-range entanglement, including in a three-dimensional toric code [10 , 17 ] . Finite-strength protocols can nevertheless retain long-range entanglement, including in a three-dimensional toric code [10 , 17 ] . For the all-plus branch, the corresponding ferromagnetic Z 2 Z_2 Z 2 gauge theory has an area law at every finite strength in two dimensions but a finite-strength deconfined perimeter-law regime in three dimensions [18 , 19 , 20 , 21 ] . Yet the survival of these nonlocal signatures or of bulk order does not by itself show that the prepared state is a useful quantum code, which must protect an unknown input and keep its logical representatives equivalent. We ask whether both operational requirements are governed by the same bulk order when stabilizer measurements are only approximate.
Approximate quantum error correction (AQEC) relates recoverability to logical information leaked to the noise environment [22 , 23 , 24 ] . Local correctability and subsystem variance connect protection to code geometry and complexity [25 , 26 ] , while studies of noisy topological phases relate quantum-memory stability to mixed-state order and reversibility [27 , 28 , 29 ] . Finite-strength preparation can obstruct recovery [30 ] and make the encoding depend on the chosen logical representative [12 ] . Inspired by earlier work on quasi-locally dressed and locally error-corrected Wilson observables [31 , 32 ] , we ask whether AQEC recoverability can similarly be restored through spatially selective control: can measurement accuracy be concentrated near selected logical supports while the bulk remains at fixed finite strength? If such selective control is sufficient, how must its strength and spatial extent scale with code size? Even then, a second question remains: does the same local control make macroscopically separated representatives act identically on a fixed encoded state, or does their equivalence require an additional condition on the intervening bulk?
We identify a local source of logical leakage in the conditioned Calderbank–Shor–Steane (CSS) encodings Figure 1 (a). A single-qubit X X X error on a selected toric support reverses only O ( 1 ) O(1) O ( 1 ) measured checks. At finite strength, each logical-basis codeword has a nonzero overlap with its error image, with a sign set by the logical input. This makes logical inputs locally distinguishable and allows independent noise to transfer information to the environment. At a fixed nonzero error rate, the leakage persists as the code grows if the incident strengths remain bounded, even in the three-dimensional deconfined phase. The same mechanism leaves nonzero leakage per selected logical qubit in CSS quantum low-density parity-check (qLDPC) encodings at uniformly bounded strength. Along a straight toric representative, d d d local signals can retain finite collective leakage even when each vanishes, requiring the common incident-check strength to grow at least logarithmically with ideal code distance d d d (Theorem 1 and Corollary 1).
Figure 1. Logical structure of the results. (a) Local errors reveal selected logical information in general CSS encodings. Along toric representatives, accumulated leakage makes logarithmic measurement strength necessary. (b) Strengthening nearby checks restores AQEC at low noise and sufficient width. (c) Once both encodings are recoverable, their representative equivalence still depends on the bulk: the Wilson overlap vanishes in 2D with macroscopic unenhanced fillings, whereas sufficiently enhanced tubes restore it to one in the strict 3D deconfined phase.
We prove that enhancing neighborhoods of selected logical representatives restores recovery while the bulk remains at fixed finite strength [Figure 1 (b)]. Increasing the measurement strength of the enhanced checks suppresses error overlaps whenever their syndrome enters the protected region, while increasing its width forces any logical crossing that leaves no protected syndrome to involve a larger support. Under sufficiently weak local-stochastic noise, without assuming independent error locations, these two mechanisms give entanglement fidelity approaching one for a fixed logical register. For straight toric representatives, enhancement strength O ( log d ) O(\log d) O ( log d ) suffices, with widths O ( log d ) O(\log d) O ( log d ) in two dimensions and O ( log d ) O(\sqrt{\log d}) O ( log d ) in three; both constructions use O ( d log d ) O(d\log d) O ( d log d ) enhanced checks, a vanishing fraction of the total. For CSS codes with good linear confinement, the sufficient width is reduced to O ( log log U ) O(\log\log U) O ( log log U ) , where U U U is the union size of the selected supports. The toric logarithmic strength matches the independent-noise lower bound in order, while the recovery construction requires no bulk phase condition (Theorem 2; Corollary 2).
With recoverability established, we compare two straight, parallel, macroscopically separated representatives of the same toric logical class in the same enhanced preparation [Figure 1 (c)]. Their equivalence is measured by the Wilson overlap w = ⟨ Z ℓ Z ℓ ′ ⟩ w=\langle Z^{\ell}Z^{\ell'}\rangle w = ⟨ Z ℓ Z ℓ ′ ⟩ , which tests whether the two physical strings act the same on a fixed encoding. In two dimensions, both fillings relating the strings retain macroscopic unenhanced area, so finite-strength factors accumulate and force w → 0 w\to0 w → 0 , even with projective strips. In the strict three-dimensional deconfined phase, full enhanced tubes suppress the dual obstructions that distinguish the representatives; logarithmic radius and strength suffice to give w → 1 w\to1 w → 1 . The same overlap directly quantifies logical action: substituting one representative for the other in an ideal controlled operation incurs the exact operator-norm error 2 ( 1 − w ) \sqrt{2(1-w)} 2 ( 1 − w ) . Thus recoverability and coherent representative equivalence are distinct requirements (Theorem 3 ; Corollary 4).
Two implications follow for code preparation and logical control:
Preparation accuracy can be concentrated near selected logical supports. For the stated toric protocols, strengthening a vanishing fraction of checks suffices for recovery at fixed logical-register size. The required accuracy scales with accumulated local leakage, rather than the location of a bulk phase boundary.
Local enhancement and three-dimensional deconfinement jointly restore representative equivalence. Once local enhancement establishes recoverability, sufficiently strong full tubes in the three-dimensional deconfined phase also restore representative equivalence. The two-dimensional comparison shows that recoverability alone does not ensure this agreement.
Results
We first identify the leakage that the enhanced checks must suppress, then bound the strength and width needed for recovery. With recoverability established, we compare the coherent actions of distant representatives.
Background: finite-strength preparation and logical representatives
Consider a Calderbank–Shor–Steane (CSS) code on n n n physical qubits, with X X X checks A μ = X a μ A_\mu= X^{a_\mu} A μ = X a μ and measured Z Z Z checks B j = Z h j B_j = Z^{h_j} B j = Z h j satisfying [ A μ , B j ] = 0 [A_\mu, B_j] = 0 [ A μ , B j ] = 0 [34 , 35 ] . The vectors h j h_j h j are the rows of the binary matrix H Z H_Z H Z ; all checks are unsigned, and redundant checks are allowed. For the edge-qubit toric codes, A v A_v A v is a vertex star and B p = ∏ e ∈ ∂ p Z e B_p = \prod_{e \in\partial p} Z_e B p = ∏ e ∈ ∂ p Z e is a plaquette check [1, 2]. Starting from ∣ + ⟩ ⊗ n |+\rangle^{\otimes n} ∣ + ⟩ ⊗ n fixes every A μ = + 1 A_\mu= +1 A μ = + 1 . Measuring the commuting B j B_j B j then imposes the remaining constraints. Finite-strength measurements replace these projective constraints by conditional filters [10 , 30 ] ; coherent deviations in syndrome extraction motivate studying this preparation error separately from subsequent noise [11 , 12 ] .
Figure 2(a) gives a circuit for one check. Prepare an ancilla in ∣ + ⟩ a |+\rangle_a ∣ + ⟩ a , apply U j ( t j ) = exp [ − i t j Z a ⊗ B j ] U_j(t_j) = \exp[-\mathrm{i}t_j Z_a \otimes B_j] U j ( t j ) = exp [ − i t j Z a ⊗ B j ] , and measure Y a Y_a Y a with result s j = ± 1 s_j = \pm1 s j = ± 1 . Here t j ∈ [ 0 , π / 4 ] t_j \in[0,\pi/4] t j ∈ [ 0 , π /4 ] is a dimensionless interaction angle. Taking the ancilla matrix element gives, up to a branch-global phase [9 , 10 , 30 ] ,
M j , s j = cos t j I + s j sin t j B j 2 = e s j β j B j / 2 2 cosh β j , (1) M_{j,s_j} = \frac{\cos t_j I + s_j \sin t_j B_j}{\sqrt{2}} = \frac{e^{s_j\beta_j B_j/2}}{\sqrt{2\cosh\beta_j}}, \tag*{(1)} M j , s j = 2 cos t j I + s j sin t j B j = 2 cosh β j e s j β j B j /2 , (1)
tanh ( β j / 2 ) = tan t j , tanh β j = sin ( 2 t j ) . \tanh(\beta_j/2) = \tan t_j,\qquad\tanh\beta_j = \sin(2t_j). tanh ( β j /2 ) = tan t j , tanh β j = sin ( 2 t j ) .
Thus β j = 2 artanh ( tan t j ) \beta_j = 2\operatorname{artanh}(\tan t_j) β j = 2 artanh ( tan t j ) is the dimensionless filter strength. At t j = 0 t_j = 0 t j = 0 (β j = 0 \beta_j = 0 β j = 0 ), M j , s j = I / 2 M_{j,s_j} = I/\sqrt{2} M j , s j = I / 2 and the readout carries no check information; at t j = π / 4 t_j = \pi/4 t j = π /4 (β j → ∞ \beta_j \to\infty β j → ∞ ), it becomes the projector ( I + s j B j ) / 2 (I + s_j B_j)/2 ( I + s j B j ) /2 . At intermediate strength, the conditional operation preserves coherence between check sectors while changing their relative amplitudes. The exact ancilla contraction and its outcome probabilities are derived in Supplementary Note 1.
We condition throughout on the all-plus record, s j = + 1 s_j = +1 s j = + 1 for every measured check. Multiplying the commuting Kraus operators and normalizing gives
∣ ψ β ⟩ = exp ( 1 2 ∑ j β j B j ) ∣ + ⟩ ⊗ n Z β , |\psi_\beta\rangle= \frac{\exp\left(\frac{1}{2}\sum_j \beta_j B_j\right)|+\rangle^{\otimes n}}{\sqrt{Z_\beta}}, ∣ ψ β ⟩ = Z β exp ( 2 1 ∑ j β j B j ) ∣ + ⟩ ⊗ n ,
Z β = ⟨ + ∣ ⊗ n exp ( ∑ j β j B j ) ∣ + ⟩ ⊗ n , (2) Z_\beta= \langle+|^{\otimes n}\exp\left(\sum_j \beta_j B_j\right)|+\rangle^{\otimes n}, \tag*{(2)} Z β = ⟨ + ∣ ⊗ n exp ( j ∑ β j B j ) ∣ + ⟩ ⊗ n , (2)
p + = Z β ∏ j 2 cosh β j . p_+ = \frac{Z_\beta}{\prod_j 2\cosh\beta_j}. p + = ∏ j 2 cosh β j Z β .
Here p + p_+ p + is the probability of obtaining that record, and β \beta β in subscripts abbreviates the full strength profile. This is a conditional pure-state preparation. Averaging over, or discarding, the readout instead defines a different quantum channel [10 , 17 ] . We do not assume that a generic outcome record can be converted to this branch at no cost; Supplementary Note 1 gives the general-record state, the postselection probability including redundant checks, and the distinction from outcome averaging.
Choose k k k ideal Z Z Z -logical representatives that are independent modulo the Z Z Z stabilizers, and set K = 2 k K = 2^k K = 2 k . They define an encoding V β : H L → H Q \mathcal{V}_\beta: \mathcal{H}_L \to\mathcal{H}_Q V β : H L → H Q , where L L L is the logical input register and Q Q Q is the n n n -qubit physical system,
V β ∣ q ⟩ L = ∣ q β ⟩ = Z ℓ ( q ) ∣ ψ β ⟩ , \mathcal{V}_\beta|q\rangle_L = |q_\beta\rangle= Z^{\ell(q)}|\psi_\beta\rangle, V β ∣ q ⟩ L = ∣ q β ⟩ = Z ℓ ( q ) ∣ ψ β ⟩ ,
ℓ ( q ) = ⨁ a = 1 k q a ℓ a , q ∈ F 2 k . (3) \ell(q) = \bigoplus_{a=1}^{k} q_a \ell_a,\qquad q \in\mathbb{F}_2^k. \tag*{(3)} ℓ ( q ) = a = 1 ⨁ k q a ℓ a , q ∈ F 2 k . (3)
Here ∣ q ⟩ L |q\rangle_L ∣ q ⟩ L is an abstract logical basis state and ∣ q β ⟩ |q_\beta\rangle ∣ q β ⟩ is its physical codeword. Complementary physical logical X X X operators distinguish these codewords, so V β † V β = I K \mathcal{V}_\beta^\dagger\mathcal{V}_\beta= I_K V β † V β = I K . Postselection has input-independent success probability, including for reference-entangled inputs (Appendix and Supplementary Note 1). Write P β = V β V β † P_\beta= \mathcal{V}_\beta\mathcal{V}_\beta^\dagger P β = V β V β † . The register can contain some or all ideal logical qubits; d d d denotes the ideal code distance, not an exact distance of the imperfect subspace. In the toric examples, qubits lie on edges of a periodic square or cubic lattice and the measured checks are plaquette boundaries. Here d = L d = L d = L , the torus side length.
The focus on the B j B_j B j sector follows from the asymmetry of this preparation. Every A μ A_\mu A μ remains an exact stabilizer of the entire encoded subspace; it stabilizes the initial state and commutes with both the filter and the chosen logical Z Z Z operators. In the toric language, the gauge charges remain absent while plaquette fluxes fluctuate, as in the finite-strength measurement construction of Ref. [10 ] . More strongly, the filtered encoding retains exact correction of any set of pure Z Z Z errors whose pairwise differences have weight below the ideal Z Z Z distance d Z d_Z d Z . Supplementary Note 1, “Why the B j B_j B j sector is the relevant recovery problem,” proves this statement from the Knill–Laflamme condition [36 ] ; it does not require the B j B_j B j to be sharp. Physical X X X errors, by contrast, anticommute with the weakened B j B_j B j checks and acquire the residual logical X X X operators studied below. We therefore analyze the quantum recovery problem for subsequent X X X noise. This sector-specific recovery problem concerns arbitrary encoded quantum inputs, including their entanglement with a reference.
The Born weights of Eq. (2) define a classical multispin model for any CSS code in this construction, with the generalized Ising form of Ref. [18 ] . For a computational-basis configuration s ∈ F 2 r s \in\mathbb{F}_2^r s ∈ F 2 r , its action and probability measure are
S β ( s ) = − ∑ j β j ( − 1 ) ( H Z s ) j , S_\beta(s) = -\sum_j \beta_j(-1)^{(H_Zs)_j}, S β ( s ) = − j ∑ β j ( − 1 ) ( H Z s ) j ,
π β ( s ) = e − S β ( s ) ∑ s ′ e − S β ( s ′ ) . (4) \pi_\beta(s) = \frac{e^{-S_\beta(s)}}{\sum_{s'}e^{-S_\beta(s')}}. \tag*{(4)} π β ( s ) = ∑ s ′ e − S β ( s ′ ) e − S β ( s ) . (4)
Diagonal quantum expectations equal averages in this measure. For toric codes, each check eigenvalue is the plaquette product ∏ e ∈ ∂ p ( − 1 ) s e \prod_{e\in\partial p}(-1)^{s_e} ∏ e ∈ ∂ p ( − 1 ) s e , so S β S_\beta S β is the ferromagnetic Z 2 \mathbb{Z}_2 Z 2 Wilson action on the square or cubic lattice. The fixed all-plus branch introduces no measurement-record disorder.
Two representatives ℓ , ℓ ′ \ell,\ell' ℓ , ℓ ′ of one logical class differ by a check product W = Z c W=Z^c W = Z c , with c = ℓ ⊕ ℓ ′ ∈ im H Z T c=\ell\oplus\ell'\in\operatorname{im}H_Z^T c = ℓ ⊕ ℓ ′ ∈ im H Z T . Their difference on the prepared state satisfies
∥ ( Z ℓ − Z ℓ ′ ) ∣ ψ β ⟩ ∥ 2 = 2 ( 1 − ⟨ W ⟩ β ) . (5) \left\Vert(Z^\ell-Z^{\ell'})|\psi_\beta\rangle\right\Vert^2=2(1-\langle W\rangle_\beta). \tag*{(5)} ( Z ℓ − Z ℓ ′ ) ∣ ψ β ⟩ 2 = 2 ( 1 − ⟨ W ⟩ β ) . (5)
The states ∣ ϕ 1 ⟩ = Z ℓ ∣ ψ β ⟩ |\phi_1\rangle=Z^\ell|\psi_\beta\rangle ∣ ϕ 1 ⟩ = Z ℓ ∣ ψ β ⟩ and ∣ ϕ 2 ⟩ = Z ℓ ′ ∣ ψ β ⟩ |\phi_2\rangle=Z^{\ell'}|\psi_\beta\rangle ∣ ϕ 2 ⟩ = Z ℓ ′ ∣ ψ β ⟩ therefore have squared fidelity F r e p = ∣ ⟨ W ⟩ β ∣ 2 F_{\rm rep}=|\langle W\rangle_\beta|^2 F rep = ∣ ⟨ W ⟩ β ∣ 2 Figure 2 (b). Expanding each check weight gives the exact general-code expression
Figure 2. Finite-strength measurement, logical representatives, and Wilson laws. (a) Ancilla-assisted measurement of a plaquette check B p B_p B p . All-plus postselection over the commuting checks prepares ∣ ψ β ⟩ |\psi_\beta\rangle ∣ ψ β ⟩ from ∣ + ⟩ ⊗ n |+\rangle^{\otimes n} ∣ + ⟩ ⊗ n [10 , 30 ] . (b) Homologous logical strings bound a filling S S S , with C = ∂ S C=\partial S C = ∂ S and W ( C ) = Z ℓ 1 Z ℓ 2 W(C)=Z^{\ell_1}Z^{\ell_2} W ( C ) = Z ℓ 1 Z ℓ 2 . Their squared state fidelity F r e p F_{\rm rep} F rep tests representative equivalence. (c) At uniform strength, 2D has an area law for every fixed 0 < β < ∞ 0<\beta<\infty 0 < β < ∞ , whereas 3D undergoes an area-to-perimeter transition at finite β c \beta_c β c [20 , 21 , 33 ] . Here A A A and P P P denote filling area and boundary length. The displayed laws give leading large-loop scaling away from criticality. A bare perimeter law still allows F r e p → 0 F_{\rm rep}\to0 F rep → 0 for growing, macroscopically separated strings and does not by itself restore their equivalence.
⟨ W ⟩ β = ∑ a H Z T a = c ∏ j : a j = 1 tanh β j ∑ a H Z T a = 0 ∏ j : a j = 1 tanh β j . (6) \langle W\rangle_\beta=
\frac{\sum_{\substack{a\\H_Z^Ta=c}}\prod_{j:a_j=1}\tanh\beta_j}
{\sum_{\substack{a\\H_Z^Ta=0}}\prod_{j:a_j=1}\tanh\beta_j}. \tag*{(6)} ⟨ W ⟩ β = ∑ a H Z T a = 0 ∏ j : a j = 1 tanh β j ∑ a H Z T a = c ∏ j : a j = 1 tanh β j . (6)
Here a a a selects a binary subset of measured checks. The numerator weights every check product implementing the change of representative; the denominator accounts for relations among the checks.
For the two-dimensional toric code, the support of c c c forms a boundary C C C . The only closed plaquette sets on the torus are the empty set and the full lattice. A filling of A A A plaquettes has a unique complementary filling of L 2 − A L^2-A L 2 − A plaquettes. At uniform strength, Eq. (6) reduces to the finite-torus expression [37 Eq. (65)]
⟨ W ( C ) ⟩ β = ( tanh β ) A + ( tanh β ) L 2 − A 1 + ( tanh β ) L 2 . (7) \langle W(C)\rangle_\beta=
\frac{(\tanh\beta)^A+(\tanh\beta)^{L^2-A}}
{1+(\tanh\beta)^{L^2}}. \tag*{(7)} ⟨ W ( C ) ⟩ β = 1 + ( tanh β ) L 2 ( tanh β ) A + ( tanh β ) L 2 − A . (7)
As both fillings grow, − log ⟨ W ( C ) ⟩ β -\log\langle W(C)\rangle_\beta − log ⟨ W ( C ) ⟩ β equals − log ( tanh β ) -\log(\tanh\beta) − log ( tanh β ) times the smaller area, up to a bounded correction. This area coefficient is positive at every fixed 0 < β < ∞ 0<\beta<\infty 0 < β < ∞ ; there is no finite-strength deconfined phase.
In three dimensions, closed plaquette surfaces include cube boundaries and noncontractible surfaces. At uniform strength, summing over all plaquette sets S S S on the cubic torus gives
⟨ W ( C ) ⟩ β = ∑ S : ∂ S = C ( tanh β ) ∣ S ∣ ∑ S : ∂ S = 0 ( tanh β ) ∣ S ∣ . (8) \langle W(C)\rangle_\beta=
\frac{\sum_{S:\,\partial S=C}(\tanh\beta)^{|S|}}
{\sum_{S:\,\partial S=0}(\tanh\beta)^{|S|}}. \tag*{(8)} ⟨ W ( C ) ⟩ β = ∑ S : ∂ S = 0 ( tanh β ) ∣ S ∣ ∑ S : ∂ S = C ( tanh β ) ∣ S ∣ . (8)
where ∣ S ∣ |S| ∣ S ∣ counts the plaquettes in S S S . The local cube constraints force plaquette defects into closed loops on the dual lattice. At sufficiently large strength these loops are dilute, and their linking with C C C produces a perimeter law. For large rectangular loops of fixed aspect ratio in the thermodynamic limit, the two phases satisfy − log ⟨ W ( C ) ⟩ β ∼ σ β A ( C ) -\log\langle W(C)\rangle_\beta\sim\sigma_\beta A(C) − log ⟨ W ( C ) ⟩ β ∼ σ β A ( C ) for 0 < β < β c 0<\beta<\beta_c 0 < β < β c and − log ⟨ W ( C ) ⟩ β ∼ τ β ∣ C ∣ -\log\langle W(C)\rangle_\beta\sim\tau_\beta|C| − log ⟨ W ( C ) ⟩ β ∼ τ β ∣ C ∣ for β > β c \beta>\beta_c β > β c , with positive coefficients and a finite β c \beta_c β c Figure 2 (c) [20 , 21 , 33 ] . Here A ( C ) A(C) A ( C ) is the enclosed area and ∣ C ∣ |C| ∣ C ∣ the perimeter. A perimeter law slows the decay of F r e p F_{\rm rep} F rep but still allows it to vanish as the boundary grows. Deconfinement therefore does not by itself restore representative equivalence, or establish recovery of a chosen encoding after noise. Appendix and Supplementary Notes 1 and 4 give the gauge mapping, the loop conditions for these asymptotic laws, and the finite-torus estimates used below.
Local logical leakage at finite measurement strength
We characterize recovery through the logical information transferred to the noise environment [30 ] . For the encoding V β V_{\beta} V β in Eq. (3) , consider the physical noise channel
N ( ρ ) = ∑ x ⊆ [ n ] P ( x ) X x ρ X x , X x = ∏ e ∈ x X e . (9) \mathcal{N}(\rho)=\sum_{x\subseteq[n]}P(x)X_x\rho X_x,\qquad X_x=\prod_{e\in x}X_e. \tag*{(9)} N ( ρ ) = x ⊆ [ n ] ∑ P ( x ) X x ρ X x , X x = e ∈ x ∏ X e . (9)
Throughout this subsection, the flips are independent, with P ( x ) = p ∣ x ∣ ( 1 − p ) n − ∣ x ∣ P(x)=p^{|x|}(1-p)^{n-|x|} P ( x ) = p ∣ x ∣ ( 1 − p ) n − ∣ x ∣ and fixed 0 < p < 1 0<p<1 0 < p < 1 . Writing its Kraus operators as E x = P ( x ) X x E_x=\sqrt{P(x)}X_x E x = P ( x ) X x , the overlaps between corrupted code states form the normalized Knill–Laflamme matrices [30 , 36 ] ,
( Λ + B ) x y , q q ′ = 1 K ⟨ q β ∣ E x † E y ∣ q β ′ ⟩ , (\Lambda+B)_{xy,qq'}=\frac{1}{K}\langle q_{\beta}|E_x^\dagger E_y|q'_{\beta}\rangle, ( Λ + B ) x y , q q ′ = K 1 ⟨ q β ∣ E x † E y ∣ q β ′ ⟩ ,
Λ x y , q q ′ = δ q q ′ K 2 ∑ u ⟨ u β ∣ E x † E y ∣ u β ⟩ . (10) \Lambda_{xy,qq'}=\frac{\delta_{qq'}}{K^2}\sum_{u}\langle u_{\beta}|E_x^\dagger E_y|u_{\beta}\rangle. \tag*{(10)} Λ x y , q q ′ = K 2 δ q q ′ u ∑ ⟨ u β ∣ E x † E y ∣ u β ⟩ . (10)
Here x , y x,y x , y label error branches and q , q ′ q,q' q , q ′ label logical inputs, with rows indexed by ( q , x ) (q,x) ( q , x ) and columns by ( q ′ , y ) (q',y) ( q ′ , y ) . The matrix Λ \Lambda Λ retains the logical average, while B B B records the dependence on the encoded state. Exact correction is possible precisely when B = 0 B=0 B = 0 .
For a logical input maximally entangled with a reference R R R , the same matrices describe correlations with the noise environment E E E [Figure 3 (a)]: Λ + B = ρ R E T \Lambda+B=\rho_{RE}^{T} Λ + B = ρ R E T and Λ = ( I R / K ⊗ ρ E ) T \Lambda=(I_R/K\otimes\rho_E)^{T} Λ = ( I R / K ⊗ ρ E ) T , where T T T denotes the full transpose in the chosen basis. The AQEC relative entropy and the trace-norm deviation therefore measure the remaining correlations between R R R and E E E :
Figure 3. AQEC framework and local logical leakage. (a) A logical input L L L is maximally entangled with a reference R R R , encoded by V β V_{\beta} V β , and subjected to noise N \mathcal{N} N . Recovery R \mathcal{R} R acts on the physical output Q Q Q . The normalized error-overlap matrix Λ + B \Lambda+B Λ + B describes the joint reference–environment state, whereas Λ \Lambda Λ retains its logical average; T T T denotes the full transpose. The AQEC relative entropy measures correlations between R R R and the noise environment E E E . (b) A single-edge error X e X_e X e on a selected toric logical support reverses its incident plaquettes. Opposite logical parities give the reduced states ρ e ± \rho_e^{\pm} ρ e ± and the single-edge subsystem variance shown. The leakage bounds assume independent X X X flips with fixed 0 < p < 1 0<p<1 0 < p < 1 . At uniform finite strength, κ β ( e ) ≥ e − s β > 0 \kappa_{\beta}(e)\ge e^{-s\beta}>0 κ β ( e ) ≥ e − s β > 0 , with s = 2 s=2 s = 2 in 2D (shown) and s = 4 s=4 s = 4 in 3D, so this local obstruction persists as the code grows.
D A Q E C = D ( Λ + B ∥ Λ ) = I ( R : E ) , D_{\mathrm{AQEC}}=D(\Lambda+B\Vert\Lambda)=I(R:E), D AQEC = D ( Λ + B ∥Λ ) = I ( R : E ) ,
ε = 1 2 ∥ B ∥ 1 = 1 2 ∥ ρ R E − I R K ⊗ ρ E ∥ 1 . (11) \varepsilon=\frac{1}{2}\lVert B\rVert_1=\frac{1}{2}\left\lVert\rho_{RE}-\frac{I_R}{K}\otimes\rho_E\right\rVert_1. \tag*{(11)} ε = 2 1 ∥ B ∥ 1 = 2 1 ρ R E − K I R ⊗ ρ E 1 . (11)
At fixed register size k k k , vanishing D A Q E C D_{\mathrm{AQEC}} D AQEC or ∥ B ∥ 1 \lVert B\rVert_1 ∥ B ∥ 1 is equivalent to F e ∗ → 1 F_e^*\to1 F e ∗ → 1 , where F e ∗ F_e^* F e ∗ is the optimal squared entanglement fidelity for recovery of the whole register [22 , 24 , 38 , 39 ] . Recovery acts on the physical output and has no access to E E E . Its definition and quantitative bounds are given in Appendix.
The prepared code can also reveal logical information through a local measurement before noise acts. For a physical region A A A , this dependence is measured by the subsystem variance [26 ] ,
ε S V ( A ) = max ρ L ∥ Tr A c ( V β ρ L V β † − P β K ) ∥ 1 . (12) \varepsilon_{\mathrm{SV}}(A)=\max_{\rho_L}\left\lVert\operatorname{Tr}_{A^c}\left(V_{\beta}\rho_LV_{\beta}^{\dagger}-\frac{P_{\beta}}{K}\right)\right\rVert_1. \tag*{(12)} ε SV ( A ) = ρ L max Tr A c ( V β ρ L V β † − K P β ) 1 . (12)
The maximization includes all logical states, including coherent superpositions, and P β / K P_{\beta}/K P β / K is the maximally mixed code state. Here ∥ ⋅ ∥ 1 \lVert\cdot\rVert_1 ∥ ⋅ ∥ 1 is the full trace norm. Thus ε S V ( A ) = 0 \varepsilon_{\mathrm{SV}}(A)=0 ε SV ( A ) = 0 means that every encoded state has the same marginal on A A A ; no measurement confined to that region can distinguish logical inputs. This is a property of the prepared subspace, while D A Q E C D_{\mathrm{AQEC}} D AQEC also depends on the noise channel.
For the encoding prepared at finite strength, both quantities are controlled by the residual overlap of physical errors. Let σ ( x ) = sup ( H Z x ) \sigma(x)=\sup(H_Zx) σ ( x ) = sup ( H Z x ) be the syndrome of x x x and g ( x ) = ( ∣ x ∣ mod 2 ) a = 1 k g(x)=(|x|\operatorname{mod}2)^k_{a=1} g ( x ) = ( ∣ x ∣ mod 2 ) a = 1 k its parity on the selected representatives. The exact logical sectors give
⟨ q β ∣ X x ∣ q β ′ ⟩ = δ q q ′ χ q ( x ) κ β ( x ) , \langle q_{\beta}|X_x|q'_{\beta}\rangle=\delta_{qq'}\chi_q(x)\kappa_{\beta}(x), ⟨ q β ∣ X x ∣ q β ′ ⟩ = δ q q ′ χ q ( x ) κ β ( x ) ,
κ β ( x ) = ⟨ ψ β ∣ X x ∣ ψ β ⟩ , χ q ( x ) = ( − 1 ) q ⋅ g ( x ) . (13) \kappa_{\beta}(x)=\langle\psi_{\beta}|X_x|\psi_{\beta}\rangle,\qquad\chi_q(x)=(-1)^{q\cdot g(x)}. \tag*{(13)} κ β ( x ) = ⟨ ψ β ∣ X x ∣ ψ β ⟩ , χ q ( x ) = ( − 1 ) q ⋅ g ( x ) . (13)
For g ( x ) ≠ 0 g(x)\ne0 g ( x ) = 0 , the sign distinguishes logical inputs. A single edge on a toric logical string already has this property [Figure 3 (b)]: opposite logical parities give reduced states ρ e ± = ( I ± κ β ( e ) X e ) / 2 \rho_e^{\pm}=(I\pm\kappa_{\beta}(e)X_e)/2 ρ e ± = ( I ± κ β ( e ) X e ) /2 . In a dilation of the independent noise channel, the environment qubit associated with this edge has the same diagonal probabilities 1 − p , p 1-p,p 1 − p , p for both inputs, but opposite off-diagonal elements ± p ( 1 − p ) κ β ( e ) \pm\sqrt{p(1-p)}\,\kappa_{\beta}(e) ± p ( 1 − p ) κ β ( e ) . Logical information is therefore carried by environment coherences even though the error probabilities are input-independent. Since X e X_e X e reverses only finitely many checks, this dependence survives at fixed finite strength as the code grows, including in the three-dimensional Wilson perimeter-law regime. A longer representative also exposes more sites carrying the same logical parity. The following bounds quantify both the information revealed at one site and its accumulation along the selected support.
Theorem 1. (Local leakage and its accumulation). For the encoding in Eq. (3) with nonnegative measurement strengths, any site with g ( e ) ≠ 0 g(e) \ne0 g ( e ) = 0 obeys
ε S V ( { e } ) ≥ κ β ( e ) ≥ exp [ − ∑ j ∈ σ ( e ) β j ] . (14) \varepsilon_{\mathrm{SV}}(\{e\}) \ge\kappa_\beta(e) \ge\exp\left[-\sum_{j\in\sigma(e)}\beta_j\right]. \tag*{(14)} ε SV ({ e }) ≥ κ β ( e ) ≥ exp − j ∈ σ ( e ) ∑ β j . (14) Under independent X X X noise with fixed 0 < p < 1 0 < p < 1 0 < p < 1 ,
∥ B ∥ 1 ≥ 2 p ( 1 − p ) κ β ( e ) , \lVert B\rVert_1 \ge2\sqrt{p(1-p)}\,\kappa_\beta(e), ∥ B ∥ 1 ≥ 2 p ( 1 − p ) κ β ( e ) , D A Q E C ≥ 2 p ( 1 − p ) κ β ( e ) 2 . (15) D_{\mathrm{AQEC}} \ge2p(1-p)\kappa_\beta(e)^2. \tag*{(15)} D AQEC ≥ 2 p ( 1 − p ) κ β ( e ) 2 . (15) If M M M sites share the same nonzero logical parity, have pairwise disjoint syndrome supports of size s s s , and every check in these supports has strength β \beta β , then
D A Q E C ≥ 2 p ( 1 − p ) M e − 2 s β 1 + 2 2 s + 2 p ( 1 − p ) M e − 2 s β . (16) D_{\mathrm{AQEC}} \ge\frac{2p(1-p)Me^{-2s\beta}}{1+2^{2s+2}p(1-p)Me^{-2s\beta}}. \tag*{(16)} D AQEC ≥ 1 + 2 2 s + 2 p ( 1 − p ) M e − 2 s β 2 p ( 1 − p ) M e − 2 s β . (16) The remaining strengths may be arbitrary and nonnegative. For straight toric representatives, M = d M=d M = d and s = 2 s=2 s = 2 or 4 4 4 ; at fixed k k k and p p p , asymptotically faithful recovery requires
β − log d 2 s → + ∞ . \beta-\frac{\log d}{2s}\to+\infty. β − 2 s log d → + ∞.
The obstruction depends only on the checks reversed by the local error. Increasing measurements elsewhere cannot remove it while these incident strengths remain bounded Figure 3 (b). In edge-qubit toric codes, a single edge reverses two plaquettes in two dimensions and four in three dimensions. For g ( e ) ≠ 0 g(e)\ne0 g ( e ) = 0 , its subsystem variance equals κ β ( e ) \kappa_\beta(e) κ β ( e ) exactly. At uniform strength and L ≥ 3 L\ge3 L ≥ 3 ,
κ β ( e ) = sech 2 β 1 + ( tanh β ) L 2 ( 2 D ) , \kappa_\beta(e)=\frac{\operatorname{sech}^{2}\beta}{1+(\tanh\beta)^{L^2}}\qquad(2\mathrm{D}), κ β ( e ) = 1 + ( tanh β ) L 2 sech 2 β ( 2 D ) ,
e − 4 β ≤ κ β ( e ) ≤ sech 4 β ( 3 D ) . (17) e^{-4\beta}\le\kappa_\beta(e)\le\operatorname{sech}^{4}\beta\qquad(3\mathrm{D}). \tag*{(17)} e − 4 β ≤ κ β ( e ) ≤ sech 4 β ( 3 D ) . (17)
Both remain bounded away from zero at every fixed finite strength. The three-dimensional deconfined phase therefore retains a local recovery obstruction despite its perimeter-law Wilson correlations.
The same mechanism applies to CSS quantum low-density parity-check (qLDPC) encodings, where bounded check incidence takes the place of lattice geometry. Physical sites can be chosen that probe independent logical parities of the selected register. Their combined contribution leaves a finite leakage per logical qubit, even when the register grows.
Corollary 1. Suppose each physical qubit belongs to at most Δ Z \Delta_Z Δ Z measured checks and all strengths satisfy β j ≤ β 0 < ∞ \beta_j\le\beta_0<\infty β j ≤ β 0 < ∞ , uniformly in size. For any nonempty selected register and fixed independent noise rate 0 < p < 1 0<p<1 0 < p < 1 ,
D A Q E C k ≥ 2 p ( 1 − p ) e − 2 Δ Z β 0 , \frac{D_{\mathrm{AQEC}}}{k}\ge2p(1-p)e^{-2\Delta_Z\beta_0}, k D AQEC ≥ 2 p ( 1 − p ) e − 2 Δ Z β 0 , 1 − F e ∗ ≥ p ( 1 − p ) 4 e − 2 Δ Z β 0 . (18) 1-F_e^*\ge\frac{p(1-p)}{4}e^{-2\Delta_Z\beta_0}. \tag*{(18)} 1 − F e ∗ ≥ 4 p ( 1 − p ) e − 2 Δ Z β 0 . (18)
This result shows the fixed-strength obstruction for general CSS qLDPC encodings. Increasing the ideal distance therefore cannot compensate for uniformly bounded preparation strength under fixed nonzero noise. The infidelity bound concerns recovery of the entire selected register and remains uniform in k k k . The entropy-density bound further shows that the obstruction cannot be diluted by storing more logical qubits.
Even when the preparation strength grows, suppressing each local signal separately need not suppress their combined information. For example, in two dimensions β = 1 4 log L \beta=\frac{1}{4}\log L β = 4 1 log L gives max e ε S V ( { e } ) ∼ 4 / L → 0 \max_e\varepsilon_{\mathrm{SV}}(\{e\})\sim4/\sqrt{L}\to0 max e ε SV ({ e }) ∼ 4/ L → 0 , while Eq. (16) leaves D A Q E C ≥ 2 p ( 1 − p ) / [ 1 + 64 p ( 1 − p ) ] > 0 D_{\mathrm{AQEC}}\ge2p(1-p)/[1+64p(1-p)]>0 D AQEC ≥ 2 p ( 1 − p ) / [ 1 + 64 p ( 1 − p )] > 0 . Thus every single-qubit marginal can become independent of the logical input in the limit, while the noise environment retains a finite amount of logical information.
For straight toric strings, the number of such signals is M = d M=d M = d . At fixed k k k and p p p , their total leakage can vanish only if
β − 1 4 log d → + ∞ ( 2 D ) , β − 1 8 log d → + ∞ ( 3 D ) . (19) \beta-\frac{1}{4}\log d\to+\infty\qquad(2\mathrm{D}),\qquad
\beta-\frac{1}{8}\log d\to+\infty\qquad(3\mathrm{D}). \tag*{(19)} β − 4 1 log d → + ∞ ( 2 D ) , β − 8 1 log d → + ∞ ( 3 D ) . (19)
The necessary strength concerns only checks incident on the selected string, suggesting that stronger preparation measurements can be concentrated near the chosen supports. Appendix and Supplementary Note 2 give the local and collective leakage derivations.
Restoring recoverability by local measurement enhancement
The preceding result requires logarithmic measurement strength near a straight toric representative. Is the scaling β ∼ log d \beta\sim\log d β ∼ log d also sufficient to make the AQEC relative entropy vanish? How large a neighborhood of the selected logical representatives must be strengthened? For a fixed logical register, logarithmic strengthening does suffice under sufficiently weak noise. Only a vanishing fraction of the toric checks need be enhanced; the required width is determined by the geometry of logical crossings.
We increase the preparation strength on a set of checks P r \mathcal{P}_r P r surrounding the selected logical supports, with r r r denoting the width of this neighborhood:
β j = { β R , j ∈ P r , β , j ∉ P r , β R ≥ β ≥ 0. (20) \beta_j=
\begin{cases}
\beta_R, & j\in\mathcal{P}_r,\\
\beta, & j\notin\mathcal{P}_r,
\end{cases}
\qquad\beta_R\ge\beta\ge0. \tag*{(20)} β j = { β R , β , j ∈ P r , j ∈ / P r , β R ≥ β ≥ 0. (20)
The bulk strength β \beta β remains fixed and finite. Figure 4 illustrates the resulting preparation around a chosen representative. Stronger checks reduce the residual overlap of short error crossings with syndrome in P r \mathcal{P}_r P r . Increasing the width catches syndrome that would lie outside a narrower region. A crossing can still escape these checks, but in the geometries considered below it must then extend across a larger support. At low noise, these extended crossings make a small contribution to logical leakage.
Figure 4. Local enhancement suppresses logical leakage. Checks P r \mathcal{P}_r P r in the shaded strip R R R are prepared at strength β R \beta_R β R ; the bulk remains at fixed β \beta β . Question marks indicate reduced logical information available to local noise. In the upper inset, stronger checks suppress the overlap of a short crossing u u u with syndrome on two enhanced checks. In the lower inset, a longer crossing has syndrome only outside the enhanced region. Widening the region raises the minimum connected support size M ( r ) M(r) M ( r ) of crossings with nonzero selected logical parity that remain silent inside, suppressing their contribution through λ M ( r ) \lambda^{M(r)} λ M ( r ) . Here λ ≤ λ 0 < 1 \lambda\leq\lambda_0 < 1 λ ≤ λ 0 < 1 accounts for noise strength and connected-support counting; colored dots denote syndrome checks.
The relevant size is M ( r ) M(r) M ( r ) , the smallest connected support of an error difference u u u with nonzero selected logical parity, g ( u ) ≠ 0 g(u) \ne0 g ( u ) = 0 , and no protected syndrome, σ ( u ) ∩ P r = ∅ \sigma(u) \cap\mathcal{P}_r = \varnothing σ ( u ) ∩ P r = ∅ . For a toric string, the parity condition is an odd intersection with the representative. The error may have syndrome outside P r \mathcal{P}_r P r . Connectedness is defined in the bounded-degree graph G Z G_Z G Z joining qubits that share a measured check; we set M ( r ) = ∞ M(r) = \infty M ( r ) = ∞ when no such support exists.
For correlated X X X noise, we assume the local-stochastic bound [40]
Pr [ A ⊆ x ] ≤ η ∣ A ∣ for every A ⊆ [ n ] . (21) \Pr[A \subseteq x] \leq\eta^{|A|} \qquad\text{for every } A \subseteq[n]. \tag*{(21)} Pr [ A ⊆ x ] ≤ η ∣ A ∣ for every A ⊆ [ n ] . (21)
There are at most μ m \mu^m μ m connected supports of size m m m through a given qubit, with μ \mu μ fixed by the local check structure. Their contributions can be summed at sufficiently small η \eta η , with λ = 4 μ η ≤ λ 0 \lambda= 4\mu\sqrt{\eta} \leq\lambda_0 λ = 4 μ η ≤ λ 0 for a fixed 0 < λ 0 < 1 0 < \lambda_0 < 1 0 < λ 0 < 1 . The number of possible starting sites is controlled by the selected-support union U = ⋃ a ℓ a U = \bigcup_a \ell_a U = ⋃ a ℓ a . We can therefore bound leakage under the full physical noise channel by the sizes of crossings near this union, even though errors may occur throughout the code.
Theorem 2 (Recovery under finite-strength local enhancement). For the preparation in Eq. (20) and noise satisfying Eq. (21) with λ ≤ λ 0 < 1 \lambda\leq\lambda_0 < 1 λ ≤ λ 0 < 1 ,
∥ B ∥ 1 ≤ exp [ C U λ M ( r ) 1 − λ ] − 1 \lVert B\rVert_1 \leq\exp\left[\frac{CU\lambda^{M(r)}}{1-\lambda}\right]-1 ∥ B ∥ 1 ≤ exp [ 1 − λ C U λ M ( r ) ] − 1 + 4 ∣ P r ∣ e − β R , (22) {}+4\sqrt{|\mathcal{P}_r|}e^{-\beta_R}, \tag*{(22)} + 4 ∣ P r ∣ e − β R , (22) where C C C depends only on the bounded local check structure and λ ∞ = 0 \lambda^\infty= 0 λ ∞ = 0 . The bound also holds for nonuniform protected strengths β j ≥ β R \beta_j \geq\beta_R β j ≥ β R and finite, nonnegative exterior strengths. Projective enhancement removes the last term.
For fixed k k k ,
U λ M ( r ) ⟶ 0 , ∣ P r ∣ e − β R ⟶ 0 (23) U\lambda^{M(r)} \longrightarrow0,\qquad\sqrt{|\mathcal{P}_r|}e^{-\beta_R} \longrightarrow0 \tag*{(23)} U λ M ( r ) ⟶ 0 , ∣ P r ∣ e − β R ⟶ 0 (23)
are sufficient for D A Q E C → 0 D_{\mathrm{AQEC}} \to0 D AQEC → 0 and F ∗ → 1 F_* \to1 F ∗ → 1 .
The two terms distinguish errors that evade the enhanced checks from the residual effect of checks measured at finite strength. The first includes crossings whose syndrome lies outside the protected region. The second follows by comparing the prepared state with its projection onto the enhanced-check sector; this projection is used only in the proof. No phase condition is imposed on the bulk. What matters for recovery is that the protected neighborhood intercepts short logical crossings and that its checks are sufficiently sharp. The bound guarantees the existence of recovery on the physical output; it does not prescribe a decoder confined to that neighborhood. The estimates are derived in Appendix and Supplementary Note 3.
In toric codes, the protected plaquette ribbons relate the crossing parities of neighboring parallel representatives. An error silent on these ribbons must cross every such string with the same parity. A full strip in two dimensions contains a number of disjoint strings proportional to its width, whereas a full transverse tube in three dimensions contains a number proportional to its cross-sectional area. With the neighborhoods specified in Appendix, this gives
M ( r ) ≥ r ( 2 D ) , M ( r ) ≥ r 2 ( 3 D ) . (24) M(r) \geq r \quad(2D),\qquad M(r) \geq r^2 \quad(3D). \tag*{(24)} M ( r ) ≥ r ( 2 D ) , M ( r ) ≥ r 2 ( 3 D ) . (24)
Thus, for a fixed set of straight representatives with U = O ( d ) U = O(d) U = O ( d ) , widths r = O ( log d ) r = O(\log d) r = O ( log d ) and r = O ( log d ) r = O(\sqrt{\log d}) r = O ( log d ) suffice in two and three dimensions, respectively. The three-dimensional gain comes from protecting the full tube; a single sheet would not enforce the same number of crossings.
For general CSS codes, linear confinement provides a different mechanism for suppressing silent crossings. Let S X = span F 2 { a μ } S_X=\operatorname{span}_{\mathbb{F}_2}\{a_\mu\} S X = span F 2 { a μ } be the support space of the X X X stabilizers. Linear confinement in the X X X -error sector means that
∣ x ∣ r e d = min s ∈ S X ∣ x ⊕ s ∣ , ∣ σ ( x ) ∣ ≥ κ ∣ x ∣ r e d for ∣ x ∣ r e d < T . (25) |x|_{\mathrm{red}}=\min_{s\in S_X}|x\oplus s|,
\qquad
|\sigma(x)|\ge\kappa|x|_{\mathrm{red}}\quad\text{for }|x|_{\mathrm{red}}<T. \tag*{(25)} ∣ x ∣ red = s ∈ S X min ∣ x ⊕ s ∣ , ∣ σ ( x ) ∣ ≥ κ ∣ x ∣ red for ∣ x ∣ red < T . (25)
Here ∣ σ ( x ) ∣ |\sigma(x)| ∣ σ ( x ) ∣ counts violated measured Z Z Z checks, and ∣ x ∣ r e d |x|_{\mathrm{red}} ∣ x ∣ red is the smallest error weight after multiplication by X X X stabilizers. The constant κ > 0 \kappa>0 κ > 0 is independent of code size, and T T T is the cutoff below which this linear syndrome growth holds. A family has good linear confinement if T ≥ c T n α T\ge c_T n^\alpha T ≥ c T n α for constants c T , α > 0 c_T,\alpha>0 c T , α > 0 independent of n n n [41, 42]. Only the X X X -error sector is required here; Supplementary Note 3 states the definition and the reduction convention explicitly.
The linear bound is the robustness property established for quantum expander codes [40, 43]. Good linear confinement was subsequently shown to suffice for single-shot correction under local stochastic noise [41], extending earlier confinement arguments for subsystem codes [44]. It also enters the final decoding step of single-shot preparation via a dimension jump [42] and has been used to construct passive quantum memories at nonzero temperature [45]. Its classical analogue has been used to establish eigenstate localization in quantum-perturbed code Hamiltonians [46].
We enhance all checks touching the radius-( r − 1 ) (r-1) ( r − 1 ) neighborhood of the selected supports in G Z G_Z G Z . To remain silent, an error must cancel the syndrome produced in each layer by extending into the next. If a qubit meets at most Δ Z \Delta_Z Δ Z measured checks, this requires
M ( r ) ≥ min { T , ( 1 + κ / Δ Z ) r } . (26) M(r)\ge\min\{T,(1+\kappa/\Delta_Z)^r\}. \tag*{(26)} M ( r ) ≥ min { T , ( 1 + κ / Δ Z ) r } . (26)
Our recovery bound needs only T ≥ c log U T\ge c\log U T ≥ c log U with c > 1 / ∣ log λ 0 ∣ c>1/|\log\lambda_0| c > 1/∣ log λ 0 ∣ , a weaker requirement than good linear confinement because U ≤ n U\le n U ≤ n . A width O ( log log U ) O(\log\log U) O ( log log U ) then suffices. Bounded degree limits the corresponding enhanced set to C U ( log U ) ν CU(\log U)^\nu C U ( log U ) ν checks, or the total check count if smaller, with C , ν C,\nu C , ν depending only on the local degree and confinement constants. This also makes logarithmic strength in U U U sufficient at fixed k k k . Combining these neighborhood counts with the finite-strength correction gives the resources in Table I.
Corollary 2 (Geometric resources for local enhancement). For a fixed nonempty logical register, uniform local constants, and λ ≤ λ 0 < 1 \lambda\le\lambda_0<1 λ ≤ λ 0 < 1 , the three constructions in Table I give ∥ B ∥ 1 → 0 \lVert B\rVert_1\to0 ∥ B ∥ 1 → 0 , D A Q E C → 0 D_{\mathrm{AQEC}}\to0 D AQEC → 0 , and F ⋆ → 1 F_\star\to1 F ⋆ → 1 at fixed finite bulk strength. The width coefficients are sufficiently large and δ > 0 \delta>0 δ > 0 is fixed. The confinement construction assumes U → ∞ U\to\infty U → ∞ and T ≥ c log U T\ge c\log U T ≥ c log U , with c > 1 / ∣ log λ 0 ∣ c>1/|\log\lambda_0| c > 1/∣ log λ 0 ∣ .
Both toric constructions use O ( d log d ) O(d\log d) O ( d log d ) enhanced checks. This is a vanishing fraction of the Θ ( d 2 ) \Theta(d^2) Θ ( d 2 ) checks in two dimensions and the Θ ( d 3 ) \Theta(d^3) Θ ( d 3 ) checks in three dimensions, as shown in Table I. Their sufficient strength, β R = ( 1 2 + δ ) log d \beta_R=(\frac{1}{2}+\delta)\log d β R = ( 2 1 + δ ) log d , has the same logarithmic order as the lower bounds from the preceding section. Those lower bounds apply to the common protected strength β R \beta_R β R under independent flips at fixed 0 < p < 1 0<p<1 0 < p < 1 . Hence logarithmic strengthening is necessary and sufficient in order for these toric protocols; equality of the coefficients is not established.
For confinement, the smaller radius does not by itself imply a vanishing enhanced fraction. The enhanced fraction vanishes when U ( log U ) ν = o ( N c h k ) U(\log U)^\nu=o(N_{\mathrm{chk}}) U ( log U ) ν = o ( N chk ) , where N c h k N_{\mathrm{chk}} N chk counts all measured Z Z Z checks. The resource estimate depends on the chosen representatives through their union U U U ; it can be written in terms of distance d d d when those representatives additionally satisfy U = O ( d ) U=O(d) U = O ( d ) . The neighborhood counts and strength estimates are derived in Appendix.
These resource savings concern a fixed number of selected logical qubits. Protecting the entire register of a constant-rate code is more demanding: a narrow neighborhood of each representative can collectively occupy an extensive part of the system. In fact, an extensive enhancement is necessary even to make the AQEC entropy per logical qubit vanish.
Corollary 3 (Enhancement cost for a growing logical register). *Suppose each qubit meets at most Δ Z \Delta_Z Δ Z measured checks and all strengths outside an enhanced set are
bounded by β 0 < ∞ \beta_0 < \infty β 0 < ∞ , uniformly in size. Let F h i F_{\mathrm{hi}} F hi be the set of physical qubits touched by the enhanced checks. For k ≥ 1 k \ge1 k ≥ 1 selected logical qubits and independent flips at fixed 0 < p < 1 0 < p < 1 0 < p < 1 ,
D A Q E C k ≥ 2 p ( 1 − p ) e − 2 Δ Z β 0 ( 1 − ∣ F h i ∣ k ) + . (27) \frac{D_{\mathrm{AQEC}}}{k} \ge2p(1-p)e^{-2\Delta_Z\beta_0}\left(1-\frac{|F_{\mathrm{hi}}|}{k}\right)_{+}. \tag*{(27)} k D AQEC ≥ 2 p ( 1 − p ) e − 2 Δ Z β 0 ( 1 − k ∣ F hi ∣ ) + . (27)
where ( z ) + = max { z , 0 } (z)_{+}=\max\{z,0\} ( z ) + = max { z , 0 } . Consequently, D A Q E C / k → 0 D_{\mathrm{AQEC}}/k \to0 D AQEC / k → 0 requires ∣ F h i ∣ ≥ [ 1 − o ( 1 ) ] k |F_{\mathrm{hi}}| \ge[1-o(1)]k ∣ F hi ∣ ≥ [ 1 − o ( 1 )] k . With uniformly bounded check weight and k = Θ ( n ) k=\Theta(n) k = Θ ( n ) , Ω ( n ) \Omega(n) Ω ( n ) enhanced checks are necessary.
Local noise on the qubits outside F h i F_{\mathrm{hi}} F hi still probes at least ( k − ∣ F h i ∣ ) + (k-|F_{\mathrm{hi}}|)_{+} ( k − ∣ F hi ∣ ) + independent logical parities. Each retains a finite leakage contribution at bounded bulk strength, giving the lower bound above (Appendix). A small protection radius can thus coexist with an extensive total enhancement when many representatives are selected. Theorem 2 remains applicable to a growing register using its full support union U U U . Its implications for total AQEC entropy, entropy density, and full-register entanglement recovery are stated separately in Appendix.
Wilson restoration and coherent logical operations
Local enhancement protects an encoded input against local noise in both two and three dimensions. Whether distant representatives act identically on that input remains a separate question. We therefore enhance two straight, parallel representatives ℓ 1 , ℓ 2 \ell_1,\ell_2 ℓ 1 , ℓ 2 of the same logical class in a single preparation, while keeping the intervening bulk at fixed finite strength β \beta β . Their separation is at least α L \alpha L α L , with fixed 0 < α < 1 / 2 0<\alpha<1/2 0 < α < 1/2 . The state ∣ ψ β ⟩ |\psi_\beta\rangle ∣ ψ β ⟩ includes both enhanced regions, as does the full strength profile abbreviated by its subscript.
Each representative defines an encoding isometry V i : H L → H Q V_i:\mathcal{H}_L\to\mathcal{H}_Q V i : H L → H Q for one logical qubit. For i = 1 , 2 i=1,2 i = 1 , 2 , its action in the logical basis of Eq. (3) is
V i ( a ∣ 0 ⟩ L + b ∣ 1 ⟩ L ) = a ∣ ψ β ⟩ + b Z ℓ i ∣ ψ β ⟩ . (28) V_i(a|0\rangle_L+b|1\rangle_L)=a|\psi_\beta\rangle+bZ^{\ell_i}|\psi_\beta\rangle. \tag*{(28)} V i ( a ∣0 ⟩ L + b ∣1 ⟩ L ) = a ∣ ψ β ⟩ + b Z ℓ i ∣ ψ β ⟩ . (28)
Thus V i V_i V i maps an arbitrary logical input into the entire n n n -qubit physical system Q Q Q ; the index i i i specifies the representative used for the image of ∣ 1 ⟩ L |1\rangle_L ∣1 ⟩ L . Both encodings share the image of ∣ 0 ⟩ L |0\rangle_L ∣0 ⟩ L . Within each encoding, the two codewords are orthogonal because they have opposite eigenvalues of a complementary physical logical X X X operator. In the projective code, V 1 = V 2 V_1=V_2 V 1 = V 2 . At finite strength, their agreement is measured by the Wilson overlap
w = ⟨ W ( C ) ⟩ β = ⟨ ψ β ∣ Z ℓ 1 Z ℓ 2 ∣ ψ β ⟩ , (29) w=\langle W(C)\rangle_\beta=\langle\psi_\beta|Z^{\ell_1}Z^{\ell_2}|\psi_\beta\rangle, \tag*{(29)} w = ⟨ W ( C ) ⟩ β = ⟨ ψ β ∣ Z ℓ 1 Z ℓ 2 ∣ ψ β ⟩ , (29)
where C = ℓ 1 ⊕ ℓ 2 C=\ell_1\oplus\ell_2 C = ℓ 1 ⊕ ℓ 2 and 0 ≤ w ≤ 1 0\le w\le1 0 ≤ w ≤ 1 . In particular, w → 1 w\to1 w → 1 makes the two physical encodings agree for every logical input (Appendix).
We now ask whether measurements confined near the two representatives can make w w w approach one while the intervening bulk remains at finite strength. Recovery of the separate encodings is possible in either dimension; the bulk determines whether the same preparation also restores their agreement.
Theorem 3 (Wilson restoration and dimensional separation). In two dimensions, suppose the enhanced strips have width r = o ( L ) r=o(L) r = o ( L ) and both complementary fillings contain at least c L 2 cL^2 c L 2 unenhanced plaquettes. At fixed 0 < β < ∞ 0<\beta<\infty 0 < β < ∞ , even projective enhancement satisfies
0 ≤ w ≤ 2 exp [ − c ∣ log tanh β ∣ L 2 ] ⟶ 0. (30) 0\le w\le2\exp[-c|\log\tanh\beta|L^2]\longrightarrow0. \tag*{(30)} 0 ≤ w ≤ 2 exp [ − c ∣ log tanh β ∣ L 2 ] ⟶ 0. (30) In three dimensions, fix the bulk in the strict deconfined phase of the uniform model [21 Theorem 5] , equivalently
p ∗ ( β ) = 2 e 2 β + 1 < p c F K ( 2 ; Z 3 ) . (31) p^*(\beta)=\frac{2}{e^{2\beta}+1}<p_c^{\mathrm{FK}}(2;\mathbb{Z}^3). \tag*{(31)} p ∗ ( β ) = e 2 β + 1 2 < p c FK ( 2 ; Z 3 ) . (31) where p c F K p_c^{\mathrm{FK}} p c FK is the critical bond parameter of the Fortuin--Kasteleyn (FK) random-cluster model with cluster weight q = 2 q=2 q = 2 [4, 48]. Full enhanced tubes around both representatives give w → 1 w\to1 w → 1 for
r = b ξ β c g log L , b > 1 , r=b\frac{\xi_\beta}{c_g}\log L,\qquad b>1, r = b c g ξ β log L , b > 1 , β R = ( 1 2 + δ ) log L , δ > 0. (32) \beta_R=\left(\frac{1}{2}+\delta\right)\log L,\qquad\delta>0. \tag*{(32)} β R = ( 2 1 + δ ) log L , δ > 0. (32) Here ξ β < ∞ \xi_\beta<\infty ξ β < ∞ is a decay length for subcritical dual clusters and c g > 0 c_g>0 c g > 0 depends on the tube geometry. Both are independent of L L L at fixed bulk strength.
The two-dimensional obstruction comes from the unenhanced filling area (Figure 5 (a)). Each unenhanced plaquette contributes a factor tanh β < 1 \tanh\beta<1 tanh β < 1 , and a macroscopic number remains in each complementary filling. Sharpening checks near the boundary cannot remove this accumulated suppression.
Figure 5. Wilson restoration and representative substitution. Two distant representatives are enhanced in the same preparation, with the bulk held at fixed finite strength. (a) In 2D, macroscopic unenhanced areas in both complementary fillings retain the suppression of the Wilson overlap, giving w → 0 w \to0 w → 0 . (b) In the strict 3D deconfined phase, a dual obstruction avoiding a full protected tube must extend over a distance of order its radius r r r . Subcritical dual-cluster tails and sufficiently strong checks give w → 1 w \to1 w → 1 for the logarithmic radius in Eq. (32) . The transverse inset shows a dual obstruction around one tube. (c) The circuit applies the distant representative to an arbitrary logical input ∣ ϕ ⟩ L |\phi\rangle_L ∣ ϕ ⟩ L encoded by V = V 1 V = V_1 V = V 1 . For control ∣ 1 ⟩ c |1\rangle_c ∣1 ⟩ c , replacing ℓ 1 \ell_1 ℓ 1 by ℓ 2 \ell_2 ℓ 2 produces code-space leakage 1 − w 2 1 - w^2 1 − w 2 . The resulting limits differ despite the separate recoverability of both encodings under the stated low-noise conditions.
In three dimensions, we adapt the random-cluster argument behind the perimeter law [21 ] to the locally enhanced preparation. The indicator coupling used in Refs. [21 , 33 ] equals 2 β p 2\beta_p 2 β p in the action convention of Eq. (4) . The plaquette random-cluster representation assigns the primal occupation parameter θ p = 1 − e − 2 β p \theta_p=1-e^{-2\beta_p} θ p = 1 − e − 2 β p to a check of strength β p \beta_p β p . A dual edge is open when its primal plaquette is absent, and its FK bond parameter is
p e ∗ = 2 ( 1 − θ p ) θ p + 2 ( 1 − θ p ) = 2 e 2 β p + 1 . (33) p_e^*=\frac{2(1-\theta_p)}{\theta_p+2(1-\theta_p)}=\frac{2}{e^{2\beta_p}+1}. \tag*{(33)} p e ∗ = θ p + 2 ( 1 − θ p ) 2 ( 1 − θ p ) = e 2 β p + 1 2 . (33)
These are random-cluster parameters, distinct from the physical noise probability p p p . Stronger checks therefore reduce the dual bond parameters. A dual obstruction avoiding a protected tube must extend over a distance of order r r r , or wind around the torus (Figure 5 (b)). At a fixed bulk point satisfying Eq. (31) , the subcritical FK theorem of Ref. [49] supplies exponential cluster tails. Combining these tails with the tube geometry and the finite-torus duality gives Theorem 3; the finite-strength and topological corrections are derived in Appendix and Supplementary Note 4. Local enhancement thus restores w → 1 w \to1 w → 1 , beyond the decay reduction implied by a bulk perimeter law.
The same preparation also restores AQEC for each encoding when the full tubes contain the ribbons used in § 1.1 B and the noise obeys λ ≤ λ 0 < 1 \lambda\le\lambda_0 < 1 λ ≤ λ 0 < 1 . With d = L d = L d = L , the enhanced checks number O ( L log 2 L ) O(L\log^2 L) O ( L log 2 L ) , a fraction O ( log 2 d / d 2 ) O(\log^2 d/d^2) O ( log 2 d / d 2 ) of all measured checks. The radius O ( log d ) O(\log d) O ( log d ) here exceeds the sufficient O ( log d ) O(\sqrt{\log d}) O ( log d ) radius for AQEC alone in Table 1 ; these are two sufficient constructions, and do not establish a necessary extra cost for Wilson restoration. In two dimensions, the encodings can likewise be separately recoverable while their Wilson overlap vanishes.
Table 1. Sufficient resources for vanishing AQEC relative entropy at fixed logical-register size and uniformly low local-stochastic X X X noise. All rows use finite enhancement strength with δ > 0 \delta>0 δ > 0 . The toric width coefficients b 2 b_2 b 2 , b 3 b_3 b 3 are sufficiently large, and d = L d=L d = L is the ideal code distance. The confinement row uses the actual support union U U U , bounded local degree, and the cutoff condition in Corollary 2; C , ν C,\nu C , ν are independent of code size. Its enhanced fraction vanishes if U ( log U ) ν = o ( N c h k ) U(\log U)^\nu=o(N_{\mathrm{chk}}) U ( log U ) ν = o ( N chk ) . The numerator and denominator count the same measured Z Z Z checks. Widths are rounded up and are sufficient, with no claim of optimality.
This distinction has a direct consequence for logical operations on an already encoded state. Fix the target encoding V = V 1 V = V_1 V = V 1 . With the basis convention of Eq. (3) , the physical string Z ℓ 1 Z^{\ell_1} Z ℓ 1 flips the logical labels, so Z ℓ 1 V = V X L Z^{\ell_1}V = V X_L Z ℓ 1 V = V X L , where X L = ∣ 0 ⟩ L ⟨ 1 ∣ L + ∣ 1 ⟩ L ⟨ 0 ∣ L X_L = |0\rangle_L\langle1|_L + |1\rangle_L\langle0|_L X L = ∣0 ⟩ L ⟨ 1 ∣ L + ∣1 ⟩ L ⟨ 0 ∣ L . For an ideal control qubit c c c , define
G j = ∣ 0 ⟩ c ⟨ 0 ∣ c ⊗ I Q + ∣ 1 ⟩ c ⟨ 1 ∣ c ⊗ Z ℓ j , j = 1 , 2. (34) G_j = |0\rangle_c\langle0|_c \otimes I_Q
+ |1\rangle_c\langle1|_c \otimes Z^{\ell_j}, \qquad j = 1,2. \tag*{(34)} G j = ∣0 ⟩ c ⟨ 0 ∣ c ⊗ I Q + ∣1 ⟩ c ⟨ 1 ∣ c ⊗ Z ℓ j , j = 1 , 2. (34)
Here I Q I_Q I Q is the identity on the physical target. Changing the representative replaces G 1 G_1 G 1 by G 2 G_2 G 2 while leaving the encoded input fixed. The relevant error is therefore the difference between these two actions on the same code space. The following identity relates that error directly to the Wilson overlap.
Corollary 4 (Replacing a representative in an encoded CNOT). With V = V 1 V = V_1 V = V 1 and I c I_c I c the identity on the control,
G 1 ( I c ⊗ V ) = ( I c ⊗ V ) C N O T c → L , G_1(I_c \otimes V) = (I_c \otimes V)\,\mathrm{CNOT}_{c\to L}, G 1 ( I c ⊗ V ) = ( I c ⊗ V ) CNOT c → L , ∥ ( G 2 − G 1 ) ( I c ⊗ V ) ∥ ∞ = 2 ( 1 − w ) , (35) \left\|(G_2-G_1)(I_c\otimes V)\right\|_\infty= \sqrt{2(1-w)}, \tag*{(35)} ∥ ( G 2 − G 1 ) ( I c ⊗ V ) ∥ ∞ = 2 ( 1 − w ) , (35) where C N O T c → L \mathrm{CNOT}_{c\to L} CNOT c → L acts on the control and logical input, and ∥ ⋅ ∥ ∞ \|\cdot\|_\infty ∥ ⋅ ∥ ∞ denotes the operator norm. For control ∣ 1 ⟩ c |1\rangle_c ∣1 ⟩ c , the second representative takes the target outside the fixed code space with probability 1 − w 2 1-w^2 1 − w 2 , for every logical target input.
In the enhanced three-dimensional preparation, either controlled string asymptotically implements the same logical CNOT on the fixed target encoding. In the two-dimensional setting of Theorem 3 , the same substitution causes leakage approaching one for control ∣ 1 ⟩ c |1\rangle_c ∣1 ⟩ c , even when the target encoding is recoverable from the stated stochastic noise. The comparison concerns ideal controlled strings on a fixed encoding; implementing them with noisy local gates requires a separate circuit analysis. The overlap also determines the optimal common recovery when the representative used for encoding is undisclosed, as derived in Appendix.
Discussion
Finite-strength preparation is often judged by how much topological or long-range structure survives away from the projective limit. Our results show that this is not enough to determine whether the prepared state functions as a quantum code. What matters for recovery is how much information about the logical input remains accessible to local errors and, ultimately, to the noise environment. This changes the preparation problem from one of making the entire state uniformly close to an ideal code state to one of controlling the parts of the encoding map that expose logical information. For a fixed logical register, that control can be highly nonuniform: strengthening measurements only near selected logical supports is sufficient for recovery while the rest of the system stays at finite strength. In this sense, the relevant preparation resource is not simply global measurement fidelity, but where accuracy is placed relative to the logical information being protected.
The logarithmic strength scale gives this point a more general meaning. A local diagnostic can become small while a growing logical support still contains enough weakly distinguishable locations to reveal a finite amount of information collectively. Local indistinguishability at any one site is therefore not, by itself, a reliable proxy for recoverability in a growing code. The same mechanism survives beyond toric geometry whenever bounded check incidence allows local errors to probe selected logical parities. This suggests a useful way to think about imperfect code preparation: scalability is controlled not only by how strongly an individual check is measured, but also by how many locations carry the same logical information. The balance between these two quantities is what produces the logarithmic preparation scale in the toric protocols.
A second lesson is that finite-strength preparation separates two properties that coincide in an ideal stabilizer code. Recovering an unknown input does not automatically restore the equivalence of different physical representatives of a logical operator. Once the local leakage has been suppressed, the bulk can still matter for whether two representatives act coherently as the same logical operation on a fixed encoding. The Wilson overlap makes this statement operational: it is not merely a diagnostic of the prepared many-body state, but directly quantifies the error made when one physical representative is substituted for another. The contrast between two and three dimensions is therefore not simply another manifestation of area-law versus perimeter-law behavior. It shows that the bulk phase can become relevant to logical operation only after the more local requirement of recoverability has been satisfied. In the three-dimensional deconfined regime, suitable local enhancement can restore both properties; in two dimensions, recoverability can be restored while representative equivalence remains lost in the setting considered here.
These distinctions open several directions in which the present framework could become more general and more directly connected to fault-tolerant architectures. The first is to move beyond the fixed all-plus branch and ask how outcome-dependent preparation records should be incorporated into recovery and logical-operator equivalence. Doing so would turn the present conditioned analysis into a theory of imperfect measurement preparation with explicit record disorder and would clarify which information must be retained by a decoder. A second direction is to identify the analogue of Wilson restoration in general CSS and qLDPC codes, where two logical representatives are related by products of checks but no Euclidean filling surface is available. There, confinement, expansion, or other code-geometric properties may determine whether coherent representative equivalence can be restored together with recovery. Finally, incorporating efficient decoding and noisy implementations of logical operations would connect these static encoding criteria to circuit-level fault tolerance. That step is important because it would determine whether the separation found here between information recovery and logical-operator equivalence remains visible once preparation, decoding, and logical control are treated as one complete protocol.
IV Appendix
A Finite-strength preparation and error overlaps
The overlap of a codeword with a physical error depends on the checks reversed by that error. This dependence can be evaluated without assuming that the measured checks are independent. The ancilla circuit in Figure 2 (a) gives the normalized Kraus operator M j , + = e β j B j / 2 / 2 cosh β j M_{j,+}=e^{\beta_j B_j/2}/\sqrt{2\cosh\beta_j} M j , + = e β j B j /2 / 2 cosh β j , with tanh ( β j / 2 ) = tan t j \tanh(\beta_j/2)=\tan t_j tanh ( β j /2 ) = tan t j , up to a global phase in the selected branch [9 , 10 ] . Supplementary Note 1 derives the ancilla contraction and its normalization.
Write F β = exp [ 1 2 ∑ j β j B j ] F_\beta=\exp\left[\frac{1}{2}\sum_j\beta_jB_j\right] F β = exp [ 2 1 ∑ j β j B j ] and let V 0 ∣ q ⟩ L = Z ℓ ( q ) ∣ ⊕ n ⟩ V_0|q\rangle_L=Z^{\ell(q)}|\oplus^n\rangle V 0 ∣ q ⟩ L = Z ℓ ( q ) ∣ ⊕ n ⟩ be the zero-strength encoding. Complementary logical X X X operators commute with F β F_\beta F β and have eigenvalues ( − 1 ) q a (-1)^{q_a} ( − 1 ) q a on its logical columns, making distinct columns orthogonal. Their norms are equal because Z ℓ ( q ) Z^{\ell(q)} Z ℓ ( q ) is unitary and commutes with the filter. Consequently,
V 0 † F β 2 V 0 = Z β I K , V β = Z β − 1 / 2 F β V 0 . (36) V_0^\dagger F_\beta^2V_0=Z_\beta I_K,\qquad V_\beta=Z_\beta^{-1/2}F_\beta V_0. \tag*{(36)} V 0 † F β 2 V 0 = Z β I K , V β = Z β − 1/2 F β V 0 . (36)
The all-plus success probability Z β / ∏ j ( 2 cosh β j ) Z_\beta/\prod_j(2\cosh\beta_j) Z β / ∏ j ( 2 cosh β j ) is independent of the logical input, including inputs entangled with a reference. Normalizing the selected branch therefore yields a linear encoding isometry.
Every physical X x X_x X x commutes with the complementary logical X X X operators, so its matrix elements between distinct logical labels vanish. Commuting it through Z ℓ ( q ) Z^{\ell(q)} Z ℓ ( q ) gives
⟨ q β ∣ X x ∣ q β ′ ⟩ = δ q q ′ χ q ( x ) κ β ( x ) , \langle q_\beta|X_x|q'_\beta\rangle=\delta_{q q'}\chi_q(x)\kappa_\beta(x), ⟨ q β ∣ X x ∣ q β ′ ⟩ = δ q q ′ χ q ( x ) κ β ( x ) ,
κ β ( x ) = ⟨ ψ β ∣ X x ∣ ψ β ⟩ . (37) \kappa_\beta(x)=\langle\psi_\beta|X_x|\psi_\beta\rangle. \tag*{(37)} κ β ( x ) = ⟨ ψ β ∣ X x ∣ ψ β ⟩ . (37)
The logical dependence is entirely in the sign χ q ( x ) \chi_q(x) χ q ( x ) ; the magnitude follows from the preparation filter. Moving X x X_x X x between its two factors cancels precisely the checks in σ ( x ) \sigma(x) σ ( x ) . With ∣ + ⟩ |+\rangle ∣ + ⟩ abbreviating ∣ + ⟩ ⊗ n |+\rangle^{\otimes n} ∣ + ⟩ ⊗ n ,
κ β ( x ) = ⟨ + ∣ e ∑ j i j σ ( x ) β j B j ∣ + ⟩ ⟨ + ∣ e ∑ j β j B j ∣ + ⟩ . (38) \kappa_\beta(x)=\frac{\langle+|e^{\sum_j i_j\sigma(x)\beta_jB_j}|+\rangle}{\langle+|e^{\sum_j\beta_jB_j}|+\rangle}. \tag*{(38)} κ β ( x ) = ⟨ + ∣ e ∑ j β j B j ∣ + ⟩ ⟨ + ∣ e ∑ j i j σ ( x ) β j B j ∣ + ⟩ . (38)
To bound this ratio, we use the high-temperature expansion of generalized Ising models [18 , 37 ] , e β j B j = e^{\beta_jB_j}= e β j B j = cosh β j ( 1 + tanh β j B j ) \cosh\beta_j(1+\tanh\beta_j B_j) cosh β j ( 1 + tanh β j B j ) . Only products whose binary check supports sum to zero survive the expectation. After extracting the hyperbolic-cosine factors, the numerator retains a subset of the denominator’s nonnegative terms. This gives the upper bound below, including all check redundancies. For the lower bound, simultaneous diagonalization bounds the denominator of Eq. (38) by its numerator times e ∑ j ∈ σ ( x ) β j e^{\sum_{j\in\sigma(x)}\beta_j} e ∑ j ∈ σ ( x ) β j :
e − ∑ j ∈ σ ( x ) β j ≤ κ β ( x ) ≤ ∏ j ∈ σ ( x ) sech β j . (39) e^{-\sum_{j\in\sigma(x)}\beta_j}\leq\kappa_\beta(x)\leq\prod_{j\in\sigma(x)}\operatorname{sech}\beta_j. \tag*{(39)} e − ∑ j ∈ σ ( x ) β j ≤ κ β ( x ) ≤ j ∈ σ ( x ) ∏ sech β j . (39)
The lower bound leaves a nonzero overlap whenever an error reverses finitely many checks of bounded strength. The upper bound quantifies its suppression when those checks are strengthened. On the square torus, the only check redundancies are the empty and full plaquette sets. Deleting the two checks incident on one edge gives the exact expression in Eq. (17) . On the cubic torus, four checks meet each edge, giving the stated three-dimensional bounds. Supplementary Note 2 gives the exact redundancy sum and the full overlap calculation.
Complementary channel and recovery
Recovery can be characterized by the logical information transferred to the noise environment [22 , 23 , 24 ] . The same overlaps determine its complementary output. For any error distribution P ( x ) P(x) P ( x ) in Eq. (9) , a Stinespring dilation [50, Sec. 2.2.2] acts as ∣ ϕ ⟩ Q ↦ ∑ x P ( x ) X x ∣ ϕ ⟩ Q ∣ x ⟩ E |\phi\rangle_Q\mapsto\sum_x\sqrt{P(x)}X_x|\phi\rangle_Q|x\rangle_E ∣ ϕ ⟩ Q ↦ ∑ x P ( x ) X x ∣ ϕ ⟩ Q ∣ x ⟩ E . Encode one half of ∣ Φ K ⟩ = K − 1 / 2 ∑ q ∣ q ⟩ R ∣ q ⟩ L |\Phi_K\rangle=K^{-1/2}\sum_q|q\rangle_R|q\rangle_L ∣ Φ K ⟩ = K − 1/2 ∑ q ∣ q ⟩ R ∣ q ⟩ L and trace out the physical output. Equation (37) gives
ρ R E = 1 K ∑ q ∣ q ⟩ R ⟨ q ∣ ⊗ ρ E q , \rho_{RE}=\frac{1}{K}\sum_q|q\rangle_R\langle q|\otimes\rho_E^q, ρ R E = K 1 q ∑ ∣ q ⟩ R ⟨ q ∣ ⊗ ρ E q ,
( ρ E q ) x y = P ( x ) P ( y ) χ q ( x ⊕ y ) κ β ( x ⊕ y ) . (40) (\rho_E^q)_{xy}=\sqrt{P(x)P(y)}\chi_q(x\mathbin{\oplus}y)\kappa_\beta(x\mathbin{\oplus}y). \tag*{(40)} ( ρ E q ) x y = P ( x ) P ( y ) χ q ( x ⊕ y ) κ β ( x ⊕ y ) . (40)
The diagonal P ( x ) P(x) P ( x ) is input-independent; logical information resides in the error-label coherences. Although this state is classical on q q q , different conditional environment states suppress coherence between logical sectors. The recovery task still concerns a quantum input. The environment here belongs to the subsequent noise and is distinct from the fixed preparation record.
With rows ( q , x ) (q,x) ( q , x ) and columns ( q ′ , y ) (q',y) ( q ′ , y ) , the normalized AQEC matrices of Ref. [30 ] , defined in Eq. (10) , equal ρ R E \rho_{RE} ρ R E and ( I R / K ⊗ ρ E ) T (I_R/K\otimes\rho_E)^{\mathsf T} ( I R / K ⊗ ρ E ) T , respectively. Full transposition preserves relative entropy and trace norm, hence
D A Q E C = I ( R : E ) , D_{\mathrm{AQEC}}=I(R:E), D AQEC = I ( R : E ) ,
ε = 1 2 ∥ B ∥ 1 = 1 2 ∥ ρ R E − I R K ⊗ ρ E ∥ 1 . (41) \varepsilon=\frac{1}{2}\lVert B\rVert_1=\frac{1}{2}\left\lVert\rho_{RE}-\frac{I_R}{K}\otimes\rho_E\right\rVert_1. \tag*{(41)} ε = 2 1 ∥ B ∥ 1 = 2 1 ρ R E − K I R ⊗ ρ E 1 . (41)
We use D ( ρ ∥ σ ) = Tr [ ρ ( log ρ − log σ ) ] D(\rho\Vert\sigma)=\operatorname{Tr}[\rho(\log\rho-\log\sigma)] D ( ρ ∥ σ ) = Tr [ ρ ( log ρ − log σ )] and natural logarithms. Recovery is possible as these correlations vanish.
Writing V β ( ρ ) = V β ρ V β † \mathcal{V}_\beta(\rho)=V_\beta\rho V_\beta^\dagger V β ( ρ ) = V β ρ V β † , the optimal squared entanglement fidelity is [51]
F e ∗ = max R ⟨ Φ K ∣ ( id R ⊗ R N V β ) ( ∣ Φ K ⟩ ⟨ Φ K ∣ ) ∣ Φ K ⟩ . (42) F_e^*=\max_{\mathcal{R}}\langle\Phi_K|(\operatorname{id}_R\otimes\mathcal{R}\mathcal{N}\mathcal{V}_\beta)(|\Phi_K\rangle\langle\Phi_K|)|\Phi_K\rangle. \tag*{(42)} F e ∗ = R max ⟨ Φ K ∣ ( id R ⊗ RN V β ) ( ∣ Φ K ⟩ ⟨ Φ K ∣ ) ∣ Φ K ⟩ . (42)
This criterion uses the maximally mixed input on the entire selected register. The recovery acts on all physical qubits and returns that register; no efficiency or worst-case channel guarantee is assumed.
Uhlmann’s theorem relates recovery to decoupling [22 , 24 ] [52]:
F e ∗ = max ξ E F ( ρ R E , I R K ⊗ ξ E ) , F_e^*=\max_{\xi_E}F\left(\rho_{RE},\frac{I_R}{K}\otimes\xi_E\right), F e ∗ = ξ E max F ( ρ R E , K I R ⊗ ξ E ) ,
F ( ρ , σ ) = ∥ ρ σ ∥ 1 2 , (43) F(\rho,\sigma)=\lVert\sqrt{\rho}\sqrt{\sigma}\rVert_1^2, \tag*{(43)} F ( ρ , σ ) = ∥ ρ σ ∥ 1 2 , (43)
where ξ E \xi_E ξ E is an environment density matrix. Choosing ξ E = ρ E \xi_E=\rho_E ξ E = ρ E and using the fidelity–trace-distance inequalities [53] gives F e ∗ ≥ ( 1 − ε ) 2 F_e^*\geq(1-\varepsilon)^2 F e ∗ ≥ ( 1 − ε ) 2 . Conversely, recovery with infidelity u u u implies ε ≤ 2 u \varepsilon\leq2\sqrt{u} ε ≤ 2 u . Together with quantum Pinsker’s inequality [50, Theorem 5.38], these give
ε 2 4 ≤ 1 − F e ∗ ≤ 2 ε , D A Q E C ≥ 2 ε 2 . (44) \frac{\varepsilon^2}{4}\leq1-F_e^*\leq2\varepsilon,\qquad D_{\mathrm{AQEC}}\geq2\varepsilon^2. \tag*{(44)} 4 ε 2 ≤ 1 − F e ∗ ≤ 2 ε , D AQEC ≥ 2 ε 2 . (44)
Supplementary Note 2 gives the full purification argument and the trace-distance estimates.
The conversion to total AQEC entropy also depends on the size of the logical register. Equation (40) makes D A Q E C D_{\mathrm{AQEC}} D AQEC a Holevo information, bounded by log K \log K log K [38]. Since the reference is classical in the same basis for ρ R E \rho_{RE} ρ R E and its product comparator, the classical–quantum conditional-entropy continuity bound gives [39 Lemma 2]
D A Q E C ≤ ε log K + ( 1 + ε ) h 2 ( ε 1 + ε ) , (45) D_{\mathrm{AQEC}}\leq\varepsilon\log K+(1+\varepsilon)h_2\left(\frac{\varepsilon}{1+\varepsilon}\right), \tag*{(45)} D AQEC ≤ ε log K + ( 1 + ε ) h 2 ( 1 + ε ε ) , (45)
h 2 ( t ) = − t log t − ( 1 − t ) log ( 1 − t ) . h_2(t)=-t\log t-(1-t)\log(1-t). h 2 ( t ) = − t log t − ( 1 − t ) log ( 1 − t ) .
At fixed k k k , vanishing D A Q E C D_{\mathrm{AQEC}} D AQEC , vanishing ε \varepsilon ε , and F e ∗ → 1 F_e^*\to1 F e ∗ → 1 are equivalent. If k ≥ 1 k\geq1 k ≥ 1 grows, ε → 0 \varepsilon\to0 ε → 0 still guarantees recovery of the entire register as D A Q E C / k → 0 D_{\mathrm{AQEC}}/k\to0 D AQEC / k → 0 , whereas k ε → 0 k\varepsilon\to0 k ε → 0 is sufficient for the total AQEC entropy to vanish. These information-theoretic conversions hold for arbitrary P ( x ) P(x) P ( x ) ; independent flips and local stochasticity enter only the respective leakage and entanglement bounds.
Local and collective leakage
The local obstruction in Theorem 1 follows directly from the error-overlap matrix. Choose a physical site e e e with g ( e ) ≠ 0 g(e)\neq0 g ( e ) = 0 . Matrix entries whose error labels differ by e e e carry the logical sign χ q ( e ) \chi_q(e) χ q ( e ) and the common overlap κ β ( e ) \kappa_\beta(e) κ β ( e ) . A signed permutation pairs ( q , x ) (q,x) ( q , x ) with ( q , x ⊕ e ) (q,x\mathbin{\oplus}e) ( q , x ⊕ e ) and cancels this sign. Its operator norm is one, so trace-norm duality gives [50, Sec. 1.1.3]
∥ B ∥ 1 ≥ κ β ( e ) ∑ x P ( x ) P ( x ⊕ e ) = 2 p ( 1 − p ) κ β ( e ) . (46) \begin{aligned}
\lVert B\rVert_1\geq\kappa_\beta(e)\sum_x\sqrt{P(x)P(x\mathbin{\oplus}e)} \\
&=2\sqrt{p(1-p)}\,\kappa_\beta(e). \tag*{(46)}
\end{aligned} ∥ B ∥ 1 ≥ κ β ( e ) x ∑ P ( x ) P ( x ⊕ e ) = 2 p ( 1 − p ) κ β ( e ) . (46)
for independent flips. The sum includes every exterior error configuration; their probabilities sum to one. Thus the bound does not require the other qubits to remain error free. The same quantum Pinsker bound then yields
D A Q E C ≥ 1 2 ∥ B ∥ 1 2 ≥ 2 p ( 1 − p ) κ β ( e ) 2 . (47) D_{\mathrm{AQEC}} \ge\frac{1}{2}\lVert B\rVert_1^2 \ge2p(1-p)\kappa_\beta(e)^2. \tag*{(47)} D AQEC ≥ 2 1 ∥ B ∥ 1 2 ≥ 2 p ( 1 − p ) κ β ( e ) 2 . (47)
If each site meets at most Δ Z \Delta_Z Δ Z measured checks and their strengths are bounded by β 0 \beta_0 β 0 , Eq. (39) gives D A Q E C ≥ 2 p ( 1 − p ) e − 2 Δ Z β 0 D_{\mathrm{AQEC}} \ge2p(1-p)e^{-2\Delta_Z\beta_0} D AQEC ≥ 2 p ( 1 − p ) e − 2 Δ Z β 0 . At fixed 0 < p < 1 0<p<1 0 < p < 1 , a single site therefore leaves a positive lower bound independent of code size. The corresponding norm bound and Eq. (44) also give the recovery obstruction in Corollary 1. Its entropy-density statement follows by choosing sites that probe independent logical parities and applying the information chain rule, as proved in Supplementary Note 2.
Subsystem variance [26 ] identifies the local information responsible for this obstruction. The expectation of X e X_e X e in a logical basis state is χ q ( e ) κ β ( e ) \chi_q(e)\kappa_\beta(e) χ q ( e ) κ β ( e ) , whereas its expectation in the maximally mixed code state is zero. Since ∥ X e ∥ ∞ = 1 \lVert X_e\rVert_\infty=1 ∥ X e ∥ ∞ = 1 , the definition in Eq. (12) gives ϵ S V ( { e } ) ≥ κ β ( e ) \epsilon_{\mathrm{SV}}(\{e\})\ge\kappa_\beta(e) ϵ SV ({ e }) ≥ κ β ( e ) . In the toric codes, exact star constraints make the Y e Y_e Y e and Z e Z_e Z e expectations vanish for every encoded input. The entire single-edge marginal is therefore
ρ e = 1 2 [ I + κ β ( e ) ∑ q ( ρ L ) q q χ q ( e ) X e ] . (48) \rho_e=\frac{1}{2}\left[I+\kappa_\beta(e)\sum_q(\rho_L)_{qq}\chi_q(e)X_e\right]. \tag*{(48)} ρ e = 2 1 [ I + κ β ( e ) q ∑ ( ρ L ) q q χ q ( e ) X e ] . (48)
The code-average marginal is I / 2 I/2 I /2 . The coefficient of X e X_e X e ranges between − κ β ( e ) -\kappa_\beta(e) − κ β ( e ) and κ β ( e ) \kappa_\beta(e) κ β ( e ) , with both endpoints attained by logical basis states, so the full trace norm gives ϵ S V ( { e } ) = κ β ( e ) \epsilon_{\mathrm{SV}}(\{e\})=\kappa_\beta(e) ϵ SV ({ e }) = κ β ( e ) . This maximization includes arbitrary logical superpositions. The same residual overlap thus quantifies local distinguishability before noise and bounds information lost to the noise environment. For a general CSS encoding the variance need not equal the overlap, so the noise bound remains stated in terms of κ β ( e ) \kappa_\beta(e) κ β ( e ) .
The replica construction for imperfectly prepared toric codes [30 Sec. V and Appendix C] extends to the selected CSS register and nonuniform strengths. The resulting traces show how local contributions enter along the selected representatives. For integer m ≥ 2 m\ge2 m ≥ 2 , multiplying the AQEC matrix around a replica cycle, with x ( m + 1 ) = x ( 1 ) x^{(m+1)}=x^{(1)} x ( m + 1 ) = x ( 1 ) , gives
K m − 1 Tr ( Λ + B ) m = ∑ x ( 1 ) , … , x ( m ) ∏ α = 1 m P ( x ( α ) ) κ β ( x ( α ) ⊕ x ( α + 1 ) ) . (49) K^{m-1}\operatorname{Tr}(\Lambda+B)^m
=
\sum_{x^{(1)},\ldots,x^{(m)}}\prod_{\alpha=1}^{m}P(x^{(\alpha)})\kappa_\beta\left(x^{(\alpha)}\oplus x^{(\alpha+1)}\right). \tag*{(49)} K m − 1 Tr ( Λ + B ) m = x ( 1 ) , … , x ( m ) ∑ α = 1 ∏ m P ( x ( α ) ) κ β ( x ( α ) ⊕ x ( α + 1 ) ) . (49)
Each error probability appears in two neighboring square roots, and the logical signs cancel around the closed cycle. The expression for K m − 1 Tr Λ m K^{m-1}\operatorname{Tr}\Lambda^m K m − 1 Tr Λ m has the same weights, with the additional constraint g ( x ( α ) ⊕ x ( α + 1 ) ) = 0 g(x^{(\alpha)}\oplus x^{(\alpha+1)})=0 g ( x ( α ) ⊕ x ( α + 1 ) ) = 0 at every step. Inserting the filter ratio in Eq. (38) gives the classical replica partition functions derived in Supplementary Note 2. The constraint excludes precisely the error differences that distinguish selected logical inputs.
Consider the first local-error configurations in this sum: one replica contains a single error e e e , and all other replicas contain no error. The two adjacent differences are both e e e , giving a factor κ β ( e ) 2 \kappa_\beta(e)^2 κ β ( e ) 2 . This contribution occurs in the full trace but is excluded from the comparator whenever g ( e ) ≠ 0 g(e)\ne0 g ( e ) = 0 . Such a configuration can be placed in any replica and at any site on the union of selected representatives. A straight toric representative supplies d d d such positions. The bounds on the local overlap involve only the two incident checks in 2D or four in 3D; at fixed finite strength, the overlap remains bounded below independently of the representative length. This counting is the origin of the accumulated leakage in the replica expansion.
At finite size, the AQEC entropy is recovered from these traces through
D A Q E C = ∂ m log Tr ( Λ + B ) m Tr Λ m ∣ m = 1 . (50) D_{\mathrm{AQEC}}=\partial_m\log\left.\frac{\operatorname{Tr}(\Lambda+B)^m}{\operatorname{Tr}\Lambda^m}\right|_{m=1}. \tag*{(50)} D AQEC = ∂ m log Tr Λ m Tr ( Λ + B ) m m = 1 . (50)
The matrix powers near one are defined spectrally. The derivative equals the relative entropy because Tr ( B log Λ ) = 0 \operatorname{Tr}(B\log\Lambda)=0 Tr ( B log Λ ) = 0 , a consequence of Tr B = 0 \operatorname{Tr}B=0 Tr B = 0 . Supplementary Note 2 evaluates the local replica coefficient while retaining the terms needed in this limit. A truncation at fixed integer m m m cannot be continued to m = 1 m=1 m = 1 by discarding those terms, nor does a finite-size dilute expansion by itself give a bound uniform in d d d .
The fixed-noise statement in Theorem 1 follows from the collective estimate in the same Supplementary Note. For the M M M sites specified there, the summed local signal is controlled by ∑ i κ β ( e i ) \sum_i\kappa_\beta(e_i) ∑ i κ β ( e i ) , while its second moment involves ∑ i ≠ j κ β ( e i ⊕ e j ) \sum_{i\ne j}\kappa_\beta(e_i\oplus e_j) ∑ i = j κ β ( e i ⊕ e j ) . Disjoint syndrome supports bound the latter through Eq. (39) , without assuming independent environment states. Data processing and Pinsker’s inequality [50, Ch. 5], together with a Cauchy–Schwarz estimate, give Eq. (16) , whose denominator accounts for repeated signals about the same logical parity. With M = d M=d M = d , this proves Eq. (19) at fixed nonzero p p p . The same argument uses only the checks incident on those sites, so it also applies to their actual common enhanced strength β R \beta_R β R when the bulk remains fixed.
Connected-component recovery bound
The finite-strength bound in Theorem 2 separates leakage from silent crossings from the error incurred by replacing the prepared state with its projection onto the enhanced-check sector. Write ∣ ψ s o f t ⟩ = ∣ ψ β ⟩ |\psi_{\mathrm{soft}}\rangle=|\psi_\beta\rangle ∣ ψ soft ⟩ = ∣ ψ β ⟩ for the finite-strength preparation. To bound the first contribution, consider
Π r = ∏ j ∈ P r I + B j 2 , ∣ ψ r ⟩ = Π r ∣ ψ s o f t ⟩ ⟨ ψ s o f t ∣ Π r ∣ ψ s o f t ⟩ . (51) \Pi_r=\prod_{j\in\mathcal{P}_r}\frac{I+B_j}{2},\qquad
|\psi_r\rangle=\frac{\Pi_r|\psi_{\mathrm{soft}}\rangle}{\sqrt{\langle\psi_{\mathrm{soft}}|\Pi_r|\psi_{\mathrm{soft}}\rangle}}. \tag*{(51)} Π r = j ∈ P r ∏ 2 I + B j , ∣ ψ r ⟩ = ⟨ ψ soft ∣ Π r ∣ ψ soft ⟩ Π r ∣ ψ soft ⟩ . (51)
This comparison state has the same exterior strengths and selected representatives as the finite-strength preparation. The projection is used only in the proof. Writing κ r ( x ) = \kappa_r(x)= κ r ( x ) = ⟨ ψ r ∣ X x ∣ ψ r ⟩ \langle\psi_r|X_x|\psi_r\rangle ⟨ ψ r ∣ X x ∣ ψ r ⟩ , the identity Π r X x Π r = 0 \Pi_rX_x\Pi_r=0 Π r X x Π r = 0 makes this overlap vanish whenever σ ( x ) ∩ P r ≠ ∅ \sigma(x)\cap P_r\ne\varnothing σ ( x ) ∩ P r = ∅ .
Decompose the error difference x = u ⊕ v x=u\oplus v x = u ⊕ v into connected components in G Z G_Z G Z . Distinct components touch disjoint sets of measured checks, so their syndromes cannot cancel. A nonzero contribution that distinguishes logical labels q , q ′ q,q' q , q ′ must therefore contain silent components with different characters χ q , χ q ′ \chi_q,\chi_{q'} χ q , χ q ′ . Denote their class by C r ( q , q ′ ) C_r(q,q') C r ( q , q ′ ) . Character multiplication gives an exact expansion of the conditional environment-state difference into families of these components:
D q , q ′ : = ρ E q − ρ E q ′ = ∑ ∅ ≠ F ⊆ C r ( q , q ′ ) F c o m p a t i b l e D F q , q ′ (52) D^{q,q'}:=\rho_E^q-\rho_E^{q'}=\sum_{\substack{\varnothing\ne\mathcal{F}\subseteq C_r(q,q')\\ \mathcal{F}\ \mathrm{compatible}}}D_{\mathcal{F}}^{q,q'} \tag*{(52)} D q , q ′ := ρ E q − ρ E q ′ = ∅ = F ⊆ C r ( q , q ′ ) F compatible ∑ D F q , q ′ (52)
Compatible families are mutually disconnected supports that can occur as complete components of one error difference. Each distinguishing component has character ratio − 1 = 1 − 2 -1=1-2 − 1 = 1 − 2 , so expanding the product gives the signed coefficient ( − 2 ) ∣ F ∣ (-2)^{|\mathcal{F}|} ( − 2 ) ∣ F ∣ in D F q , q ′ D_{\mathcal{F}}^{q,q'} D F q , q ′ . Supplementary Note 3 gives the full matrix elements.
For a fixed family with union S S S , the two error labels must be complementary on S S S and equal on its external vertex boundary. After fixing these assignments, the remaining matrix is, up to signs and the family coefficient, a cross-Gram matrix of the vectors p P ( u ) X u ∣ ψ r ⟩ \sqrt{p_P(u)}X_u|\psi_r\rangle p P ( u ) X u ∣ ψ r ⟩ . The Schatten inequality ∥ A † C ∥ 1 ≤ ∥ A ∥ 2 ∥ C ∥ 2 \lVert A^\dagger C\rVert_1\le\lVert A\rVert_2\lVert C\rVert_2 ∥ A † C ∥ 1 ≤ ∥ A ∥ 2 ∥ C ∥ 2 [50, Sec. 1.1.3] bounds its trace norm by the square root of the corresponding row and column probabilities. These probabilities already sum over exterior errors. Summing the common boundary assignments by Cauchy–Schwarz, then applying local stochasticity to the complementary occupations of S S S , gives a factor η ∣ S ∣ / 2 \eta^{|S|/2} η ∣ S ∣/2 . The 2 ∣ S ∣ 2^{|S|} 2 ∣ S ∣ assignments and 2 ∣ F ∣ ≤ 2 ∣ S ∣ 2^{|\mathcal{F}|}\le2^{|S|} 2 ∣ F ∣ ≤ 2 ∣ S ∣ family coefficient consequently yield
∥ D F q , q ′ ∥ 1 ≤ ( 4 η ) ∣ S ∣ = ∏ Γ ∈ F ( 4 η ) ∣ Γ ∣ . (53) \lVert D_{\mathcal{F}}^{q,q'}\rVert_1\le(4\sqrt{\eta})^{|S|}=\prod_{\Gamma\in\mathcal{F}}(4\sqrt{\eta})^{|\Gamma|}. \tag*{(53)} ∥ D F q , q ′ ∥ 1 ≤ ( 4 η ) ∣ S ∣ = Γ ∈ F ∏ ( 4 η ) ∣Γ∣ . (53)
This estimate retains arbitrary noise correlations allowed by Eq. (21) ; it is a bound on the full operator, rather than on individual matrix elements.
Compatibility can now be dropped when summing these nonnegative norm bounds. Every distinguishing component intersects the selected support union, with U U U possible anchors, and has size at least M ( r ) M(r) M ( r ) . The exponential bound on the number of connected supports in a graph of bounded degree [40] therefore gives
∑ Γ ∈ C r ( q , q ′ ) ( 4 η ) ∣ Γ ∣ ≤ C U λ M ( r ) 1 − λ , λ = 4 μ η < 1. (54) \sum_{\Gamma\in C_r(q,q')}(4\sqrt{\eta})^{|\Gamma|}\le\frac{CU\lambda^{M(r)}}{1-\lambda},\qquad\lambda=4\mu\sqrt{\eta}<1. \tag*{(54)} Γ ∈ C r ( q , q ′ ) ∑ ( 4 η ) ∣Γ∣ ≤ 1 − λ C U λ M ( r ) , λ = 4 μ η < 1. (54)
The family sum is bounded by the exponential of this single-component sum minus one. Averaging over logical labels then gives
∥ B h a r d ∥ 1 ≤ K − 2 ∑ q , q ′ ∥ D q , q ′ ∥ 1 ≤ exp [ C U λ M ( r ) 1 − λ ] − 1. (55) \begin{aligned}
\lVert B_{\mathrm{hard}}\rVert_1\le K^{-2}\sum_{q,q'}\lVert D^{q,q'}\rVert_1 \\
&\le\exp\left[\frac{CU\lambda^{M(r)}}{1-\lambda}\right]-1. \tag*{(55)}
\end{aligned} ∥ B hard ∥ 1 ≤ K − 2 q , q ′ ∑ ∥ D q , q ′ ∥ 1 ≤ exp [ 1 − λ C U λ M ( r ) ] − 1. (55)
The average introduces no factor growing with the logical dimension K K K . This proves the projective part of Theorem 2.
Finite-strength enhancement and resources
For the unsigned CSS checks and nonnegative preparation strengths considered here, the probability of violating a measured check is at most e − 2 β j e^{-2\beta_j} e − 2 β j . To obtain this inequality, remove that check’s coupling: the expansion in products of checks has nonnegative weights, hence a nonnegative expectation of B j B_j B j . Restoring its coupling reweights the two check eigenvalues by e ± β j e^{\pm\beta_j} e ± β j . A union bound therefore gives
Pr ( B j = − 1 ) ≤ e − 2 β j , \Pr(B_j=-1)\le e^{-2\beta_j}, Pr ( B j = − 1 ) ≤ e − 2 β j ,
q r : = 1 − ⟨ Π r ⟩ s o f t ≤ ∣ P r ∣ e − 2 β R . (56) q_r:=1-\langle\Pi_r\rangle_{\mathrm{soft}}\le|\mathcal{P}_r|e^{-2\beta_R}. \tag*{(56)} q r := 1 − ⟨ Π r ⟩ soft ≤ ∣ P r ∣ e − 2 β R . (56)
No independence of check violations is required. Within the protected sector, each enhanced filter acts as a scalar, so the normalized projection is precisely the hard state with the exterior strengths unchanged.
The preparation states have overlap 1 − q r \sqrt{1-q_r} 1 − q r . Their purifications with the same maximally entangled logical input have this overlap as well, because the selected representatives are unchanged and commute with Π r \Pi_r Π r . The pure-state trace-distance formula [53] gives distance q r \sqrt{q_r} q r between these purifications. A common noise dilation and partial trace cannot increase this distance [50, Sec. 3], so the reference-environment states and their environment marginals obey the same bound. Thus
ε s o f t ≤ ε h a r d + 2 q r , (57) \varepsilon_{\mathrm{soft}}\le\varepsilon_{\mathrm{hard}}+2\sqrt{q_r}, \tag*{(57)} ε soft ≤ ε hard + 2 q r , (57)
and ∥ B ∥ 1 = 2 ε \lVert B\rVert_1=2\varepsilon ∥ B ∥ 1 = 2 ε gives
∥ B s o f t ∥ 1 ≤ ∥ B h a r d ∥ 1 + 4 ∣ P r ∣ e − β R . (58) \lVert B_{\mathrm{soft}}\rVert_1\le\lVert B_{\mathrm{hard}}\rVert_1+4\sqrt{|\mathcal{P}_r|}\,e^{-\beta_R}. \tag*{(58)} ∥ B soft ∥ 1 ≤ ∥ B hard ∥ 1 + 4 ∣ P r ∣ e − β R . (58)
Together with Eq. (55) , this proves the finite-strength bound in Theorem 2. The correction vanishes when ∣ P r ∣ e − β R → 0 \sqrt{|\mathcal{P}_r|}e^{-\beta_R}\to0 ∣ P r ∣ e − β R → 0 , including the case of an empty protected set.
A complementary proof keeps finite measurement strength throughout the component expansion. For a fixed family, extracting the overlap factors associated with its syndrome leaves a positive Gram kernel, as shown in Supplementary Note 3. The cross-Gram argument then gives
∥ D β R , F q , q ′ ∥ 1 ≤ ∏ Γ ∈ F ( 4 η ) ∣ Γ ∣ ( sech β R ) ∣ σ ( Γ ) ∩ P r ∣ . (59) \lVert D_{\beta_R,\mathcal{F}}^{q,q'}\rVert_1\le\prod_{\Gamma\in\mathcal{F}}(4\sqrt{\eta})^{|\Gamma|}(\operatorname{sech}\beta_R)^{|\sigma(\Gamma)\cap\mathcal{P}_r|}. \tag*{(59)} ∥ D β R , F q , q ′ ∥ 1 ≤ Γ ∈ F ∏ ( 4 η ) ∣Γ∣ ( sech β R ) ∣ σ ( Γ ) ∩ P r ∣ . (59)
Here D β R , F q , q ′ D_{\beta_R,\mathcal{F}}^{q,q'} D β R , F q , q ′ is the same family contribution evaluated in the finite-strength state. Every component smaller than M ( r ) M(r) M ( r ) has a protected syndrome; the remaining components are bounded by size. Summing gives
∥ B s o f t ∥ 1 ≤ exp [ C U 1 − λ ( λ sech β R + λ M ( r ) ) ] − 1. (60) \lVert B_{\mathrm{soft}}\rVert_1\le\exp\left[\frac{CU}{1-\lambda}\left(\lambda\operatorname{sech}\beta_R+\lambda^{M(r)}\right)\right]-1. \tag*{(60)} ∥ B soft ∥ 1 ≤ exp [ 1 − λ C U ( λ sech β R + λ M ( r ) ) ] − 1. (60)
The two terms describe suppression by measurement strength and by crossing size. Its simple sufficient strength condition is U sech β R → 0 U\operatorname{sech}\beta_R \to0 U sech β R → 0 . The coefficients in Table I follow from the sharper soft–hard comparison, Eq. (58) .
Protection geometry. For toric codes, take an integer r ≥ 1 r \ge1 r ≥ 1 , set r 0 = ⌊ r / 2 ⌋ r_0 = \lfloor r/2\rfloor r 0 = ⌊ r /2 ⌋ , and assume 2 r 0 + 1 ≤ L 2r_0+1 \le L 2 r 0 + 1 ≤ L . Include parallel translates of each selected straight string with transverse coordinates between − r 0 -r_0 − r 0 and r 0 r_0 r 0 , and protect the plaquette ribbons joining neighboring translates. Silence on a ribbon A y A_y A y implies
∣ Γ ∩ ℓ y ∣ + ∣ Γ ∩ ℓ y + 1 ∣ = ∣ σ ( Γ ) ∩ A y ∣ = 0 ( m o d 2 ) . (61) |\Gamma\cap\ell_y| + |\Gamma\cap\ell_{y+1}| = |\sigma(\Gamma) \cap A_y| = 0 \pmod{2}. \tag*{(61)} ∣Γ ∩ ℓ y ∣ + ∣Γ ∩ ℓ y + 1 ∣ = ∣ σ ( Γ ) ∩ A y ∣ = 0 ( mod 2 ) . (61)
Odd crossing of the selected string propagates to every translate. The 2 r 0 + 1 2r_0+1 2 r 0 + 1 strings in a two-dimensional strip are disjoint, giving ∣ Γ ∣ ≥ r |\Gamma| \ge r ∣Γ∣ ≥ r . A full three-dimensional tube contains ( 2 r 0 + 1 ) 2 (2r_0+1)^2 ( 2 r 0 + 1 ) 2 such strings, giving ∣ Γ ∣ ≥ r 2 |\Gamma| \ge r^2 ∣Γ∣ ≥ r 2 . This proves Eq. (24) .
For a family with linear confinement in the sense of Eq. (25) [41, 42], protect every check touching the radius-( r − 1 ) (r-1) ( r − 1 ) neighborhood of the selected union. Consider an error with no protected syndrome and nonzero selected logical parity. If its weight is at least T T T , it already satisfies Eq. (26) . For a lighter error, first minimize its full support modulo X X X stabilizers, preserving syndrome and logical parity, and then choose a distinguishing component Γ \Gamma Γ . Every subset of a minimum-weight support is itself reduced, since a stabilizer reducing that subset would also reduce the full support. Let Γ t \Gamma_t Γ t be its intersection with the radius-t t t neighborhood of the selected union. These supports have weight below T T T , so Eq. (25) gives ∣ σ ( Γ t ) ∣ ≥ κ ∣ Γ t ∣ |\sigma(\Gamma_t)| \ge\kappa|\Gamma_t| ∣ σ ( Γ t ) ∣ ≥ κ ∣ Γ t ∣ . For 0 ≤ t < r 0 \le t < r 0 ≤ t < r , each such check is protected. Silence of the full component forces it to meet the next layer, where each added qubit can cancel at most Δ Z \Delta_Z Δ Z checks. Hence
∣ Γ t + 1 ∣ ≥ ( 1 + κ / Δ Z ) ∣ Γ t ∣ . (62) |\Gamma_{t+1}| \ge(1+\kappa/\Delta_Z)|\Gamma_t|. \tag*{(62)} ∣ Γ t + 1 ∣ ≥ ( 1 + κ / Δ Z ) ∣ Γ t ∣. (62)
Nonzero logical parity gives ∣ Γ 0 ∣ ≥ 1 |\Gamma_0| \ge1 ∣ Γ 0 ∣ ≥ 1 , and iteration proves Eq. (26) , with growth capped at T T T . Supplementary Note 3 gives both geometric proofs.
Resources at fixed logical-register size. For fixed straight toric registers, U = O ( d ) U=O(d) U = O ( d ) . Choose a fixed envelope 0 < λ 0 < 1 0<\lambda_0<1 0 < λ 0 < 1 with λ ≤ λ 0 \lambda\le\lambda_0 λ ≤ λ 0 . Taking r ≥ ( 1 + δ ) log d / ∣ log λ 0 ∣ r\ge(1+\delta)\log d/|\log\lambda_0| r ≥ ( 1 + δ ) log d /∣ log λ 0 ∣ in two dimensions, or r 2 r^2 r 2 at least that large in three, gives λ M ( r ) ≤ d − 1 − δ \lambda^{M(r)}\le d^{-1-\delta} λ M ( r ) ≤ d − 1 − δ and hard leakage O ( d − δ ) O(d^{-\delta}) O ( d − δ ) . Strips contain O ( d r ) O(dr) O ( d r ) checks and full tubes O ( d r 2 ) O(dr^2) O ( d r 2 ) , so both constructions use O ( d log d ) O(d\log d) O ( d log d ) enhanced checks. With β R = ( 1 2 + δ ) log d \beta_R=(\frac12+\delta)\log d β R = ( 2 1 + δ ) log d , Eq. (58) adds O ( d − δ log d ) O(d^{-\delta}\sqrt{\log d}) O ( d − δ log d ) . Dividing by the total measured-check count, Θ ( d 2 ) \Theta(d^2) Θ ( d 2 ) or Θ ( d 3 ) \Theta(d^3) Θ ( d 3 ) , gives the first two enhanced fractions in Table I. These choices work for every fixed finite bulk strength; widths are rounded up.
For confinement, write γ = 1 + κ / Δ Z > 1 \gamma=1+\kappa/\Delta_Z>1 γ = 1 + κ / Δ Z > 1 . Assume U → ∞ U\to\infty U → ∞ , choose c > 1 / ∣ log γ ∣ c>1/|\log\gamma| c > 1/∣ log γ ∣ , and require T > c log U T>c\log U T > c log U . This cutoff follows from good linear confinement as defined in Results, since U ≤ n U\le n U ≤ n and a uniform polynomial cutoff eventually exceeds c log U c\log U c log U [41, 42]. The radius
r = ⌈ log ( c log U ) log γ ⌉ (63) r=\left\lceil\frac{\log(c\log U)}{\log\gamma}\right\rceil \tag*{(63)} r = ⌈ log γ log ( c log U ) ⌉ (63)
then gives M ( r ) ≥ c log U M(r)\ge c\log U M ( r ) ≥ c log U and U λ M ( r ) → 0 U\lambda^{M(r)}\to0 U λ M ( r ) → 0 . If D G ≥ 2 D_G\ge2 D G ≥ 2 bounds the degree of G Z G_Z G Z , the neighborhood contains at most C U ( D G + 1 ) r CU(D_G+1)^r C U ( D G + 1 ) r qubits. Bounded check incidence therefore gives
∣ P r ∣ ≤ min { N c h k , C U ( log U ) ν } , ν = log ( D G + 1 ) log γ . (64) |\mathcal{P}_r|\le\min\{N_{\mathrm{chk}},CU(\log U)^\nu\},
\qquad
\nu=\frac{\log(D_G+1)}{\log\gamma}. \tag*{(64)} ∣ P r ∣ ≤ min { N chk , C U ( log U ) ν } , ν = log γ log ( D G + 1 ) . (64)
At β R = ( 1 2 + δ ) log U \beta_R=(\frac12+\delta)\log U β R = ( 2 1 + δ ) log U , the finite-strength correction is O ( U − δ ( log U ) ν / 2 ) → 0 O(U^{-\delta}(\log U)^{\nu/2})\to0 O ( U − δ ( log U ) ν /2 ) → 0 . This proves the confinement row of Table I and completes Corollary 2. The enhanced fraction vanishes under the additional condition U ( log U ) ν = o ( N c h k ) U(\log U)^\nu=o(N_{\mathrm{chk}}) U ( log U ) ν = o ( N chk ) ; a small radius alone does not ensure a small fraction.
For the toric protocols, logarithmic strength is also necessary under independent flips at fixed 0 < p < 1 0<p<1 0 < p < 1 : all checks incident on the selected straight line lie in its protected neighborhood, so the collective bound of Theorem 1 applies to their common actual strength β R \beta_R β R . This necessity statement uses the independent-noise model, whereas the sufficient construction allows uniformly weak local-stochastic noise. It concerns the actual strength used in the protocol; a lower bound β j ≥ β R \beta_j\ge\beta_R β j ≥ β R alone cannot establish necessity for its parameter β R \beta_R β R .
Recovery and entropy for a growing logical register. The cost of protecting many logical qubits is constrained by the physical sites left outside the enhanced checks. The binary matrix with rows ℓ 1 , … , ℓ k \ell_1,\ldots,\ell_k ℓ 1 , … , ℓ k has rank k k k . Deleting columns in F h i F_{h_i} F h i , touched by the enhanced checks, leaves rank at least ( k − ∣ F h i ∣ ) + (k-|F_{h_i}|)_+ ( k − ∣ F h i ∣ ) + . Choose sites e a e_a e a with independent remaining columns g ( e a ) g(e_a) g ( e a ) . For a uniformly mixed logical register, the bits g ( e a ) ⋅ q g(e_a)\cdot q g ( e a ) ⋅ q are independent and uniform. Let E e a E_{e_a} E e a be the environment qubit recording the absence or presence of a physical flip at e a e_a e a . Under independent flips at fixed 0 < p < 1 0<p<1 0 < p < 1 , the information chain rule and data processing [50, Ch. 5], together with the single-site environment bound in Supplementary Note 2, give
D A Q E C ≥ ∑ a I ( g ( e a ) ⋅ q : E e a ) D_{\mathrm{AQEC}}\ge\sum_a I(g(e_a)\cdot q:E_{e_a}) D AQEC ≥ a ∑ I ( g ( e a ) ⋅ q : E e a )
≥ 2 p ( 1 − p ) ∑ a κ β ( e a ) 2 . (65) \ge2p(1-p)\sum_a\kappa_\beta(e_a)^2. \tag*{(65)} ≥ 2 p ( 1 − p ) a ∑ κ β ( e a ) 2 . (65)
If all exterior strengths are at most β 0 \beta_0 β 0 and each qubit touches at most Δ Z \Delta_Z Δ Z measured checks, then κ β ( e a ) ≥ e − Δ Z β 0 \kappa_\beta(e_a)\ge e^{-\Delta_Z\beta_0} κ β ( e a ) ≥ e − Δ Z β 0 . Substitution proves the footprint bound in Corollary 3. Vanishing entropy density consequently requires ∣ F h i ∣ ≥ [ 1 − o ( 1 ) ] k |F_{h_i}|\ge[1-o(1)]k ∣ F h i ∣ ≥ [ 1 − o ( 1 )] k ; bounded check weight makes an extensive number of enhanced checks necessary when k = Θ ( n ) k=\Theta(n) k = Θ ( n ) .
The upper bound in Theorem 2 also applies to a growing register, using its actual support union U U U and uniform local constants. If its right-hand side tends to zero, Eqs. (44) and (45) give full-register entanglement recovery and D A Q E C / k → 0 D_{\mathrm{AQEC}}/k\to0 D AQEC / k → 0 . To make the total entropy vanish, the continuity bound gives the stronger sufficient conditions
k U λ M ( r ) ⟶ 0 , k ∣ P r ∣ e − β R ⟶ 0. (66) kU\lambda^{M(r)}\longrightarrow0,\qquad k\sqrt{|\mathcal{P}_r|}\,e^{-\beta_R}\longrightarrow0. \tag*{(66)} k U λ M ( r ) ⟶ 0 , k ∣ P r ∣ e − β R ⟶ 0. (66)
Without the factors k k k , these conditions already suffice for recovery and vanishing entropy density. The extra factors ensure vanishing total entropy and are not claimed necessary for recovery. Supplementary Note 3 gives the register-dependent estimates.
Wilson restoration on a finite torus
Theorem 3 compares two representatives enhanced in the same preparation. Their Wilson overlap depends on the checks between them as well as those near their supports. Because W ( C ) W(C) W ( C ) is diagonal, this dependence can be evaluated in the Born measure of the prepared state. Write Z e = ( − 1 ) s e Z_e=(-1)^{s_e} Z e = ( − 1 ) s e with s e ∈ F 2 s_e\in\mathbb{F}_2 s e ∈ F 2 ; the plaquette parity ( δ s ) p (\delta s)_p ( δ s ) p is the sum of these bits around p p p . Squaring the preparation amplitudes gives
π β ( s ) = exp [ ∑ p β p ( − 1 ) ( δ s ) p ] ∑ s ′ exp [ ∑ p β p ( − 1 ) ( δ s ′ ) p ] . (67) \pi_\beta(s)=\frac{\exp\left[\sum_p\beta_p(-1)^{(\delta s)_p}\right]}{\sum_{s'}\exp\left[\sum_p\beta_p(-1)^{(\delta s')_p}\right]}. \tag*{(67)} π β ( s ) = ∑ s ′ exp [ ∑ p β p ( − 1 ) ( δ s ′ ) p ] exp [ ∑ p β p ( − 1 ) ( δ s ) p ] . (67)
Each unoriented plaquette is counted once, and all chains and intersections below have F 2 \mathbb{F}_2 F 2 coefficients.
In two dimensions, the high-temperature expansion [37 Sec. 7] , e β p B p = cosh β p ( 1 + u p B p ) e^{\beta_pB_p}=\cosh\beta_p(1+u_pB_p) e β p B p = cosh β p ( 1 + u p B p ) , and the sum over edge bits retain plaquette sets with boundary zero in the denominator and boundary C C C in the numerator. The only fillings are the two complementary regions S S S and S c S^c S c , giving
w = ∏ p ∈ S u p + ∏ p ∈ S c u p 1 + ∏ p u p , u p = tanh β p . (68) w=\frac{\prod_{p\in S}u_p+\prod_{p\in S^c}u_p}{1+\prod_pu_p},\qquad u_p=\tanh\beta_p. \tag*{(68)} w = 1 + ∏ p u p ∏ p ∈ S u p + ∏ p ∈ S c u p , u p = tanh β p . (68)
The denominator is at least one, and every enhanced-plaquette factor is at most one. Hence w w w is bounded by ( tanh β ) ∣ S ∖ P r ∣ + ( tanh β ) ∣ S c ∖ P r ∣ (\tanh\beta)^{|S\setminus\mathcal{P}_r|}+(\tanh\beta)^{|S^c\setminus\mathcal{P}_r|} ( tanh β ) ∣ S ∖ P r ∣ + ( tanh β ) ∣ S c ∖ P r ∣ . When each filling retains at least c L 2 cL^2 c L 2 unenhanced plaquettes, this proves the two-dimensional part of Theorem 3, including the limit of projective enhancement.
In three dimensions, we adapt the random-cluster and dual-linking arguments used by Duncan and Schweinhart to prove the perimeter law [21 Secs. 5, 7, and 9] . Our setting has nonuniform couplings and two winding representatives on a finite torus. The plaquette coupling of Ref. [21 Sec. 5.1] gives a direct representation in this setting: conditional on the open plaquette set A A A , the gauge configurations are uniform subject to B p = + 1 B_p=+1 B p = + 1 for p ∈ A p\in A p ∈ A . Their Wilson character averages to one when C C C bounds a chain of open plaquettes and to zero otherwise. Thus w w w is the filling probability. Supplementary Note 4 derives this identity for the present couplings and bounds the obstructions that remain after local enhancement.
The parameter conventions matter in applying this representation. The identity e β p B p = e − β p e 2 β p 1 { B p = + 1 } e^{\beta_pB_p}=e^{-\beta_p}e^{2\beta_p\mathbf{1}_{\{B_p=+1\}}} e β p B p = e − β p e 2 β p 1 { B p = + 1 } shows that the indicator-action coupling of Refs. [21 , 33 ] is 2 β p 2\beta_p 2 β p . The primal plaquette parameter is θ p = 1 − e − 2 β p \theta_p=1-e^{-2\beta_p} θ p = 1 − e − 2 β p . A dual edge is open when its primal plaquette is absent, and its parameter is
p e ∗ = 2 ( 1 − θ p ) θ p + 2 ( 1 − θ p ) = 2 e 2 β p + 1 . p_e^*=\frac{2(1-\theta_p)}{\theta_p+2(1-\theta_p)}=\frac{2}{e^{2\beta_p}+1}. p e ∗ = θ p + 2 ( 1 − θ p ) 2 ( 1 − θ p ) = e 2 β p + 1 2 .
These are random-cluster weight parameters with cluster weight q = 2 q=2 q = 2 , distinct from the physical noise probability. Stronger checks decrease p e ∗ p_e^* p e ∗ .
On the dual torus, write D D D for the set of open edges. The Fortuin–Kasteleyn (FK) random-cluster measure is a probability distribution on these configurations [47, 48], with weights
ϕ L , p ∗ , 2 ( D ) ∝ 2 c ( D ) ∏ e ∈ D p e ∗ ∏ e ∉ D ( 1 − p e ∗ ) . (69) \phi_{L,p^*,2}(D)\propto2^{c(D)}\prod_{e\in D}p_e^*\prod_{e\notin D}(1-p_e^*). \tag*{(69)} ϕ L , p ∗ , 2 ( D ) ∝ 2 c ( D ) e ∈ D ∏ p e ∗ e ∈ / D ∏ ( 1 − p e ∗ ) . (69)
Here c ( D ) c(D) c ( D ) counts connected components, including isolated vertices, on the graph with periodic boundaries; the weights are normalized over all D D D . Each component contributes a factor two, corresponding to its two spin assignments in the Ising representation. This cluster factor correlates the edge occupations, so p e ∗ p_e^* p e ∗ is a weight parameter rather than the marginal probability that e e e is open.
Plaquette-to-edge duality on a torus also retains winding information [33 Sec. 4.3] . Keeping the ordinary gauge weights, the rank calculation in Supplementary Note 4 gives the induced dual law for our inhomogeneous couplings:
ν ∗ ( D ) = 2 − h ( D ) ϕ L , p ∗ , 2 ( D ) E ϕ L , p ∗ , 2 [ 2 − h ( D ) ] . (70) \nu^*(D)=\frac{2^{-h(D)}\phi_{L,p^*,2}(D)}{\mathbb{E}_{\phi_{L,p^*,2}}[2^{-h(D)}]}. \tag*{(70)} ν ∗ ( D ) = E ϕ L , p ∗ , 2 [ 2 − h ( D ) ] 2 − h ( D ) ϕ L , p ∗ , 2 ( D ) . (70)
where h ( D ) h(D) h ( D ) counts the independent winding classes carried by D D D , with coefficients in F 2 \mathbb{F}_2 F 2 . The factor 2 − h ( D ) 2^{-h(D)} 2 − h ( D ) is the topological correction to the standard FK distribution. Since the torus has three independent periodic directions, 0 ≤ h ( D ) ≤ 3 0\le h(D)\le3 0 ≤ h ( D ) ≤ 3 ; both this factor and its FK average lie in [ 1 / 8 , 1 ] [1/8,1] [ 1/8 , 1 ] . Thus any event probability under ν ∗ \nu^* ν ∗ is at most eight times its probability under periodic FK. This comparison allows the standard random-cluster estimates to be used for the gauge model.
At the fixed bulk point in Eq. (31) , the dual FK model is subcritical. Theorem 1.2 of Ref. [49] gives exponential decay of connections from the center to the wired boundary of a finite graph-distance ball. This is also the probability estimate used in the perimeter-law proof [21 Theorem 36] . For unit bond couplings, Ref. [49] uses an inverse-temperature parameter β F K \beta_{\mathrm{FK}} β FK with bond parameter 1 − e − β F K 1-e^{-\beta_{\mathrm{FK}}} 1 − e − β FK , so our bulk parameter corresponds to
β F K = − log [ 1 − p ∗ ( β ) ] = − log tanh β . (71) \beta_{\mathrm{FK}}=-\log[1-p^*(\beta)]=-\log\tanh\beta. \tag*{(71)} β FK = − log [ 1 − p ∗ ( β )] = − log tanh β . (71)
To use this theorem on the inhomogeneous torus, condition periodic FK on the edges outside a ball that lifts to Z 3 \mathbb{Z}^3 Z 3 . Exterior connections identify some boundary vertices. The fully wired condition identifies all of them and therefore gives the largest connection probabilities. Together with monotonicity in the edge parameters [48, Chs. 3–4], the inequality p e ∗ ≤ p ∗ ( β ) p_e^* \le p^*(\beta) p e ∗ ≤ p ∗ ( β ) bounds the conditional connection probability by that of the homogeneous wired ball. Averaging over the exterior and including the factor-eight comparison yields
ν ∗ { v ↔ ∂ B R ( v ) } ≤ C β e − R / ξ β , 1 ≤ R ≤ L / 4 , (72) \nu^*\{v \leftrightarrow\partial B_R(v)\} \le C_\beta e^{-R/\xi_\beta}, \qquad1 \le R \le L/4, \tag*{(72)} ν ∗ { v ↔ ∂ B R ( v )} ≤ C β e − R / ξ β , 1 ≤ R ≤ L /4 , (72)
where B R ( v ) B_R(v) B R ( v ) is the graph-distance ball about a dual vertex v v v . A union over vertices also bounds the probability of a cluster with periodic diameter at least L / 4 L/4 L /4 by C β L 3 e − c β L C_\beta L^3 e^{-c_\beta L} C β L 3 e − c β L . The constants are uniform in the enhanced parameters and system size at fixed bulk strength; no uniformity at the critical point is assumed.
Protect the plaquettes whose dual edges lie within cell-incidence distance r r r of C = ℓ 1 ⊕ ℓ 2 C=\ell_1\oplus\ell_2 C = ℓ 1 ⊕ ℓ 2 . For 1 ≤ r ≤ c α L 1\le r\le c_\alpha L 1 ≤ r ≤ c α L , with c α < min { α / 8 , 1 / 16 } c_\alpha<\min\{\alpha/8,1/16\} c α < min { α /8 , 1/16 } , these full tubes contain ∣ P r ∣ = O ( L ( r + 1 ) 2 ) |\mathcal{P}_r|=O(L(r+1)^2) ∣ P r ∣ = O ( L ( r + 1 ) 2 ) checks. Projective enhancement closes every dual edge in the tubes. If an open filling fails to exist, linear duality supplies an Eulerian dual chain with odd intersection with a filling S S S . Exclude the macroscopic-cluster event just bounded. Each remaining cluster lifts to a contractible box, and at least one Eulerian part γ \gamma γ has odd intersection with S S S . It bounds a dual surface A γ A_\gamma A γ inside that box, so
1 = γ ⋅ S = ( ∂ A γ ) ⋅ S = A γ ⋅ C . (73) 1=\gamma\cdot S=(\partial A_\gamma)\cdot S=A_\gamma\cdot C. \tag*{(73)} 1 = γ ⋅ S = ( ∂ A γ ) ⋅ S = A γ ⋅ C . (73)
The surface must meet C C C , and the cluster therefore approaches C C C within a distance proportional to its diameter. A cluster avoiding the protected tube must consequently have diameter at least c g r c_g r c g r . At diameter scale s s s , it has an anchor among O ( L ( s + 1 ) 2 ) O(L(s+1)^2) O ( L ( s + 1 ) 2 ) edges near C C C . Applying Eq. (72) at these anchors and summing over dyadic scales bounds the remaining filling-failure probability by C β L ( r + 1 ) 2 e − c g r / ξ β C_\beta L(r+1)^2e^{-c_g r/\xi_\beta} C β L ( r + 1 ) 2 e − c g r / ξ β , after adjusting geometric constants.
For finite-strength enhancement, a protected edge has conditional open probability at most p e ∗ ≤ 2 e − 2 β R p_e^*\le2e^{-2\beta_R} p e ∗ ≤ 2 e − 2 β R under ordinary FK. If E r E_r E r is the event that any protected dual edge is open, the factor-eight comparison and a union bound give
ν ∗ ( E r ) ≤ 16 ∣ P r ∣ e − 2 β R . (74) \nu^*(E_r)\le16|\mathcal{P}_r|e^{-2\beta_R}. \tag*{(74)} ν ∗ ( E r ) ≤ 16∣ P r ∣ e − 2 β R . (74)
Conditioning all protected edges to be closed makes their local weight factors constant and reproduces the projectively enhanced dual law exactly, including its topological weight. The Wilson deficit is thus bounded by the projective deficit plus ν ∗ ( E r ) \nu^*(E_r) ν ∗ ( E r ) :
1 − w ≤ C β L ( r + 1 ) 2 e − c g r / ξ β + C β L 3 e − c β L + C β L r 2 e − 2 β R . (75) 1-w\le C_\beta L(r+1)^2e^{-c_g r/\xi_\beta}+C_\beta L^3e^{-c_\beta L}+C_\beta Lr^2e^{-2\beta_R}. \tag*{(75)} 1 − w ≤ C β L ( r + 1 ) 2 e − c g r / ξ β + C β L 3 e − c β L + C β L r 2 e − 2 β R . (75)
The three terms respectively control local obstructing clusters, macroscopic clusters, and residual open edges in the protected tubes. For the choices in Eq. (32) , with r r r rounded upward, they are O ( L 1 − b log 2 L ) O(L^{1-b}\log^2 L) O ( L 1 − b log 2 L ) , O ( L 3 e − c β L ) O(L^3e^{-c_\beta L}) O ( L 3 e − c β L ) , and O ( L − 2 b log 2 L ) O(L^{-2b}\log^2 L) O ( L − 2 b log 2 L ) . All vanish, while r ≤ c α L r\le c_\alpha L r ≤ c α L and β R ≥ β \beta_R\ge\beta β R ≥ β hold for sufficiently large L L L .
The same preparation restores AQEC under the separate low-noise assumption λ ≤ λ 0 < 1 \lambda\le\lambda_0<1 λ ≤ λ 0 < 1 . The cell-incidence and qubit check-graph distances differ only by fixed factors, so the full tubes contain the AQEC ribbons at width proportional to r r r . Their minimum silent distinguishing size is then proportional to r 2 r^2 r 2 , making the projective activity O ( L e − c ( log L ) 2 ) O(L e^{-c(\log L)^2}) O ( L e − c ( l o g L ) 2 ) for some c > 0 c>0 c > 0 . The O ( L log 2 L ) O(L\log^2 L) O ( L log 2 L ) enhanced checks give a soft-hard AQEC correction O ( L − δ log L ) O(L^{-\delta}\log L) O ( L − δ log L ) . Both contributions vanish for either chosen encoding. Supplementary Notes 3 and 4 give the geometric comparison and the complete recovery and Wilson estimates.
Controlled strings and common recovery
The operation in Corollary 4 acts on the fixed target encoding V = V 1 V=V_1 V = V 1 ; only its physical representative is changed. The two encodings share their zero codeword, while their one codewords have overlap w w w . Opposite eigenvalues of the complementary physical logical X X X symmetry make the cross-sector overlaps vanish. Thus, in the ordered logical basis ( ∣ 0 ⟩ L , ∣ 1 ⟩ L ) (|0\rangle_L,|1\rangle_L) ( ∣0 ⟩ L , ∣1 ⟩ L ) ,
V 1 † V 2 = diag ( 1 , w ) , ∥ V 1 − V 2 ∥ ∞ = 2 ( 1 − w ) . (76) V_1^\dagger V_2=\operatorname{diag}(1,w), \qquad\lVert V_1-V_2\rVert_\infty=\sqrt{2(1-w)}. \tag*{(76)} V 1 † V 2 = diag ( 1 , w ) , ∥ V 1 − V 2 ∥ ∞ = 2 ( 1 − w ) . (76)
The norm follows from ( V 1 − V 2 ) † ( V 1 − V 2 ) = diag ( 0 , 2 ( 1 − w ) ) (V_1-V_2)^\dagger(V_1-V_2)=\operatorname{diag}(0,2(1-w)) ( V 1 − V 2 ) † ( V 1 − V 2 ) = diag ( 0 , 2 ( 1 − w )) . In particular, w → 1 w\to1 w → 1 makes the encodings agree on arbitrary logical inputs, including inputs entangled with a reference.
For the fixed encoding, Z ℓ V = V X L Z^{\ell}V=VX_L Z ℓ V = V X L . Commutativity of the two physical strings and the same logical symmetry give
V † Z ℓ 2 V = w X L , V † Z ℓ 1 Z ℓ 2 V = w I L . (77) V^\dagger Z^{\ell_2}V=wX_L,\qquad V^\dagger Z^{\ell_1}Z^{\ell_2}V=wI_L. \tag*{(77)} V † Z ℓ 2 V = w X L , V † Z ℓ 1 Z ℓ 2 V = w I L . (77)
Consequently, [ ( Z ℓ 2 − Z ℓ 1 ) V ] † [ ( Z ℓ 2 − Z ℓ 1 ) V ] = 2 ( 1 − w ) I L [(Z^{\ell_2}-Z^{\ell_1})V]^\dagger[(Z^{\ell_2}-Z^{\ell_1})V]=2(1-w)I_L [( Z ℓ 2 − Z ℓ 1 ) V ] † [( Z ℓ 2 − Z ℓ 1 ) V ] = 2 ( 1 − w ) I L . Since replacing G 1 G_1 G 1 by G 2 G_2 G 2 changes only the control-one block, its largest singular value gives Eq. (35) . The projection onto the fixed code space is V V † VV^\dagger V V † , and
V † Z ℓ 2 V V † Z ℓ 2 V = w 2 I L . V^\dagger Z^{\ell_2}VV^\dagger Z^{\ell_2}V=w^2I_L. V † Z ℓ 2 V V † Z ℓ 2 V = w 2 I L .
For control ∣ 1 ⟩ C |1\rangle_C ∣1 ⟩ C , the target therefore remains in the code with probability w 2 w^2 w 2 for every logical input. This proves the leakage statement 1 − w 2 1-w^2 1 − w 2 . Tensoring the operator identities with a reference identity leaves these conclusions unchanged.
The overlap also constrains recovery when the encoder choice is undisclosed. With no subsequent noise, the optimal worst-choice fidelity of a single recovery R \mathcal{R} R from the full physical system Q Q Q to the logical qubit is
max R min i = 1 , 2 F e ( R V i ) = 1 + ( 1 + w ) / 2 2 , (78) \max_{\mathcal{R}}\min_{i=1,2}F_e(\mathcal{R}V_i)=\frac{1+\sqrt{(1+w)/2}}{2}, \tag*{(78)} R max i = 1 , 2 min F e ( R V i ) = 2 1 + ( 1 + w ) /2 , (78)
where V i ( ρ L ) = V i ρ L V i † \mathcal{V}_i(\rho_L)=V_i\rho_LV_i^\dagger V i ( ρ L ) = V i ρ L V i † and F e F_e F e is the squared entanglement fidelity for a maximally mixed logical qubit [51]. The equal mixture of encoders retains coherence ( 1 + w ) / 2 \sqrt{(1+w)/2} ( 1 + w ) /2 between their shared zero codeword and the normalized symmetric combination of their one codewords. Cauchy–Schwarz bounds the recovered coherence, and a recovery using the symmetric and antisymmetric combinations attains the displayed fidelity for both encoders (Supplementary Note 4). Subsequent common noise cannot improve this optimum, since its composition with a recovery is already among the channels optimized above. Conversely, w → 1 w \to1 w → 1 transfers any successful fixed recovery under the same noise between the two encodings by channel contractivity [50, Sec. 3.3]. The common-recovery result relies on the shared zero codeword and concerns an unknown encoder choice; the controlled-string result keeps the encoding fixed throughout.
Acknowledge
This work was supported by Quantum Science and Technology-National Science and Technology Major Project (Grant No. 2021ZD0302400), Shanghai Municipal Science and Technology (Grant No. 25LZ2600200), and National Natural Science Foundation of China (Grants No. 92365111).
Detailed Derivations and Proofs of the Theorems
FINITE-STRENGTH PREPARATION AND BARE WILSON LOOPS
The finite-strength measurement instrument fixes both the conditional encoding and the classical measure used to evaluate its Wilson observables. We derive this connection for general CSS codes before specializing to the square and cubic toric codes. The two-dimensional Wilson expression is exact at finite volume. In three dimensions, high- and low-temperature expansions establish the area and perimeter mechanisms, and a sharp-transition theorem extends the corresponding laws to the full strict phases. Wilson restoration by local enhancement is proved in Supplementary Note 4.
Ancilla circuit and measurement strength
Let B j B_j B j be a Hermitian Pauli check, B j 2 = I B_j^2=I B j 2 = I , and let a a a denote a fresh ancilla prepared in ∣ + ⟩ a = ( ∣ 0 ⟩ a + ∣ 1 ⟩ a ) / 2 |+\rangle_a=(|0\rangle_a+|1\rangle_a)/\sqrt{2} ∣ + ⟩ a = ( ∣0 ⟩ a + ∣1 ⟩ a ) / 2 . The circuit in the main text is the standard Pauli finite-strength measurement construction [1, 2]:
U j ( t j ) = e − i t j Z a ⊗ B j = cos t j I − i sin t j Z a ⊗ B j , 0 ≤ t j ≤ π / 4. (79) U_j(t_j)=e^{-it_j Z_a\otimes B_j}=\cos t_j I-i\sin t_j Z_a\otimes B_j,\qquad0\leq t_j\leq\pi/4. \tag*{(79)} U j ( t j ) = e − i t j Z a ⊗ B j = cos t j I − i sin t j Z a ⊗ B j , 0 ≤ t j ≤ π /4. (79)
The angle t j t_j t j is the integrated coupling in units of ℏ \hbar ℏ ; for an interaction Hamiltonian H j ( τ ) = g j ( τ ) Z a ⊗ B j H_j(\tau)=g_j(\tau)Z_a\otimes B_j H j ( τ ) = g j ( τ ) Z a ⊗ B j , t j = ℏ − 1 ∫ g j ( τ ) d τ t_j=\hbar^{-1}\int g_j(\tau)\,\mathrm{d}\tau t j = ℏ − 1 ∫ g j ( τ ) d τ . It is not a duration by itself. The displayed multiqubit gate specifies the measurement instrument; no assumption of native multiqubit hardware is needed for the recovery results.
For an arbitrary data state ∣ ϕ ⟩ |\phi\rangle ∣ ϕ ⟩ , the joint state is
U j ( t j ) ∣ + ⟩ a ∣ ϕ ⟩ = ∣ 0 ⟩ a e − i t j B j ∣ ϕ ⟩ + ∣ 1 ⟩ a e i t j B j ∣ ϕ ⟩ 2 . (80) U_j(t_j)|+\rangle_a|\phi\rangle=\frac{|0\rangle_a e^{-it_jB_j}|\phi\rangle+|1\rangle_a e^{it_jB_j}|\phi\rangle}{\sqrt{2}}. \tag*{(80)} U j ( t j ) ∣ + ⟩ a ∣ ϕ ⟩ = 2 ∣0 ⟩ a e − i t j B j ∣ ϕ ⟩ + ∣1 ⟩ a e i t j B j ∣ ϕ ⟩ . (80)
Measure Y a Y_a Y a and label its eigenstate ∣ y s ⟩ = ( ∣ 0 ⟩ + i s ∣ 1 ⟩ ) / 2 |y_s\rangle=(|0\rangle+is|1\rangle)/\sqrt{2} ∣ y s ⟩ = ( ∣0 ⟩ + i s ∣1 ⟩) / 2 , s = ± 1 s=\pm1 s = ± 1 . The unnormalized data branch follows by taking the ancilla inner product:
⟨ y s ∣ U j ( t j ) ∣ + ⟩ a = e − i t j B j − i s e i t j B j 2 = 1 − i s 2 ( cos t j I + s sin t j B j ) = e − i s π / 4 cos t j I + s sin t j B j 2 . (81) \begin{aligned}
\langle y_s|U_j(t_j)|+\rangle_a
&=\frac{e^{-it_jB_j}-is e^{it_jB_j}}{2}\\
&=\frac{1-is}{2}\left(\cos t_j I+s\sin t_j B_j\right)\\
&=e^{-is\pi/4}\frac{\cos t_j I+s\sin t_j B_j}{\sqrt{2}}. \tag*{(81)}
\end{aligned} ⟨ y s ∣ U j ( t j ) ∣ + ⟩ a = 2 e − i t j B j − i s e i t j B j = 2 1 − i s ( cos t j I + s sin t j B j ) = e − i s π /4 2 cos t j I + s sin t j B j . (81)
The phase e − i s π / 4 e^{-is\pi/4} e − i s π /4 is global within a fixed readout branch and has no physical effect. Thus s s s in the positive filter below is precisely the Y a Y_a Y a eigenvalue, with the same sign convention as the basic Pauli circuit in Ref. [2].
Define β j ≥ 0 \beta_j\geq0 β j ≥ 0 by tanh ( β j / 2 ) = tan t j \tanh(\beta_j/2)=\tan t_j tanh ( β j /2 ) = tan t j . Using e s β j B j / 2 = cosh ( β j / 2 ) I + s sinh ( β j / 2 ) B j e^{s\beta_jB_j/2}=\cosh(\beta_j/2)I+s\sinh(\beta_j/2)B_j e s β j B j /2 = cosh ( β j /2 ) I + s sinh ( β j /2 ) B j and cosh β j = cosh 2 ( β j / 2 ) + sinh 2 ( β j / 2 ) \cosh\beta_j=\cosh^2(\beta_j/2)+\sinh^2(\beta_j/2) cosh β j = cosh 2 ( β j /2 ) + sinh 2 ( β j /2 ) yields
M j , s = cos t j I + s sin t j B j 2 = e s β j B j / 2 2 cosh β j . (82) M_{j,s}=\frac{\cos t_j I+s\sin t_j B_j}{\sqrt{2}}=\frac{e^{s\beta_jB_j/2}}{\sqrt{2\cosh\beta_j}}. \tag*{(82)} M j , s = 2 cos t j I + s sin t j B j = 2 cosh β j e s β j B j /2 . (82)
The useful equivalent parameter conversions are
β j = 2 artanh ( tan t j ) = log tan ( π / 4 + t j ) , tanh β j = sin ( 2 t j ) , sech β j = cos ( 2 t j ) . (83) \beta_j=2\operatorname{artanh}(\tan t_j)=\log\tan(\pi/4+t_j),\qquad\tanh\beta_j=\sin(2t_j),\qquad\operatorname{sech}\beta_j=\cos(2t_j). \tag*{(83)} β j = 2 artanh ( tan t j ) = log tan ( π /4 + t j ) , tanh β j = sin ( 2 t j ) , sech β j = cos ( 2 t j ) . (83)
These identities hold for 0 ≤ t j < π / 4 0\leq t_j<\pi/4 0 ≤ t j < π /4 and extend by a limit at π / 4 \pi/4 π /4 . In particular, the amplitude exponent is β j B j / 2 \beta_jB_j/2 β j B j /2 , whereas the squared-amplitude (classical gauge) exponent is β j B j \beta_jB_j β j B j . The expansion weights used below are denoted by u j = tanh β j u_j=\tanh\beta_j u j = tanh β j to keep them distinct from the interaction angles t j t_j t j .
The POVM effects and probabilities for a normalized data density matrix ρ \rho ρ are
M j , s † M j , s = I + s tanh β j B j 2 , ∑ s = ± 1 M j , s † M j , s = I , (84) M_{j,s}^{\dagger}M_{j,s}=\frac{I+s\tanh\beta_j B_j}{2},\qquad\sum_{s=\pm1}M_{j,s}^{\dagger}M_{j,s}=I, \tag*{(84)} M j , s † M j , s = 2 I + s tanh β j B j , s = ± 1 ∑ M j , s † M j , s = I , (84)
p j ( s ∣ ρ ) = 1 + s tanh β j Tr ( B j ρ ) 2 , ρ j , s = M j , s ρ M j , s † p j ( s ∣ ρ ) . (85) p_j(s\mid\rho)=\frac{1+s\tanh\beta_j\operatorname{Tr}(B_j\rho)}{2},\qquad\rho_{j,s}=\frac{M_{j,s}\rho M_{j,s}^{\dagger}}{p_j(s\mid\rho)}. \tag*{(85)} p j ( s ∣ ρ ) = 2 1 + s tanh β j Tr ( B j ρ ) , ρ j , s = p j ( s ∣ ρ ) M j , s ρ M j , s † . (85)
The conditional state is defined when its probability is nonzero. At t j = 0 t_j=0 t j = 0 both outcomes have probability 1 / 2 1/2 1/2 and leave the state unchanged. At t j ↑ π / 4 t_j\uparrow\pi/4 t j ↑ π /4 , M j , s → ( I + s B j ) / 2 M_{j,s}\to(I+sB_j)/2 M j , s → ( I + s B j ) /2 . For a B j B_j B j eigenstate with eigenvalue b = ± 1 b=\pm1 b = ± 1 , an incorrect readout has probability ( 1 − sin 2 t j ) / 2 (1-\sin2t_j)/2 ( 1 − sin 2 t j ) /2 . For a superposition, however, the finite measurement also changes amplitudes coherently. It is not equivalent to a projective measurement followed only by a classical readout error [5 ] [3, 4].
Postselection, normalization, and logical inputs
For the CSS preparation, the measured checks are unsigned Z Z Z operators, B j = Z h j B_j = Z^{h_j} B j = Z h j , where the binary vectors h j h_j h j are the rows of H Z ∈ F 2 m Z × n H_Z \in\mathbb{F}_2^{m_Z \times n} H Z ∈ F 2 m Z × n . They commute, and their rows need not be independent. For a complete record s = ( s 1 , … , s m Z ) s = (s_1,\ldots,s_{m_Z}) s = ( s 1 , … , s m Z ) , discarding only the irrelevant global phase gives
K s = ∏ j M j , s j = F β , s ∏ j 2 cosh β j , F β , s = exp ( 1 2 ∑ j s j β j B j ) . (86) K_s = \prod_j M_{j,s_j} = \frac{F_{\beta,s}}{\sqrt{\prod_j 2\cosh\beta_j}}, \qquad F_{\beta,s} = \exp\left(\frac{1}{2}\sum_j s_j\beta_j B_j\right). \tag*{(86)} K s = j ∏ M j , s j = ∏ j 2 cosh β j F β , s , F β , s = exp ( 2 1 j ∑ s j β j B j ) . (86)
Hence Pr ( s ∣ ρ ) = Tr ( K s ρ K s † ) \Pr(s \mid\rho) = \operatorname{Tr}(K_s\rho K_s^\dagger) Pr ( s ∣ ρ ) = Tr ( K s ρ K s † ) for an arbitrary input. For ∣ + ⟩ ⊗ n |+\rangle^{\otimes n} ∣ + ⟩ ⊗ n , define
Z β , s = ⟨ + ∣ ⊗ n F β , s 2 ∣ + ⟩ ⊗ n , ∣ ψ β , s ⟩ = Z β , s − 1 / 2 F β , s ∣ + ⟩ ⊗ n . (87) Z_{\beta,s} = \langle+|^{\otimes n}F_{\beta,s}^2|+\rangle^{\otimes n}, \qquad|\psi_{\beta,s}\rangle= Z_{\beta,s}^{-1/2}F_{\beta,s}|+\rangle^{\otimes n}. \tag*{(87)} Z β , s = ⟨ + ∣ ⊗ n F β , s 2 ∣ + ⟩ ⊗ n , ∣ ψ β , s ⟩ = Z β , s − 1/2 F β , s ∣ + ⟩ ⊗ n . (87)
A fixed record therefore gives a pure state. Postselection in this paper means retaining precisely s j = + 1 s_j = +1 s j = + 1 for all j j j . Writing F β F_\beta F β and Z β Z_\beta Z β for this record,
K + ∣ + ⟩ ⊗ n = F β ∣ + ⟩ ⊗ n ∏ j 2 cosh β j , ∣ ψ β ⟩ = F β ∣ + ⟩ ⊗ n Z β , p + = Z β ∏ j 2 cosh β j . (88) K_+|+\rangle^{\otimes n} = \frac{F_\beta|+\rangle^{\otimes n}}{\sqrt{\prod_j 2\cosh\beta_j}}, \qquad|\psi_\beta\rangle= \frac{F_\beta|+\rangle^{\otimes n}}{\sqrt{Z_\beta}}, \qquad p_+ = \frac{Z_\beta}{\prod_j 2\cosh\beta_j}. \tag*{(88)} K + ∣ + ⟩ ⊗ n = ∏ j 2 cosh β j F β ∣ + ⟩ ⊗ n , ∣ ψ β ⟩ = Z β F β ∣ + ⟩ ⊗ n , p + = ∏ j 2 cosh β j Z β . (88)
No independence of the measured checks has been assumed. To make the postselection cost explicit even for redundant checks, expand e β j B j = cosh β j ( I + tanh β j B j ) e^{\beta_jB_j} = \cosh\beta_j(I+\tanh\beta_jB_j) e β j B j = cosh β j ( I + tanh β j B j ) . A check product has a nonzero expectation in ∣ + ⟩ ⊗ n |+\rangle^{\otimes n} ∣ + ⟩ ⊗ n exactly when its binary support vanishes. Consequently,
p + = 2 − m Z ∑ a ∈ ker H Z T ∏ j : a j = 1 tanh β j . (89) p_+ = 2^{-m_Z}\sum_{a\in\ker H_Z^{\mathsf T}}\prod_{j:a_j=1}\tanh\beta_j. \tag*{(89)} p + = 2 − m Z a ∈ k e r H Z T ∑ j : a j = 1 ∏ tanh β j . (89)
Independent rows give p + = 2 − m Z p_+ = 2^{-m_Z} p + = 2 − m Z ; in the fully projective limit, p + → 2 − rank H Z p_+ \to2^{-\operatorname{rank}H_Z} p + → 2 − rank H Z . Thus the selected branch need not have size-independent probability. Our enhanced-check counts quantify control within this conditioned protocol, not the expected number of preparation attempts. At finite strengths all records have nonzero probability; in the projective limit, inconsistent records imposed by check redundancies have zero probability.
Discarding the readout produces a different channel [6 ] [1, 2]. For one check,
∑ s M j , s ρ M j , s † = cos 2 t j ρ + sin 2 t j B j ρ B j . (90) \sum_s M_{j,s}\rho M_{j,s}^\dagger= \cos^2 t_j\,\rho+ \sin^2 t_j\,B_j\rho B_j. \tag*{(90)} s ∑ M j , s ρ M j , s † = cos 2 t j ρ + sin 2 t j B j ρ B j . (90)
For all checks the unconditioned output is ∑ s K s ρ K s † \sum_s K_s\rho K_s^\dagger ∑ s K s ρ K s † ; retaining a classical record instead gives ∑ s ∣ s ⟩ ⟨ s ∣ ⊗ K s ρ K s † \sum_s |s\rangle\langle s|\otimes K_s\rho K_s^\dagger ∑ s ∣ s ⟩ ⟨ s ∣ ⊗ K s ρ K s † . Neither object is the all-plus pure branch. Other fixed records put signs s j β j s_j\beta_j s j β j in the gauge action; the Born-weighted ensemble is the disordered problem studied in Ref. [2]. The present all-plus branch has a nonnegative, disorder-free gauge coupling.
The same conditioned operation encodes an unknown logical input. Choose k k k ideal logical Z Z Z representatives ℓ a \ell_a ℓ a that are independent modulo row H Z H_Z H Z , put K = 2 k K=2^k K = 2 k , and write ℓ ( q ) = ⨁ a = 1 k q a ℓ a \ell(q)=\bigoplus_{a=1}^k q_a\ell_a ℓ ( q ) = ⨁ a = 1 k q a ℓ a for q ∈ F 2 k q\in\mathbb{F}_2^k q ∈ F 2 k . Here H L \mathcal{H}_L H L is the K K K -dimensional logical input space and H Q = ( C 2 ) ⊗ n \mathcal{H}_Q=(\mathbb{C}^2)^{\otimes n} H Q = ( C 2 ) ⊗ n is the full physical space. The map V 0 : H L → H Q V_0:\mathcal{H}_L\to\mathcal{H}_Q V 0 : H L → H Q defined by V 0 ∣ q ⟩ L = Z ℓ ( q ) ∣ + ⟩ ⊗ n V_0|q\rangle_L=Z^{\ell(q)}|+\rangle^{\otimes n} V 0 ∣ q ⟩ L = Z ℓ ( q ) ∣ + ⟩ ⊗ n is an isometry; V 0 V_0 V 0 denotes the zero-strength encoding. Since F β 2 F_\beta^2 F β 2 expands entirely into Z Z Z stabilizer products,
L ⟨ q ∣ V 0 † F β 2 V 0 ∣ q ′ ⟩ L = ⟨ + ∣ ⊗ n Z ℓ ( q ⊕ q ′ ) F β 2 ∣ + ⟩ ⊗ n = Z β δ q q ′ . (91) {}_L\langle q|V_0^\dagger F_\beta^2V_0|q'\rangle_L = \langle+|^{\otimes n}Z^{\ell(q\oplus q')}F_\beta^2|+\rangle^{\otimes n}=Z_\beta\delta_{qq'}. \tag*{(91)} L ⟨ q ∣ V 0 † F β 2 V 0 ∣ q ′ ⟩ L = ⟨ + ∣ ⊗ n Z ℓ ( q ⊕ q ′ ) F β 2 ∣ + ⟩ ⊗ n = Z β δ q q ′ . (91)
For q ≠ q ′ q\ne q' q = q ′ , the selected logical class cannot be canceled by a check product; for q = q ′ q=q' q = q ′ , commutation makes the norm independent of q q q . Therefore K + V 0 = p + V β K_+V_0=\sqrt{p_+}\,V_\beta K + V 0 = p + V β with V β = Z β − 1 / 2 F β V 0 V_\beta=Z_\beta^{-1/2}F_\beta V_0 V β = Z β − 1/2 F β V 0 and V β † V β = I K V_\beta^\dagger V_\beta=I_K V β † V β = I K . Its physical codewords are ∣ q β ⟩ = V β ∣ q ⟩ L = Z ℓ ( q ) ∣ ψ β ⟩ |q_\beta\rangle=V_\beta|q\rangle_L=Z^{\ell(q)}|\psi_\beta\rangle ∣ q β ⟩ = V β ∣ q ⟩ L = Z ℓ ( q ) ∣ ψ β ⟩ . Tensoring this identity with I R I_R I R shows that success is input independent even for a logical input entangled with a reference R R R . The normalized successful operation is a linear encoding isometry. The same argument applies to every nonzero-probability fixed record; only the all-plus branch is used below. Supplementary Lemma 2.1 relates this encoding to the X X X -error overlap kernel.
Why the B j B_j B j sector is the relevant recovery problem
CSS codes separate the two Pauli syndrome sectors [7 , 8 ] . Our preparation treats them asymmetrically. Every X X X check A μ A_\mu A μ stabilizes ∣ + ⟩ ⊗ n |+\rangle^{\otimes n} ∣ + ⟩ ⊗ n and commutes with both F β F_\beta F β and Z ℓ ( q ) Z^{\ell(q)} Z ℓ ( q ) , so
A μ V β = V β . (92) A_\mu V_\beta= V_\beta. \tag*{(92)} A μ V β = V β . (92)
For toric codes this fixes every vertex charge, while the weakened plaquette constraints allow flux fluctuations [9 , 10 ] [2]. Equation (S14) detects a Z Z Z operator with nonzero X X X syndrome. Exact correction also requires that errors with the same syndrome reveal no logical information, which follows from the matrix elements below.
Let d Z d_Z d Z be the minimum weight of a nontrivial ideal Z Z Z logical operator, and let E Z = { Z z : z ∈ S } \mathcal{E}_Z = \{Z^z : z \in S\} E Z = { Z z : z ∈ S } be any error set for which ∣ z ⊕ z ′ ∣ < d Z |z \oplus z'| < d_Z ∣ z ⊕ z ′ ∣ < d Z for every pair z , z ′ ∈ S z,z' \in S z , z ′ ∈ S . For v = z ⊕ z ′ v = z \oplus z' v = z ⊕ z ′ , commutation of Z v Z^v Z v with the filter gives
⟨ q β ∣ Z v ∣ q β ′ ⟩ = ⟨ + ∣ ⊗ n Z v ⊕ ℓ ( q ⊕ q ′ ) F β 2 ∣ + ⟩ ⊗ n Z β . (93) \langle q_\beta|Z^v|q'_\beta\rangle=
\frac{\langle+|^{\otimes n} Z^{v\oplus\ell(q\oplus q')}F_\beta^2|+\rangle^{\otimes n}}{Z_\beta}. \tag*{(93)} ⟨ q β ∣ Z v ∣ q β ′ ⟩ = Z β ⟨ + ∣ ⊗ n Z v ⊕ ℓ ( q ⊕ q ′ ) F β 2 ∣ + ⟩ ⊗ n . (93)
When q ≠ q ′ q \ne q' q = q ′ , a nonzero numerator would require v ⊕ ℓ ( q ⊕ q ′ ) ∈ row H Z v \oplus\ell(q \oplus q') \in\operatorname{row} H_Z v ⊕ ℓ ( q ⊕ q ′ ) ∈ row H Z , because F β 2 F_\beta^2 F β 2 contains only products of measured checks. It would make v v v a nontrivial ideal logical Z Z Z operator, contradicting ∣ v ∣ < d Z |v| < d_Z ∣ v ∣ < d Z . For q = q ′ q = q' q = q ′ , the result is independent of q q q . Hence
V β † ( Z v ) † Z v ′ V β = ⟨ ψ β ∣ Z v ⊕ v ′ ∣ ψ β ⟩ I K . (94) V_\beta^\dagger(Z^{v})^\dagger Z^{v'} V_\beta= \langle\psi_\beta|Z^{v\oplus v'}|\psi_\beta\rangle I_K. \tag*{(94)} V β † ( Z v ) † Z v ′ V β = ⟨ ψ β ∣ Z v ⊕ v ′ ∣ ψ β ⟩ I K . (94)
This is the exact Knill–Laflamme condition [11 ] for E Z \mathcal{E}_Z E Z . In particular, all pure Z Z Z errors of weight at most ⌊ ( d Z − 1 ) / 2 ⌋ \lfloor(d_Z - 1)/2\rfloor ⌊( d Z − 1 ) /2 ⌋ remain exactly correctable, for every finite strength profile and in the projective limit. The result also covers channels with Kraus operators in the linear span of this error set.
A product of B j B_j B j may change the physical state at finite strength; Eq. (94) asserts a scalar compression , not that every B j B_j B j acts as the identity on the imperfect subspace. Residual deformations from degenerate Z Z Z errors nevertheless carry no logical information in this error set and can be removed by an unrestricted recovery. A usual syndrome-dependent Pauli correction alone need not return the state to the original imperfect subspace.
The X X X sector has the opposite behavior. An X x X_x X x error changes the signs of the B j B_j B j it anticommutes with; finite-strength filtering leaves a nonzero overlap between error branches. As shown in Supplementary Lemma 2.1, that overlap is
⟨ q β ∣ X x ∣ q β ′ ⟩ = δ q q ′ ( − 1 ) q ⋅ g ( x ) κ β ( x ) \langle q_\beta|X_x|q'_\beta\rangle= \delta_{qq'}(-1)^{q\cdot g(x)}\kappa_\beta(x) ⟨ q β ∣ X x ∣ q β ′ ⟩ = δ q q ′ ( − 1 ) q ⋅ g ( x ) κ β ( x )
and can depend on the logical input even for a single-site error [4]. This identifies the sector in which finite-strength preparation creates the recovery obstruction studied here. Restricting the subsequent channel to X X X noise still tests preservation of arbitrary quantum superpositions and their reference entanglement. Combining the sector-specific results into a threshold for joint X / Z X/Z X / Z noise, or allowing noisy syndrome-extraction circuits, is a further task and is not assumed in the present bounds.
Gauge measure and comparison of representatives
For a general CSS code, the probability of a computational-basis string s ∈ F 2 n s \in\mathbb{F}_2^n s ∈ F 2 n in ∣ ψ β ⟩ |\psi_\beta\rangle ∣ ψ β ⟩ is
S β ( s ) = − ∑ j β j ( − 1 ) ( H Z s ) j , π β ( s ) = e − S β ( s ) 2 n Z β . (95) \mathcal{S}_\beta(s)=-\sum_j\beta_j(-1)^{(H_Zs)_j},\qquad
\pi_\beta(s)=\frac{e^{-\mathcal{S}_\beta(s)}}{2^n Z_\beta}. \tag*{(95)} S β ( s ) = − j ∑ β j ( − 1 ) ( H Z s ) j , π β ( s ) = 2 n Z β e − S β ( s ) . (95)
The normalization is ∑ s e − S β ( s ) = 2 n Z β \sum_s e^{-\mathcal{S}_\beta(s)}=2^nZ_\beta ∑ s e − S β ( s ) = 2 n Z β ; Z β Z_\beta Z β is the quantum normalization defined in Eq. (88) . Diagonal quantum expectations therefore equal averages in this classical multispin measure, whose generalized Ising form is described in Ref. [12 ] .
For representatives with difference c = ℓ 1 ⊕ ℓ 2 ∈ im H Z T c=\ell_1\oplus\ell_2\in\operatorname{im}H_Z^T c = ℓ 1 ⊕ ℓ 2 ∈ im H Z T , put W = Z c W=Z^c W = Z c . Expanding the check factors as in the high-temperature expansion [13 Sec. 4] and using 2 − n ∑ s ( − 1 ) v ⋅ s = 1 { v = 0 } 2^{-n}\sum_s(-1)^{v\cdot s}=\mathbf{1}_{\{v=0\}} 2 − n ∑ s ( − 1 ) v ⋅ s = 1 { v = 0 } gives
⟨ W ⟩ β = ∑ a ∈ F 2 m H Z T a = c ∏ j : a j = 1 u j ∑ a ∈ F 2 m H Z T a = 0 ∏ j : a j = 1 u j , u j = tanh β j . (96) \langle W\rangle_\beta=
\frac{\sum_{\substack{a\in\mathbb{F}_2^m\\H_Z^Ta=c}}\prod_{j:a_j=1}u_j}
{\sum_{\substack{a\in\mathbb{F}_2^m\\H_Z^Ta=0}}\prod_{j:a_j=1}u_j},
\qquad u_j=\tanh\beta_j. \tag*{(96)} ⟨ W ⟩ β = ∑ a ∈ F 2 m H Z T a = 0 ∏ j : a j = 1 u j ∑ a ∈ F 2 m H Z T a = c ∏ j : a j = 1 u j , u j = tanh β j . (96)
The numerator sums over all products of measured checks implementing the change of representative; the denominator sums over their relations. This identity is exact for any check matrix, including redundant rows, and any nonnegative strength profile. It also shows ⟨ W ⟩ β ≥ 0 \langle W\rangle_\beta\ge0 ⟨ W ⟩ β ≥ 0 for the all-plus branch. For a signed record, the same expansion replaces u j u_j u j by s j u j s_j u_j s j u j ; its weights need not be nonnegative.
For toric codes this multispin action becomes the Z 2 \mathbb{Z}_2 Z 2 lattice gauge action [12 , 14 ] . All chains, cochains, homology groups, and intersection numbers below have coefficients in F 2 \mathbb{F}_2 F 2 . The boundary of a chain is denoted by ∂ \partial ∂ , and the coboundary of a cochain by δ \delta δ . On the cubic torus T L 3 T_L^3 T L 3 , a computational-basis configuration is an edge cochain s ∈ C 1 ( T L 3 ; F 2 ) s \in C^1(T_L^3;\mathbb{F}_2) s ∈ C 1 ( T L 3 ; F 2 ) , with Z e = ( − 1 ) s e Z_e=(-1)^{s_e} Z e = ( − 1 ) s e . Thus the plaquette operator B p = ∏ e ∈ ∂ p Z e B_p=\prod_{e\in\partial p} Z_e B p = ∏ e ∈ ∂ p Z e has eigenvalue ( − 1 ) ( δ s ) p (-1)^{(\delta s)_p} ( − 1 ) ( δ s ) p . The squared amplitudes of ∣ ψ β ⟩ ∝ exp [ 1 2 ∑ p β p B p ] ∣ + ⟩ ⊗ n |\psi_\beta\rangle\propto\exp[\frac{1}{2}\sum_p\beta_pB_p]|+\rangle^{\otimes n} ∣ ψ β ⟩ ∝ exp [ 2 1 ∑ p β p B p ] ∣ + ⟩ ⊗ n give the probability measure
π β ( s ) = exp [ ∑ p β p ( − 1 ) ( δ s ) p ] ∑ s ′ exp [ ∑ p β p ( − 1 ) ( δ s ′ ) p ] . (97) \pi_\beta(s)=\frac{\exp\left[\sum_p\beta_p(-1)^{(\delta s)_p}\right]}{\sum_{s'}\exp\left[\sum_p\beta_p(-1)^{(\delta s')_p}\right]}. \tag*{(97)} π β ( s ) = ∑ s ′ exp [ ∑ p β p ( − 1 ) ( δ s ′ ) p ] exp [ ∑ p β p ( − 1 ) ( δ s ) p ] . (97)
Every diagonal Wilson expectation in the quantum state is therefore an expectation in this classical gauge measure. We count each unoriented plaquette once. The action β p B p \beta_pB_p β p B p corresponds to β ~ p 1 { B p = + 1 } \widetilde{\beta}_p\mathbf{1}_{\{B_p=+1\}} β p 1 { B p = + 1 } with β ~ p = 2 β p \widetilde{\beta}_p=2\beta_p β p = 2 β p in the Potts-gauge convention of Refs. [15 , 16 ] . The convention that counts both plaquette orientations in Ref. [17 ] instead has coupling β F V = β / 2 \beta_{\mathrm{FV}}=\beta/2 β FV = β /2 for our uniform action.
Supplementary Lemma 1.1. (Representative identity and comparison of encodings). Let ℓ 1 \ell_1 ℓ 1 and ℓ 2 \ell_2 ℓ 2 be two Z \mathbb{Z} Z representatives such that ℓ 1 ⊕ ℓ 2 = ∂ S \ell_1\oplus\ell_2=\partial S ℓ 1 ⊕ ℓ 2 = ∂ S , and put C = ℓ 1 ⊕ ℓ 2 C=\ell_1\oplus\ell_2 C = ℓ 1 ⊕ ℓ 2 . Then
Z ℓ 1 Z ℓ 2 = ∏ p ∈ S B p = : W ( C ) , ∥ ( Z ℓ 1 − Z ℓ 2 ) ∣ φ ⟩ ∥ 2 = 2 − 2 ⟨ W ( C ) ⟩ φ (98) Z^{\ell_1}Z^{\ell_2}=\prod_{p\in S}B_p=:W(C),\qquad\left\|(Z^{\ell_1}-Z^{\ell_2})|\varphi\rangle\right\|^2=2-2\langle W(C)\rangle_\varphi \tag*{(98)} Z ℓ 1 Z ℓ 2 = p ∈ S ∏ B p =: W ( C ) , ( Z ℓ 1 − Z ℓ 2 ) ∣ φ ⟩ 2 = 2 − 2 ⟨ W ( C ) ⟩ φ (98) for every normalized state ∣ φ ⟩ |\varphi\rangle ∣ φ ⟩ . Suppose that a normalized state ∣ ψ ⟩ |\psi\rangle ∣ ψ ⟩ is a + 1 +1 + 1 eigenstate of an X X X logical operator that anticommutes with both Z ℓ i Z^{\ell_i} Z ℓ i . Define V i ∣ 0 ⟩ L = ∣ ψ ⟩ V_i|0\rangle_L=|\psi\rangle V i ∣0 ⟩ L = ∣ ψ ⟩ and V i ∣ 1 ⟩ L = Z ℓ i ∣ ψ ⟩ V_i|1\rangle_L=Z^{\ell_i}|\psi\rangle V i ∣1 ⟩ L = Z ℓ i ∣ ψ ⟩ . Both V i : H L → H Q V_i:\mathcal{H}_L\to\mathcal{H}_Q V i : H L → H Q are isometries. In the common ordered logical basis ( ∣ 0 ⟩ L , ∣ 1 ⟩ L ) (|0\rangle_L,|1\rangle_L) ( ∣0 ⟩ L , ∣1 ⟩ L ) , with w = ⟨ ψ ∣ W ( C ) ∣ ψ ⟩ w=\langle\psi|W(C)|\psi\rangle w = ⟨ ψ ∣ W ( C ) ∣ ψ ⟩ , we have
V 1 † V 2 = ( 1 0 0 w ) , ∥ V 1 − V 2 ∥ ∞ 2 = 2 ( 1 − w ) . (99) V_1^\dagger V_2=\begin{pmatrix}1&0\\0&w\end{pmatrix},\qquad\|V_1-V_2\|_\infty^2=2(1-w). \tag*{(99)} V 1 † V 2 = ( 1 0 0 w ) , ∥ V 1 − V 2 ∥ ∞ 2 = 2 ( 1 − w ) . (99)
Proof. Multiplying the two strings cancels their common edges. Multiplying the plaquette boundaries in S S S cancels every interior edge, leaving precisely ∂ S \partial S ∂ S . This proves the operator identity in Eq. (98) . Since the two strings are commuting Hermitian unitaries, the square of their difference is 2 I − 2 W ( C ) 2I-2W(C) 2 I − 2 W ( C ) .
Let X ‾ \overline{X} X be the logical operator in the hypothesis. The identities X ‾ ∣ ψ ⟩ = ∣ ψ ⟩ \overline{X}|\psi\rangle=|\psi\rangle X ∣ ψ ⟩ = ∣ ψ ⟩ and X ‾ Z ℓ i X ‾ = − Z ℓ i \overline{X}Z^{\ell_i}\overline{X}=-Z^{\ell_i} X Z ℓ i X = − Z ℓ i imply ⟨ ψ ∣ Z ℓ i ∣ ψ ⟩ = 0 \langle\psi|Z^{\ell_i}|\psi\rangle=0 ⟨ ψ ∣ Z ℓ i ∣ ψ ⟩ = 0 . The two columns of each V i V_i V i are therefore orthonormal. The same identities eliminate the off-diagonal entries of V 1 † V 2 V_1^\dagger V_2 V 1 † V 2 ; its remaining entries are 1 and w w w . Consequently, ( V 1 − V 2 ) † ( V 1 − V 2 ) = diag ( 0 , 2 − 2 w ) (V_1-V_2)^\dagger(V_1-V_2)=\operatorname{diag}(0,2-2w) ( V 1 − V 2 ) † ( V 1 − V 2 ) = diag ( 0 , 2 − 2 w ) , proving Eq. (99) . □ \square □
For the states prepared with finite-strength measurements in this work, the specified X X X logical commutes with every preparation filter and stabilizes ∣ + ⟩ ⊗ n |+\rangle^{\otimes n} ∣ + ⟩ ⊗ n . Equation (S21) thus applies before and after the boundary enhancement. It compares the two encoding maps in a common logical basis. In particular, its operator-norm bound also holds after tensoring the maps with an identity on any reference system. Recovery after noise is a separate question, governed by the reference-environment state.
The background figure also expresses this comparison as a state fidelity. For ∣ ϕ i ⟩ = Z ℓ i ∣ ψ β ⟩ |\phi_i\rangle=Z^{\ell_i}|\psi_\beta\rangle ∣ ϕ i ⟩ = Z ℓ i ∣ ψ β ⟩ , the squared fidelity is
F r e p : = ∣ ⟨ ϕ 1 ∣ ϕ 2 ⟩ ∣ 2 = ∣ ⟨ W ( C ) ⟩ β ∣ 2 . (100) F_{\mathrm{rep}}:=|\langle\phi_1|\phi_2\rangle|^2=|\langle W(C)\rangle_\beta|^2. \tag*{(100)} F rep := ∣ ⟨ ϕ 1 ∣ ϕ 2 ⟩ ∣ 2 = ∣ ⟨ W ( C ) ⟩ β ∣ 2 . (100)
An area or perimeter law for the Wilson overlap therefore gives the corresponding decay of this squared fidelity. For parallel winding strings of length L L L separated by R = Θ ( L ) R=\Theta(L) R = Θ ( L ) on the square or cubic torus, the smaller filling has area A min = L min ( R , L − R ) = Θ ( L 2 ) A_{\min}=L\min(R,L-R)=\Theta(L^2) A m i n = L min ( R , L − R ) = Θ ( L 2 ) , while the total boundary length is P = 2 L P=2L P = 2 L , provided both complementary separations are proportional to L L L . Thus a perimeter law can still give F r e p → 0 F_{\mathrm{rep}}\to0 F rep → 0 as L → ∞ L\to\infty L → ∞ at fixed finite strength. The background figure separates the representative comparison in panel (b) from the bare large-loop laws in panel (c). At fixed L L L , changing R R R changes the filling area but leaves P = 2 L P=2L P = 2 L unchanged; it is distinct from increasing L L L at fixed R / L R/L R / L . A perimeter law alone does not remove finite-separation corrections or specify the full R R R dependence.
Exact area law in two dimensions
In two dimensions, a closed boundary has only two plaquette fillings on the torus. This simple geometry makes the Wilson expectation exactly calculable and exposes the obstruction to repairing it near the boundary alone. The uniform finite-torus expression is given in Ref. [13 Eq. (65)] ; we retain distinct plaquette couplings below.
Supplementary Lemma 1.2 (Exact inhomogeneous Wilson expectation in two dimensions). On the connected square two-torus, let β p ≥ 0 \beta_p \ge0 β p ≥ 0 and u p = tanh β p u_p = \tanh\beta_p u p = tanh β p . For a null-homologous one-cycle C = ∂ S C = \partial S C = ∂ S ,
⟨ W ( C ) ⟩ β = ∏ p ∈ S u p + ∏ p ∈ S c u p 1 + ∏ p u p . (101) \langle W(C)\rangle_{\beta} = \frac{\prod_{p\in S}u_p+\prod_{p\in S^c}u_p}{1+\prod_p u_p}. \tag*{(101)} ⟨ W ( C ) ⟩ β = 1 + ∏ p u p ∏ p ∈ S u p + ∏ p ∈ S c u p . (101) Here S c S^c S c is the complementary plaquette chain. The formula includes a disconnected pair of homologous winding strings. If every plaquette outside a protected set P r \mathcal{P}_r P r has the same coupling 0 < β < ∞ 0<\beta<\infty 0 < β < ∞ , then arbitrary nonnegative couplings inside P r \mathcal{P}_r P r obey
0 ≤ ⟨ W ( C ) ⟩ β ≤ ( tanh β ) ∣ S ∖ P r ∣ + ( tanh β ) ∣ S c ∖ P r ∣ . (102) 0\le\langle W(C)\rangle_{\beta}\le(\tanh\beta)^{|S\setminus\mathcal{P}_r|}+(\tanh\beta)^{|S^c\setminus\mathcal{P}_r|}. \tag*{(102)} 0 ≤ ⟨ W ( C ) ⟩ β ≤ ( tanh β ) ∣ S ∖ P r ∣ + ( tanh β ) ∣ S c ∖ P r ∣ . (102) If each unprotected filling contains at least c L 2 cL^2 c L 2 plaquettes, (S24) gives the exponentially small upper bound
2 exp [ − c ∣ log tanh β ∣ L 2 ] , 2\exp[-c|\log\tanh\beta|L^2], 2 exp [ − c ∣ log tanh β ∣ L 2 ] , also in the hard-protection limit.
Proof. Expand every local factor as e β p B p = cosh β p ( 1 + u p B p ) e^{\beta_pB_p}=\cosh\beta_p(1+u_pB_p) e β p B p = cosh β p ( 1 + u p B p ) , then sum over the edge spins. A product of plaquettes survives this sum precisely when its boundary vanishes. In the numerator with the insertion W ( C ) W(C) W ( C ) , it survives precisely when its boundary is C C C . On the connected two-torus the space of two-cycles consists of the empty chain and the full surface. Hence the denominator has exactly these two surviving terms, and the numerator has the two chains S S S and S c S^c S c . Their common factors cancel, giving Eq. (S23).
The denominator is at least one and each protected factor satisfies 0 ≤ u p ≤ 1 0\le u_p\le1 0 ≤ u p ≤ 1 . Dropping these factors yields Eq. (S24). Taking u p ↑ 1 u_p\uparrow1 u p ↑ 1 on protected plaquettes proves the same bound for hard protection. At background β = 0 \beta=0 β = 0 , the expectation is exactly zero whenever each complementary filling contains an unprotected plaquette.
For two straight representatives separated by a distance proportional to L L L , tubes of width r = o ( L ) r=o(L) r = o ( L ) leave order L 2 L^2 L 2 plaquettes in both complementary fillings. This unprotected-area condition is part of the statement: a protected set covering an entire filling is not covered by the no-go bound with extensive unprotected area.
For uniform strength on the two-torus, Eq. (S23) reduces to
⟨ W ( C ) ⟩ β = u ∣ S ∣ + u L 2 − ∣ S ∣ 1 + u L 2 , u = tanh β . (103) \langle W(C)\rangle_{\beta}=\frac{u^{|S|}+u^{L^2-|S|}}{1+u^{L^2}},\qquad u=\tanh\beta. \tag*{(103)} ⟨ W ( C ) ⟩ β = 1 + u L 2 u ∣ S ∣ + u L 2 − ∣ S ∣ , u = tanh β . (103)
In a simply connected planar box there are no nonempty two-cycles. The same expansion consequently gives
⟨ W ( C ) ⟩ β = ( tanh β ) ∣ S ∣ . (104) \langle W(C)\rangle_{\beta}=(\tanh\beta)^{|S|}. \tag*{(104)} ⟨ W ( C ) ⟩ β = ( tanh β ) ∣ S ∣ . (104)
This remains true in the infinite-plane limit for any fixed finite loop. For every 0 < β < ∞ 0<\beta<\infty 0 < β < ∞ , its area tension is − log tanh β > 0 -\log\tanh\beta>0 − log tanh β > 0 . On the torus the smaller of the two complementary areas controls the exponent. In particular, loops with both areas proportional to L 2 L^2 L 2 decay exponentially in L 2 L^2 L 2 . There is no finite β \beta β at which the bare two-dimensional Wilson observable acquires a perimeter law. These statements concern fixed β \beta β ; their conclusion must not be interchanged with a limit in which β \beta β grows with L L L .
Three dimensions: the high-temperature area bound
The gauge representation also shows why a different behavior is possible in three dimensions. There are now many plaquette surfaces with the same boundary. On a finite cubic box with free boundary conditions, or on a finite cubic torus for a null-homologous C C C , expansion of each factor e β B p e^{\beta B_p} e β B p gives the exact formula
⟨ W ( C ) ⟩ β = ∑ S : ∂ S = C u ∣ S ∣ ∑ T : ∂ T = 0 u ∣ T ∣ , u = tanh β . (105) \langle W(C)\rangle_{\beta}
=
\frac{\sum_{S:\,\partial S=C}u^{|S|}}
{\sum_{T:\,\partial T=0}u^{|T|}},
\qquad u=\tanh\beta. \tag*{(105)} ⟨ W ( C ) ⟩ β = ∑ T : ∂ T = 0 u ∣ T ∣ ∑ S : ∂ S = C u ∣ S ∣ , u = tanh β . (105)
The common factors 2 n ∏ p cosh β 2^{n}\prod_{p}\cosh\beta 2 n ∏ p cosh β cancel. The sums are over binary plaquette chains; surfaces need not be embedded or orientable.
Supplementary Lemma 1.3 (Area bounds at sufficiently small strength). Let C C C be a simple loop in a cubic box and let A ( C ) A(C) A ( C ) be the smallest number of plaquettes in a filling of C C C . There are numerical constants c 0 , c 1 c_0,c_1 c 0 , c 1 , independent of the box and the loop, such that, when c 1 tanh β < 1 c_1\tanh\beta<1 c 1 tanh β < 1 ,
( tanh β ) A ( C ) ≤ ⟨ W ( C ) ⟩ β ≤ c 0 [ c 1 tanh β ] A ( C ) 1 − c 1 tanh β . (106) (\tanh\beta)^{A(C)}\leq\langle W(C)\rangle_\beta\leq\frac{c_0[c_1\tanh\beta]^{A(C)}}{1-c_1\tanh\beta}. \tag*{(106)} ( tanh β ) A ( C ) ≤ ⟨ W ( C ) ⟩ β ≤ 1 − c 1 tanh β c 0 [ c 1 tanh β ] A ( C ) . (106)
For a planar rectangle in the infinite lattice, A ( C ) A(C) A ( C ) is its usual area.
Proof. Choose a minimum filling S 0 S_0 S 0 . Symmetric difference with S 0 S_0 S 0 bijects closed chains T T T with chains having boundary C C C . For 0 < u < 1 0<u<1 0 < u < 1 , the inequality ∣ S 0 ⊕ T ∣ ≤ ∣ S 0 ∣ + ∣ T ∣ |S_0\mathbin{\oplus}T|\leq|S_0|+|T| ∣ S 0 ⊕ T ∣ ≤ ∣ S 0 ∣ + ∣ T ∣ implies that the numerator of Eq. (105) is at least u A ( C ) u^{A(C)} u A ( C ) times its denominator. This proves the lower bound without an independence assumption.
For the upper bound, make plaquettes adjacent when they share an edge. Each connected component of a plaquette chain has a boundary that is a cycle, and distinct components have disjoint boundary edges. Each component boundary is therefore a subcycle of C C C . Since a simple loop has no nonempty proper subcycle, exactly one component has boundary C C C ; every other component is closed. Denote the former component by S 1 S_1 S 1 . Removing it leaves a closed chain disjoint from S 1 S_1 S 1 . Dropping this disjointness restriction bounds the numerator by
( ∑ S 1 c o n n e c t e d : ∂ S 1 = C u ∣ S 1 ∣ ) ( ∑ T : ∂ T = 0 u ∣ T ∣ ) . \left(\sum_{\substack{S_1\ \mathrm{connected}:\\ \partial S_1=C}}u^{|S_1|}\right)\left(\sum_{\substack{T:\ \partial T=0}}u^{|T|}\right). S 1 connected : ∂ S 1 = C ∑ u ∣ S 1 ∣ T : ∂ T = 0 ∑ u ∣ T ∣ .
The denominator cancels. A component S 1 S_1 S 1 must contain a plaquette incident on a fixed edge of C C C . The plaquette adjacency graph has bounded degree, so the number of such connected sets of size m m m is at most c 0 c 1 m c_0c_1^m c 0 c 1 m . One way to obtain this bound is to encode a spanning tree by a depth-first walk of length at most 2 m 2m 2 m on this graph. Summing over m ≥ A ( C ) m\geq A(C) m ≥ A ( C ) gives Eq. (106) ; the case β = 0 \beta=0 β = 0 follows by continuity. Projection onto the plane of a rectangular loop shows that every filling uses at least one plaquette above each unit square of the rectangle, proving the final assertion.
The surface sum establishes an area-law interval: producing a Wilson boundary costs a number of plaquettes proportional to its minimum filling area. Uniformity in the containing box carries this conclusion to the thermodynamic limit. The constants do not locate the transition.
Three dimensions: the low-temperature perimeter bound
At large β \beta β , it is more useful to count violated plaquettes. Mark a dual edge whenever its primal plaquette has B p = − 1 B_p=-1 B p = − 1 . The product of the six plaquettes around every cube is one, so an even number of marked dual edges meets each dual vertex. The marked edges therefore form closed networks. Each violated plaquette carries the relative weight e − 2 β e^{-2\beta} e − 2 β .
The elementary expansion below is taken in simply connected boxes with flat exterior boundary conditions, followed by the infinite-volume limit. Equivalently, it is the gas of finite closed dual networks with a cutoff on their support. The gauge-to-defect map has fibers of equal size, and every such network has a filling in a simply connected box; hence these weights give the gauge measure. Winding networks on a finite torus impose an additional global constraint. Their contribution is treated in Supplementary Note 4, rather than dropped here.
Supplementary Lemma 1.4 (Perimeter bounds at sufficiently large strength). For all sufficiently large fixed β \beta β , there are constants 0 < c β ≤ C β < ∞ 0<c_\beta\leq C_\beta<\infty 0 < c β ≤ C β < ∞ such that every simple finite loop in the infinite cubic lattice satisfies
e − C β ∣ C ∣ ≤ ⟨ W ( C ) ⟩ β ≤ e − c β ∣ C ∣ . (107) e^{-C_\beta|C|}\leq\langle W(C)\rangle_\beta\leq e^{-c_\beta|C|}. \tag*{(107)} e − C β ∣ C ∣ ≤ ⟨ W ( C ) ⟩ β ≤ e − c β ∣ C ∣ . (107)
The constants can be chosen with c β , C β c_\beta,C_\beta c β , C β both of order e − 8 β e^{-8\beta} e − 8 β as β → ∞ \beta\to\infty β → ∞ .
Proof. First we prove the upper bound directly in the spin variables. Choose a subset J J J of the edges of C C C such that no plaquette contains two edges of J J J . An edge shares a plaquette with at most twelve other edges, so a greedy choice gives ∣ J ∣ ≥ ∣ C ∣ / 13 |J|\geq|C|/13 ∣ J ∣ ≥ ∣ C ∣/13 . Conditional on all other spins, the spins in J J J are independent. Each sees a field of magnitude at most 4 β 4\beta 4 β , because an edge belongs to four plaquettes. Consequently,
∣ ⟨ W ( C ) ⟩ β ∣ ≤ ( tanh 4 β ) ∣ J ∣ ≤ ( tanh 4 β ) ∣ C ∣ / 13 . |\langle W(C)\rangle_\beta|\leq(\tanh4\beta)^{|J|}\leq(\tanh4\beta)^{|C|/13}. ∣ ⟨ W ( C ) ⟩ β ∣ ≤ ( tanh 4 β ) ∣ J ∣ ≤ ( tanh 4 β ) ∣ C ∣/13 .
The Wilson expectation is nonnegative: expansion in plaquette subsets has nonnegative coefficients both for free boundaries and for exterior edge spins fixed to + 1 +1 + 1 . Thus one may take c β = − 1 13 log tanh 4 β c_\beta=-\frac{1}{13}\log\tanh4\beta c β = − 13 1 log tanh 4 β .
For the lower bound, call a nonempty connected closed dual network γ \gamma γ a polymer. Distinct polymers are compatible when they share no dual vertex, and their activities are z ( γ ) = e − 2 β ∣ γ ∣ z(\gamma)=e^{-2\beta|\gamma|} z ( γ ) = e − 2 β ∣ γ ∣ . There is a constant a 0 a_0 a 0 such that at most a 0 m a_0^m a 0 m polymers of m m m edges contain a given dual vertex. Every polymer has at least four edges and at most ∣ γ ∣ |\gamma| ∣ γ ∣ vertices. For sufficiently large β \beta β ,
∑ γ ′ : γ ′ incompatible with γ ∣ z ( γ ′ ) ∣ e ∣ γ ′ ∣ ≤ ∣ γ ∣ ∑ m ≥ 4 ( a 0 e 1 − 2 β ) m ≤ ∣ γ ∣ . (108) \sum_{\substack{\gamma':\,\gamma'\\\text{incompatible with }\gamma}} |z(\gamma')|e^{|\gamma'|}
\leq|\gamma|\sum_{m\geq4}(a_0e^{1-2\beta})^m\leq|\gamma|. \tag*{(108)} γ ′ : γ ′ incompatible with γ ∑ ∣ z ( γ ′ ) ∣ e ∣ γ ′ ∣ ≤ ∣ γ ∣ m ≥ 4 ∑ ( a 0 e 1 − 2 β ) m ≤ ∣ γ ∣. (108)
This is the standard polymer convergence criterion [18 ] . It guarantees absolute convergence of the connected-cluster expansion of the logarithm of the partition function; the sum of absolute cluster weights containing a specified polymer γ \gamma γ is bounded by ∣ z ( γ ) ∣ e ∣ γ ∣ |z(\gamma)|e^{|\gamma|} ∣ z ( γ ) ∣ e ∣ γ ∣ . The same estimate holds if any activities are multiplied by signs.
Choose a plaquette filling S S S of C C C . Inserting W ( C ) W(C) W ( C ) multiplies each polymer activity by ( − 1 ) ∣ γ ∩ S ∗ ∣ (-1)^{|\gamma\cap S^*|} ( − 1 ) ∣ γ ∩ S ∗ ∣ , where S ∗ S^* S ∗ denotes the dual edges crossing S S S . Thus the logarithm of the Wilson expectation is the difference of two convergent cluster expansions. A contributing cluster must contain a polymer with odd intersection with S ∗ S^* S ∗ . Marking one such polymer and applying the preceding bound gives
∣ log ⟨ W ( C ) ⟩ β ∣ ≤ 2 ∑ γ : ∣ γ ∩ S ∗ ∣ odd e − ( 2 β − 1 ) ∣ γ ∣ . (109) |\log\langle W(C)\rangle_\beta|
\leq2\sum_{\substack{\gamma:\,|\gamma\cap S^*|\\\text{odd}}}e^{-(2\beta-1)|\gamma|}. \tag*{(109)} ∣ log ⟨ W ( C ) ⟩ β ∣ ≤ 2 γ : ∣ γ ∩ S ∗ ∣ odd ∑ e − ( 2 β − 1 ) ∣ γ ∣ . (109)
At β = ∞ \beta=\infty β = ∞ the Wilson expectation is one. Absolute convergence makes the twisted partition function nonzero, so continuity fixes a real logarithm throughout this large-β \beta β interval.
An odd-intersection polymer of size m m m must approach C C C within distance of order m m m . Indeed, its coordinate bounding box has diameter at most of order m m m . If this box were disjoint from C C C , the closed network would bound inside the box, and its intersection with any filling of C C C would be even. There are at most a 1 ∣ C ∣ ( m + 1 ) 3 a_1|C|(m+1)^3 a 1 ∣ C ∣ ( m + 1 ) 3 possible root vertices within this distance of C C C . Counting the polymers from these vertices in Eq. (109) therefore gives
∣ log ⟨ W ( C ) ⟩ β ∣ ≤ 2 a 1 ∣ C ∣ ∑ m ≥ 4 ( m + 1 ) 3 ( a 0 e 1 − 2 β ) m . |\log\langle W(C)\rangle_\beta|
\leq2a_1|C|\sum_{m\geq4}(m+1)^3(a_0e^{1-2\beta})^m. ∣ log ⟨ W ( C ) ⟩ β ∣ ≤ 2 a 1 ∣ C ∣ m ≥ 4 ∑ ( m + 1 ) 3 ( a 0 e 1 − 2 β ) m .
The series converges under a sufficiently strong version of Eq. (108) . It supplies C β C_\beta C β and is of order e − 8 β e^{-8\beta} e − 8 β . The direct upper bound has the same order since − log tanh 4 β ∼ 2 e − 8 β -\log\tanh4\beta\sim2e^{-8\beta} − log tanh 4 β ∼ 2 e − 8 β . Uniform convergence permits the infinite-volume limit.
This argument displays the geometric reason for the perimeter: a small closed defect network can link a filling only near its boundary. An intersection deep in the interior requires a larger network. More precise low-temperature coefficients, including corner and bottleneck corrections, are supplied by Forström and Viklund [17 ] . We use only the coarse bounds above; an expansion controlled at large β \beta β is not being extrapolated to the transition.
The full strict phases and the dual parameter
The expansions explain the two decay mechanisms but leave an intermediate range of measurement strengths unresolved. The sharp-transition theorem supplies the laws throughout both strict phases. Let p c F K ( Z 3 , 2 ) p_c^{\mathrm{FK}}(\mathbb{Z}^3,2) p c FK ( Z 3 , 2 ) be the critical edge parameter of the Fortuin–Kasteleyn random-cluster model on Z 3 \mathbb{Z}^3 Z 3 with cluster weight two [19 , 20 ] , and define β c \beta_c β c by
2 e 2 β c + 1 = p c F K ( Z 3 , 2 ) . (110) \frac{2}{e^{2\beta_c}+1}=p_c^{\mathrm{FK}}(\mathbb{Z}^3,2). \tag*{(110)} e 2 β c + 1 2 = p c FK ( Z 3 , 2 ) . (110)
For axis-parallel rectangular loops C j C_j C j whose two side lengths diverge and whose minimum side divided by the logarithm of their maximum side tends to infinity, Theorem 5 of Duncan and Schweinhart [15 ] implies
− log ⟨ W ( C j ) ⟩ β A ( C j ) ⟶ σ β > 0 , 0 < β < β c , − log ⟨ W ( C j ) ⟩ β ∣ C j ∣ ⟶ τ β > 0 , β > β c . (111) \begin{aligned}
-\frac{\log\langle W(C_j)\rangle_\beta}{A(C_j)}&\longrightarrow\sigma_\beta>0,\qquad0<\beta<\beta_c,\\
-\frac{\log\langle W(C_j)\rangle_\beta}{|C_j|}&\longrightarrow\tau_\beta>0,\qquad\beta>\beta_c. \tag*{(111)}
\end{aligned} − A ( C j ) log ⟨ W ( C j ) ⟩ β − ∣ C j ∣ log ⟨ W ( C j ) ⟩ β ⟶ σ β > 0 , 0 < β < β c , ⟶ τ β > 0 , β > β c . (111)
Here the expectations are thermodynamic gauge expectations. The q = 2 q=2 q = 2 statement is unconditional; the dual Ising surface-tension threshold equals its critical threshold, as used in the cited theorem. No assertion at β = β c \beta=\beta_c β = β c is required. For clarity, the parameter conversion follows directly from
e β B p = e − β e 2 β 1 { B p = + 1 } . e^{\beta B_p}=e^{-\beta}e^{2\beta1_{\{B_p=+1\}}}. e β B p = e − β e 2 β 1 { B p = + 1 } .
Thus the Potts-gauge coupling in the reference is 2 β 2\beta 2 β . The primal plaquette occupation parameter is 1 − e − 2 β 1-e^{-2\beta} 1 − e − 2 β , and duality sends it to 2 / ( e 2 β + 1 ) 2/(e^{2\beta}+1) 2/ ( e 2 β + 1 ) , giving Eq. (110) . Increasing the measurement strength decreases the dual parameter. The strict perimeter-law interval is consequently the subcritical interval of the dual random-cluster model.
Equation (S33) concerns rectangular loops in an infinite-volume measure. The two distant winding representatives in our restoration result live on a finite torus and form a different loop family. Their estimate is proved with the finite-torus dual law and its topological correction in Supplementary Note 4. This is the step that allows the full strict condition β > β c \beta>\beta_c β > β c to be used in the restoration theorem.
AQEC theory and finite-strength logical leakage
Finite-strength preparation leaves a residual overlap between a codeword and a local error acting on it. This overlap makes local reduced states depend on the logical input and allows subsequent noise to transfer logical information to its environment. We first derive the overlap and the corresponding single-site leakage. Bounds for several sites then show why vanishing leakage requires increasing measurement strength even when each single-site signal tends to zero.
All logarithms are natural. There are n n n physical qubits and m Z m_Z m Z measured Z Z Z checks, with binary check matrix H Z H_Z H Z . Physical error supports are identified with vectors in F 2 n \mathbb{F}_2^n F 2 n , and addition of supports means symmetric difference. For selected independent logical representatives ℓ 1 , … , ℓ k \ell_1,\ldots,\ell_k ℓ 1 , … , ℓ k , set
σ ( x ) = supp ( H Z x ) , g ( x ) = ( ∣ x ∩ ℓ 1 ∣ , … , ∣ x ∩ ℓ k ∣ ) ( m o d 2 ) , χ q ( x ) = ( − 1 ) q ⋅ g ( x ) . (112) \sigma(x)=\operatorname{supp}(H_Zx),\qquad g(x)=\left(|x\cap\ell_1|,\ldots,|x\cap\ell_k|\right)\pmod2,\qquad\chi_q(x)=(-1)^{q\cdot g(x)}. \tag*{(112)} σ ( x ) = supp ( H Z x ) , g ( x ) = ( ∣ x ∩ ℓ 1 ∣ , … , ∣ x ∩ ℓ k ∣ ) ( mod 2 ) , χ q ( x ) = ( − 1 ) q ⋅ g ( x ) . (112)
Thus σ ( x ) \sigma(x) σ ( x ) records the measured checks that anticommute with X x X_x X x , whereas g ( x ) g(x) g ( x ) records its commutation signs with the selected logical representatives. Here q ∈ F 2 k q\in\mathbb{F}_2^k q ∈ F 2 k labels a logical basis state and K = 2 k K=2^k K = 2 k is the logical Hilbert-space dimension. System subscripts L L L , Q Q Q , R R R , and E E E denote the logical input, the physical qubits, an untouched reference, and the physical-noise environment, respectively. In subscripts, β \beta β abbreviates the full profile β = ( β j ) \beta=(\beta_j) β = ( β j ) . A scalar β \beta β in a bound is the common strength of the checks specified in that statement. Unless stated otherwise, all strengths are finite and nonnegative. The toric examples use the standard periodic square or cubic lattice with side length L ≥ 3 L\ge3 L ≥ 3 .
The recovery task and normalized AQEC matrices
An encoding isometry V : H L → H Q V:\mathcal{H}_L\to\mathcal{H}_Q V : H L → H Q , with H L ≃ C K \mathcal{H}_L\simeq\mathbb{C}^K H L ≃ C K and H Q = ( C 2 ) ⊗ n \mathcal{H}_Q=(\mathbb{C}^2)^{\otimes n} H Q = ( C 2 ) ⊗ n , maps an arbitrary logical input into the full physical system and fixes its code subspace. For our construction V = V β V=V_\beta V = V β . Write N ( ρ ) = ∑ x E x ρ E x † \mathcal{N}(\rho)=\sum_x E_x\rho E_x^\dagger N ( ρ ) = ∑ x E x ρ E x † . A physical channel N ( ρ ) = ∑ x E x ρ E x † \mathcal{N}(\rho)=\sum_x E_x\rho E_x^\dagger N ( ρ ) = ∑ x E x ρ E x † is exactly correctable when a completely positive trace-preserving recovery R \mathcal{R} R satisfies R N = id L \mathcal{R}\mathcal{N}=\operatorname{id}_L RN = id L . This equality protects arbitrary logical inputs and their correlations with an untouched reference. The Knill–Laflamme condition [11 ] is
V † E x † E y V = λ x y I K . (113) V^\dagger E_x^\dagger E_yV=\lambda_{xy}I_K. \tag*{(113)} V † E x † E y V = λ x y I K . (113)
Thus overlaps between error branches may reveal an error syndrome, but must not depend on the unknown logical input. The condition includes both equality of diagonal logical matrix elements and vanishing of the off-diagonal elements.
For a physical input ∣ ϕ ⟩ Q |\phi\rangle_Q ∣ ϕ ⟩ Q , a Stinespring dilation of the noise [21 Sec. 2.2.2] ∣ ϕ ⟩ Q ↦ ∑ x E x ∣ ϕ ⟩ Q ∣ x ⟩ E |\phi\rangle_Q\mapsto\sum_x E_x|\phi\rangle_Q|x\rangle_E ∣ ϕ ⟩ Q ↦ ∑ x E x ∣ ϕ ⟩ Q ∣ x ⟩ E has complementary output
[ N ^ ( ρ ) ] x y = Tr ( ρ V † E y † E x V ) . (114) [\widehat{\mathcal{N}}(\rho)]_{xy}=\operatorname{Tr}(\rho V^\dagger E_y^\dagger E_xV). \tag*{(114)} [ N ( ρ ) ] x y = Tr ( ρ V † E y † E x V ) . (114)
Equation (S35) is therefore equivalent to an environment state independent of ρ \rho ρ . This is the information-disturbance formulation of correction [22 , 23 , 24 , 25 ] . The environment here belongs to the subsequent physical noise. It is separate from the fixed all-plus preparation record.
To quantify approximate correction, prepare ∣ Φ K ⟩ = K − 1 / 2 ∑ q ∣ q ⟩ R ∣ q ⟩ L |\Phi_K\rangle=K^{-1/2}\sum_q|q\rangle_R|q\rangle_L ∣ Φ K ⟩ = K − 1/2 ∑ q ∣ q ⟩ R ∣ q ⟩ L and let ρ R E \rho_{RE} ρ R E be its reference–environment state after encoding and noise. Its reference marginal is I R / K I_R/K I R / K . Following Ref. [4], define normalized matrices on the product basis ∣ q ⟩ R ∣ x ⟩ E |q\rangle_R|x\rangle_E ∣ q ⟩ R ∣ x ⟩ E by
( Λ + B ) x y , q q ′ = 1 K ⟨ q ∣ V x † E x † E y V ∣ q ′ ⟩ L , (\Lambda+B)_{xy,qq'}=\frac{1}{K}\langle q|V_x^\dagger E_x^\dagger E_yV|q'\rangle_L, ( Λ + B ) x y , q q ′ = K 1 ⟨ q ∣ V x † E x † E y V ∣ q ′ ⟩ L ,
Λ x y , q q ′ = δ q q ′ K 2 Tr ( V † E x † E y V ) . (115) \Lambda_{xy,qq'}=\frac{\delta_{qq'}}{K^2}\operatorname{Tr}(V^\dagger E_x^\dagger E_yV). \tag*{(115)} Λ x y , q q ′ = K 2 δ q q ′ Tr ( V † E x † E y V ) . (115)
The subscript x y , q q ′ xy,qq' x y , q q ′ denotes row ( q , x ) (q,x) ( q , x ) and column ( q ′ , y ) (q',y) ( q ′ , y ) . Direct partial trace shows that these matrices are ρ R E T \rho_{RE}^{T} ρ R E T and ( I R / K ⊗ ρ E ) T (I_R/K\otimes\rho_E)^{T} ( I R / K ⊗ ρ E ) T , respectively. The transpose is a full transpose in this product basis; it preserves positivity, relative entropy, and trace norm. In particular, Λ + B \Lambda+B Λ + B and Λ \Lambda Λ are density matrices, and B B B is their traceless Hermitian difference. In particular, both partial traces of B B B vanish. The normalization in Eq. (115) is retained when the number of selected logical qubits grows.
The standard AQEC entropy and its accompanying trace-distance defect are
D A Q E C = D ( Λ + B ∥ Λ ) = I ( R : E ) , ε = 1 2 ∥ B ∥ 1 . (116) D_{\mathrm{AQEC}}=D(\Lambda+B\Vert\Lambda)=I(R:E),\qquad\varepsilon=\frac{1}{2}\lVert B\rVert_1. \tag*{(116)} D AQEC = D ( Λ + B ∥Λ ) = I ( R : E ) , ε = 2 1 ∥ B ∥ 1 . (116)
Here D ( ρ ∥ σ ) = Tr [ ρ ( log ρ − log σ ) ] D(\rho\Vert\sigma)=\operatorname{Tr}[\rho(\log\rho-\log\sigma)] D ( ρ ∥ σ ) = Tr [ ρ ( log ρ − log σ )] and I ( R : E ) = S ( ρ R ) + S ( ρ E ) − S ( ρ R E ) I(R:E)=S(\rho_R)+S(\rho_E)-S(\rho_{RE}) I ( R : E ) = S ( ρ R ) + S ( ρ E ) − S ( ρ R E ) . If Q Q Q is the physical output, purity of R Q E RQE R QE also gives
D A Q E C = log K + S ( ρ E ) − S ( ρ Q ) = log K − I c ( I K / K , N V ) , (117) D_{\mathrm{AQEC}}=\log K+S(\rho_E)-S(\rho_Q)=\log K-I_c(I_K/K,\mathcal{N}V), \tag*{(117)} D AQEC = log K + S ( ρ E ) − S ( ρ Q ) = log K − I c ( I K / K , N V ) , (117)
where the coherent information is I c = S ( ρ Q ) − S ( ρ E ) I_c=S(\rho_Q)-S(\rho_E) I c = S ( ρ Q ) − S ( ρ E ) . The AQEC entropy is the coherent information lost relative to its noiseless value [22 , 23 ] ; it is not the Shannon entropy of the error distribution.
Our recovery criterion is the optimal entanglement fidelity for the maximally mixed input on the entire selected register [26 ] :
F e ⋆ = sup R ⟨ Φ K ∣ ( id R ⊗ R N V ) ( ∣ Φ K ⟩ ⟨ Φ K ∣ ) ∣ Φ K ⟩ . (118) F_e^\star=\sup_{\mathcal{R}}\langle\Phi_K|(\operatorname{id}_R\otimes\mathcal{R}\mathcal{N}V)(|\Phi_K\rangle\langle\Phi_K|)|\Phi_K\rangle. \tag*{(118)} F e ⋆ = R sup ⟨ Φ K ∣ ( id R ⊗ RN V ) ( ∣ Φ K ⟩ ⟨ Φ K ∣ ) ∣ Φ K ⟩ . (118)
The recovery acts on all of Q Q Q , outputs the logical register, and may depend on the code and noise but not on the logical input. Purification relates this task to environmental decoupling:
F e ⋆ = max ξ E F ( ρ R E , I R K ⊗ ξ E ) , F ( ρ , σ ) = ∥ ρ σ ∥ 1 2 . (119) F_e^\star=\max_{\xi_E}F\left(\rho_{RE},\frac{I_R}{K}\otimes\xi_E\right),\qquad F(\rho,\sigma)=\lVert\sqrt{\rho}\sqrt{\sigma}\rVert_1^2. \tag*{(119)} F e ⋆ = ξ E max F ( ρ R E , K I R ⊗ ξ E ) , F ( ρ , σ ) = ∥ ρ σ ∥ 1 2 . (119)
The maximization is over environment density matrices, and fidelity is squared throughout. A proof of both directions and quantitative trace-distance bounds is given in Supplementary Lemma 2.9. This establishes the existence of recovery, without assuming an efficient decoding algorithm or a recovery confined to the enhanced region.
The lower bounds below use independent X X X noise with fixed 0 < p < 1 0<p<1 0 < p < 1 . The recovery upper bounds in Supplementary Note 3 require only Pr [ A ⊆ supp ( x ) ] ≤ η ∣ A ∣ \Pr[A\subseteq\operatorname{supp}(x)]\leq\eta^{|A|} Pr [ A ⊆ supp ( x )] ≤ η ∣ A ∣ and therefore allow correlated local stochastic noise. A probability upper bound alone also allows the identity channel; it cannot imply nonzero leakage. This is why the lower and upper bounds have different noise assumptions.
The encoding and its overlap kernel
The preparation filter commutes with complementary logical X X X operators. Their exact eigenvalues keep distinct logical basis states orthogonal, although the measured Z Z Z checks are no longer sharp. Consequently every X X X -error overlap reduces to a logical sign times one scalar kernel.
Supplementary Lemma 2.1 (Encoding isometry and logical matrix elements). Let ℓ 1 , … , ℓ k \ell_1,\ldots,\ell_k ℓ 1 , … , ℓ k represent Z Z Z logical operators that are independent modulo the Z Z Z -stabilizer space. The map
V β ∣ q ⟩ L = ∣ q β ⟩ : = Z ( ℓ ( q ) ) ∣ ψ β ⟩ , ℓ ( q ) = ⨁ a = 1 k q a ℓ a , q ∈ F 2 k (120) V_\beta|q\rangle_L=|q_\beta\rangle:=Z^{(\ell(q))}|\psi_\beta\rangle,\qquad\ell(q)=\bigoplus_{a=1}^{k}q_a\ell_a,\qquad q\in\mathbb{F}_2^k \tag*{(120)} V β ∣ q ⟩ L = ∣ q β ⟩ := Z ( ℓ ( q )) ∣ ψ β ⟩ , ℓ ( q ) = a = 1 ⨁ k q a ℓ a , q ∈ F 2 k (120) is an isometry from a register of dimension K = 2 k K=2^k K = 2 k . Applying the preparation filter to V 0 V_0 V 0 has success probability independent of the logical input, including coherent superpositions and inputs entangled with a reference. For every physical X X X error,
⟨ q β ∣ X x ∣ q β ′ ⟩ = δ q q ′ χ q ( x ) κ β ( x ) , κ β ( x ) = ⟨ ψ β ∣ X x ∣ ψ β ⟩ . (121) \langle q_\beta|X_x|q'_\beta\rangle=\delta_{qq'}\chi_q(x)\kappa_\beta(x),\qquad\kappa_\beta(x)=\langle\psi_\beta|X_x|\psi_\beta\rangle. \tag*{(121)} ⟨ q β ∣ X x ∣ q β ′ ⟩ = δ q q ′ χ q ( x ) κ β ( x ) , κ β ( x ) = ⟨ ψ β ∣ X x ∣ ψ β ⟩ . (121)
Proof. Choose complementary logical operators X ‾ a \overline{X}_a X a whose binary commutation pairing with the selected Z Z Z logicals is δ a b \delta_{ab} δ ab . Each X a X_a X a commutes with all measured checks and stabilizes ∣ + ⟩ ⊗ n |+\rangle^{\otimes n} ∣ + ⟩ ⊗ n . It therefore has eigenvalue ( − 1 ) q a (-1)^{q_a} ( − 1 ) q a on ∣ q ⟩ |q\rangle ∣ q ⟩ . Different logical labels belong to orthogonal joint eigenspaces. All codewords have the same norm because Z ℓ ( q ) Z^{\ell(q)} Z ℓ ( q ) is unitary, proving the isometry claim. Since X x X_x X x commutes with every X ‾ a \overline{X}_a X a , it cannot connect different logical labels. Commuting it through Z ℓ ( q ) Z^{\ell(q)} Z ℓ ( q ) gives the diagonal matrix element in Eq. (121) .
To check the normalization of postselection, define
F β = exp ( 1 2 ∑ j β j B j ) , V 0 ∣ q ⟩ L = Z ℓ ( q ) ∣ + ⟩ ⊗ n , Z β = ⟨ + ∣ ⊗ n F β † F β ∣ + ⟩ ⊗ n . (122) F_\beta= \exp\left(\frac{1}{2}\sum_j \beta_j B_j\right), \qquad V_0|q\rangle_L = Z^{\ell(q)}|+\rangle^{\otimes n}, \qquad Z_\beta= \langle+|^{\otimes n}F_\beta^\dagger F_\beta|+\rangle^{\otimes n}. \tag*{(122)} F β = exp ( 2 1 j ∑ β j B j ) , V 0 ∣ q ⟩ L = Z ℓ ( q ) ∣ + ⟩ ⊗ n , Z β = ⟨ + ∣ ⊗ n F β † F β ∣ + ⟩ ⊗ n . (122)
Expanding F β † F β F_\beta^\dagger F_\beta F β † F β into products of checks gives
V 0 † F β † F β V 0 = Z β I K , V β = Z β − 1 / 2 F β V 0 . (123) V_0^\dagger F_\beta^\dagger F_\beta V_0 = Z_\beta I_K, \qquad V_\beta= Z_\beta^{-1/2}F_\beta V_0. \tag*{(123)} V 0 † F β † F β V 0 = Z β I K , V β = Z β − 1/2 F β V 0 . (123)
Every diagonal entry is Z β Z_\beta Z β . For q ≠ q ′ q \ne q' q = q ′ , the off-diagonal entry vanishes: ℓ ( q ⊕ q ′ ) \ell(q \oplus q') ℓ ( q ⊕ q ′ ) is a nontrivial logical class, so no product of Z Z Z checks can cancel it inside an expectation in ∣ + ⟩ ⊗ n |+\rangle^{\otimes n} ∣ + ⟩ ⊗ n . Consequently a physically normalized Kraus operator c F β cF_\beta c F β succeeds with probability ∣ c ∣ 2 Z β |c|^2Z_\beta ∣ c ∣ 2 Z β on every logical density operator. Equation (S45), tensored with the identity on a reference, also proves the assertion for entangled inputs. □
The choice of representatives specifies the encoding map itself. For example, when two representatives of one ideal logical class use the same ∣ ψ β ⟩ |\psi_\beta\rangle ∣ ψ β ⟩ as their logical zero state, their encoding maps obey
V 1 † V 2 = diag ( 1 , ⟨ Z ℓ 1 Z ℓ 2 ⟩ β ) . (124) V_1^\dagger V_2 = \operatorname{diag}\left(1,\langle Z^{\ell_1}Z^{\ell_2}\rangle_\beta\right). \tag*{(124)} V 1 † V 2 = diag ( 1 , ⟨ Z ℓ 1 Z ℓ 2 ⟩ β ) . (124)
Both maps act on the same logical input and have the same full physical output space, but their image subspaces need not coincide at finite strength. Recovery is therefore specified relative to a chosen encoding map.
Supplementary Lemma 2.2 (Exact overlap kernel). Let R Z = ker H Z T R_Z = \ker H_Z^T R Z = ker H Z T be the space of check redundancies and put u j = tanh β j u_j = \tanh\beta_j u j = tanh β j . Then
κ β ( x ) = ( ∏ j ∈ σ ( x ) sech β j ) ∑ a ∈ R Z supp ( a ) ∩ σ ( x ) = ∅ ∏ j : a j = 1 u j ∑ a ∈ R Z ∏ j : a j = 1 u j , (125) \kappa_\beta(x) =
\left(\prod_{j\in\sigma(x)} \operatorname{sech}\beta_j\right)
\frac{\sum_{\substack{a\in R_Z\\ \operatorname{supp}(a)\cap\sigma(x)=\varnothing}}\prod_{j:a_j=1}u_j}
{\sum_{a\in R_Z}\prod_{j:a_j=1}u_j}, \tag*{(125)} κ β ( x ) = j ∈ σ ( x ) ∏ sech β j ∑ a ∈ R Z ∏ j : a j = 1 u j ∑ a ∈ R Z supp ( a ) ∩ σ ( x ) = ∅ ∏ j : a j = 1 u j , (125) and
exp ( − ∑ j ∈ σ ( x ) β j ) ≤ κ β ( x ) ≤ ∏ j ∈ σ ( x ) sech β j . (126) \exp\left(-\sum_{j\in\sigma(x)}\beta_j\right)\le\kappa_\beta(x)\le\prod_{j\in\sigma(x)}\operatorname{sech}\beta_j. \tag*{(126)} exp − j ∈ σ ( x ) ∑ β j ≤ κ β ( x ) ≤ j ∈ σ ( x ) ∏ sech β j . (126) For a single edge in the standard periodic toric codes of side length L ≥ 3 L\ge3 L ≥ 3 , uniform preparation gives
κ β 2 D ( e ) = sech 2 β 1 + ( tanh β ) L 2 , e − 4 β ≤ κ β 3 D ( e ) ≤ sech 4 β . (127) \kappa_\beta^{2\mathrm{D}}(e)=\frac{\operatorname{sech}^2\beta}{1+(\tanh\beta)^{L^2}}, \qquad e^{-4\beta}\le\kappa_\beta^{3\mathrm{D}}(e)\le\operatorname{sech}^4\beta. \tag*{(127)} κ β 2 D ( e ) = 1 + ( tanh β ) L 2 sech 2 β , e − 4 β ≤ κ β 3 D ( e ) ≤ sech 4 β . (127) Proof. Write ∣ + ⟩ = ∣ + ⟩ ⊗ n |+\rangle=|+\rangle^{\otimes n} ∣ + ⟩ = ∣ + ⟩ ⊗ n for this proof. Moving X x X_x X x through the preparation filter reverses the signs of precisely the checks in σ ( x ) \sigma(x) σ ( x ) . Those checks cancel between the two half-strength filters; the others add. Since X x ∣ + ⟩ = ∣ + ⟩ X_x|+\rangle=|+\rangle X x ∣ + ⟩ = ∣ + ⟩ ,
κ β ( x ) = ⟨ + ∣ exp ( ∑ j ∉ σ ( x ) β j B j ) ∣ + ⟩ ⟨ + ∣ exp ( ∑ j β j B j ) ∣ + ⟩ . (128) \kappa_\beta(x)=\frac{\langle+|\exp\left(\sum_{j\notin\sigma(x)}\beta_jB_j\right)|+\rangle}{\langle+|\exp\left(\sum_j\beta_jB_j\right)|+\rangle}. \tag*{(128)} κ β ( x ) = ⟨ + ∣ exp ( ∑ j β j B j ) ∣ + ⟩ ⟨ + ∣ exp ( ∑ j ∈ / σ ( x ) β j B j ) ∣ + ⟩ . (128) In the expansion e β j B j = cosh β j ( 1 + u j B j ) e^{\beta_jB_j}=\cosh\beta_j(1+u_jB_j) e β j B j = cosh β j ( 1 + u j B j ) , a product of checks has nonzero expectation exactly when its binary indicator belongs to ker H Z T \ker H_Z^T ker H Z T . This proves Eq. (125) . All redundancy weights are nonnegative, and restricting the redundancy sum can only decrease it, proving the upper bound.
For the lower bound, simultaneous diagonalization of the commuting checks gives
exp ( ∑ j β j B j ) ≤ exp ( ∑ j ∈ σ ( x ) β j ) exp ( ∑ j ∉ σ ( x ) β j B j ) . (129) \exp\left(\sum_j\beta_jB_j\right)\le\exp\left(\sum_{j\in\sigma(x)}\beta_j\right)\exp\left(\sum_{j\notin\sigma(x)}\beta_jB_j\right). \tag*{(129)} exp ( j ∑ β j B j ) ≤ exp j ∈ σ ( x ) ∑ β j exp j ∈ / σ ( x ) ∑ β j B j . (129) Taking the ∣ + ⟩ |+\rangle ∣ + ⟩ expectation and using Eq. (128) proves the lower bound. In two dimensions the plaquette redundancies are the empty set and the entire torus. Removing the two plaquettes incident on an edge leaves only the empty redundancy. In three dimensions an edge meets four plaquettes, so Eq. (126) directly gives Eq. (127) . □
Standard subsystem variance
A local observer can detect imperfect encoding when the reduced state depends on the logical input. Subsystem variance measures the largest such dependence relative to the code average; it is a property of the encoding before a noise channel is specified.
Subsystem variance uses the reduced maximally mixed code state as its comparator [27 ] . For a fixed physical region A A A , the definition in the main text can equivalently be written
ε S V ( A ) = max ρ L ∥ Tr A c ( V β ρ L V β † ) − Tr A c ( P β / K ) ∥ 1 . (130) \varepsilon_{\mathrm{SV}}(A)=\max_{\rho_L}\left\lVert\operatorname{Tr}_{A^c}\left(V_\beta\rho_LV_\beta^\dagger\right)-\operatorname{Tr}_{A^c}\left(P_\beta/K\right)\right\rVert_1. \tag*{(130)} ε SV ( A ) = ρ L max Tr A c ( V β ρ L V β † ) − Tr A c ( P β / K ) 1 . (130)
Convexity makes maximization over pure inputs equivalent to maximization over density matrices. Here P β = V β V β † P_\beta=V_\beta V_\beta^\dagger P β = V β V β † is the code projector, as in the main text. The convention uses the full trace norm, without 1 / 2 1/2 1/2 . Maximizing this quantity over the connected regions of a prescribed maximum size gives the subsystem-variance parameter of Ref. [27 ] . At fixed A A A , it measures the input dependence of that region, including dependence on logical superpositions. If it vanishes exactly, the channel to A A A is constant. Since this channel is complementary to erasure of A A A , exact correction from A c A^c A c follows from Eq. (114) . Approximate bounds must retain the specified noise channel: erasure and stochastic X X X noise sample the encoded state differently.
Supplementary Lemma 2.3 (Subsystem variance of a physical site). For a site e e e with g ( e ) ≠ 0 g(e)\ne0 g ( e ) = 0 , the finite-strength CSS encoding satisfies
ε S V ( { e } ) ≥ κ β ( e ) . (131) \varepsilon_{\mathrm{SV}}(\{e\})\geq\kappa_\beta(e). \tag*{(131)} ε SV ({ e }) ≥ κ β ( e ) . (131) For the edge-qubit toric codes, equality holds for every nonnegative strength profile. In particular, at uniform finite strength,
ε S V 2 D ( { e } ) = sech 2 β 1 + tanh 2 β ≥ e − 2 β , (132) \varepsilon_{\mathrm{SV}}^{2\mathrm{D}}(\{e\})=\frac{\operatorname{sech}^{2}\beta}{1+\tanh^{2}\beta}\geq e^{-2\beta}, \tag*{(132)} ε SV 2 D ({ e }) = 1 + tanh 2 β sech 2 β ≥ e − 2 β , (132) e − 4 β ≤ ε S V 3 D ( { e } ) = κ β 3 D ( e ) ≤ sech 4 β . (133) e^{-4\beta}\leq\varepsilon_{\mathrm{SV}}^{3\mathrm{D}}(\{e\})=\kappa_\beta^{3\mathrm{D}}(e)\leq\operatorname{sech}^{4}\beta. \tag*{(133)} e − 4 β ≤ ε SV 3 D ({ e }) = κ β 3 D ( e ) ≤ sech 4 β . (133) For a toric edge with g ( e ) = 0 g(e)=0 g ( e ) = 0 , its subsystem variance is zero.
Proof. The character χ q ( e ) \chi_q(e) χ q ( e ) takes both signs equally often when g ( e ) ≠ 0 g(e)\ne0 g ( e ) = 0 . Equation (121) consequently gives
V β † X e V β = κ β ( e ) ∑ q χ q ( e ) ∣ q ⟩ L ⟨ q ∣ L , Tr ( X e P β / K ) = 0. (134) V_\beta^\dagger X_eV_\beta=\kappa_\beta(e)\sum_q\chi_q(e)|q\rangle_L\langle q|_L,\qquad\operatorname{Tr}(X_eP_\beta/K)=0. \tag*{(134)} V β † X e V β = κ β ( e ) q ∑ χ q ( e ) ∣ q ⟩ L ⟨ q ∣ L , Tr ( X e P β / K ) = 0. (134)
Choose a logical basis state with χ q ( e ) = 1 \chi_q(e)=1 χ q ( e ) = 1 . The expectation of X e X_e X e in the difference entering Eq. (130) is κ β ( e ) \kappa_\beta(e) κ β ( e ) . Since ∥ X e ∥ ∞ = 1 \lVert X_e\rVert_\infty=1 ∥ X e ∥ ∞ = 1 , trace-norm duality proves the general CSS lower bound.
For toric codes, every codeword is stabilized by every X X X star. A star incident on e e e anticommutes with Y e Y_e Y e and Z e Z_e Z e , so their compressions vanish on the entire code subspace:
V β † Y e V β = V β † Z e V β = 0. (135) V_\beta^\dagger Y_eV_\beta=V_\beta^\dagger Z_eV_\beta=0. \tag*{(135)} V β † Y e V β = V β † Z e V β = 0. (135)
This statement holds for arbitrary superpositions, not only for basis codewords. For any logical density matrix it gives
Tr { e } c ( V β ρ L V β † ) = 1 2 [ I + κ β ( e ) ∑ q ( ρ L ) q q χ q ( e ) X e ] . (136) \operatorname{Tr}_{\{e\}^c}\left(V_\beta\rho_LV_\beta^\dagger\right)=\frac{1}{2}\left[I+\kappa_\beta(e)\sum_q(\rho_L)_{qq}\chi_q(e)X_e\right]. \tag*{(136)} Tr { e } c ( V β ρ L V β † ) = 2 1 [ I + κ β ( e ) q ∑ ( ρ L ) q q χ q ( e ) X e ] . (136)
The code-average marginal is I / 2 I/2 I /2 when g ( e ) ≠ 0 g(e)\ne0 g ( e ) = 0 , and ∑ q ( ρ L ) q q χ q ( e ) \sum_q(\rho_L)_{qq}\chi_q(e) ∑ q ( ρ L ) q q χ q ( e ) ranges over [ − 1 , 1 ] [-1,1] [ − 1 , 1 ] . The trace norm of the difference from I / 2 I/2 I /2 is therefore at most κ β ( e ) \kappa_\beta(e) κ β ( e ) , with equality at either endpoint. Equation (127) gives the displayed uniform-strength results. If g ( e ) = 0 g(e)=0 g ( e ) = 0 , the sum is always one, so every input has the same reduced state and the variance vanishes. □
The complementary channel and replica identities
The complementary channel determines how the local input dependence enters the environment. For stochastic X X X noise, its matrix elements are given directly by the same overlap kernel. This form also yields exact replica partition functions at finite size. The construction below extends the toric-code replica calculation of Ref. [4, Sec. V and Appendix C] to the present CSS register and nonuniform strengths.
For stochastic X X X noise with distribution P ( x ) P(x) P ( x ) , use the Stinespring isometry
∣ ϕ ⟩ Q ⟼ ∑ x P ( x ) X x ∣ ϕ ⟩ Q ∣ x ⟩ E . (137) |\phi\rangle_Q \longmapsto\sum_x \sqrt{P(x)}X_x|\phi\rangle_Q|x\rangle_E. \tag*{(137)} ∣ ϕ ⟩ Q ⟼ x ∑ P ( x ) X x ∣ ϕ ⟩ Q ∣ x ⟩ E . (137)
Encoding half of the maximally entangled state of a K K K -dimensional register gives
∣ Ω ⟩ = 1 K ∑ q , x P ( x ) ∣ q ⟩ R X x ∣ q β ⟩ Q ∣ x ⟩ E . (138) |\Omega\rangle= \frac{1}{\sqrt{K}}\sum_{q,x}\sqrt{P(x)}|q\rangle_R X_x|q_\beta\rangle_Q|x\rangle_E. \tag*{(138)} ∣Ω ⟩ = K 1 q , x ∑ P ( x ) ∣ q ⟩ R X x ∣ q β ⟩ Q ∣ x ⟩ E . (138)
Tracing out the physical output Q Q Q and using Eq. (121) makes ρ R E \rho_{RE} ρ R E classical on R R R in the q q q basis. We use the normalized AQEC matrices and entropy from Eqs. (115) and (116) . Their entries, with the logical indices last, are
( Λ + B ) x y , q q ′ = δ q q ′ K P ( x ) P ( y ) χ q ( x ⊕ y ) κ β ( x ⊕ y ) , (139) (\Lambda+B)_{xy,qq'}=\frac{\delta_{qq'}}{K}\sqrt{P(x)P(y)}\chi_q(x\oplus y)\kappa_\beta(x\oplus y), \tag*{(139)} ( Λ + B ) x y , q q ′ = K δ q q ′ P ( x ) P ( y ) χ q ( x ⊕ y ) κ β ( x ⊕ y ) , (139)
Λ x y , q q ′ = δ q q ′ K P ( x ) P ( y ) κ β ( x ⊕ y ) 1 { g ( x ⊕ y ) = 0 } . (140) \Lambda_{xy,qq'}=\frac{\delta_{qq'}}{K}\sqrt{P(x)P(y)}\kappa_\beta(x\oplus y)\mathbf{1}_{\{g(x\oplus y)=0\}}. \tag*{(140)} Λ x y , q q ′ = K δ q q ′ P ( x ) P ( y ) κ β ( x ⊕ y ) 1 { g ( x ⊕ y ) = 0 } . (140)
The second line uses orthogonality of the characters under uniform q q q . In particular Tr R B = 0 \operatorname{Tr}_R B=0 Tr R B = 0 . Equivalently,
ρ R E = 1 K ∑ q ∣ q ⟩ R ⟨ q ∣ R ⊗ ρ E q , ρ E = 1 K ∑ q ρ E q , D A Q E C = S ( ρ E ) − 1 K ∑ q S ( ρ E q ) . (141) \rho_{RE}=\frac{1}{K}\sum_q|q\rangle_R\langle q|_R\otimes\rho_E^q,\qquad\rho_E=\frac{1}{K}\sum_q\rho_E^q,\qquad D_{\mathrm{AQEC}}=S(\rho_E)-\frac{1}{K}\sum_qS(\rho_E^q). \tag*{(141)} ρ R E = K 1 q ∑ ∣ q ⟩ R ⟨ q ∣ R ⊗ ρ E q , ρ E = K 1 q ∑ ρ E q , D AQEC = S ( ρ E ) − K 1 q ∑ S ( ρ E q ) . (141)
This is the Holevo information of the conditional environment ensemble [28 ] , hence D A Q E C ≤ log K D_{\mathrm{AQEC}}\le\log K D AQEC ≤ log K for this model. The error-label diagonal is P ( x ) P(x) P ( x ) for every q q q , and all logical dependence resides in the off-diagonal entries. Classicality on R R R does not make the recovery task classical: different ρ E q \rho_E^q ρ E q suppress coherence between the corresponding physical logical sectors. Although the physical channel is a stochastic mixture of Pauli errors, its complementary output retains coherences between error labels. Dephasing those labels would discard the very information measured by these bounds.
Supplementary Lemma 2.4 (Replica identities at finite size). For an integer m ≥ 2 m\ge2 m ≥ 2 , put x ( m + 1 ) = x ( 1 ) x^{(m+1)}=x^{(1)} x ( m + 1 ) = x ( 1 ) and d α = x ( α ) ⊕ x ( α + 1 ) d_\alpha=x^{(\alpha)}\oplus x^{(\alpha+1)} d α = x ( α ) ⊕ x ( α + 1 ) . Then
K m − 1 Tr ( Λ + B ) m = ∑ x ( 1 ) , … , x ( m ) ∏ α = 1 m P ( x ( α ) ) κ β ( d α ) , (142) K^{m-1}\operatorname{Tr}(\Lambda+B)^m=\sum_{x^{(1)},\ldots,x^{(m)}}\prod_{\alpha=1}^{m}P(x^{(\alpha)})\kappa_\beta(d_\alpha), \tag*{(142)} K m − 1 Tr ( Λ + B ) m = x ( 1 ) , … , x ( m ) ∑ α = 1 ∏ m P ( x ( α ) ) κ β ( d α ) , (142) K m − 1 Tr Λ m = ∑ x ( 1 ) , … , x ( m ) ∏ α = 1 m P ( x ( α ) ) κ β ( d α ) 1 { g ( d α ) = 0 } . (143) K^{m-1}\operatorname{Tr}\Lambda^m=\sum_{x^{(1)},\ldots,x^{(m)}}\prod_{\alpha=1}^{m}P(x^{(\alpha)})\kappa_\beta(d_\alpha)\mathbf{1}_{\{g(d_\alpha)=0\}}. \tag*{(143)} K m − 1 Tr Λ m = x ( 1 ) , … , x ( m ) ∑ α = 1 ∏ m P ( x ( α ) ) κ β ( d α ) 1 { g ( d α ) = 0 } . (143) At every finite system size,
D A Q E C = d d m log Tr ( Λ + B ) m Tr Λ m ∣ m = 1 . (144) D_{\mathrm{AQEC}}=\left.\frac{d}{dm}\log\frac{\operatorname{Tr}(\Lambda+B)^m}{\operatorname{Tr}\Lambda^m}\right|_{m=1}. \tag*{(144)} D AQEC = d m d log Tr Λ m Tr ( Λ + B ) m m = 1 . (144)
Proof. Multiply the matrix blocks around the replica cycle. Their character product equals one because ⨁ α d α = 0 \bigoplus_\alpha d_\alpha=0 ⨁ α d α = 0 . The logical sum then contributes K K K , and every error probability appears as two square roots, giving one P ( x ( α ) ) P(x^{(\alpha)}) P ( x ( α ) ) per replica. This proves the integer identities.
For real m m m near one, define the powers through the spectra of the finite matrices. Since both matrices have trace one, the derivative in Eq. (144) is Tr [ ( Λ + B ) log ( Λ + B ) ] − Tr ( Λ log Λ ) \operatorname{Tr}[(\Lambda+B)\log(\Lambda+B)]-\operatorname{Tr}(\Lambda\log\Lambda) Tr [( Λ + B ) log ( Λ + B )] − Tr ( Λ log Λ ) . The support of ρ R E \rho_{RE} ρ R E lies in H R ⊗ supp ρ E \mathcal{H}_R\otimes\operatorname{supp}\rho_E H R ⊗ supp ρ E . On this support, log Λ = − ( log K ) I + I R ⊗ log ρ E T \log\Lambda=-(\log K)I+I_R\otimes\log\rho_E^T log Λ = − ( log K ) I + I R ⊗ log ρ E T . Both Tr B = 0 \operatorname{Tr}B=0 Tr B = 0 and Tr R B = 0 \operatorname{Tr}_R B=0 Tr R B = 0 , so Tr ( B log Λ ) = 0 \operatorname{Tr}(B\log\Lambda)=0 Tr ( B log Λ ) = 0 . The derivative is therefore exactly the relative entropy in Eq. (116) . This spectral definition supplies the real-m m m derivative; the integer replica sums by themselves do not justify an analytic continuation after a thermodynamic or dilute-noise limit. □
The replica sums have a local classical-spin representation. For z = ( z 1 , … , z n ) ∈ { + 1 } n z=(z_1,\ldots,z_n)\in\{+1\}^n z = ( z 1 , … , z n ) ∈ { + 1 } n , let b j ( z ) = ∏ i : ( H Z ) j i = 1 z i b_j(z)=\prod_{i:(H_Z)_{ji}=1}z_i b j ( z ) = ∏ i : ( H Z ) j i = 1 z i be the classical value of check j j j . The normalization in Eq. (122) is Z β = 2 − n ∑ z exp [ ∑ j β j b j ( z ) ] Z_\beta=2^{-n}\sum_z\exp[\sum_j\beta_jb_j(z)] Z β = 2 − n ∑ z exp [ ∑ j β j b j ( z )] . Since a check survives in Eq. (128) precisely when ( H Z d α ) j = 0 (H_Zd_\alpha)_j=0 ( H Z d α ) j = 0 , Eq. (142) becomes
K m − 1 Tr ( Λ + B ) m = 1 ( 2 n Z β ) m ∑ { x ( α ) , z ( α ) } ∏ α = 1 m P ( x ( α ) ) K^{m-1}\operatorname{Tr}(\Lambda+B)^m=\frac{1}{(2^nZ_\beta)^m}\sum_{\{x^{(\alpha)},z^{(\alpha)}\}}\prod_{\alpha=1}^{m}P(x^{(\alpha)}) K m − 1 Tr ( Λ + B ) m = ( 2 n Z β ) m 1 { x ( α ) , z ( α ) } ∑ α = 1 ∏ m P ( x ( α ) )
× exp { ∑ α , j β j 2 [ 1 + ( − 1 ) ( H Z d α ) j ] b j ( z ( α ) ) } . (145) {}\times\exp\left\{\sum_{\alpha,j}\frac{\beta_j}{2}\left[1+(-1)^{(H_Zd_\alpha)_j}\right]b_j(z^{(\alpha)})\right\}. \tag*{(145)} × exp { α , j ∑ 2 β j [ 1 + ( − 1 ) ( H Z d α ) j ] b j ( z ( α ) ) } . (145)
The expression for K m − 1 Tr Λ m K^{m-1}\operatorname{Tr}\Lambda^m K m − 1 Tr Λ m inserts ∏ α 1 { g ( d α ) = 0 } \prod_\alpha\mathbf{1}_{\{g(d_\alpha)=0\}} ∏ α 1 { g ( d α ) = 0 } into the same sum. For independent noise, P ( x ( α ) ) P(x^{(\alpha)}) P ( x ( α ) ) also factors over physical sites. Thus Eq. (145) is a classical partition function with local check interactions between adjacent replicas. In toric codes, the variables z i z_i z i live on edges and b j b_j b j is a plaquette product, giving the gauge-theory form. This is the direct extension of the replica mapping in Ref. [4] to the check structure and nonuniform strengths considered here. The two partition functions share all local weights; the product comparator additionally forbids replica differences with nonzero logical parity.
The simplest such contribution has a single flip e e e in one replica and no errors in the others. The two adjacent error differences then both equal e e e , giving the weight P ( e ) P ( 0 ) m − 1 κ β ( e ) 2 P(e)P(0)^{m-1}\kappa_\beta(e)^2 P ( e ) P ( 0 ) m − 1 κ β ( e ) 2 in Eq. (142) . If g ( e ) ≠ 0 g(e)\ne0 g ( e ) = 0 , this configuration is excluded from Eq. (143) . Such terms occur at every site in ⋃ a S ℓ a \bigcup_a \mathcal{S}_{\ell_a} ⋃ a S ℓ a , so the replica difference includes contributions from local errors along the entire selected support. The single-flag calculation below determines the local dilute coefficient while retaining the terms needed for the m → 1 m\to1 m → 1 limit. Accumulation at fixed nonzero noise strength is established by the subsequent collective bounds, with correlations between different sites included.
A local witness and the dilute coefficient
A lower bound does not require diagonalizing the environment state. It is enough to identify an observable whose mean depends on the logical input. The following witness compares error branches differing by a fixed support; its expectation separates the probability supplied by the noise from the overlap left by finite-strength preparation.
Supplementary Lemma 2.5 (A norm-one witness for leakage). If g ( u ) ≠ 0 g(u)\ne0 g ( u ) = 0 , define the Hellinger overlap of an error distribution with its translate [29 ] by
H P ( u ) = ∑ x P ( x ) P ( x ⊕ u ) . (146) H_P(u)=\sum_x\sqrt{P(x)P(x\mathbin{\oplus}u)}. \tag*{(146)} H P ( u ) = x ∑ P ( x ) P ( x ⊕ u ) . (146) Then
∥ B ∥ 1 ≥ H P ( u ) κ β ( u ) , D A Q E C ≥ 1 2 H P ( u ) 2 κ β ( u ) 2 . (147) \lVert B\rVert_1\ge H_P(u)\kappa_\beta(u),\qquad D_{\mathrm{AQEC}}\ge\frac{1}{2}H_P(u)^2\kappa_\beta(u)^2. \tag*{(147)} ∥ B ∥ 1 ≥ H P ( u ) κ β ( u ) , D AQEC ≥ 2 1 H P ( u ) 2 κ β ( u ) 2 . (147) For independent bit flips of probability p p p , H P ( u ) = [ 2 p ( 1 − p ) ] ∣ u ∣ H_P(u)=[2\sqrt{p(1-p)}]^{|u|} H P ( u ) = [ 2 p ( 1 − p ) ] ∣ u ∣ .
Proof. On the matrix-label space, the operator defined by W u ( ∣ q ⟩ R ∣ x ⟩ E ) = χ q ( u ) ∣ q ⟩ R ∣ x ⊕ u ⟩ E W_u(|q\rangle_R|x\rangle_E)=\chi_q(u)|q\rangle_R|x\mathbin{\oplus}u\rangle_E W u ( ∣ q ⟩ R ∣ x ⟩ E ) = χ q ( u ) ∣ q ⟩ R ∣ x ⊕ u ⟩ E is unitary and has operator norm one. Since g ( u ) ≠ 0 g(u)\ne0 g ( u ) = 0 , the corresponding entries of Λ \Lambda Λ vanish. Equation (S60) gives
Tr ( W u B ) = 1 K ∑ q , x χ q ( u ) 2 P ( x ) P ( x ⊕ u ) κ β ( u ) = H P ( u ) κ β ( u ) . (148) \operatorname{Tr}(W_uB)=\frac{1}{K}\sum_{q,x}\chi_q(u)^2\sqrt{P(x)P(x\mathbin{\oplus}u)}\kappa_\beta(u)=H_P(u)\kappa_\beta(u). \tag*{(148)} Tr ( W u B ) = K 1 q , x ∑ χ q ( u ) 2 P ( x ) P ( x ⊕ u ) κ β ( u ) = H P ( u ) κ β ( u ) . (148)
Trace-norm duality [21 Sec. 1.1.3] gives the first bound, and quantum Pinsker with natural logarithms [21 Theorem 5.38] gives D A Q E C ≥ ∥ B ∥ 1 2 / 2 D_{\mathrm{AQEC}}\ge\lVert B\rVert_1^2/2 D AQEC ≥ ∥ B ∥ 1 2 /2 . For independent noise, the overlap factors over the sites of u u u ; sites outside u u u contribute one.
For a toric edge, Supplementary Lemmas 2.3 and 2.5 give the direct connection to the complete noise channel:
∥ B ∥ 1 ≥ 2 p ( 1 − p ) ϵ S V ( { e } ) , D A Q E C ≥ 2 p ( 1 − p ) ϵ S V ( { e } ) 2 . (149) \lVert B\rVert_1\ge2\sqrt{p(1-p)}\epsilon_{\mathrm{SV}}(\{e\}),\qquad D_{\mathrm{AQEC}}\ge2p(1-p)\epsilon_{\mathrm{SV}}(\{e\})^2. \tag*{(149)} ∥ B ∥ 1 ≥ 2 p ( 1 − p ) ϵ SV ({ e }) , D AQEC ≥ 2 p ( 1 − p ) ϵ SV ({ e } ) 2 . (149)
For a general CSS encoding the common lower bounds remain those in terms of κ β ( e ) \kappa_\beta(e) κ β ( e ) ; an upper bound on subsystem variance would be needed to replace κ β ( e ) \kappa_\beta(e) κ β ( e ) by the variance itself in that more general case. For bounded check incidence and uniformly bounded strengths, any site with g ( e ) ≠ 0 g(e)\ne0 g ( e ) = 0 has a positive, size-independent kernel lower bound. Equations (S48) and (S68) then give positive nontrivial error defects at every fixed 0 < p < 1 0<p<1 0 < p < 1 . The norm defect obstructs recovery directly through Supplementary Lemma 2.9, even when the number of selected logical qubits grows.
An exact environmental calculation makes the dilute expansion transparent. For g ( e ) ≠ 0 g(e)\ne0 g ( e ) = 0 , let S = g ( e ) ⋅ q S=g(e)\cdot q S = g ( e ) ⋅ q be the corresponding uniformly random reference parity and retain the environment flag E e E_e E e . This is a partial trace of the environment of the original full noise channel. In the error-flag basis { ∣ 0 ⟩ E e , ∣ 1 ⟩ E e } \{|0\rangle_{E_e},|1\rangle_{E_e}\} { ∣0 ⟩ E e , ∣1 ⟩ E e } , the conditional density matrices are
ρ E e ± = ( 1 − p ± p ( 1 − p ) κ β ( e ) ± p ( 1 − p ) κ β ( e ) p ) . (150) \rho_{E_e}^{\pm}=
\begin{pmatrix}
1-p & \pm\sqrt{p(1-p)}\kappa_\beta(e)\\
\pm\sqrt{p(1-p)}\kappa_\beta(e) & p
\end{pmatrix}. \tag*{(150)} ρ E e ± = ( 1 − p ± p ( 1 − p ) κ β ( e ) ± p ( 1 − p ) κ β ( e ) p ) . (150)
The superscripts + + + and − - − correspond to S = 0 S=0 S = 0 and S = 1 S=1 S = 1 , respectively; the flag basis records absence or presence of the physical flip and is not a logical basis. Tracing all other flags sets their two error indices equal and sums their probabilities to one. The average is diag ( 1 − p , p ) \operatorname{diag}(1-p,p) diag ( 1 − p , p ) . Diagonalizing these two matrices and applying data processing [21 Theorem 5.35] gives
D A Q E C ≥ I ( S : E e ) = h 2 ( p ) − h 2 ( 1 − ( 1 − 2 p ) 2 + 4 p ( 1 − p ) κ β ( e ) 2 2 ) ≥ 2 p ( 1 − p ) κ β ( e ) 2 , (151) D_{\mathrm{AQEC}} \ge I(S:E_e)=h_2(p)-h_2\left(\frac{1-\sqrt{(1-2p)^2+4p(1-p)\kappa_\beta(e)^2}}{2}\right)\ge2p(1-p)\kappa_\beta(e)^2, \tag*{(151)} D AQEC ≥ I ( S : E e ) = h 2 ( p ) − h 2 ( 2 1 − ( 1 − 2 p ) 2 + 4 p ( 1 − p ) κ β ( e ) 2 ) ≥ 2 p ( 1 − p ) κ β ( e ) 2 , (151)
where h 2 ( z ) = − z log z − ( 1 − z ) log ( 1 − z ) h_2(z)=-z\log z-(1-z)\log(1-z) h 2 ( z ) = − z log z − ( 1 − z ) log ( 1 − z ) . All entropy bounds refer to the same D A Q E C D_{\mathrm{AQEC}} D AQEC of the full channel; I ( S : E e ) I(S:E_e) I ( S : E e ) is only the intermediate mutual information produced by discarding other systems. The final inequality follows from Pinsker applied to each conditional matrix and its average.
This calculation also fixes the local coefficient in a dilute replica expansion. At fixed 0 < p < 1 / 2 0<p<1/2 0 < p < 1/2 , expansion in κ β ( e ) \kappa_\beta(e) κ β ( e ) gives
I ( S : E e ) = p ( 1 − p ) 1 − 2 p log 1 − p p κ β ( e ) 2 + O p ( κ β ( e ) 4 ) . (152) I(S:E_e)=\frac{p(1-p)}{1-2p}\log\frac{1-p}{p}\,\kappa_\beta(e)^2+O_p(\kappa_\beta(e)^4). \tag*{(152)} I ( S : E e ) = 1 − 2 p p ( 1 − p ) log p 1 − p κ β ( e ) 2 + O p ( κ β ( e ) 4 ) . (152)
Indeed, the two conditional eigenvalues are 1 − p + p ( 1 − p ) 1 − 2 p κ β ( e ) 2 + O p ( κ β ( e ) 4 ) 1-p+\frac{p(1-p)}{1-2p}\kappa_\beta(e)^2+O_p(\kappa_\beta(e)^4) 1 − p + 1 − 2 p p ( 1 − p ) κ β ( e ) 2 + O p ( κ β ( e ) 4 ) and p − p ( 1 − p ) 1 − 2 p κ β ( e ) 2 + O p ( κ β ( e ) 4 ) p-\frac{p(1-p)}{1-2p}\kappa_\beta(e)^2+O_p(\kappa_\beta(e)^4) p − 1 − 2 p p ( 1 − p ) κ β ( e ) 2 + O p ( κ β ( e ) 4 ) . The corresponding replica entropy difference is ( m − 1 ) − 1 log { Tr [ ( ρ E e + ) m ] / [ ( 1 − p ) m + p m ] } (m-1)^{-1}\log\{\operatorname{Tr}[(\rho_{E_e}^{+})^m]/[(1-p)^m+p^m]\} ( m − 1 ) − 1 log { Tr [( ρ E e + ) m ] / [( 1 − p ) m + p m ]} . Its coefficient of κ β ( e ) 2 \kappa_\beta(e)^2 κ β ( e ) 2 is
m p ( 1 − p ) ( m − 1 ) ( 1 − 2 p ) ( 1 − p ) m − 1 − p m − 1 ( 1 − p ) m + p m → m → 1 p ( 1 − p ) 1 − 2 p log 1 − p p . (153) \frac{mp(1-p)}{(m-1)(1-2p)}\frac{(1-p)^{m-1}-p^{m-1}}{(1-p)^m+p^m}\xrightarrow{m\to1}\frac{p(1-p)}{1-2p}\log\frac{1-p}{p}. \tag*{(153)} ( m − 1 ) ( 1 − 2 p ) m p ( 1 − p ) ( 1 − p ) m + p m ( 1 − p ) m − 1 − p m − 1 m → 1 1 − 2 p p ( 1 − p ) log p 1 − p . (153)
Both p p p and p m p^m p m occur before the m → 1 m\to1 m → 1 limit. Dropping p m p^m p m at fixed integer m > 1 m>1 m > 1 creates a spurious m / ( m − 1 ) m/(m-1) m / ( m − 1 ) singularity. Keeping both terms gives the finite coefficient proportional to p log ( 1 / p ) p\log(1/p) p log ( 1/ p ) . Equivalently, at fixed overlap,
I ( S : E e ) = p κ β ( e ) 2 log ( 1 / p ) + p { κ β ( e ) 2 + [ 1 − κ β ( e ) 2 ] log [ 1 − κ β ( e ) 2 ] } + O ( p 2 log ( 1 / p ) ) . (154) \begin{aligned}
I(S:E_e)&=p\kappa_\beta(e)^2\log(1/p)\\
&\quad+p\{\kappa_\beta(e)^2+[1-\kappa_\beta(e)^2]\log[1-\kappa_\beta(e)^2]\}+O(p^2\log(1/p)). \tag*{(154)}
\end{aligned} I ( S : E e ) = p κ β ( e ) 2 log ( 1/ p ) + p { κ β ( e ) 2 + [ 1 − κ β ( e ) 2 ] log [ 1 − κ β ( e ) 2 ]} + O ( p 2 log ( 1/ p )) . (154)
We use 0 log 0 = 0 0\log0=0 0 log 0 = 0 . At p = 1 / 2 p=1/2 p = 1/2 , the quadratic coefficient in Eq. (152) is its continuous value 1 / 2 1/2 1/2 . A local expansion of this kind identifies the leakage mechanism but cannot be summed over an increasing number of positions without controlling correlations and information saturation. The following collective bounds perform that step nonperturbatively.
Independent parities and collective anchors
There are two ways for local leakage to accumulate. Different locations may reveal independent logical parities, or they may provide repeated evidence about the same parity. The first allows an additive information bound. The second requires a collective witness that retains correlations and respects the finite information available in each logical degree of freedom.
Supplementary Lemma 2.6 (Independent logical parities). For a physical set A A A , choose e 1 , … , e r ∈ A e_1,\ldots,e_r\in A e 1 , … , e r ∈ A whose columns g ( e a ) g(e_a) g ( e a ) form a basis of the span of { g ( e ) : e ∈ A } \{g(e):e\in A\} { g ( e ) : e ∈ A } . Under independent bit-flip noise,
D A Q E C ≥ 2 p ( 1 − p ) ∑ a = 1 r κ β ( e a ) 2 . (155) D_{\mathrm{AQEC}}\ge2p(1-p)\sum_{a=1}^{r}\kappa_\beta(e_a)^2. \tag*{(155)} D AQEC ≥ 2 p ( 1 − p ) a = 1 ∑ r κ β ( e a ) 2 . (155) If k ≥ 1 k\ge1 k ≥ 1 , every physical qubit meets at most Δ Z \Delta_Z Δ Z measured checks, and 0 ≤ β j ≤ β 0 0\le\beta_j\le\beta_0 0 ≤ β j ≤ β 0 for all j j j , choosing A = [ n ] A=[n] A = [ n ] gives
D A Q E C k ≥ 2 p ( 1 − p ) e − 2 Δ Z β 0 . (156) \frac{D_{\mathrm{AQEC}}}{k}\ge2p(1-p)e^{-2\Delta_Z\beta_0}. \tag*{(156)} k D AQEC ≥ 2 p ( 1 − p ) e − 2 Δ Z β 0 . (156) Proof. The bits S a = g ( e a ) ⋅ q S_a=g(e_a)\cdot q S a = g ( e a ) ⋅ q are independent and uniform because the columns are linearly independent and q q q is uniform. The mutual-information chain rule, data processing, and nonnegativity of conditional mutual information [21 Ch. 5] give
I ( R : E ) ≥ I ( S 1 ⋯ S r : E ) = ∑ a [ I ( S a : E ) + I ( S a : S 1 ⋯ S a − 1 ∣ E ) ] ≥ ∑ a I ( S a : E e a ) . (157) \begin{aligned}
I(R:E)&\ge I(S_1\cdots S_r:E)\\
&=\sum_a[I(S_a:E)+I(S_a:S_1\cdots S_{a-1}\mid E)]\ge\sum_a I(S_a:E_{e_a}). \tag*{(157)}
\end{aligned} I ( R : E ) ≥ I ( S 1 ⋯ S r : E ) = a ∑ [ I ( S a : E ) + I ( S a : S 1 ⋯ S a − 1 ∣ E )] ≥ a ∑ I ( S a : E e a ) . (157)
Here independence removes I ( S a : S 1 ⋯ S a − 1 ) I(S_a : S_1 \cdots S_{a-1}) I ( S a : S 1 ⋯ S a − 1 ) , and the conditional mutual information is nonnegative. Equation (151) bounds each term. The environment flags themselves need not be independent.
The physical binary rows ℓ a \ell_a ℓ a are linearly independent: a relation giving the empty physical support would contradict independence of their logical classes. The full matrix with columns g ( e ) g(e) g ( e ) therefore has rank k k k . For A = [ n ] A=[n] A = [ n ] one may choose r = k r=k r = k , and Eq. (126) gives κ β ( e a ) ≥ exp [ − ∑ j ∈ σ ( e a ) β j ] ≥ e − Δ Z β 0 \kappa_\beta(e_a) \ge\exp[-\sum_{j\in\sigma(e_a)}\beta_j] \ge e^{-\Delta_Z\beta_0} κ β ( e a ) ≥ exp [ − ∑ j ∈ σ ( e a ) β j ] ≥ e − Δ Z β 0 . The same argument applies to any subset whose incident checks obey the strength bound, as used for the enhanced-footprint converse in Eq. (255) .
Supplementary Lemma 2.7 (Many anchors for one logical parity). Suppose M ≥ 1 M \ge1 M ≥ 1 sites e 1 , … , e M e_1,\ldots,e_M e 1 , … , e M have the same nonzero column g ( e i ) = g 0 g(e_i)=g_0 g ( e i ) = g 0 , their syndrome supports are pairwise disjoint, and each syndrome contains s s s checks. Suppose all checks in these syndromes have strength β \beta β ; checks elsewhere may have arbitrary nonnegative strengths. Under independent noise with a = 2 p ( 1 − p ) a=2\sqrt{p(1-p)} a = 2 p ( 1 − p ) ,
D A Q E C ≥ a 2 M e − 2 s β 2 [ 1 + 2 s a 2 M e − 2 s β ] (158) D_{\mathrm{AQEC}} \ge\frac{a^2 M e^{-2s\beta}}{2[1+2s a^2 M e^{-2s\beta}]} \tag*{(158)} D AQEC ≥ 2 [ 1 + 2 s a 2 M e − 2 s β ] a 2 M e − 2 s β (158)
The constants are independent of system size, logical dimension, and preparation strength.
Proof. Measure Pauli X on the selected environment flags, obtaining Y i = ± 1 Y_i=\pm1 Y i = ± 1 , and set S = ( − 1 ) g 0 ⋅ q S=(-1)^{g_0\cdot q} S = ( − 1 ) g 0 ⋅ q . Because g 0 ≠ 0 g_0\ne0 g 0 = 0 and q q q is uniform, the two signs of S S S are equally probable. In this proof, ⟨ ⋅ ⟩ \langle\cdot\rangle ⟨ ⋅ ⟩ denotes a classical average over the joint measurement distribution P ( q , Y ) = K − 1 P ( Y ∣ q ) P(q,Y)=K^{-1}P(Y\mid q) P ( q , Y ) = K − 1 P ( Y ∣ q ) , while ⟨ f ( Y ) ⟩ q = ∑ Y P ( Y ∣ q ) f ( Y ) \langle f(Y)\rangle_q=\sum_Y P(Y\mid q)f(Y) ⟨ f ( Y ) ⟩ q = ∑ Y P ( Y ∣ q ) f ( Y ) fixes the logical label. For any subset A A A of the selected physical sites, direct summation of Eq. (139) yields
⟨ ∏ e i ∈ A Y i ⟩ q = a ∣ A ∣ χ q ( A ) κ β ( A ) . (159) \left\langle\prod_{e_i\in A}Y_i\right\rangle_q=a^{|A|}\chi_q(A)\kappa_\beta(A). \tag*{(159)} ⟨ e i ∈ A ∏ Y i ⟩ q = a ∣ A ∣ χ q ( A ) κ β ( A ) . (159)
Each flipped environment index contributes 2 p ( 1 − p ) 2\sqrt{p(1-p)} 2 p ( 1 − p ) ; every unflipped index contributes one. With T = ∑ i Y i T=\sum_iY_i T = ∑ i Y i ,
⟨ S T ⟩ = a ∑ i κ β ( e i ) ≥ a M e − s β , (160) \langle ST\rangle=a\sum_i\kappa_\beta(e_i)\ge aM e^{-s\beta}, \tag*{(160)} ⟨ S T ⟩ = a i ∑ κ β ( e i ) ≥ a M e − s β , (160)
⟨ T 2 ⟩ = M + a 2 ∑ i ≠ j κ β ( e i ⊕ e j ) ≤ M + a 2 M ( M − 1 ) sech 2 s β . (161) \langle T^2\rangle=M+a^2\sum_{i\ne j}\kappa_\beta(e_i\oplus e_j)\le M+a^2M(M-1)\operatorname{sech}^{2s}\beta. \tag*{(161)} ⟨ T 2 ⟩ = M + a 2 i = j ∑ κ β ( e i ⊕ e j ) ≤ M + a 2 M ( M − 1 ) sech 2 s β . (161)
The pair bound follows from disjoint syndrome supports and Eq. (126) ; it does not factor the joint environment state. The uniform binary prior gives
I ( S : Y ) = ∑ Y P ( Y ) [ log 2 − h 2 ( P ( S = 1 ∣ Y ) ) ] ≥ 1 2 ∑ Y P ( Y ) [ P ( S = 1 ∣ Y ) − P ( S = − 1 ∣ Y ) ] 2 ≥ ⟨ S T ⟩ 2 2 ⟨ T 2 ⟩ . (162) \begin{aligned}
I(S:Y)=\sum_Y P(Y)[\log2-h_2(P(S=1\mid Y))] \\
&\ge\frac{1}{2}\sum_Y P(Y)[P(S=1\mid Y)-P(S=-1\mid Y)]^2 \\
&\ge\frac{\langle ST\rangle^2}{2\langle T^2\rangle}. \tag*{(162)}
\end{aligned} I ( S : Y ) = Y ∑ P ( Y ) [ log 2 − h 2 ( P ( S = 1 ∣ Y ))] ≥ 2 1 Y ∑ P ( Y ) [ P ( S = 1 ∣ Y ) − P ( S = − 1 ∣ Y ) ] 2 ≥ 2 ⟨ T 2 ⟩ ⟨ S T ⟩ 2 . (162)
The first inequality is the binary Pinsker bound [21 Theorem 5.38] . The second is Cauchy–Schwarz applied to ⟨ S T ⟩ = ∑ Y T ( Y ) [ P ( S = 1 , Y ) − P ( S = − 1 , Y ) ] \langle ST\rangle=\sum_Y T(Y)[P(S=1,Y)-P(S=-1,Y)] ⟨ S T ⟩ = ∑ Y T ( Y ) [ P ( S = 1 , Y ) − P ( S = − 1 , Y )] . Data processing gives D A Q E C ≥ I ( S : Y ) D_{\mathrm{AQEC}}\ge I(S:Y) D AQEC ≥ I ( S : Y ) . Substitution of the two moments, followed by sech β ≤ 2 e − β \operatorname{sech}\beta\le2e^{-\beta} sech β ≤ 2 e − β and M − 1 ≤ M M-1\le M M − 1 ≤ M , proves the bound.
For standard straight toric representatives, distinct collinear edges do not share a plaquette. Thus M = L M=L M = L , with s = 2 s=2 s = 2 in two dimensions and s = 4 s=4 s = 4 in three dimensions. The same anchors work when all two or three toric logical qubits are selected; the other coordinate-direction representatives do not use these physical edges. At any fixed 0 < p < 1 0<p<1 0 < p < 1 , Eq. (158) shows that D A Q E C → 0 D_{\mathrm{AQEC}}\to0 D AQEC → 0 requires respectively
β − 1 4 log L ⟶ + ∞ ( 2 D ) , β − 1 8 log L ⟶ + ∞ ( 3 D ) . (163) \beta-\frac{1}{4}\log L\longrightarrow+\infty\quad(2\mathrm{D}),\qquad
\beta-\frac{1}{8}\log L\longrightarrow+\infty\quad(3\mathrm{D}). \tag*{(163)} β − 4 1 log L ⟶ + ∞ ( 2 D ) , β − 8 1 log L ⟶ + ∞ ( 3 D ) . (163)
These are necessary conditions; the coefficients are not asserted to be optimal. The denominator in Eq. (158) is essential: many anchors observe the same binary variable, whose information cannot grow without bound.
The same argument supplies a necessary strength for local enhancement. If each selected edge lies on the enhanced representative and all checks incident on these edges have strength β R \beta_R β R , then both Eq. (160) and Eq. (161) use only β R \beta_R β R . The kernel bounds on these syndromes do not involve checks in the bulk. Taking the same M = L M=L M = L disjoint collinear anchors gives
β R − 1 4 log L ⟶ + ∞ ( 2 D ) , β R − 1 8 log L ⟶ + ∞ ( 3 D ) (164) \beta_R-\frac{1}{4}\log L\longrightarrow+\infty\quad(2\mathrm{D}),\qquad\beta_R-\frac{1}{8}\log L\longrightarrow+\infty\quad(3\mathrm{D}) \tag*{(164)} β R − 4 1 log L ⟶ + ∞ ( 2 D ) , β R − 8 1 log L ⟶ + ∞ ( 3 D ) (164)
as necessary conditions for vanishing total AQEC entropy, even when the bulk strength is kept fixed. Thus the necessary and sufficient local enhancement strengths agree at logarithmic order; no matching of their leading coefficients is asserted.
A concrete example explains why the collective estimate is necessary. For the two-dimensional code, take β ( L ) = 1 4 log L \beta(L)=\frac{1}{4}\log L β ( L ) = 4 1 log L . Equation (S54) gives
max e ε S V ( { e } ) = sech 2 β ( L ) 1 + tanh L 2 β ( L ) ∼ 4 L ⟶ 0. (165) \max_e\varepsilon_{\mathrm{SV}}(\{e\})=\frac{\operatorname{sech}^{2}\beta(L)}{1+\tanh^{L^{2}}\beta(L)}\sim\frac{4}{\sqrt{L}}\longrightarrow0. \tag*{(165)} e max ε SV ({ e }) = 1 + tanh L 2 β ( L ) sech 2 β ( L ) ∼ L 4 ⟶ 0. (165)
Nevertheless L e − 4 β ( L ) = 1 Le^{-4\beta(L)}=1 L e − 4 β ( L ) = 1 , so Equation (S79) gives
D A Q E C ≥ 2 p ( 1 − p ) 1 + 64 p ( 1 − p ) > 0 ( 0 < p < 1 fixed ) . (166) D_{\mathrm{AQEC}}\ge\frac{2p(1-p)}{1+64p(1-p)}>0\qquad(0<p<1\text{ fixed}). \tag*{(166)} D AQEC ≥ 1 + 64 p ( 1 − p ) 2 p ( 1 − p ) > 0 ( 0 < p < 1 fixed ) . (166)
Thus no size-independent upper bound on the full AQEC entropy can be a function only of single-site subsystem variance that vanishes when that variance vanishes. Many individually weak local signals can leave a finite total correlation with the environment. This example applies also at any sufficiently small fixed noise probability, within the regime where recovery upper bounds are relevant.
Supplementary Lemma 2.8 (Many anchors at arbitrary logical dimension). The common-column requirement can be removed for a trace-distance bound. For M ≥ 1 M\ge1 M ≥ 1 sites satisfying g ( e i ) ≠ 0 g(e_i)\ne0 g ( e i ) = 0 , pairwise disjoint syndrome supports, and ∣ σ ( e i ) ∣ ≤ Δ Z |\sigma(e_i)|\le\Delta_Z ∣ σ ( e i ) ∣ ≤ Δ Z , uniform strength and independent noise give
ε = 1 2 ∥ B ∥ 1 ≥ a 2 M e − 2 Δ Z β 4 [ 1 + 2 2 Δ Z a 2 M e − 2 Δ Z β ] , a = 2 p ( 1 − p ) . (167) \varepsilon=\frac{1}{2}\lVert B\rVert_1\ge\frac{a^{2}Me^{-2\Delta_Z\beta}}{4[1+2^{2\Delta_Z}a^{2}Me^{-2\Delta_Z\beta}]},\qquad a=2\sqrt{p(1-p)}. \tag*{(167)} ε = 2 1 ∥ B ∥ 1 ≥ 4 [ 1 + 2 2 Δ Z a 2 M e − 2 Δ Z β ] a 2 M e − 2 Δ Z β , a = 2 p ( 1 − p ) . (167)
Proof. Let P ( q , Y ) P(q,Y) P ( q , Y ) denote the joint classical distribution obtained by measuring the reference in its classical basis and the chosen environment flags in the X X X basis. Its product comparator is Q ( q , Y ) = K − 1 P ( Y ) Q(q,Y)=K^{-1}P(Y) Q ( q , Y ) = K − 1 P ( Y ) , where P ( Y ) = ∑ q P ( q , Y ) P(Y)=\sum_q P(q,Y) P ( Y ) = ∑ q P ( q , Y ) . Use the statistic
F ( q , Y ) = ∑ i χ q ( e i ) Y i , b i = e − β ∣ σ ( e i ) ∣ . (168) F(q,Y)=\sum_i\chi_q(e_i)Y_i,\qquad b_i=e^{-\beta|\sigma(e_i)|}. \tag*{(168)} F ( q , Y ) = i ∑ χ q ( e i ) Y i , b i = e − β ∣ σ ( e i ) ∣ . (168)
We write ⟨ F ⟩ P = ∑ q , Y P ( q , Y ) F ( q , Y ) \langle F\rangle_P=\sum_{q,Y}P(q,Y)F(q,Y) ⟨ F ⟩ P = ∑ q , Y P ( q , Y ) F ( q , Y ) , and similarly for F 2 F^2 F 2 and the distribution Q Q Q . The first moments and character orthogonality give
⟨ F ⟩ P = a ∑ i κ β ( e i ) ≥ a ∑ i b i , ⟨ F ⟩ Q = 0. (169) \langle F\rangle_P=a\sum_i\kappa_\beta(e_i)\ge a\sum_i b_i,\qquad\langle F\rangle_Q=0. \tag*{(169)} ⟨ F ⟩ P = a i ∑ κ β ( e i ) ≥ a i ∑ b i , ⟨ F ⟩ Q = 0. (169)
For the second moments,
⟨ F 2 ⟩ Q = M + a 2 ∑ i ≠ j g ( e i ) = g ( e j ) κ β ( e i ⊕ e j ) ≤ ⟨ F 2 ⟩ P = M + a 2 ∑ i ≠ j κ β ( e i ⊕ e j ) ≤ M + 2 2 Δ Z a 2 ( ∑ i b i ) 2 . (170) \begin{aligned}
\langle F^2\rangle_Q=M+a^2\sum_{\substack{i\ne j\\g(e_i)=g(e_j)}}\kappa_\beta(e_i\oplus e_j) \\
&\le\langle F^2\rangle_P=M+a^2\sum_{i\ne j}\kappa_\beta(e_i\oplus e_j)\le M+2^{2\Delta_Z}a^2\left(\sum_i b_i\right)^2. \tag*{(170)}
\end{aligned} ⟨ F 2 ⟩ Q = M + a 2 i = j g ( e i ) = g ( e j ) ∑ κ β ( e i ⊕ e j ) ≤ ⟨ F 2 ⟩ P = M + a 2 i = j ∑ κ β ( e i ⊕ e j ) ≤ M + 2 2 Δ Z a 2 ( i ∑ b i ) 2 . (170)
Nonnegativity of the kernel proves the first inequality. Disjoint syndromes and sech β ≤ 2 e − β \operatorname{sech}\beta\le2e^{-\beta} sech β ≤ 2 e − β bound each pair kernel by 2 2 Δ Z b i b j 2^{2\Delta_Z}b_ib_j 2 2 Δ Z b i b j .
For any two distributions and real statistic, Cauchy–Schwarz with weight ∣ P − Q ∣ |P-Q| ∣ P − Q ∣ gives
∣ ⟨ F ⟩ P − ⟨ F ⟩ Q ∣ 2 ≤ 2 TV ( P , Q ) ( ⟨ F 2 ⟩ P + ⟨ F 2 ⟩ Q ) . (171) \left|\langle F\rangle_P-\langle F\rangle_Q\right|^2\le2\operatorname{TV}(P,Q)\left(\langle F^2\rangle_P+\langle F^2\rangle_Q\right). \tag*{(171)} ∣ ⟨ F ⟩ P − ⟨ F ⟩ Q ∣ 2 ≤ 2 TV ( P , Q ) ( ⟨ F 2 ⟩ P + ⟨ F 2 ⟩ Q ) . (171)
Here TV ( P , Q ) = 1 2 ∑ q , Y ∣ P ( q , Y ) − Q ( q , Y ) ∣ \operatorname{TV}(P,Q)=\frac{1}{2}\sum_{q,Y}|P(q,Y)-Q(q,Y)| TV ( P , Q ) = 2 1 ∑ q , Y ∣ P ( q , Y ) − Q ( q , Y ) ∣ . Trace-distance contractivity under the measurement channel [21 Sec. 3.3] then yields
ε ≥ TV ( P , Q ) ≥ a 2 ( ∑ i b i ) 2 4 [ M + 2 2 Δ Z a 2 ( ∑ i b i ) 2 ] . (172) \varepsilon\ge\operatorname{TV}(P,Q)\ge\frac{a^2\left(\sum_i b_i\right)^2}{4\left[M+2^{2\Delta_Z}a^2\left(\sum_i b_i\right)^2\right]}. \tag*{(172)} ε ≥ TV ( P , Q ) ≥ 4 [ M + 2 2 Δ Z a 2 ( ∑ i b i ) 2 ] a 2 ( ∑ i b i ) 2 . (172)
The right-hand side increases with ( ∑ i b i ) 2 \left(\sum_i b_i\right)^2 ( ∑ i b i ) 2 . Substituting ∑ i b i ≥ M e − Δ Z β \sum_i b_i\ge Me^{-\Delta_Z\beta} ∑ i b i ≥ M e − Δ Z β proves the claim. ∎
Let G Z G_Z G Z join physical qubits sharing a measured check, and suppose its maximum degree is D D D . Set U = ∣ ⋃ a ℓ a ∣ U = \left|\bigcup_a \ell_a\right| U = ∣ ⋃ a ℓ a ∣ , as in the main text. The induced graph on the union of selected representatives has an independent set of size at least U / ( D + 1 ) U/(D+1) U / ( D + 1 ) . Its vertices have disjoint syndrome supports, and all have nonzero logical column. For fixed 0 < p < 1 0 < p < 1 0 < p < 1 and size-independent bounds D D D and Δ Z > 0 \Delta_Z > 0 Δ Z > 0 , vanishing ε \varepsilon ε therefore requires
β − log U 2 Δ Z ⟶ + ∞ . (173) \beta- \frac{\log U}{2\Delta_Z} \longrightarrow+\infty. \tag*{(173)} β − 2 Δ Z log U ⟶ + ∞. (173)
This argument makes no fixed-k k k assumption. At constant encoding rate, U ≥ k = Θ ( n ) U \ge k = \Theta(n) U ≥ k = Θ ( n ) , so the necessary strength is logarithmic in n n n . For general selected encodings, U U U cannot be replaced by n n n : adding ancillary qubits outside the logical supports does not add vulnerable anchors. This obstruction concerns full-register trace-distance decoupling; it is not a necessary condition for vanishing entropy density alone.
Conversion to recovery and entropy density
Trace-distance decoupling controls recovery of the entire selected register uniformly in its dimension. Conversion to total AQEC entropy depends on that dimension, so total entropy and entropy per logical qubit must be kept distinct when k k k grows.
Supplementary Lemma 2.9 (Recovery and entropy at arbitrary rate). Let F e ∗ F_e^* F e ∗ be the optimal entanglement fidelity for the maximally mixed input on the entire selected register, with squared fidelity convention. The entropy-density statements assume k ≥ 1 k \ge1 k ≥ 1 . For an empty register, K = 1 K = 1 K = 1 and B = D A Q E C = 0 B = D_{\mathrm{AQEC}} = 0 B = D AQEC = 0 , so no density is defined. The fidelity obeys the fixed-input duality in equation (119) and the bounds
ε 2 4 ≤ 1 − F e ∗ ≤ 2 ε − ε 2 ≤ 2 ε . (174) \frac{\varepsilon^2}{4} \le1-F_e^* \le2\varepsilon-\varepsilon^2 \le2\varepsilon. \tag*{(174)} 4 ε 2 ≤ 1 − F e ∗ ≤ 2 ε − ε 2 ≤ 2 ε . (174)
For the classical–quantum reference–environment state of this model,
D A Q E C ≤ ε log K + ( 1 + ε ) h 2 ( ε 1 + ε ) . (175) D_{\mathrm{AQEC}} \le\varepsilon\log K + (1+\varepsilon)h_2\left(\frac{\varepsilon}{1+\varepsilon}\right). \tag*{(175)} D AQEC ≤ ε log K + ( 1 + ε ) h 2 ( 1 + ε ε ) . (175)
Consequently ε → 0 \varepsilon\to0 ε → 0 is equivalent to F e ∗ → 1 F_e^* \to1 F e ∗ → 1 and implies D A Q E C / k → 0 D_{\mathrm{AQEC}}/k \to0 D AQEC / k → 0 . The stronger condition k ε → 0 k\varepsilon\to0 k ε → 0 implies D A Q E C → 0 D_{\mathrm{AQEC}} \to0 D AQEC → 0 . For fixed k k k , vanishing AQEC entropy, vanishing ε \varepsilon ε , and F e ∗ → 1 F_e^* \to1 F e ∗ → 1 are therefore equivalent.
Proof. First prove equation (119) . Fix an environment state ξ E \xi_E ξ E and a purification ∣ χ ⟩ F E |\chi\rangle_{FE} ∣ χ ⟩ F E . Uhlmann’s theorem [30 ] optimizes the overlap of a purification of ρ R E \rho_{RE} ρ R E with ∣ Φ K ⟩ R L ∣ χ ⟩ F E |\Phi_K\rangle_{RL}|\chi\rangle_{FE} ∣ Φ K ⟩ R L ∣ χ ⟩ F E by an isometry from the physical output Q Q Q to L F LF L F , enlarging F F F when necessary. Regard this isometry as a recovery dilation. The resulting entanglement fidelity, which sums over all states of F E FE F E , is at least this one squared overlap. Maximizing over ξ E \xi_E ξ E proves the ≥ \ge ≥ direction. Conversely, for any recovery dilation let ∣ Ψ ⟩ R L F E |\Psi\rangle_{RLFE} ∣Ψ ⟩ R L F E be its pure output and put ∣ v ⟩ F E = ( ⟨ Φ K ∣ R L ⊗ I F E ) ∣ Ψ ⟩ |v\rangle_{FE}=(\langle\Phi_K|_{RL}\otimes I_{FE})|\Psi\rangle ∣ v ⟩ F E = (⟨ Φ K ∣ R L ⊗ I F E ) ∣Ψ ⟩ . Its squared norm is that recovery’s entanglement fidelity. If it is nonzero, normalize ∣ v ⟩ |v\rangle ∣ v ⟩ to ∣ χ ⟩ |\chi\rangle ∣ χ ⟩ and set ξ E = Tr F ∣ χ ⟩ ⟨ χ ∣ \xi_E=\operatorname{Tr}_F|\chi\rangle\langle\chi| ξ E = Tr F ∣ χ ⟩ ⟨ χ ∣ . The overlap with ∣ Φ K ⟩ ∣ χ ⟩ |\Phi_K\rangle|\chi\rangle ∣ Φ K ⟩ ∣ χ ⟩ is then exactly ∥ v ∥ \|v\| ∥ v ∥ , so Uhlmann’s theorem bounds this recovery fidelity by the right-hand side of equation (119) . A zero fidelity is immediate. Taking the supremum over recoveries proves equality.
For squared fidelity F ( ρ , σ ) = ∥ ρ σ ∥ 1 2 F(\rho,\sigma)=\|\sqrt{\rho}\sqrt{\sigma}\|_1^2 F ( ρ , σ ) = ∥ ρ σ ∥ 1 2 , the Fuchs–van de Graaf inequality [29 ] gives
F ( ρ R E , I R K ⊗ ρ E ) ≥ ( 1 − ε ) 2 . (176) F\left(\rho_{RE},\frac{I_R}{K}\otimes\rho_E\right)\ge(1-\varepsilon)^2. \tag*{(176)} F ( ρ R E , K I R ⊗ ρ E ) ≥ ( 1 − ε ) 2 . (176)
Choosing ξ E = ρ E \xi_E=\rho_E ξ E = ρ E in Eq. (119) gives F e ∗ ≥ ( 1 − ε ) 2 F_e^* \ge(1-\varepsilon)^2 F e ∗ ≥ ( 1 − ε ) 2 and hence the upper infidelity bound.
Conversely, let a recovery have entanglement infidelity r r r . Uhlmann’s theorem applied to its output and the desired maximally entangled state gives some environment state ξ E \xi_E ξ E for which
1 2 ∥ ρ R E − I R K ⊗ ξ E ∥ 1 ≤ r , 1 2 ∥ ρ E − ξ E ∥ 1 ≤ r . (177) \frac{1}{2}\left\|\rho_{RE}-\frac{I_R}{K}\otimes\xi_E\right\|_1\le\sqrt{r},\qquad
\frac{1}{2}\|\rho_E-\xi_E\|_1\le\sqrt{r}. \tag*{(177)} 2 1 ρ R E − K I R ⊗ ξ E 1 ≤ r , 2 1 ∥ ρ E − ξ E ∥ 1 ≤ r . (177)
The second inequality follows by partial trace. The triangle inequality then gives ε ≤ 2 r \varepsilon\le2\sqrt{r} ε ≤ 2 r , proving the lower bound on optimal infidelity.
Finally, both ρ R E \rho_{RE} ρ R E and I R / K ⊗ ρ E I_R/K\otimes\rho_E I R / K ⊗ ρ E are classical on the same K K K -dimensional reference. The classical–quantum case of Lemma 2 in Ref. [31 ] , with its systems A = R A=R A = R and B = E B=E B = E , bounds their conditional-entropy difference by ε log K + ( 1 + ε ) h 2 ( ε / ( 1 + ε ) ) \varepsilon\log K+(1+\varepsilon)h_2(\varepsilon/(1+\varepsilon)) ε log K + ( 1 + ε ) h 2 ( ε / ( 1 + ε )) . Using I ( R : E ) = log K − S ( R ∣ E ) I(R:E)=\log K-S(R|E) I ( R : E ) = log K − S ( R ∣ E ) gives Eq. (175) . Here the classical reference permits the coefficient log K \log K log K , rather than the 2 log K 2\log K 2 log K of the general quantum bound. Since log K = k log 2 \log K=k\log2 log K = k log 2 , division by k k k proves the density statement; k ε → 0 k\varepsilon\to0 k ε → 0 controls the total entropy as well. ∎
These bounds concern entanglement fidelity for the full maximally mixed logical input; the proof does not replace it with a worst-case recovery criterion. Entropy density alone is weaker even than this full-register criterion. For example, complete dephasing of one logical qubit and the identity channel on the other k − 1 k-1 k − 1 qubits give
D A Q E C = log 2 , D A Q E C k ⟶ 0 , F e ⋆ = 1 2 . (178) D_{\mathrm{AQEC}}=\log2,\qquad\frac{D_{\mathrm{AQEC}}}{k}\longrightarrow0,\qquad F_{\mathrm{e}}^\star=\frac{1}{2}. \tag*{(178)} D AQEC = log 2 , k D AQEC ⟶ 0 , F e ⋆ = 2 1 . (178)
This example explains why the rate-dependent statements retain separate conditions for entropy density, total leakage, and recovery of the entire logical register.
Local enhancement: operator estimates and geometry
Local enhancement suppresses the local leakage identified in Supplementary Note 2. Errors with syndrome on an enhanced check have a small residual overlap, while silent errors with nonzero logical parity must grow across the protected neighborhood. We first bound their contributions to the environment state under projective enhancement, and then treat finite measurement strength. The estimates apply to correlated noise; the code geometry enters through the minimum size of a silent component.
All qubit supports are identified with vectors over F 2 \mathbb{F}_2 F 2 ; addition of supports is symmetric difference, denoted by ⊕ \oplus ⊕ . The probability of the physical X X X -error support x x x is P ( x ) P(x) P ( x ) . We assume local stochasticity [32 ] ,
Pr [ A ⊆ x ] ≤ η ∣ A ∣ for every fixed qubit set A , 0 ≤ η ≤ 1. (179) \Pr[A\subseteq x]\leq\eta^{|A|}\qquad\text{for every fixed qubit set }A,\qquad0\leq\eta\leq1. \tag*{(179)} Pr [ A ⊆ x ] ≤ η ∣ A ∣ for every fixed qubit set A , 0 ≤ η ≤ 1. (179)
The distribution P P P may contain correlations. Local stochasticity controls the probability of a prescribed support while allowing arbitrary dependence among errors elsewhere.
Let G Z G_Z G Z be the graph whose vertices are physical qubits, with an edge between two distinct qubits whenever they belong to a common measured Z Z Z check. We use connectedness and graph distance in this graph. Its degree is bounded uniformly for the code families considered here. Write
σ ( x ) = supp ( H Z x ) , g ( x ) = ( ∣ x ∩ ℓ 1 ∣ , … , ∣ x ∩ ℓ k ∣ ) ( m o d 2 ) , χ q ( x ) = ( − 1 ) q ⋅ g ( x ) . (180) \sigma(x)=\operatorname{supp}(H_Zx),\qquad g(x)=(|x\cap\ell_1|,\ldots,|x\cap\ell_k|)\pmod2,\qquad\chi_q(x)=(-1)^{q\cdot g(x)}. \tag*{(180)} σ ( x ) = supp ( H Z x ) , g ( x ) = ( ∣ x ∩ ℓ 1 ∣ , … , ∣ x ∩ ℓ k ∣ ) ( mod 2 ) , χ q ( x ) = ( − 1 ) q ⋅ g ( x ) . (180)
Here ℓ 1 , … , ℓ k \ell_1,\ldots,\ell_k ℓ 1 , … , ℓ k are the selected independent Z Z Z -logical representatives, q ∈ F 2 k q\in\mathbb{F}_2^k q ∈ F 2 k , K = 2 k K=2^k K = 2 k , and ℓ ( q ) = ⨁ a q a ℓ a \ell(q)=\bigoplus_a q^a\ell_a ℓ ( q ) = ⨁ a q a ℓ a . The protected measured checks form a set P r \mathcal{P}_r P r . A support is silent in the protected region when σ ( x ) ∩ P r = ∅ \sigma(x)\cap\mathcal{P}_r=\varnothing σ ( x ) ∩ P r = ∅ . This condition need not make its syndrome vanish outside that region.
Projection and the exact connected-component expansion
Hard projection isolates the enhancement mechanism in its simplest form: any error with a protected syndrome has zero overlap within the encoded subspace. Decomposing the remaining errors into connected components then separates those that can carry logical information from those that cannot.
Supplementary Lemma 3.1 (Hard projection and logical sectors). For the finite-strength state ∣ ψ β ⟩ |\psi_\beta\rangle ∣ ψ β ⟩ , define
Π r = ∏ j ∈ P r I + B j 2 , ∣ ψ ⟩ r = Π r ∣ ψ β ⟩ ⟨ ψ β ∣ Π r ∣ ψ β ⟩ , ∣ q ⟩ r = Z ℓ ( q ) ∣ ψ ⟩ r . (181) \Pi_r=\prod_{j\in\mathcal{P}_r}\frac{I+B_j}{2},\qquad
|\psi\rangle_r=\frac{\Pi_r|\psi_\beta\rangle}{\sqrt{\langle\psi_\beta|\Pi_r|\psi_\beta\rangle}},\qquad
|q\rangle_r=Z^{\ell(q)}|\psi\rangle_r. \tag*{(181)} Π r = j ∈ P r ∏ 2 I + B j , ∣ ψ ⟩ r = ⟨ ψ β ∣ Π r ∣ ψ β ⟩ Π r ∣ ψ β ⟩ , ∣ q ⟩ r = Z ℓ ( q ) ∣ ψ ⟩ r . (181) These states are well defined and the states ∣ q ⟩ r |q\rangle_r ∣ q ⟩ r are orthonormal. With κ r ( x ) = ⟨ ψ r ∣ X x ∣ ψ r ⟩ \kappa_r(x)=\langle\psi_r|X_x|\psi_r\rangle κ r ( x ) = ⟨ ψ r ∣ X x ∣ ψ r ⟩ ,
r ⟨ q ∣ X x ∣ q ′ ⟩ r = δ q , q ′ χ q ( x ) κ r ( x ) , (182) {}_r\langle q|X_x|q'\rangle_r=\delta_{q,q'}\chi_q(x)\kappa_r(x), \tag*{(182)} r ⟨ q ∣ X x ∣ q ′ ⟩ r = δ q , q ′ χ q ( x ) κ r ( x ) , (182) σ ( x ) ∩ P r ≠ ∅ ⟹ κ r ( x ) = 0. (183) \sigma(x)\cap\mathcal{P}_r\ne\varnothing\quad\Longrightarrow\quad\kappa_r(x)=0. \tag*{(183)} σ ( x ) ∩ P r = ∅ ⟹ κ r ( x ) = 0. (183)
Proof. All measured checks are positive-phase Z Z Z operators. The computational-basis state ∣ 0 ⟩ ⊗ n |0\rangle^{\otimes n} ∣0 ⟩ ⊗ n therefore has eigenvalue + 1 +1 + 1 for every B j B_j B j . Its amplitude in the finite-strength state is nonzero, so the denominator in Eq. (181) is strictly positive. On the range of Π r \Pi_r Π r , each protected filter exp ( β j B j / 2 ) \exp(\beta_jB_j/2) exp ( β j B j /2 ) acts as the scalar exp ( β j / 2 ) \exp(\beta_j/2) exp ( β j /2 ) . Consequently ∣ ψ ⟩ r |\psi\rangle_r ∣ ψ ⟩ r depends only on the exterior measurement strengths and the protected projectors, not on the finite strengths used on protected checks before projection.
Choose logical X X X operators dual to the selected Z Z Z representatives. They commute with every Z Z Z -check filter and with Π r \Pi_r Π r , and they stabilize the initial state ∣ + ⟩ ⊗ n |+\rangle^{\otimes n} ∣ + ⟩ ⊗ n . Their eigenvalues on ∣ q ⟩ r |q\rangle_r ∣ q ⟩ r are ( − 1 ) q a (-1)^{q_a} ( − 1 ) q a . Distinct logical labels thus lie in orthogonal joint eigenspaces. Every physical X x X_x X x commutes with these logical X X X operators, so its matrix elements between distinct q q q sectors vanish. For the diagonal matrix element, commute X x X_x X x through Z ℓ ( q ) Z^{\ell(q)} Z ℓ ( q ) to obtain the sign χ q ( x ) \chi_q(x) χ q ( x ) in (S103).
The same argument covers a selected subset of all logical qubits. The unnormalized filtered states belonging to its basis labels have the same norm, because Z ℓ ( q ) Z^{\ell(q)} Z ℓ ( q ) is unitary and commutes with the filter and projector. Their orthogonality then makes the branch success probability independent of an arbitrary logical superposition. Hence the normalized encoding is an isometry, rather than an input-dependent postselection map.
Finally, if j ∈ σ ( x ) ∩ P r j \in\sigma(x) \cap\mathcal{P}_r j ∈ σ ( x ) ∩ P r , then B j B_j B j anticommutes with X x X_x X x , and
( I + B j ) X x ( I + B j ) = 0. (184) (I+B_j)X_x(I+B_j)=0. \tag*{(184)} ( I + B j ) X x ( I + B j ) = 0. (184)
All protected projectors commute. Their product therefore satisfies Π r X x Π r = 0 \Pi_rX_x\Pi_r=0 Π r X x Π r = 0 , which proves (S104).
Supplementary Lemma 3.2 (Syndrome separation of connected components). If x = ⨆ i Γ i x=\bigsqcup_i\Gamma_i x = ⨆ i Γ i is its connected-component decomposition in G Z G_Z G Z , then
σ ( x ) = ⨆ i σ ( Γ i ) , χ q ( x ) = ∏ i χ q ( Γ i ) . (185) \sigma(x)=\bigsqcup_i\sigma(\Gamma_i),\qquad\chi_q(x)=\prod_i\chi_q(\Gamma_i). \tag*{(185)} σ ( x ) = i ⨆ σ ( Γ i ) , χ q ( x ) = i ∏ χ q ( Γ i ) . (185) Thus a protected-syndrome-silent error has only silent components. If it distinguishes q q q from q ′ q' q ′ , an odd number of its components distinguish that same pair.
Proof. A measured check cannot touch two distinct components: a qubit of one and a qubit of the other would then be adjacent in G Z G_Z G Z . The sets of checks touched by distinct components are consequently disjoint, so their syndromes cannot cancel each other. This proves the first identity. Intersection parity adds on disjoint supports, proving the second. In particular, χ q ( x ) χ q ′ ( x ) \chi_q(x)\chi_{q'}(x) χ q ( x ) χ q ′ ( x ) is the product of a factor − 1 -1 − 1 for each component that distinguishes the two labels and a factor + 1 +1 + 1 for each other component.
For a logical pair q , q ′ q,q' q , q ′ , define
C ( q , q ′ ) = { Γ : Γ is connected in G Z , χ q ( Γ ) ≠ χ q ′ ( Γ ) } , (186) C(q,q')=\{\Gamma:\Gamma\text{ is connected in }G_Z,\ \chi_q(\Gamma)\ne\chi_{q'}(\Gamma)\}, \tag*{(186)} C ( q , q ′ ) = { Γ : Γ is connected in G Z , χ q ( Γ ) = χ q ′ ( Γ )} , (186)
C r ( q , q ′ ) = { Γ ∈ C ( q , q ′ ) : σ ( Γ ) ∩ P r = ∅ } . (187) C_r(q,q')=\{\Gamma\in C(q,q'):\sigma(\Gamma)\cap\mathcal{P}_r=\varnothing\}. \tag*{(187)} C r ( q , q ′ ) = { Γ ∈ C ( q , q ′ ) : σ ( Γ ) ∩ P r = ∅ } . (187)
A family of connected supports is compatible if its members are pairwise disjoint and no edge of G Z G_Z G Z joins distinct members. This is precisely the condition for its members to occur as complete connected components of one error difference. We write Comp ( x ) \operatorname{Comp}(x) Comp ( x ) for the set of connected components of x x x .
Supplementary Lemma 3.3 (Parity-family identity). Define the pairwise environment matrix
( D r q , q ′ ) u v = P ( u ) P ( v ) [ χ q ( u ⊕ v ) − χ q ′ ( u ⊕ v ) ] κ r ( u ⊕ v ) . (188) (D_r^{q,q'})_{uv}=\sqrt{P(u)P(v)}[\chi_q(u\oplus v)-\chi_{q'}(u\oplus v)]\kappa_r(u\oplus v). \tag*{(188)} ( D r q , q ′ ) uv = P ( u ) P ( v ) [ χ q ( u ⊕ v ) − χ q ′ ( u ⊕ v )] κ r ( u ⊕ v ) . (188) For each compatible, nonempty family F ⊆ C r ( q , q ′ ) F\subseteq C_r(q,q') F ⊆ C r ( q , q ′ ) , define
( D r , F q , q ′ ) u v = ( − 2 ) ∣ F ∣ P ( u ) P ( v ) χ q ′ ( u ⊕ v ) κ r ( u ⊕ v ) 1 { F ⊆ Comp ( u ⊕ v ) } . (189) (D_{r,F}^{q,q'})_{uv}=(-2)^{|F|}\sqrt{P(u)P(v)}\,\chi_{q'}(u\oplus v)\kappa_r(u\oplus v)\mathbf{1}\{F\subseteq\operatorname{Comp}(u\oplus v)\}. \tag*{(189)} ( D r , F q , q ′ ) uv = ( − 2 ) ∣ F ∣ P ( u ) P ( v ) χ q ′ ( u ⊕ v ) κ r ( u ⊕ v ) 1 { F ⊆ Comp ( u ⊕ v )} . (189) Then the following is an exact matrix identity:
D r q , q ′ = ∑ ∅ ≠ F ⊆ C r ( q , q ′ ) F c o m p a t i b l e D r , F q , q ′ . (190) D_r^{q,q'}=\sum_{\substack{\varnothing\ne F\subseteq C_r(q,q')\\F\ \mathrm{compatible}}}D_{r,F}^{q,q'}. \tag*{(190)} D r q , q ′ = ∅ = F ⊆ C r ( q , q ′ ) F compatible ∑ D r , F q , q ′ . (190)
Proof. For a fixed error difference x x x , let A ( x ) A(x) A ( x ) be the components of x x x that distinguish q q q from q ′ q' q ′ . By character multiplication,
χ q ( x ) − χ q ′ ( x ) = χ q ′ ( x ) [ ∏ Γ ∈ A ( x ) ( 1 − 2 ) − 1 ] (191) \chi_q(x)-\chi_{q'}(x)=\chi_{q'}(x)\left[\prod_{\Gamma\in A(x)}(1-2)-1\right] \tag*{(191)} χ q ( x ) − χ q ′ ( x ) = χ q ′ ( x ) Γ ∈ A ( x ) ∏ ( 1 − 2 ) − 1 (191)
= χ q ′ ( x ) ∑ ∅ ≠ F ⊆ A ( x ) ( − 2 ) ∣ F ∣ . (192) =\chi_{q'}(x)\sum_{\substack{\varnothing\ne F\subseteq A(x)}}(-2)^{|F|}. \tag*{(192)} = χ q ′ ( x ) ∅ = F ⊆ A ( x ) ∑ ( − 2 ) ∣ F ∣ . (192)
When κ r ( x ) ≠ 0 \kappa_r(x)\ne0 κ r ( x ) = 0 , Lemmas 3.1 and 3.2 imply that all components of x x x are silent in P r \mathcal{P}_r P r . Every member of A ( x ) A(x) A ( x ) then belongs to C r ( q , q ′ ) C_r(q,q') C r ( q , q ′ ) . Actual components are compatible, and the indicator in (S110) selects exactly the nonempty subsets of A ( x ) A(x) A ( x ) . Multiplying (S113) by P ( u ) P ( v ) κ r ( x ) \sqrt{P(u)P(v)}\kappa_r(x) P ( u ) P ( v ) κ r ( x ) with x = u ⊕ v x=u\oplus v x = u ⊕ v proves the claimed identity entry by entry. If κ r ( x ) = 0 \kappa_r(x)=0 κ r ( x ) = 0 , both sides vanish. The signed inclusion–exclusion coefficients are retained throughout this step.
A trace-norm estimate for each family
For each family, a cross-Gram representation bounds the trace norm after summing over all exterior errors. Local stochasticity then controls the probability of the prescribed support, without conditioning the noise on the exterior configuration.
Supplementary Lemma 3.4 (Cross-Gram estimate). Every compatible, nonempty family satisfies
∥ D r , F q , q ′ ∥ 1 ≤ ∏ Γ ∈ F ( C 0 η ) ∣ Γ ∣ , C 0 = 4. (193) \lVert D_{r,F}^{q,q'} \rVert_1 \le\prod_{\Gamma\in F}(C_0\sqrt{\eta})^{|\Gamma|}, \qquad C_0=4. \tag*{(193)} ∥ D r , F q , q ′ ∥ 1 ≤ Γ ∈ F ∏ ( C 0 η ) ∣Γ∣ , C 0 = 4. (193) The constant is independent of system size and logical dimension.
Proof. For η = 0 \eta=0 η = 0 , (S100) implies that the empty error has probability one, and the claim is immediate. Assume η > 0 \eta>0 η > 0 . Put S = ⋃ Γ ∈ F Γ S=\bigcup_{\Gamma\in F}\Gamma S = ⋃ Γ ∈ F Γ and define its external vertex boundary by
∂ G S = { e ∉ S : e has a neighbor in S } . (194) \partial_G S=\{e\notin S:e\text{ has a neighbor in }S\}. \tag*{(194)} ∂ G S = { e ∈ / S : e has a neighbor in S } . (194) Compatibility gives the exact equivalence
F ⊆ Comp ( u ⊕ v ) ⟺ ( u ⊕ v ) ∣ S = 1 , ( u ⊕ v ) ∣ ∂ G S = 0. (195) F\subseteq\operatorname{Comp}(u\oplus v)\quad\Longleftrightarrow\quad(u\oplus v)|_S=1,\qquad(u\oplus v)|_{\partial_G S}=0. \tag*{(195)} F ⊆ Comp ( u ⊕ v ) ⟺ ( u ⊕ v ) ∣ S = 1 , ( u ⊕ v ) ∣ ∂ G S = 0. (195) No condition is imposed outside S ∪ ∂ G S S\cup\partial_G S S ∪ ∂ G S .
Partition the permitted matrix entries into assignments
u ∣ S = z , v ∣ S = 1 − z , u ∣ ∂ G S = v ∣ ∂ G S = w . (196) u|_S=z,\qquad v|_S=1-z,\qquad u|_{\partial_G S}=v|_{\partial_G S}=w. \tag*{(196)} u ∣ S = z , v ∣ S = 1 − z , u ∣ ∂ G S = v ∣ ∂ G S = w . (196) For fixed z , w z,w z , w , denote the allowed row labels by U z , w \mathcal{U}_{z,w} U z , w and column labels by V z , w \mathcal{V}_{z,w} V z , w . The character χ q ′ ( u ⊕ v ) = χ q ′ ( u ) χ q ′ ( v ) \chi_{q'}(u\oplus v)=\chi_{q'}(u)\chi_{q'}(v) χ q ′ ( u ⊕ v ) = χ q ′ ( u ) χ q ′ ( v ) merely multiplies rows and columns by signs and therefore does not change singular values. After removing these signs, the submatrix is ( − 2 ) ∣ F ∣ A † C (-2)^{|F|}A^\dagger C ( − 2 ) ∣ F ∣ A † C , where
A = ∑ u ∈ U z , w P ( u ) ∣ X u ψ r ⟩ ⟨ u ∣ , C = ∑ v ∈ V z , w P ( v ) ∣ X v ψ r ⟩ ⟨ v ∣ . (197) A=\sum_{u\in\mathcal{U}_{z,w}}\sqrt{P(u)}|X_u\psi_r\rangle\langle u|,\qquad C=\sum_{v\in\mathcal{V}_{z,w}}\sqrt{P(v)}|X_v\psi_r\rangle\langle v|. \tag*{(197)} A = u ∈ U z , w ∑ P ( u ) ∣ X u ψ r ⟩ ⟨ u ∣ , C = v ∈ V z , w ∑ P ( v ) ∣ X v ψ r ⟩ ⟨ v ∣. (197) Indeed, κ r ( u ⊕ v ) = ⟨ X u ψ r ∣ X v ψ r ⟩ \kappa_r(u\oplus v)=\langle X_u\psi_r|X_v\psi_r\rangle κ r ( u ⊕ v ) = ⟨ X u ψ r ∣ X v ψ r ⟩ . The vectors in these operators have unit norm, giving ∥ A ∥ 2 2 = Pr [ x ∈ U z , w ] \lVert A\rVert_2^2=\Pr[x\in\mathcal{U}_{z,w}] ∥ A ∥ 2 2 = Pr [ x ∈ U z , w ] and ∥ C ∥ 2 2 = Pr [ x ∈ V z , w ] \lVert C\rVert_2^2=\Pr[x\in\mathcal{V}_{z,w}] ∥ C ∥ 2 2 = Pr [ x ∈ V z , w ] . Schatten Hölder [21 Sec. 1.1.3] thus yields
∥ ( − 2 ) ∣ F ∣ A † C ∥ 1 ≤ 2 ∣ F ∣ Pr [ x ∈ U z , w ] Pr [ x ∈ V z , w ] . (198) \lVert(-2)^{|F|}A^\dagger C\rVert_1\le2^{|F|}\sqrt{\Pr[x\in\mathcal{U}_{z,w}]\Pr[x\in\mathcal{V}_{z,w}]}. \tag*{(198)} ∥( − 2 ) ∣ F ∣ A † C ∥ 1 ≤ 2 ∣ F ∣ Pr [ x ∈ U z , w ] Pr [ x ∈ V z , w ] . (198) These probabilities already include the sum over every exterior error configuration.
Sum the boundary assignments before applying local stochasticity. For fixed z z z , Cauchy–Schwarz gives
∑ w Pr [ x ∈ U z , w ] Pr [ x ∈ V z , w ] ≤ Pr [ x ∣ S = z ] Pr [ x ∣ S = 1 − z ] ≤ η ∣ z ∣ η ∣ S ∣ − ∣ z ∣ = η ∣ S ∣ / 2 . (199) \sum_w\sqrt{\Pr[x\in\mathcal{U}_{z,w}]\Pr[x\in\mathcal{V}_{z,w}]}
\le\sqrt{\Pr[x|_S=z]\Pr[x|_S=1-z]}
\le\sqrt{\eta^{|z|}\eta^{|S|-|z|}}=\eta^{|S|/2}. \tag*{(199)} w ∑ Pr [ x ∈ U z , w ] Pr [ x ∈ V z , w ] ≤ Pr [ x ∣ S = z ] Pr [ x ∣ S = 1 − z ] ≤ η ∣ z ∣ η ∣ S ∣ − ∣ z ∣ = η ∣ S ∣/2 . (199) Here the event x ∣ S = z x|_S=z x ∣ S = z requires occupation of the ∣ z ∣ |z| ∣ z ∣ specified sites, so (S100) bounds its probability by η ∣ z ∣ \eta^{|z|} η ∣ z ∣ ; requiring additional unoccupied sites can only reduce that probability. The complementary assignment is treated in the same way. This argument neither conditions on the boundary nor assumes that boundary and interior errors are independent.
Use the triangle inequality to sum the assignment submatrices, sum all 2 ∣ S ∣ 2^{|S|} 2 ∣ S ∣ choices of z z z , and use ∣ F ∣ ≤ ∣ S ∣ |F|\le|S| ∣ F ∣ ≤ ∣ S ∣ :
∥ D r , F q , q ′ ∥ 1 ≤ 2 ∣ F ∣ + ∣ S ∣ η ∣ S ∣ / 2 ≤ ( 4 η ) ∣ S ∣ . (200) \lVert D_{r,F}^{q,q'}\rVert_1\le2^{|F|+|S|}\eta^{|S|/2}\le(4\sqrt{\eta})^{|S|}. \tag*{(200)} ∥ D r , F q , q ′ ∥ 1 ≤ 2 ∣ F ∣ + ∣ S ∣ η ∣ S ∣/2 ≤ ( 4 η ) ∣ S ∣ . (200) The supports in F F F are disjoint, so ∣ S ∣ = ∑ Γ ∈ F ∣ Γ ∣ |S|=\sum_{\Gamma\in F}|\Gamma| ∣ S ∣ = ∑ Γ ∈ F ∣Γ∣ , which proves (S114). Summing w w w by (S120) is also why no factor exponential in the boundary size is necessary. □ \square □
Supplementary Lemma 3.5 (Summing the families). Set
Δ r ( q , q ′ ) = ∑ Γ ∈ C r ( q , q ′ ) ( C 0 η ) ∣ Γ ∣ . (201) \Delta_r(q,q')=\sum_{\Gamma\in\mathcal{C}_r(q,q')}(C_0\sqrt{\eta})^{|\Gamma|}. \tag*{(201)} Δ r ( q , q ′ ) = Γ ∈ C r ( q , q ′ ) ∑ ( C 0 η ) ∣Γ∣ . (201)
Then
∥ D r q , q ′ ∥ 1 ≤ e Δ r ( q , q ′ ) − 1 , 1 2 K 2 ∥ B r ∥ 1 ≤ 1 2 K 2 ∑ q , q ′ [ e Δ r ( q , q ′ ) − 1 ] . (202) \lVert D_r^{q,q'} \rVert_1 \le e^{\Delta_r(q,q')} - 1,\qquad\frac{1}{2K^2}\lVert B_r \rVert_1 \le\frac{1}{2K^2}\sum_{q,q'}[e^{\Delta_r(q,q')} - 1]. \tag*{(202)} ∥ D r q , q ′ ∥ 1 ≤ e Δ r ( q , q ′ ) − 1 , 2 K 2 1 ∥ B r ∥ 1 ≤ 2 K 2 1 q , q ′ ∑ [ e Δ r ( q , q ′ ) − 1 ] . (202)
Here B r B_r B r is the normalized AQEC matrix deviation for the hard-enhanced encoding and noise distribution P P P .
Proof. Apply the triangle inequality to the exact operator decomposition (S111) and then use Lemma 3.4 :
∥ D r q , q ′ ∥ 1 ≤ ∑ ∅ ≠ F ⊆ C r ( q , q ′ ) F c o m p a t i b l e ∏ Γ ∈ F ( C 0 η ) ∣ Γ ∣ ≤ ∏ Γ ∈ C r ( q , q ′ ) [ 1 + ( C 0 η ) ∣ Γ ∣ ] − 1 ≤ e Δ r ( q , q ′ ) − 1. (203) \begin{aligned}
\lVert D_r^{q,q'} \rVert_1 \le\sum_{\substack{\varnothing\ne F \subseteq C_r(q,q')\\F\ \mathrm{compatible}}}\prod_{\Gamma\in F}(C_0\sqrt{\eta})^{|\Gamma|} \\
&\le\prod_{\Gamma\in C_r(q,q')}\left[1+(C_0\sqrt{\eta})^{|\Gamma|}\right]-1 \le e^{\Delta_r(q,q')}-1. \tag*{(203)}
\end{aligned} ∥ D r q , q ′ ∥ 1 ≤ ∅ = F ⊆ C r ( q , q ′ ) F compatible ∑ Γ ∈ F ∏ ( C 0 η ) ∣Γ∣ ≤ Γ ∈ C r ( q , q ′ ) ∏ [ 1 + ( C 0 η ) ∣Γ∣ ] − 1 ≤ e Δ r ( q , q ′ ) − 1. (203)
Compatibility is dropped only in the second line, after the contribution of each compatible family has been bounded in trace norm. The final inequality is 1 + t ≤ e t 1+t\le e^t 1 + t ≤ e t for t ≥ 0 t\ge0 t ≥ 0 .
For completeness, let ρ E q \rho_E^q ρ E q be the conditional environment state at logical label q q q . Equation (S103) makes the reference–environment state block diagonal in q q q , with conditional matrices
( ρ E q ) u v = P ( u ) P ( v ) χ q ( u ⊕ v ) κ r ( u ⊕ v ) . (204) (\rho_E^q)_{uv}=\sqrt{P(u)P(v)}\,\chi_q(u\oplus v)\kappa_r(u\oplus v). \tag*{(204)} ( ρ E q ) uv = P ( u ) P ( v ) χ q ( u ⊕ v ) κ r ( u ⊕ v ) . (204)
Thus D r q , q ′ = ρ E q − ρ E q ′ D_r^{q,q'}=\rho_E^q-\rho_E^{q'} D r q , q ′ = ρ E q − ρ E q ′ . Since ρ E = K − 1 ∑ q ′ ρ E q ′ \rho_E=K^{-1}\sum_{q'}\rho_E^{q'} ρ E = K − 1 ∑ q ′ ρ E q ′ , the deviation from the product of the marginals is
B r = ⨁ q 1 K ( ρ E q − ρ E ) = ⨁ q 1 K 2 ∑ q ′ D r q , q ′ . (205) B_r=\bigoplus_q\frac{1}{K}(\rho_E^q-\rho_E)=\bigoplus_q\frac{1}{K^2}\sum_{q'}D_r^{q,q'}. \tag*{(205)} B r = q ⨁ K 1 ( ρ E q − ρ E ) = q ⨁ K 2 1 q ′ ∑ D r q , q ′ . (205)
The full-transpose convention for the AQEC matrices leaves all trace norms unchanged; here the conditional matrices are also real symmetric. Trace-norm additivity over the logical blocks, followed by the triangle inequality within each block, proves the second bound in (S123). It involves an average over pairs of logical labels, not a multiplicative penalty growing with K K K .
For the geometric estimates, write the actual union of selected representatives and its size as
R = ⋃ a = 1 k ℓ a , U = ∣ R ∣ . (206) R=\bigcup_{a=1}^{k}\ell_a,\qquad U=|R|. \tag*{(206)} R = a = 1 ⋃ k ℓ a , U = ∣ R ∣. (206)
Suppose at most μ m \mu^m μ m connected qubit sets of size m m m contain a fixed qubit, uniformly in system size. Such an exponential bound follows from bounded degree [32 ] , with a constant μ \mu μ chosen for the family. Set
λ = μ C 0 η < 1. (207) \lambda=\mu C_0\sqrt{\eta}<1. \tag*{(207)} λ = μ C 0 η < 1. (207)
The operator estimates themselves do not require λ < 1 \lambda<1 λ < 1 ; this restriction makes the subsequent sum over connected supports converge uniformly.
Every component distinguishing q q q from q ′ q' q ′ intersects ℓ ( q ⊕ q ′ ) ⊆ R \ell(q\oplus q')\subseteq R ℓ ( q ⊕ q ′ ) ⊆ R . If every such protected-silent component has size at least M ( r ) M(r) M ( r ) , anchoring it at a qubit of R R R therefore gives
Δ r ( q , q ′ ) ≤ U ∑ m ≥ M ( r ) λ m ≤ U 1 − λ λ M ( r ) . (208) \Delta_r(q,q')\le U\sum_{m\ge M(r)}\lambda^m\le\frac{U}{1-\lambda}\lambda^{M(r)}. \tag*{(208)} Δ r ( q , q ′ ) ≤ U m ≥ M ( r ) ∑ λ m ≤ 1 − λ U λ M ( r ) . (208)
The sum runs over positive integers and is empty when M ( r ) = ∞ M(r)=\infty M ( r ) = ∞ . Substituting into (S123) gives the projective part of the main-text enhancement theorem.
Finite measurement strength and the residual kernel
Finite enhancement replaces exact cancellation by a small residual overlap. The same component expansion applies because the residual kernel admits a Gram representation. Each protected syndrome check contributes a suppression factor to the family bound.
Let every strength obey 0 ≤ β j < ∞ 0\le\beta_j<\infty 0 ≤ β j < ∞ , with β j ≥ β R ≥ 0 \beta_j\ge\beta_R\ge0 β j ≥ β R ≥ 0 on P r \mathcal{P}_r P r .
Supplementary Lemma 3.6 (Residual kernel and finite-strength bound). Define D β q , q ′ D_{\beta}^{q,q'} D β q , q ′ and D β , F q , q ′ D_{\beta,F}^{q,q'} D β , F q , q ′ by (S109) and (S110), replacing κ r \kappa_r κ r by κ β \kappa_\beta κ β and allowing compatible nonempty F ⊆ C ( q , q ′ ) F \subseteq C(q,q') F ⊆ C ( q , q ′ ) . Then
∥ D β , F q , q ′ ∥ 1 ≤ ∏ Γ ∈ F [ ( C 0 η ) ∣ Γ ∣ ( sech β R ) ∣ σ ( Γ ) ∩ P r ∣ ] . (209) \left\|D_{\beta,F}^{q,q'}\right\|_1 \le\prod_{\Gamma\in F}\left[(C_0\sqrt{\eta})^{|\Gamma|}(\operatorname{sech}\beta_R)^{|\sigma(\Gamma)\cap P_r|}\right]. \tag*{(209)} D β , F q , q ′ 1 ≤ Γ ∈ F ∏ [ ( C 0 η ) ∣Γ∣ ( sech β R ) ∣ σ ( Γ ) ∩ P r ∣ ] . (209) Suppose every silent distinguishing component has size at least M ( r ) > 0 M(r)>0 M ( r ) > 0 . If no such component exists, take M ( r ) = ∞ M(r)=\infty M ( r ) = ∞ and λ M ( r ) = 0 \lambda^{M(r)}=0 λ M ( r ) = 0 . Then
Δ r , β R ( q , q ′ ) : = ∑ Γ ∈ C ( q , q ′ ) ( C 0 η ) ∣ Γ ∣ ( sech β R ) ∣ σ ( Γ ) ∩ P r ∣ \Delta_{r,\beta_R}(q,q') := \sum_{\Gamma\in C(q,q')} (C_0\sqrt{\eta})^{|\Gamma|}(\operatorname{sech}\beta_R)^{|\sigma(\Gamma)\cap P_r|} Δ r , β R ( q , q ′ ) := Γ ∈ C ( q , q ′ ) ∑ ( C 0 η ) ∣Γ∣ ( sech β R ) ∣ σ ( Γ ) ∩ P r ∣ ≤ U 1 − λ [ λ sech β R + λ M ( r ) ] . (210) \le\frac{U}{1-\lambda}\left[\lambda\operatorname{sech}\beta_R+\lambda^{M(r)}\right]. \tag*{(210)} ≤ 1 − λ U [ λ sech β R + λ M ( r ) ] . (210) The bounds in (S123) hold for the finite-strength state with Δ r \Delta_r Δ r replaced by Δ r , β R \Delta_{r,\beta_R} Δ r , β R .
Proof. The parity expansion (S113) holds without projection. It gives the exact family identity over all of C ( q , q ′ ) C(q,q') C ( q , q ′ ) . It remains to justify the factors involving sech β R \operatorname{sech}\beta_R sech β R in the operator norm bound, rather than infer them from an elementwise estimate.
Fix a compatible family F F F , and write
S = ⋃ Γ ∈ F Γ , J = ⨆ Γ ∈ F σ ( Γ ) . (211) S=\bigcup_{\Gamma\in F}\Gamma,\qquad J=\bigsqcup_{\Gamma\in F}\sigma(\Gamma). \tag*{(211)} S = Γ ∈ F ⋃ Γ , J = Γ ∈ F ⨆ σ ( Γ ) . (211)
Lemma 3.2 makes the syndrome union disjoint. Fix local assignments on S ∪ ∂ G S S\cup\partial_G S S ∪ ∂ G S as in (S117). Remove those assignments from u u u and v v v , leaving supports u ˉ , v ˉ \bar{u},\bar{v} u ˉ , v ˉ outside S ∪ ∂ G S S\cup\partial_G S S ∪ ∂ G S . No remaining qubit shares a check with S S S . Moreover the prescribed boundary assignments cancel in u ⊕ v u\oplus v u ⊕ v . Consequently
σ ( u ⊕ v ) = J ⊔ T , T = σ ( u ˉ ⊕ v ˉ ) , T ∩ J = ∅ . (212) \sigma(u\oplus v)=J\sqcup T,\qquad T=\sigma(\bar{u}\oplus\bar{v}),\qquad T\cap J=\varnothing. \tag*{(212)} σ ( u ⊕ v ) = J ⊔ T , T = σ ( u ˉ ⊕ v ˉ ) , T ∩ J = ∅ . (212)
In fact each of σ ( u ˉ ) \sigma(\bar{u}) σ ( u ˉ ) and σ ( v ˉ ) \sigma(\bar{v}) σ ( v ˉ ) separately avoids all checks touching S S S .
The finite-strength-state overlap depends only on the syndrome. Let k β ( s ) k_\beta(s) k β ( s ) denote its expression for a check subset s s s , so κ β ( x ) = k β ( σ ( x ) ) \kappa_\beta(x)=k_\beta(\sigma(x)) κ β ( x ) = k β ( σ ( x )) for physical supports x x x . Put u j = tanh β j u_j=\tanh\beta_j u j = tanh β j and R Z = ker H Z T R_Z=\ker H_Z^T R Z = ker H Z T . The redundancy expansion gives
k β ( s ) = ( ∏ j ∈ s sech β j ) ∑ a ∈ R Z supp ( a ) ∩ s = ∅ ∏ j u j a j ∑ a ∈ R Z ∏ j u j a j . (213) k_\beta(s)=\left(\prod_{j\in s}\operatorname{sech}\beta_j\right)
\frac{\sum_{\substack{a\in R_Z\\ \operatorname{supp}(a)\cap s=\varnothing}}\prod_j u_j^{a_j}}
{\sum_{a\in R_Z}\prod_j u_j^{a_j}}. \tag*{(213)} k β ( s ) = ( j ∈ s ∏ sech β j ) ∑ a ∈ R Z ∏ j u j a j ∑ a ∈ R Z supp ( a ) ∩ s = ∅ ∏ j u j a j . (213)
Commuting X x X_x X x between the two amplitude filters gives this formula. A check in σ ( x ) \sigma(x) σ ( x ) cancels out of their combined exponent, whereas every other check retains its full coefficient β j \beta_j β j . Expanding e β j B j = cosh β j ( I + u j B j ) e^{\beta_j B_j}=\cosh\beta_j(I+u_jB_j) e β j B j = cosh β j ( I + u j B j ) against ∣ + ⟩ ⊗ n \lvert+\rangle^{\otimes n} ∣ + ⟩ ⊗ n retains only products with H Z T a = 0 H_Z^T a=0 H Z T a = 0 . Dividing by the normalization gives (S134). Both sums are positive because they contain the empty redundancy, including when some u j u_j u j vanish.
For fixed J J J , define on all check subsets outside J J J
f J ( T ) : = k β ( J ⊔ T ) k β ( J ) = ( ∏ j ∈ T sech β j ) ∑ a ∈ R Z supp ( a ) ∩ J = ∅ ( ∏ j u j a j ) 1 { supp ( a ) ∩ T = ∅ } ∑ a ∈ R Z supp ( a ) ∩ J = ∅ ∏ j u j a j . (214) \begin{aligned}
f_J(T):=\frac{k_\beta(J\sqcup T)}{k_\beta(J)} \\
&=\left(\prod_{j\in T}\operatorname{sech}\beta_j\right)
\frac{\sum_{\substack{a\in R_Z\\ \operatorname{supp}(a)\cap J=\varnothing}}\left(\prod_j u_j^{a_j}\right)\mathbf{1}\{\operatorname{supp}(a)\cap T=\varnothing\}}
{\sum_{\substack{a\in R_Z\\ \operatorname{supp}(a)\cap J=\varnothing}}\prod_j u_j^{a_j}}. \tag*{(214)}
\end{aligned} f J ( T ) := k β ( J ) k β ( J ⊔ T ) = j ∈ T ∏ sech β j ∑ a ∈ R Z supp ( a ) ∩ J = ∅ ∏ j u j a j ∑ a ∈ R Z supp ( a ) ∩ J = ∅ ( ∏ j u j a j ) 1 { supp ( a ) ∩ T = ∅ } . (214)
This defines an extension to the full Boolean group of those check subsets, even if some subsets are not attainable physical syndromes. The function has nonnegative Fourier coefficients on that group. Indeed, for b ∈ { 0 , 1 } b\in\{0,1\} b ∈ { 0 , 1 } and c = sech β j ∈ ( 0 , 1 ] c=\operatorname{sech}\beta_j\in(0,1] c = sech β j ∈ ( 0 , 1 ] ,
c b = 1 + c 2 + 1 − c 2 ( − 1 ) b , 1 { supp ( a ) ∩ T = ∅ } = ∏ j ∈ supp ( a ) 1 + ( − 1 ) 1 { j ∈ T } 2 . (215) c^b=\frac{1+c}{2}+\frac{1-c}{2}(-1)^b,\qquad
\mathbf{1}\{\operatorname{supp}(a)\cap T=\varnothing\}=\prod_{j\in\operatorname{supp}(a)}\frac{1+(-1)^{\mathbf{1}\{j\in T\}}}{2}. \tag*{(215)} c b = 2 1 + c + 2 1 − c ( − 1 ) b , 1 { supp ( a ) ∩ T = ∅ } = j ∈ supp ( a ) ∏ 2 1 + ( − 1 ) 1 { j ∈ T } . (215)
Every displayed coefficient is nonnegative. Products correspond to convolution of Fourier coefficients, and the weights ∏ j u j a j \prod_j u_j^{a_j} ∏ j u j a j in (S135) are nonnegative. Products and weighted sums thus preserve nonnegativity. After division by the positive denominator, write
f J ( T ) = ∑ b c b ( − 1 ) b ⋅ T , c b ≥ 0 , ∑ b c b = f J ( ∅ ) = 1. (216) f_J(T)=\sum_b c_b(-1)^{b\cdot T},\qquad c_b\geq0,\qquad\sum_b c_b=f_J(\varnothing)=1. \tag*{(216)} f J ( T ) = b ∑ c b ( − 1 ) b ⋅ T , c b ≥ 0 , b ∑ c b = f J ( ∅ ) = 1. (216)
The index b b b runs over binary strings on the measured checks outside J J J . The vectors with components c b ( − 1 ) b ⋅ s \sqrt{c_b}(-1)^{b\cdot s} c b ( − 1 ) b ⋅ s have unit norm and inner product f J ( s ⊕ s ′ ) f_J(s\oplus s') f J ( s ⊕ s ′ ) . Use the vectors indexed by s = σ ( u ˉ ) s=\sigma(\bar{u}) s = σ ( u ˉ ) and s ′ = σ ( v ˉ ) s'=\sigma(\bar{v}) s ′ = σ ( v ˉ ) in place of ∣ X u ψ r ⟩ |X_u\psi_r\rangle ∣ X u ψ r ⟩ and ∣ X v ψ r ⟩ |X_v\psi_r\rangle ∣ X v ψ r ⟩ in (S118). This expresses each residual assignment submatrix as a cross-Gram matrix of unit vectors. The same Schatten estimate and boundary sum apply, with the single additional scalar k β ( J ) k_\beta(J) k β ( J ) . From (S134),
0 < k β ( J ) ≤ ∏ j ∈ J sech β j ≤ ( sech β R ) ∣ J ∩ P r ∣ . (217) 0<k_\beta(J)\leq\prod_{j\in J}\operatorname{sech}\beta_j\leq(\operatorname{sech}\beta_R)^{|J\cap\mathcal{P}_r|}. \tag*{(217)} 0 < k β ( J ) ≤ j ∈ J ∏ sech β j ≤ ( sech β R ) ∣ J ∩ P r ∣ . (217)
The numerator in the redundancy ratio is a sub-sum of the denominator. Disjointness of the syndromes in (S132) makes ∣ J ∩ P r ∣ = ∣ σ ( Γ ) ∩ P r ∣ |J\cap\mathcal{P}_r|=|\sigma(\Gamma)\cap\mathcal{P}_r| ∣ J ∩ P r ∣ = ∣ σ ( Γ ) ∩ P r ∣ . Together with (S121), this proves (S130).
For the geometric sum, each distinguishing component intersects ℓ ( q ⊕ q ′ ) ⊆ R \ell(q\oplus q')\subseteq R ℓ ( q ⊕ q ′ ) ⊆ R , and therefore has an anchor in the actual union R R R . There are at most U μ m U\mu^m U μ m such connected supports of size m m m . If m < M ( r ) m<M(r) m < M ( r ) , the component cannot be silent, so its weight contains at least one sech β R \operatorname{sech}\beta_R sech β R factor. For m ≥ M ( r ) m\geq M(r) m ≥ M ( r ) we simply bound all such factors by one. Hence
Δ r , β R ( q , q ′ ) ≤ U [ sech β R ∑ 1 ≤ m < M ( r ) λ m + ∑ m ≥ M ( r ) λ m ] ≤ U 1 − λ [ λ sech β R + λ M ( r ) ] . (218) \begin{aligned}
\Delta_{r,\beta_R}(q,q')\leq U\left[\operatorname{sech}\beta_R\sum_{1\leq m<M(r)}\lambda^m+\sum_{m\geq M(r)}\lambda^m\right] \\
&\leq\frac{U}{1-\lambda}\left[\lambda\operatorname{sech}\beta_R+\lambda^{M(r)}\right]. \tag*{(218)}
\end{aligned} Δ r , β R ( q , q ′ ) ≤ U sech β R 1 ≤ m < M ( r ) ∑ λ m + m ≥ M ( r ) ∑ λ m ≤ 1 − λ U [ λ sech β R + λ M ( r ) ] . (218)
Both sums run over positive integers. If M ( r ) M(r) M ( r ) is not an integer, the second sum starts at ⌈ M ( r ) ⌉ \lceil M(r)\rceil ⌈ M ( r )⌉ , and λ ⌈ M ( r ) ⌉ ≤ λ M ( r ) \lambda^{\lceil M(r)\rceil}\leq\lambda^{M(r)} λ ⌈ M ( r )⌉ ≤ λ M ( r ) . The same family resummation and logical-block average as in Lemma 3.5 finish the proof.
The two terms in (S131) express different mechanisms. Increasing β R \beta_R β R suppresses crossings that leave a protected syndrome. Increasing r r r raises the minimum size of a silent crossing in the geometries below, reducing its contribution through the probability of its support.
The minimum size of a silent component
The operator estimate is common to all the code families considered here. Their differences enter through geometry: a silent component that changes a logical parity must cross enough parallel representatives, or satisfy enough successive syndrome constraints, to acquire a large support.
Supplementary Lemma 3.7 (Toric-code tube geometry). Choose a fixed number of straight, minimum-weight Z Z Z -logical strings in the two- or three-dimensional edge-qubit toric code of linear size L L L . Let r r r be a positive integer, set r 0 = ⌊ r / 2 ⌋ r_0=\lfloor r/2\rfloor r 0 = ⌊ r /2 ⌋ , and assume 2 r 0 + 1 ≤ L 2r_0+1\leq L 2 r 0 + 1 ≤ L . Around each chosen string include parallel translates with displacement at most r 0 r_0 r 0 in each transverse coordinate, and protect all plaquette ribbons joining neighboring translates in this array. Every protected-syndrome-silent distinguishing component satisfies
∣ Γ ∣ ≥ 2 r 0 + 1 ≥ r in two dimensions, (219) |\Gamma|\geq2r_0+1\geq r \qquad\text{in two dimensions,} \tag*{(219)} ∣Γ∣ ≥ 2 r 0 + 1 ≥ r in two dimensions, (219) ∣ Γ ∣ ≥ ( 2 r 0 + 1 ) 2 ≥ r 2 in three dimensions. (220) |\Gamma|\geq(2r_0+1)^2\geq r^2 \qquad\text{in three dimensions.} \tag*{(220)} ∣Γ∣ ≥ ( 2 r 0 + 1 ) 2 ≥ r 2 in three dimensions. (220) The protected ribbons lie in a check tube of radius O ( r ) O(r) O ( r ) around each selected string. Protecting further checks in that tube preserves both bounds.
Proof. A distinguishing component has odd intersection with at least one selected string. Indeed, χ q ( Γ ) ≠ χ q ′ ( Γ ) \chi_q(\Gamma)\ne\chi_{q'}(\Gamma) χ q ( Γ ) = χ q ′ ( Γ ) requires a nonzero parity among the selected representatives. Fix one such string.
In two dimensions, label its parallel translates by ℓ y \ell_y ℓ y , − r 0 ≤ y ≤ r 0 -r_0\leq y\leq r_0 − r 0 ≤ y ≤ r 0 . Let A y A_y A y be the plaquette ribbon joining neighboring translates. The stabilizer relation ℓ y ⊕ ℓ y + 1 = H 1 Z A y \ell_y\oplus\ell_{y+1}=H^Z_1A_y ℓ y ⊕ ℓ y + 1 = H 1 Z A y implies
∣ Γ ∩ ℓ y ∣ + ∣ Γ ∩ ℓ y + 1 ∣ = ∣ σ ( Γ ) ∩ A y ∣ = 0 ( m o d 2 ) . (221) |\Gamma\cap\ell_y|+|\Gamma\cap\ell_{y+1}|=|\sigma(\Gamma)\cap A_y|=0\pmod2. \tag*{(221)} ∣Γ ∩ ℓ y ∣ + ∣Γ ∩ ℓ y + 1 ∣ = ∣ σ ( Γ ) ∩ A y ∣ = 0 ( mod 2 ) . (221)
The last equality holds because every plaquette in the ribbon is protected. Odd intersection with the central string therefore propagates to every one of its 2 r 0 + 1 2r_0+1 2 r 0 + 1 translates. They are edge-disjoint, so Γ \Gamma Γ must contain at least one qubit on each, proving (S140).
In three dimensions label the parallel translates by ℓ y , z \ell_{y,z} ℓ y , z with − r 0 ≤ y , z ≤ r 0 -r_0\le y,z\le r_0 − r 0 ≤ y , z ≤ r 0 . Apply the same parity relation to the protected ribbons in each transverse direction. The transverse square is connected by these neighboring relations, so odd intersection propagates throughout its ( 2 r 0 + 1 ) 2 (2r_0+1)^2 ( 2 r 0 + 1 ) 2 edge-disjoint strings. This proves (S141). The transverse square has ℓ ∞ \ell_\infty ℓ ∞ radius r 0 r_0 r 0 and transverse graph distance at most 2 r 0 ≤ r 2r_0\le r 2 r 0 ≤ r from its center; the ribbons therefore lie in a tube of radius O ( r ) O(r) O ( r ) . The condition 2 r 0 + 1 ≤ L 2r_0+1\le L 2 r 0 + 1 ≤ L ensures that the listed translates are distinct on the torus. If tubes around different selected strings overlap, the argument is unchanged: all ribbons used for the chosen central string remain protected.
For a fixed number of straight representatives, U ≤ C L U\le CL U ≤ C L for a constant C C C . Using the minimum sizes in Lemma 3.7 and counting through anchors therefore yields, uniformly in the logical pair,
Δ r 2 D ( q , q ′ ) ≤ C L 1 − λ λ 2 r 0 + 1 , Δ r 3 D ( q , q ′ ) ≤ C L 1 − λ λ ( 2 r 0 + 1 ) 2 . (222) \Delta_r^{2\mathrm{D}}(q,q')\le\frac{CL}{1-\lambda}\lambda^{2r_0+1},\qquad\Delta_r^{3\mathrm{D}}(q,q')\le\frac{CL}{1-\lambda}\lambda^{(2r_0+1)^2}. \tag*{(222)} Δ r 2 D ( q , q ′ ) ≤ 1 − λ C L λ 2 r 0 + 1 , Δ r 3 D ( q , q ′ ) ≤ 1 − λ C L λ ( 2 r 0 + 1 ) 2 . (222)
For finite-strength enhancement, substitute these same minimum sizes into (S131). The separate Wilson-restoration conditions are proved in Supplementary Note 4.
For more general CSS codes, a confinement inequality can replace the ribbon construction. The linear syndrome bound is the robustness property used in quantum expander decoding [32 , 33 ] . Confinement entered single-shot correction through subsystem codes [34 ] ; good linear confinement was subsequently shown to suffice under local stochastic noise [35 ] . It also enters dimension-jump state preparation [36 ] and passive quantum memory at nonzero temperature [37 ] . Its classical counterpart has been used to establish eigenstate localization in quantum-perturbed code Hamiltonians [38 ] . Here we use the inequality to bound the growth of an error that remains silent throughout the enhanced region. The cutoff in the following definition specifies the range of error weights to which the argument applies.
Supplementary Definition 1 (Linear confinement in the X X X -error sector). Let A μ = X a μ A_\mu=X^{a_\mu} A μ = X a μ be the X X X checks and S X = span F 2 { a μ } S_X=\operatorname{span}_{\mathbb{F}_2}\{a_\mu\} S X = span F 2 { a μ } their support space. For an error X e X^e X e , its measured syndrome is the set σ ( e ) = supp ( H Z e ) \sigma(e)=\operatorname{supp}(H_Ze) σ ( e ) = supp ( H Z e ) , and its reduced weight is
∣ e ∣ r e d = min s ∈ S X ∣ e ⊕ s ∣ . (223) |e|_{\mathrm{red}}=\min_{s\in S_X}|e\oplus s|. \tag*{(223)} ∣ e ∣ red = s ∈ S X min ∣ e ⊕ s ∣. (223) The code has linear confinement up to cutoff T > 0 T>0 T > 0 if
∣ σ ( e ) ∣ ≥ κ ∣ e ∣ r e d whenever ∣ e ∣ r e d < T , (224) |\sigma(e)|\ge\kappa|e|_{\mathrm{red}}\qquad\text{whenever }|e|_{\mathrm{red}}<T, \tag*{(224)} ∣ σ ( e ) ∣ ≥ κ ∣ e ∣ red whenever ∣ e ∣ red < T , (224) for some κ > 0 \kappa>0 κ > 0 . A family of n n n -qubit codes has good linear confinement in this sector if the inequality holds with κ ≥ κ 0 > 0 \kappa\ge\kappa_0>0 κ ≥ κ 0 > 0 and T ≥ c T n α T\ge c_Tn^\alpha T ≥ c T n α , where κ 0 , c T , α > 0 \kappa_0,c_T,\alpha>0 κ 0 , c T , α > 0 are independent of n n n [35 , 36 ] .
Reduced weight here is taken only modulo stabilizers, preserving logical parity as well as syndrome. Errors with the same syndrome but different logical actions are not identified. The adjective “good” specifies the uniform polynomial cutoff; it imposes no code-rate assumption. We require confinement only for X X X errors detected by the measured Z Z Z checks. The constant κ \kappa κ is distinct from the preparation overlap κ β ( x ) \kappa_\beta(x) κ β ( x ) .
Supplementary Lemma 3.8 (Growth of a silent core under confinement). Assume Eq. (224) with constants κ > 0 \kappa>0 κ > 0 and T > 0 T>0 T > 0 . Let N t ( R ) N_t(R) N t ( R ) be the radius-t t t qubit neighborhood of the actual union R R R in G Z G_Z G Z , t ≥ 1 t\ge1 t ≥ 1 be an integer, and assume P r \mathcal{P}_r P r contains every measured check touching N r − 1 ( R ) N_{r-1}(R) N r − 1 ( R ) . If each qubit touches at most Δ Z > 0 \Delta_Z>0 Δ Z > 0 measured checks, then every protected-syndrome-silent error with g ( e ) ≠ 0 g(e)\ne0 g ( e ) = 0 satisfies
∣ e ∣ ≥ min { T , ( 1 + κ Δ Z ) r } . (225) |e|\ge\min\left\{T,\left(1+\frac{\kappa}{\Delta_Z}\right)^r\right\}. \tag*{(225)} ∣ e ∣ ≥ min { T , ( 1 + Δ Z κ ) r } . (225) In particular the right-hand side is a valid choice of the minimum-size lower bound M ( r ) M(r) M ( r ) in Lemma 3.6 .
Proof. Every subset of a minimum-weight representative modulo S X S_X S X is itself minimum weight in its stabilizer coset. Indeed, let a ⊆ e a\subseteq e a ⊆ e with e e e minimum weight modulo S X S_X S X . If some s ∈ S X s\in S_X s ∈ S X reduced the weight of a a a , then
∣ a ⊕ s ∣ < ∣ a ∣ ⟺ ∣ a ∩ s ∣ > ∣ s ∣ 2 . (226) |a\oplus s|<|a|\quad\Longleftrightarrow\quad|a\cap s|>\frac{|s|}{2}. \tag*{(226)} ∣ a ⊕ s ∣ < ∣ a ∣ ⟺ ∣ a ∩ s ∣ > 2 ∣ s ∣ . (226)
It would follow that ∣ e ∩ s ∣ > ∣ s ∣ / 2 |e \cap s| > |s|/2 ∣ e ∩ s ∣ > ∣ s ∣/2 and hence ∣ e ⊕ s ∣ < ∣ e ∣ |e \oplus s| < |e| ∣ e ⊕ s ∣ < ∣ e ∣ , a contradiction. If ∣ e ∣ ≥ T |e| \ge T ∣ e ∣ ≥ T , (S146) is immediate. Otherwise choose a minimum-weight representative e min = e ⊕ s e_{\min} = e \oplus s e m i n = e ⊕ s modulo S X S_X S X , so ∣ e min ∣ ≤ ∣ e ∣ < T |e_{\min}| \le|e| < T ∣ e m i n ∣ ≤ ∣ e ∣ < T . Multiplication by an X X X stabilizer changes neither the Z Z Z syndrome nor intersection parity with any selected Z Z Z logical: H Z s = 0 H_Zs = 0 H Z s = 0 by the CSS commutation relation, and s ⋅ ℓ a = 0 s \cdot\ell_a = 0 s ⋅ ℓ a = 0 because ℓ a \ell_a ℓ a commutes with every X X X stabilizer. Thus e min e_{\min} e m i n is still silent in P r \mathcal{P}_r P r and has g ( e min ) ≠ 0 g(e_{\min}) \ne0 g ( e m i n ) = 0 .
By Lemma 3.2 , each component of e min e_{\min} e m i n is silent in P r \mathcal{P}_r P r , and at least one component Γ \Gamma Γ has g ( Γ ) ≠ 0 g(\Gamma) \ne0 g ( Γ ) = 0 . Being a subset of the reduced support e min e_{\min} e m i n , Γ \Gamma Γ and each of its subsets are reduced. In particular, ∣ Γ ∣ < T |\Gamma| < T ∣Γ∣ < T . Set
Γ t = Γ ∩ N t ( R ) , 0 ≤ t ≤ r . (227) \Gamma_t = \Gamma\cap N_t(R), \qquad0 \le t \le r. \tag*{(227)} Γ t = Γ ∩ N t ( R ) , 0 ≤ t ≤ r . (227)
Nonzero logical parity forces Γ \Gamma Γ to intersect R R R , so ∣ Γ 0 ∣ ≥ 1 |\Gamma_0| \ge1 ∣ Γ 0 ∣ ≥ 1 . For 0 ≤ t ≤ r − 1 0 \le t \le r - 1 0 ≤ t ≤ r − 1 , reducedness and the cutoff allow (S145) to be applied to Γ t \Gamma_t Γ t :
∣ σ ( Γ t ) ∣ ≥ κ ∣ Γ t ∣ . (228) |\sigma(\Gamma_t)| \ge\kappa|\Gamma_t|. \tag*{(228)} ∣ σ ( Γ t ) ∣ ≥ κ ∣ Γ t ∣. (228)
Every check in this syndrome touches N t ( R ) ⊆ N r − 1 ( R ) N_t(R) \subseteq N_{r-1}(R) N t ( R ) ⊆ N r − 1 ( R ) and is therefore protected. The full Γ \Gamma Γ is silent on that check, whereas Γ t \Gamma_t Γ t has odd overlap with it. The check must consequently also touch a qubit of Γ ∖ Γ t \Gamma\setminus\Gamma_t Γ ∖ Γ t to cancel that parity. It shares the check with a qubit of Γ t \Gamma_t Γ t , so common-check adjacency places the added qubit in Γ t + 1 ∖ Γ t \Gamma_{t+1} \setminus\Gamma_t Γ t + 1 ∖ Γ t . Each added qubit can account for at most Δ Z \Delta_Z Δ Z such checks. Counting check–qubit incidences gives
Δ Z ∣ Γ t + 1 ∖ Γ t ∣ ≥ ∣ σ ( Γ t ) ∣ ≥ κ ∣ Γ t ∣ , ∣ Γ t + 1 ∣ ≥ ( 1 + κ Δ Z ) ∣ Γ t ∣ . (229) \Delta_Z|\Gamma_{t+1} \setminus\Gamma_t| \ge|\sigma(\Gamma_t)| \ge\kappa|\Gamma_t|,\qquad|\Gamma_{t+1}| \ge\left(1+\frac{\kappa}{\Delta_Z}\right)|\Gamma_t|. \tag*{(229)} Δ Z ∣ Γ t + 1 ∖ Γ t ∣ ≥ ∣ σ ( Γ t ) ∣ ≥ κ ∣ Γ t ∣ , ∣ Γ t + 1 ∣ ≥ ( 1 + Δ Z κ ) ∣ Γ t ∣. (229)
Iterating r r r times and using ∣ Γ 0 ∣ ≥ 1 |\Gamma_0| \ge1 ∣ Γ 0 ∣ ≥ 1 yields
∣ e ∣ ≥ ∣ e min ∣ ≥ ∣ Γ r ∣ ≥ ( 1 + κ Δ Z ) r . (230) |e| \ge|e_{\min}| \ge|\Gamma_r| \ge\left(1+\frac{\kappa}{\Delta_Z}\right)^r. \tag*{(230)} ∣ e ∣ ≥ ∣ e m i n ∣ ≥ ∣ Γ r ∣ ≥ ( 1 + Δ Z κ ) r . (230)
This proves the claim in the case ∣ e ∣ < T |e| < T ∣ e ∣ < T , and together with the first case proves (S146). Reducing the full error before choosing its component is essential: the confinement hypothesis controls reduced weight, whereas the connected-family sum ranges over arbitrary physical supports.
Combining the appropriate minimum-size bound with Lemma 3.6 gives the width and measurement-strength conditions used in the main text. For uniform asymptotic claims, the graph-counting constant, confinement constant, check incidence, and cutoff must be controlled at the stated scales. Growth in (S146) is guaranteed only up to the cutoff T T T . Throughout, the anchor factor is the actual union size U = ∣ ⋃ a ℓ a ∣ U = |\bigcup_a \ell_a| U = ∣ ⋃ a ℓ a ∣ ; it should not be replaced by the size of a single representative when the selected logical register grows.
A soft–hard comparison for finite-strength CSS filters
A second route to finite-strength recovery compares the prepared state directly with its projection onto the protected sector. This separates the quality of that projection from the geometric recovery argument. The comparison uses nonnegative couplings and unsigned Z Z Z checks, and remains valid when the protected checks are dependent.
Supplementary Lemma 3.9 (Protected-sector probability and AQEC comparison). Consider the normalized CSS-filter state
∣ ψ s o f t ⟩ ∝ exp ( 1 2 ∑ j β j B j ) ∣ + ⟩ ⊗ n , B j = Z h j , β j ≥ 0. (231) |\psi_{\mathrm{soft}}\rangle\propto\exp\left(\frac{1}{2}\sum_j \beta_j B_j\right)|+\rangle^{\otimes n},\qquad B_j = Z^{h_j},\quad\beta_j \ge0. \tag*{(231)} ∣ ψ soft ⟩ ∝ exp ( 2 1 j ∑ β j B j ) ∣ + ⟩ ⊗ n , B j = Z h j , β j ≥ 0. (231) Let P r \mathcal{P}_r P r be a set of checks with β j ≥ β R \beta_j \ge\beta_R β j ≥ β R , and set Π r = ∏ j ∈ P r ( I + B j ) / 2 \Pi_r = \prod_{j\in\mathcal{P}_r}(I+B_j)/2 Π r = ∏ j ∈ P r ( I + B j ) /2 . The probability of a violation of at least one protected check satisfies
q r : = 1 − ⟨ ψ s o f t ∣ Π r ∣ ψ s o f t ⟩ ≤ ∑ j ∈ P r 1 1 + e 2 β j ≤ ∣ P r ∣ e − 2 β R . (232) q_r := 1-\langle\psi_{\mathrm{soft}}|\Pi_r|\psi_{\mathrm{soft}}\rangle\le\sum_{j\in\mathcal{P}_r}\frac{1}{1+e^{2\beta_j}} \le|\mathcal{P}_r|e^{-2\beta_R}. \tag*{(232)} q r := 1 − ⟨ ψ soft ∣ Π r ∣ ψ soft ⟩ ≤ j ∈ P r ∑ 1 + e 2 β j 1 ≤ ∣ P r ∣ e − 2 β R . (232) Define ∣ ψ h a r d ⟩ = Π r ∣ ψ s o f t ⟩ / 1 − q r |\psi_{\mathrm{hard}}\rangle= \Pi_r|\psi_{\mathrm{soft}}\rangle/\sqrt{1-q_r} ∣ ψ hard ⟩ = Π r ∣ ψ soft ⟩ / 1 − q r . Construct the soft and hard logical bases from the same selected Z Z Z representatives, under the logical orthonormality assumptions of the encoding construction, and apply the same physical noise channel. Then their reference–environment states obey
1 2 ∥ ρ R E s o f t − ρ R E h a r d ∥ 1 ≤ q r , ε s o f t ≤ ε h a r d + 2 q r , (233) \frac{1}{2}\left\|\rho_{RE}^{\mathrm{soft}}-\rho_{RE}^{\mathrm{hard}}\right\|_1 \le\sqrt{q_r},\qquad\varepsilon_{\mathrm{soft}} \le\varepsilon_{\mathrm{hard}} + 2\sqrt{q_r}, \tag*{(233)} 2 1 ρ R E soft − ρ R E hard 1 ≤ q r , ε soft ≤ ε hard + 2 q r , (233) where ε = 1 2 ∥ ρ R E − ( I R / K ) ⊗ ρ E ∥ 1 \varepsilon= \frac{1}{2}\left\|\rho_{RE} - (I_R/K)\otimes\rho_E\right\|_1 ε = 2 1 ∥ ρ R E − ( I R / K ) ⊗ ρ E ∥ 1 is the normalized AQEC trace-distance deviation and K K K is the logical dimension.
Proof. Remove only the coupling β j \beta_j β j , and let m j = ⟨ B j ⟩ β j = 0 m_j=\langle B_j\rangle_{\beta_j=0} m j = ⟨ B j ⟩ β j = 0 in the resulting Born measure. Its high-temperature expansion has nonnegative coefficients: each coefficient is a product of nonnegative tanh β i \tanh\beta_i tanh β i , and
⟨ + ∣ ⊗ n B j ∏ i ∈ A B i ∣ + ⟩ ⊗ n ∈ { 0 , 1 } . (234) \langle+|^{\otimes n}B_j\prod_{i\in A}B_i|+\rangle^{\otimes n}\in\{0,1\}. \tag*{(234)} ⟨ + ∣ ⊗ n B j i ∈ A ∏ B i ∣ + ⟩ ⊗ n ∈ { 0 , 1 } . (234)
The denominator has the same positivity and contains an empty term. Therefore m j ≥ 0 m_j\ge0 m j ≥ 0 . Restoring β j \beta_j β j reweights the two values of B j B_j B j by e ± β j e^{\pm\beta_j} e ± β j , giving the exact formula
Pr ( B j = − 1 ) = 1 − m j e 2 β j ( 1 + m j ) + ( 1 − m j ) ≤ 1 1 + e 2 β j . (235) \Pr(B_j=-1)=\frac{1-m_j}{e^{2\beta_j}(1+m_j)+(1-m_j)}\le\frac{1}{1+e^{2\beta_j}}. \tag*{(235)} Pr ( B j = − 1 ) = e 2 β j ( 1 + m j ) + ( 1 − m j ) 1 − m j ≤ 1 + e 2 β j 1 . (235)
The event 1 − Π r 1-\Pi_r 1 − Π r is the union of the protected-check violation events, so a union bound proves Eq. (232) . No independence assumption is involved. Since all checks are unsigned, the simultaneous + 1 +1 + 1 sector has nonzero weight; the normalized projection is well-defined. On this sector each protected filter is scalar. Consequently ∣ ψ h a r d ⟩ |\psi_{\mathrm{hard}}\rangle ∣ ψ hard ⟩ is precisely the hard-protected state with all outside couplings unchanged.
The overlap of the soft and hard preparation states is 1 − q r \sqrt{1-q_r} 1 − q r . The map
∣ φ ⟩ ⟼ 1 K ∑ q ∣ q ⟩ R Z ℓ ( q ) ∣ φ ⟩ (236) |\varphi\rangle\longmapsto\frac{1}{\sqrt{K}}\sum_q|q\rangle_R Z^{\ell(q)}|\varphi\rangle \tag*{(236)} ∣ φ ⟩ ⟼ K 1 q ∑ ∣ q ⟩ R Z ℓ ( q ) ∣ φ ⟩ (236)
is an isometry: its squared norm is the average of K K K identical squared norms, because the reference labels are orthogonal and every Z ℓ ( q ) Z^{\ell(q)} Z ℓ ( q ) is unitary. Each representative also commutes with Π r \Pi_r Π r . The two reference–code purifications therefore have overlap 1 − q r \sqrt{1-q_r} 1 − q r , and the pure-state trace-distance formula [29 ] gives distance q r \sqrt{q_r} q r . Apply a common Stinespring dilation of the noise and trace out the physical output. Contractivity [21 Sec. 3.3] gives the first inequality in Eq. (233) , and the same bound holds for the environment marginals.
Finally, the triangle inequality gives
ϵ s o f t − ϵ h a r d ≤ 1 2 ∥ ρ R E s o f t − ρ R E h a r d ∥ 1 + 1 2 ∥ ρ E s o f t − ρ E h a r d ∥ 1 ≤ 2 q r . (237) \begin{aligned}
\epsilon_{\mathrm{soft}}-\epsilon_{\mathrm{hard}}
&\le\frac{1}{2}\left\|\rho_{RE}^{\mathrm{soft}}-\rho_{RE}^{\mathrm{hard}}\right\|_1+\frac{1}{2}\left\|\rho_E^{\mathrm{soft}}-\rho_E^{\mathrm{hard}}\right\|_1\\
&\le2\sqrt{q_r}. \tag*{(237)}
\end{aligned} ϵ soft − ϵ hard ≤ 2 1 ρ R E soft − ρ R E hard 1 + 2 1 ρ E soft − ρ E hard 1 ≤ 2 q r . (237)
A full transpose, if used to identify these states with the AQEC matrix convention, leaves all trace norms unchanged. □
Combining Lemma 3.9 with (S123) and (S129), and using ∥ B ∥ 1 = 2 ϵ \|B\|_1=2\epsilon ∥ B ∥ 1 = 2 ϵ , gives
∥ B s o f t ∥ 1 ≤ exp [ U λ M ( r ) 1 − λ ] − 1 + 4 ∣ P r ∣ e − β R . (238) \|B_{\mathrm{soft}}\|_1\le\exp\left[\frac{U\lambda^{M(r)}}{1-\lambda}\right]-1+4\sqrt{|\mathcal{P}_r|}e^{-\beta_R}. \tag*{(238)} ∥ B soft ∥ 1 ≤ exp [ 1 − λ U λ M ( r ) ] − 1 + 4 ∣ P r ∣ e − β R . (238)
This proves the main-text finite-strength enhancement theorem. The counting convention used here absorbs the local constant into μ \mu μ . At fixed k k k , U λ M ( r ) → 0 U\lambda^{M(r)}\to0 U λ M ( r ) → 0 and ∣ P r ∣ e − β R → 0 \sqrt{|\mathcal{P}_r|}e^{-\beta_R}\to0 ∣ P r ∣ e − β R → 0 imply both D A Q E C → 0 D_{\mathrm{AQEC}}\to0 D AQEC → 0 and F e ⋆ → 1 F_e^\star\to1 F e ⋆ → 1 by Lemma 2.9. The direct estimate in Lemma 3.6 remains a separate bound. Its sufficient condition U s e c h β R → 0 U\mathop{\mathrm{sech}}\beta_R\to0 U sech β R → 0 is stronger than the soft–hard condition for the toric neighborhoods below, and does not supply their stated strength coefficient for every δ > 0 \delta>0 δ > 0 . The dependence on a growing logical register is treated at the end of this Note.
Resource estimates and the logarithmic strength scale
The two suppression mechanisms lead to separate choices of protection width and measurement strength. Their combination gives the main-text corollary on toric resources and logarithmic strength. The distance d d d in these statements is the distance of the ideal code; in the two toric families under consideration, d = L d=L d = L . The exact Knill–Laflamme distance of the finite-strength subspace is not being used as a growing parameter. Take a fixed number of straight representatives, so U ≤ C d U\le Cd U ≤ C d , and choose a fixed positive noise envelope 0 < λ 0 < 1 0<\lambda_0<1 0 < λ 0 < 1 with λ ≤ λ 0 \lambda\le\lambda_0 λ ≤ λ 0 . This also includes zero physical noise and keeps the explicit width formulas below well defined.
Fix δ > 0 \delta>0 δ > 0 . Lemma 3.7 permits the choices
r 2 D = ⌈ ( 1 + δ ) log d ∣ log λ 0 ∣ ⌉ , r 3 D = ⌈ ( 1 + δ ) log d ∣ log λ 0 ∣ ⌉ . (239) \begin{aligned}
r_{2\mathrm{D}}&=\left\lceil\frac{(1+\delta)\log d}{|\log\lambda_0|}\right\rceil,\\
r_{3\mathrm{D}}&=\left\lceil\sqrt{\frac{(1+\delta)\log d}{|\log\lambda_0|}}\right\rceil. \tag*{(239)}
\end{aligned} r 2 D r 3 D = ⌈ ∣ log λ 0 ∣ ( 1 + δ ) log d ⌉ , = ⌈ ∣ log λ 0 ∣ ( 1 + δ ) log d ⌉ . (239)
The corresponding minimum silent sizes obey λ M ( r ) ≤ d − 1 − δ \lambda^{M(r)} \le d^{-1-\delta} λ M ( r ) ≤ d − 1 − δ . Therefore (S143) and Lemma 3.5 give
ε h a r d = O ( d − δ ) . (240) \varepsilon_{\mathrm{hard}} = O(d^{-\delta}). \tag*{(240)} ε hard = O ( d − δ ) . (240)
All tube widths are sublinear in L L L , so the distinct-translate assumption of Lemma 3.7 holds for sufficiently large d d d . A two-dimensional strip contains O ( d r ) O(dr) O ( d r ) protected plaquettes. A three-dimensional tube contains O ( d r 2 ) O(dr^2) O ( d r 2 ) : there are O ( r 2 ) O(r^2) O ( r 2 ) transverse positions along a string of length d d d , and each position touches a bounded number of plaquettes. Both choices in (S160) consequently use
∣ P r ∣ = O ( d log d ) . (241) |\mathcal{P}_r| = O(d\log d). \tag*{(241)} ∣ P r ∣ = O ( d log d ) . (241)
With a common enhanced strength
β R = ( 1 2 + δ ) log d , (242) \beta_R = \left(\frac{1}{2}+\delta\right)\log d, \tag*{(242)} β R = ( 2 1 + δ ) log d , (242)
Lemma 3.9 gives
q r = O ( d − 2 δ log d ) , ε s o f t = O ( d − δ log d ) ⟶ 0. (243) q_r = O(d^{-2\delta}\log d), \qquad\varepsilon_{\mathrm{soft}} = O\left(d^{-\delta}\sqrt{\log d}\right) \longrightarrow0. \tag*{(243)} q r = O ( d − 2 δ log d ) , ε soft = O ( d − δ log d ) ⟶ 0. (243)
The recovery and entropy bounds in Lemma 2.9 then imply F e ∗ → 1 F_e^* \to1 F e ∗ → 1 and D A Q E C → 0 D_{\mathrm{AQEC}} \to0 D AQEC → 0 for fixed k k k . The bulk measurement strength can remain fixed and finite throughout this construction. The widths differ, although the two constructions have the same asymptotic enhanced-check count. Counting all measured plaquette checks gives N c h k = d 2 N_{\mathrm{chk}} = d^2 N chk = d 2 in two dimensions and N c h k = 3 d 3 N_{\mathrm{chk}} = 3d^3 N chk = 3 d 3 in three, including the checks related by stabilizer redundancies. Hence
∣ P r ∣ N c h k = { O ( log d / d ) , two dimensions , O ( log d / d 2 ) , three dimensions . (244) \frac{|\mathcal{P}_r|}{N_{\mathrm{chk}}}
=
\begin{cases}
O(\log d/d), & \text{two dimensions},\\
O(\log d/d^2), & \text{three dimensions}.
\end{cases} \tag*{(244)} N chk ∣ P r ∣ = { O ( log d / d ) , O ( log d / d 2 ) , two dimensions , three dimensions . (244)
These are the first two fractions in the main-text resource table.
The logarithmic strength scale is also necessary within these protocols. To compare the lower and upper statements, take every enhanced check to have strength exactly β R \beta_R β R and keep the exterior strength fixed. Along one straight selected representative, its d d d edges have the same nonzero logical membership column and pairwise disjoint syndromes. Every syndrome lies inside the protected tube and contains two checks in two dimensions or four in three dimensions. The proof of Lemma 2.7 thus applies with the common incident strength β R \beta_R β R : it uses the lower kernel bound for a single edge and the upper kernel bound for a pair of these edges, and both involve only their incident checks. The exterior couplings do not enter these bounds. Consequently, vanishing total AQEC entropy requires
β R − 1 4 log d ⟶ + ∞ in two dimensions , \beta_R - \frac{1}{4}\log d \longrightarrow+\infty\quad\text{in two dimensions}, β R − 4 1 log d ⟶ + ∞ in two dimensions ,
β R − 1 8 log d ⟶ + ∞ in three dimensions . (245) \beta_R - \frac{1}{8}\log d \longrightarrow+\infty\quad\text{in three dimensions}. \tag*{(245)} β R − 8 1 log d ⟶ + ∞ in three dimensions . (245)
For fixed k k k , these are necessary for asymptotically faithful recovery as well, by Lemma 2.9. Together with (S163), they establish necessity and sufficiency of the logarithmic order. They do not identify an optimal leading coefficient or a necessary protection width. If the upper-bound hypothesis is stated only as β j ≥ β R \beta_j \ge\beta_R β j ≥ β R , its lower parameter β R \beta_R β R alone does not constrain the actual strengths from above; the necessity statement concerns the common strength used by the specified protocol.
For a family satisfying the confinement hypothesis of Lemma 3.8 , write γ = 1 + κ / Δ z > 1 \gamma= 1 + \kappa/\Delta_z > 1 γ = 1 + κ / Δ z > 1 . The sufficient hard-state condition is
∣ log λ 0 ∣ min ( T , γ r ) − log U ⟶ + ∞ . (246) |\log\lambda_0| \min(T,\gamma^r) - \log U \longrightarrow+\infty. \tag*{(246)} ∣ log λ 0 ∣ min ( T , γ r ) − log U ⟶ + ∞. (246)
Assume U → ∞ U \to\infty U → ∞ . It suffices to have T ≥ c log U T \ge c\log U T ≥ c log U and γ r ≥ c log U \gamma^r \ge c\log U γ r ≥ c log U , with c > 1 / ∣ log λ 0 ∣ c > 1/|\log\lambda_0| c > 1/∣ log λ 0 ∣ . Good linear confinement provides this cutoff because U ≤ n U \le n U ≤ n and its polynomial cutoff eventually exceeds c log U c\log U c log U . Taking
r = ⌈ log ( c log U ) log γ ⌉ (247) r = \left\lceil\frac{\log(c\log U)}{\log\gamma} \right\rceil \tag*{(247)} r = ⌈ log γ log ( c log U ) ⌉ (247)
gives r = O ( log log U ) r = O(\log\log U) r = O ( log log U ) and U λ M ( r ) ≤ U 1 − c ∣ log λ 0 ∣ → 0 U\lambda^{M(r)} \le U^{1-c|\log\lambda_0|} \to0 U λ M ( r ) ≤ U 1 − c ∣ l o g λ 0 ∣ → 0 .
The enhanced-check count follows from the bounded degree of G Z G_Z G Z . Let D G ≥ 2 D_G \ge2 D G ≥ 2 be a uniform degree bound. The union of radius-r r r neighborhoods about the U U U anchor qubits contains at most C U ( D G + 1 ) r C_U(D_G+1)^r C U ( D G + 1 ) r qubits. Each qubit touches at most Δ Z \Delta_Z Δ Z measured checks, so enhancing exactly the checks touching N r − 1 ( R ) N_{r-1}(R) N r − 1 ( R ) requires
∣ P r ∣ ≤ min { N c h k , C U ( log U ) ν } , ν = log ( D G + 1 ) log γ . (248) |\mathcal{P}_r| \le\min\{N_{\mathrm{chk}},C_U(\log U)^\nu\}, \qquad
\nu= \frac{\log(D_G+1)}{\log\gamma}. \tag*{(248)} ∣ P r ∣ ≤ min { N chk , C U ( log U ) ν } , ν = log γ log ( D G + 1 ) . (248)
Here C C C is independent of code size, and N c h k N_{\mathrm{chk}} N chk counts the same measured Z Z Z checks as the numerator. For β R = ( 1 2 + δ ) log U \beta_R = \left(\frac{1}{2}+\delta\right)\log U β R = ( 2 1 + δ ) log U , the soft–hard correction is bounded by
4 ∣ P r ∣ e − β R = O ( U − δ ( log U ) ν / 2 ) ⟶ 0. (249) 4\sqrt{|\mathcal{P}_r|}e^{-\beta_R}=O\left(U^{-\delta}(\log U)^{\nu/2}\right)\longrightarrow0. \tag*{(249)} 4 ∣ P r ∣ e − β R = O ( U − δ ( log U ) ν /2 ) ⟶ 0. (249)
Together with the hard-state estimate, this proves the confinement row in the main-text resource table for fixed k k k . Its enhanced fraction vanishes under the additional condition U ( log U ) ν = o ( N c h k ) U(\log U)^\nu=o(N_{\mathrm{chk}}) U ( log U ) ν = o ( N chk ) .
More generally, for finite strengths and nonempty protected sets, Lemma 3.9 adds the condition
β R − 1 2 log ∣ P r ∣ ⟶ + ∞ . (250) \beta_R-\frac{1}{2}\log|\mathcal{P}_r|\longrightarrow+\infty. \tag*{(250)} β R − 2 1 log ∣ P r ∣ ⟶ + ∞. (250)
If the protected set is empty, the soft–hard correction is exactly zero; the equivalent product condition ∣ P r ∣ e − β R → 0 \sqrt{|\mathcal{P}_r|}e^{-\beta_R}\to0 ∣ P r ∣ e − β R → 0 remains well defined in either case. Alternatively, the direct finite-strength bound in Lemma 3.6 uses β R − log U → + ∞ \beta_R-\log U\to+\infty β R − log U → + ∞ . Which estimate is smaller depends on the actual enhanced-check count. The confinement inequality controls the minimum silent support only up to T T T ; it gives neither an unrestricted exponential growth law nor a small aggregate footprint by itself.
A growing logical register
Protecting a growing register requires accounting for all vulnerable logical supports and for the distinction between total leakage and leakage density. The recovery theorem also applies when the selected number k k k and the union size U U U grow, provided the local constants remain uniform. The average over logical pairs in (S123) prevents an additional factor K = 2 k K=2^k K = 2 k from appearing in the trace-norm estimate. In particular, the direct finite-strength bound gives
ε ≤ 1 2 { exp [ U 1 − λ ( λ sech β R + λ M ( r ) ) ] − 1 } . (251) \varepsilon\le\frac{1}{2}\left\{\exp\left[\frac{U}{1-\lambda}\left(\lambda\operatorname{sech}\beta_R+\lambda^{M(r)}\right)\right]-1\right\}. \tag*{(251)} ε ≤ 2 1 { exp [ 1 − λ U ( λ sech β R + λ M ( r ) ) ] − 1 } . (251)
With a fixed 0 < λ 0 < 1 0<\lambda_0<1 0 < λ 0 < 1 and λ ≤ λ 0 \lambda\le\lambda_0 λ ≤ λ 0 , it therefore suffices that
β R − log U ⟶ + ∞ , ∣ log λ 0 ∣ M ( r ) − log U ⟶ + ∞ (252) \beta_R-\log U\longrightarrow+\infty,\qquad|\log\lambda_0|M(r)-\log U\longrightarrow+\infty \tag*{(252)} β R − log U ⟶ + ∞ , ∣ log λ 0 ∣ M ( r ) − log U ⟶ + ∞ (252)
to obtain ε → 0 \varepsilon\to0 ε → 0 and faithful entanglement recovery of the entire selected register. Its entropy per logical qubit vanishes as well. Indeed, Lemma 2.9 implies
D A Q E C k ≤ ε log 2 + 1 + ε k h 2 ( ε 1 + ε ) ⟶ 0. (253) \frac{D_{\mathrm{AQEC}}}{k}\le\varepsilon\log2+\frac{1+\varepsilon}{k}h_2\left(\frac{\varepsilon}{1+\varepsilon}\right)\longrightarrow0. \tag*{(253)} k D AQEC ≤ ε log 2 + k 1 + ε h 2 ( 1 + ε ε ) ⟶ 0. (253)
To force the total entropy to vanish, it suffices instead that k ε → 0 k\varepsilon\to0 k ε → 0 . Applied to (S172), this amounts to replacing U U U by k U kU k U in both sufficient conditions (S173). Under confinement, the minimum M ( r ) ≥ min ( T , γ r ) M(r)\ge\min(T,\gamma^r) M ( r ) ≥ min ( T , γ r ) is used in either version; the cutoff must reach the logarithm of U U U or of k U kU k U , respectively.
The soft–hard route has a corresponding distinction. Full-register recovery follows from a vanishing hard-state bound and ∣ P r ∣ e − β R → 0 \sqrt{|\mathcal{P}_r|}e^{-\beta_R}\to0 ∣ P r ∣ e − β R → 0 . A sufficient condition for vanishing total entropy is
k U λ M ( r ) → 0 , k ∣ P r ∣ e − β R → 0. (254) kU\lambda^{M(r)}\to0,\qquad k\sqrt{|\mathcal{P}_r|}e^{-\beta_R}\to0. \tag*{(254)} k U λ M ( r ) → 0 , k ∣ P r ∣ e − β R → 0. (254)
These explicit factors matter when many representatives are protected. The anchor count is the size of their actual union, not the length of one representative.
Finally, consider the aggregate lower bound under independent noise at fixed 0 < p < 1 0<p<1 0 < p < 1 . Suppose checks outside an enhanced set P h i \mathcal{P}_{hi} P hi have strength at most fixed β 0 \beta_0 β 0 , and let F h i = ⋃ j ∈ P h i supp B j F_{hi}=\bigcup_{j\in\mathcal{P}_{hi}}\operatorname{supp} B_j F hi = ⋃ j ∈ P hi supp B j be the qubits touched by that set. The binary matrix with rows ℓ 1 , … , ℓ k \ell_1,\ldots,\ell_k ℓ 1 , … , ℓ k has rank k k k . Deleting its columns in F h i F_{hi} F hi reduces rank by at most ∣ F h i ∣ |F_{hi}| ∣ F hi ∣ . Among the remaining columns, one can therefore select at least ( k − ∣ F h i ∣ ) + (k-|F_{\mathrm{hi}}|)_+ ( k − ∣ F hi ∣ ) + independent logical parities. Their qubits touch only checks with strength at most β 0 \beta_0 β 0 , giving κ β ( e ) ≥ e − β 0 Δ Z \kappa_\beta(e) \ge e^{-\beta_0\Delta_Z} κ β ( e ) ≥ e − β 0 Δ Z . Applying Lemma 2.6 to these independent parities yields
D A Q E C k ≥ 2 p ( 1 − p ) e − 2 β 0 Δ Z ( 1 − ∣ F h i ∣ k ) + . (255) \frac{D_{\mathrm{AQEC}}}{k} \ge2p(1-p)e^{-2\beta_0\Delta_Z}\left(1-\frac{|F_{\mathrm{hi}}|}{k}\right)_+ . \tag*{(255)} k D AQEC ≥ 2 p ( 1 − p ) e − 2 β 0 Δ Z ( 1 − k ∣ F hi ∣ ) + . (255)
Here ( x ) + = max { x , 0 } (x)_+ = \max\{x,0\} ( x ) + = max { x , 0 } . If each measured check has weight at most a constant w Z w_Z w Z , then ∣ F h i ∣ ≤ w Z ∣ P h i ∣ |F_{\mathrm{hi}}| \le w_Z|\mathcal{P}_{\mathrm{hi}}| ∣ F hi ∣ ≤ w Z ∣ P hi ∣ . Vanishing entropy density requires ∣ F h i ∣ ≥ [ 1 − o ( 1 ) ] k |F_{\mathrm{hi}}| \ge[1-o(1)]k ∣ F hi ∣ ≥ [ 1 − o ( 1 )] k . At constant encoding rate k = Θ ( n ) k=\Theta(n) k = Θ ( n ) , this entails Ω ( n ) \Omega(n) Ω ( n ) enhanced checks, as stated in the main-text corollary on the cost of protecting a growing logical register. An O ( log log n ) O(\log\log n) O ( log log n ) protection radius can coexist with this extensive footprint. Conversely, a vanishing encoding rate alone does not ensure that the union of selected representatives is subextensive.
Supplementary Note 4: Enhanced Wilson Loops and Logical Operations
Recovering each selected encoding does not yet establish that two distant representatives implement the same logical action. This Supplementary Note addresses that remaining question through the main-text Wilson-restoration theorem and the encoded-CNOT corollary. The bulk coupling remains fixed while only checks near the representatives are strengthened. In three dimensions, the finite-torus gauge measure in Eq. (97) turns Wilson restoration into a bound on dual clusters that obstruct a plaquette filling. All chains and intersection numbers have coefficients in F 2 \mathbb{F}_2 F 2 ; ∂ \partial ∂ and δ \delta δ denote boundary and coboundary.
The two-dimensional obstruction
Take two straight representatives separated by a a a , with α L ≤ a ≤ ( 1 − α ) L \alpha L \le a \le(1-\alpha)L α L ≤ a ≤ ( 1 − α ) L for fixed 0 < α < 1 / 2 0<\alpha<1/2 0 < α < 1/2 . Enhanced strips of width r = o ( L ) r=o(L) r = o ( L ) leave order L 2 L^2 L 2 unprotected plaquettes in each complementary filling. Lemma 1.2 therefore gives
0 ≤ ⟨ W ⟩ ≤ 2 exp [ − c ∣ log tanh β ∣ L 2 ] ⟶ 0 0 \le\langle W\rangle\le2\exp[-c|\log\tanh\beta|L^2] \longrightarrow0 0 ≤ ⟨ W ⟩ ≤ 2 exp [ − c ∣ log tanh β ∣ L 2 ] ⟶ 0
for fixed 0 < β < ∞ 0<\beta<\infty 0 < β < ∞ , whatever the nonnegative couplings inside the tubes, including infinite couplings. At β = 0 \beta=0 β = 0 , the expectation is exactly zero as long as each complementary filling contains an unprotected plaquette. This proves the two-dimensional part of the main-text Wilson-restoration theorem; its condition on unprotected area is essential.
Finite-torus duality and the subcritical probability bound
The three-dimensional argument adapts the random-cluster and dual-linking methods used in the perimeter-law proof of Duncan and Schweinhart [15 Secs. 5, 7, and 9] . Here we need finite-volume estimates for two winding representatives with enhanced couplings near their supports. Since W ( C ) W(C) W ( C ) is diagonal in the computational basis, its expectation in the prepared quantum state is exactly its average in the gauge measure of Eq. (97) . The plaquette random-cluster representation expresses this average as the probability that an open filling of C C C exists. We derive this identity below from the coupling construction of Ref. [15 Sec. 5.1] , allowing each plaquette its own coupling. On the torus, winding dual clusters modify the induced dual measure through a bounded topological weight.
The local gauge action satisfies β p β p ∗ = 2 β p β p ∗ + 1 − β p \beta_p\beta_{p^*}=2\beta_p\beta_{p^*+1}-\beta_p β p β p ∗ = 2 β p β p ∗ + 1 − β p , so the indicator-action coupling used in Refs. [15 , 16 ] is 2 β p 2\beta_p 2 β p . The resulting plaquette random-cluster model has cluster weight q = 2 q=2 q = 2 . Let A A A consist of its open primal plaquettes and the full primal one-skeleton. A dual edge e = p ∗ e=p^* e = p ∗ is open in the complementary graph D D D precisely when the corresponding primal plaquette p p p is absent from A A A . The graph D D D includes all dual vertices, including isolated ones. The configuration D D D describes the auxiliary statistical model used to evaluate the prepared-state overlap.
Supplementary Lemma 4.1 (Exact dual law and the Wilson filling event). For each primal plaquette p p p and its dual edge e = p ∗ e=p^* e = p ∗ , set
θ p = 1 − e − 2 β p , p e ∗ = 2 ( 1 − θ p ) θ p + 2 ( 1 − θ p ) = 2 e 2 β p + 1 . (256) \theta_p=1-e^{-2\beta_p},\qquad p_e^*=\frac{2(1-\theta_p)}{\theta_p+2(1-\theta_p)}=\frac{2}{e^{2\beta_p}+1}. \tag*{(256)} θ p = 1 − e − 2 β p , p e ∗ = θ p + 2 ( 1 − θ p ) 2 ( 1 − θ p ) = e 2 β p + 1 2 . (256)
Here θ p \theta_p θ p and p e ∗ p_e^* p e ∗ are parameters of the primal plaquette and dual edge weights, respectively. They are not the marginal occupation probabilities: the cluster weights correlate the occupations. Neither is the physical noise probability used in the AQEC calculation. The Fortuin–Kasteleyn (FK) random-cluster measure on the periodic dual graph is the probability distribution [19 , 20 ]
ϕ L , p ∗ , 2 ( D ) ∝ 2 c ( D ) ∏ e ∈ D p e ∗ ∏ e ∉ D ( 1 − p e ∗ ) , (257) \phi_{L,p^*,2}(D) \propto2^{c(D)} \prod_{e\in D} p_e^* \prod_{e\notin D} (1-p_e^*), \tag*{(257)} ϕ L , p ∗ , 2 ( D ) ∝ 2 c ( D ) e ∈ D ∏ p e ∗ e ∈ / D ∏ ( 1 − p e ∗ ) , (257)
normalized by summing these weights over all edge configurations. Here c ( D ) c(D) c ( D ) counts the connected components of the periodic graph, including isolated vertices. The cluster weight is q = 2 q=2 q = 2 ; summing over the two Ising spin assignments on each component gives the factor 2 c ( D ) 2^{c(D)} 2 c ( D ) . For a dual configuration D D D , define
h ( D ) = rank im [ H 1 ( D ) ⟶ H 1 ( T L 3 ) ] ∈ { 0 , 1 , 2 , 3 } . (258) h(D)=\operatorname{rank}\operatorname{im}\left[H_1(D)\longrightarrow H_1(\mathbb{T}_L^3)\right]\in\{0,1,2,3\}. \tag*{(258)} h ( D ) = rank im [ H 1 ( D ) ⟶ H 1 ( T L 3 ) ] ∈ { 0 , 1 , 2 , 3 } . (258)
The induced dual law is
ν ∗ ( D ) = 2 − h ( D ) ϕ L , p ∗ , 2 ( D ) ∑ D ′ 2 − h ( D ′ ) ϕ L , p ∗ , 2 ( D ′ ) . (259) \nu^*(D)=\frac{2^{-h(D)}\phi_{L,p^*,2}(D)}{\sum_{D'}2^{-h(D')}\phi_{L,p^*,2}(D')}. \tag*{(259)} ν ∗ ( D ) = ∑ D ′ 2 − h ( D ′ ) ϕ L , p ∗ , 2 ( D ′ ) 2 − h ( D ) ϕ L , p ∗ , 2 ( D ) . (259)
For every C = ∂ S C=\partial S C = ∂ S ,
⟨ W ( C ) ⟩ = Pr { C is the boundary of a chain of open primal plaquettes } . (260) \langle W(C)\rangle=\Pr\{C\text{ is the boundary of a chain of open primal plaquettes}\}. \tag*{(260)} ⟨ W ( C )⟩ = Pr { C is the boundary of a chain of open primal plaquettes } . (260)
The probability is taken over A A A in the gauge/random-cluster coupling. In particular, for every dual-edge event E E E ,
1 8 ϕ L , p ∗ , 2 ( E ) ≤ ν ∗ ( E ) ≤ 8 ϕ L , p ∗ , 2 ( E ) , (261) \frac{1}{8}\phi_{L,p^*,2}(E)\leq\nu^*(E)\leq8\phi_{L,p^*,2}(E), \tag*{(261)} 8 1 ϕ L , p ∗ , 2 ( E ) ≤ ν ∗ ( E ) ≤ 8 ϕ L , p ∗ , 2 ( E ) , (261)
uniformly in the local couplings.
Proof. The finite-volume coupling in Ref. [15 Sec. 5.1] is stated for uniform parameters. Its product construction gives the inhomogeneous coupling used here, as the calculation below shows. For torus duality, Sec. 4.3 of Ref. [16 ] introduces a homologically reweighted, or balanced, measure and proves its exact duality in Theorem 18. We retain the ordinary gauge weights and compute their dual explicitly for q = 2 q=2 q = 2 , obtaining the factor 2 − h ( D ) 2^{-h(D)} 2 − h ( D ) in Eq. (259) .
The identity e β p B p = e β p [ ( 1 − θ p ) + θ p 1 { B p = + 1 } ] e^{\beta_pB_p}=e^{\beta_p}[(1-\theta_p)+\theta_p1_{\{B_p=+1\}}] e β p B p = e β p [( 1 − θ p ) + θ p 1 { B p = + 1 } ] gives the marginal weight
Pr ( A ) ∝ ∏ p ∈ A θ p ∏ p ∉ A ( 1 − θ p ) 2 b 1 ( A ) , (262) \Pr(A)\propto\prod_{p\in A}\theta_p\prod_{p\notin A}(1-\theta_p)2^{b_1(A)}, \tag*{(262)} Pr ( A ) ∝ p ∈ A ∏ θ p p ∈ / A ∏ ( 1 − θ p ) 2 b 1 ( A ) , (262)
where b i ( A ) = dim H i ( A ; F 2 ) b_i(A)=\dim H_i(A;\mathbb{F}_2) b i ( A ) = dim H i ( A ; F 2 ) . To see the power of two, write N e N_e N e and N v N_v N v for the numbers of primal edges and vertices. The number of edge cochains satisfying the checks selected by A A A is 2 N e − rank ( ∂ 2 ∣ A ) = 2 N v − 1 + b 1 ( A ) 2^{N_e-\operatorname{rank}(\partial_2|_A)}=2^{N_v-1+b_1(A)} 2 N e − rank ( ∂ 2 ∣ A ) = 2 N v − 1 + b 1 ( A ) . The factor 2 N v − 1 2^{N_v-1} 2 N v − 1 is independent of A A A . Conditional on A A A , the allowed cochains are uniform in this linear space. The character ( − 1 ) s ( C ) (-1)^{s(C)} ( − 1 ) s ( C ) averages to one if C ∈ im ( ∂ 2 ∣ A ) C\in\operatorname{im}(\partial_2|_A) C ∈ im ( ∂ 2 ∣ A ) and to zero otherwise. This is linear duality over F 2 \mathbb{F}_2 F 2 , and proves Eq. (260) .
Put N = L 3 N=L^3 N = L 3 and retain c ( D ) c(D) c ( D ) for the number of dual connected components. The ranks needed to dualize Eq. (262) are
b 2 ( A ) = c ( D ) − 1 + 3 − h ( D ) , b 1 ( A ) = ∣ D ∣ − N + c ( D ) + 3 − h ( D ) . (263) \begin{aligned}
b_2(A)&=c(D)-1+3-h(D),\\
b_1(A)&=|D|-N+c(D)+3-h(D). \tag*{(263)}
\end{aligned} b 2 ( A ) b 1 ( A ) = c ( D ) − 1 + 3 − h ( D ) , = ∣ D ∣ − N + c ( D ) + 3 − h ( D ) . (263)
For the first identity, dualize primal two-cycles to dual one-cocycles and restrict them to D D D . Exact dual one-cocycles, restricted to this graph, form a space of dimension N − c ( D ) N-c(D) N − c ( D ) . Modulo this space, the restrictions of the three global cohomology classes have rank h ( D ) h(D) h ( D ) , by the homology–cohomology pairing. The full space of primal two-cycles has dimension ( N − 1 ) + 3 (N-1)+3 ( N − 1 ) + 3 . The kernel of restriction therefore has dimension ( N − 1 ) + 3 − [ N − c ( D ) + h ( D ) ] (N-1)+3-[N-c(D)+h(D)] ( N − 1 ) + 3 − [ N − c ( D ) + h ( D )] . This kernel consists precisely of the two-cycles supported in A A A . Since the complex A A A has no three-cells, its dimension is b 2 ( A ) b_2(A) b 2 ( A ) , proving the first identity. The full one-skeleton is connected. Applying Euler’s identity with N e = 3 N N_e=3N N e = 3 N and ∣ A ∣ = 3 N − ∣ D ∣ |A|=3N-|D| ∣ A ∣ = 3 N − ∣ D ∣ gives the second identity.
Substituting Eq. (263) into Eq. (262) produces the factor 2 ∣ D ∣ 2 c ( D ) 2 − h ( D ) 2^{|D|}2^{c(D)}2^{-h(D)} 2 ∣ D ∣ 2 c ( D ) 2 − h ( D ) , apart from a constant. The first factor changes the local odds according to
p e ∗ 1 − p e ∗ = 2 ( 1 − θ p ) θ p , e = p ∗ . (264) \frac{p_e^*}{1-p_e^*}=\frac{2(1-\theta_p)}{\theta_p},\qquad e=p^*. \tag*{(264)} 1 − p e ∗ p e ∗ = θ p 2 ( 1 − θ p ) , e = p ∗ . (264)
After normalization, the remaining product is the ordinary FK weight 2 c ( D ) ∏ e ∈ D p e ∗ ∏ e ∉ D ( 1 − p e ∗ ) 2^{c(D)}\prod_{e\in D}p_e^*\prod_{e\notin D}(1-p_e^*) 2 c ( D ) ∏ e ∈ D p e ∗ ∏ e ∈ / D ( 1 − p e ∗ ) multiplied by 2 − h ( D ) 2^{-h(D)} 2 − h ( D ) . This proves Eq. (259) . Both the tilt and its normalizing sum belong to [ 1 / 8 , 1 ] [1/8,1] [ 1/8 , 1 ] , proving Eq. (261) . The calculation initially assumes finite positive couplings; zero and infinite couplings follow by continuity on the finite torus. ∎
The factor 2 − h ( D ) 2^{-h(D)} 2 − h ( D ) accounts for the winding cycles of the dual graph. Its bounded size allows estimates from periodic FK to be transferred to the gauge model with a uniform constant.
Supplementary Lemma 4.2 (Uniform subcritical cluster tail). Fix a background coupling satisfying
p ∗ ( β ) = 2 e 2 β + 1 < p c F K ( 2 ; Z 3 ) . (265) p^*(\beta)=\frac{2}{e^{2\beta}+1}<p_c^{\mathrm{FK}}(2;\mathbb{Z}^3). \tag*{(265)} p ∗ ( β ) = e 2 β + 1 2 < p c FK ( 2 ; Z 3 ) . (265) Here p c F K ( 2 ; Z 3 ) p_c^{\mathrm{FK}}(2;\mathbb{Z}^3) p c FK ( 2 ; Z 3 ) is the critical bond parameter of the homogeneous random-cluster model on the infinite cubic lattice with cluster weight q = 2 q=2 q = 2 . For every inhomogeneous profile 0 ≤ p e ∗ ≤ p ∗ ( β ) 0\le p^*_e\le p^*(\beta) 0 ≤ p e ∗ ≤ p ∗ ( β ) , there are positive constants C β , ξ β , c β C_\beta,\xi_\beta,c_\beta C β , ξ β , c β , independent of L L L and of the profile, such that for every dual vertex v v v and 1 ≤ R ≤ L / 4 1\le R\le L/4 1 ≤ R ≤ L /4 ,
ν ∗ { v ⟷ ∂ B R ( v ) } ≤ C β e − R / ξ β , (266) \nu^*\{v\longleftrightarrow\partial B_R(v)\}\le C_\beta e^{-R/\xi_\beta}, \tag*{(266)} ν ∗ { v ⟷ ∂ B R ( v )} ≤ C β e − R / ξ β , (266) ν ∗ { some cluster has periodic diameter at least L / 4 } ≤ C β L 3 e − c β L . (267) \nu^*\{\text{some cluster has periodic diameter at least }L/4\}\le C_\beta L^3e^{-c_\beta L}. \tag*{(267)} ν ∗ { some cluster has periodic diameter at least L /4 } ≤ C β L 3 e − c β L . (267) Here B R ( v ) B_R(v) B R ( v ) is a ball in the graph distance of the dual cubic lattice; for the stated range of R R R it lifts to a ball in Z 3 \mathbb{Z}^3 Z 3 . The connection event means an open path from v v v to the boundary of this ball. Periodic diameter refers to the ambient periodic lattice distance, not the internal path length of a cluster.
Proof. First work with the ordinary periodic FK measure. Conditioning on all edges outside the lifted ball B R ( v ) B_R(v) B R ( v ) induces a partition of its boundary vertices according to their connections through the exterior. For q = 2 q=2 q = 2 , the fully wired boundary condition, which identifies all boundary vertices, dominates every such partition. These comparisons and monotonicity in the edge parameters are standard random-cluster properties [20 Chs. 3–4] . They bound the conditional connection probability by that of the homogeneous wired ball with parameter p ∗ ( β ) p^*(\beta) p ∗ ( β ) .
The finite wired-volume estimate of Theorem 1.2 in Ref. [39] is also used as Theorem 36 in the perimeter-law proof of Ref. [15]. To apply it, take unit nearest-neighbor couplings J e = 1 J_e=1 J e = 1 . Its inverse-temperature parameter, denoted b F K b_{\mathrm{FK}} b FK here, enters the FK edge odds as e b F K − 1 e^{b_{\mathrm{FK}}}-1 e b FK − 1 . Hence the conversion is
b F K = − log [ 1 − p ∗ ( β ) ] = − log tanh β . b_{\mathrm{FK}}=-\log[1-p^*(\beta)]=-\log\tanh\beta. b FK = − log [ 1 − p ∗ ( β )] = − log tanh β .
Condition (S186) places b F K b_{\mathrm{FK}} b FK below the critical value − log [ 1 − p c F K ( 2 ; Z 3 ) ] -\log[1-p_c^{\mathrm{FK}}(2;\mathbb{Z}^3)] − log [ 1 − p c FK ( 2 ; Z 3 )] . The cited theorem gives exponential decay of the center-to-boundary connection probability in each finite wired graph-distance ball. The comparison above makes the bound uniform in the exterior configuration and in the enhanced edge parameters. This is the part of the cited theorem used here; the conversion to the inhomogeneous torus and the Wilson estimate require the additional arguments in this note. Averaging over the exterior and applying the factor-eight comparison in Eq. (261) proves Eq. (266) for the actual dual gauge law ν ∗ \nu^* ν ∗ .
A cluster of periodic diameter at least L / 4 L/4 L /4 contains a vertex with an open path reaching periodic distance at least ⌊ L / 8 ⌋ \lfloor L/8\rfloor ⌊ L /8 ⌋ . A union bound over the L 3 L^3 L 3 vertices and Eq. (266) prove Eq. (267) . Bounded small values of L L L are absorbed in the constants.
Condition (S186) is the strict perimeter-law side of the transition identified in Ref. [15], after converting the action convention. We require fixed background coupling on this side; the constants need not remain bounded as the transition is approached. The finite wired-volume estimate in Lemma 4.2 supplies the input needed for distant winding representatives. Infinite-volume rectangular-loop asymptotics, including the low-temperature expansion of Ref. [17], do not by themselves give this estimate on a finite torus.
Tube geometry and boundary restoration
The cluster tail becomes useful once a geometric obstruction is identified. A small dual cluster can prevent a filling only when it approaches the Wilson boundary. Hardening a tube around that boundary therefore leaves only large, exponentially unlikely obstructions. The linking criterion in Ref. [15, Sec. 7] is formulated in the lattice and box setting. For the two winding representatives considered here, we construct the obstruction directly on the torus and apply the intersection argument after lifting each small cluster to a local box.
Take straight x x x -directed representatives at ( y , z ) = ( 0 , 0 ) (y,z)=(0,0) ( y , z ) = ( 0 , 0 ) and ( a , 0 ) (a,0) ( a , 0 ) , where α L ≤ a ≤ ( 1 − α ) L \alpha L\le a\le(1-\alpha)L α L ≤ a ≤ ( 1 − α ) L for fixed 0 < α ≤ 1 / 2 0<\alpha\le1/2 0 < α ≤ 1/2 . Let C C C be their union, ∣ C ∣ = 2 L |C|=2L ∣ C ∣ = 2 L , and let S S S be the intervening flat strip, so that ∂ S = C \partial S=C ∂ S = C . Choose a fixed local periodic cell-coincidence distance L L L between primal and dual cells and define
T r ∗ ( C ) = { e ∗ : dist L ( e ∗ , C ) < r } , P r = { p : p e ∗ ∈ T r ∗ ( C ) } . (268) T_r^*(C)=\{e^*:\operatorname{dist}_L(e^*,C)<r\},\qquad\mathcal{P}_r=\{p:p^*_e\in T_r^*(C)\}. \tag*{(268)} T r ∗ ( C ) = { e ∗ : dist L ( e ∗ , C ) < r } , P r = { p : p e ∗ ∈ T r ∗ ( C )} . (268)
Changing to an equivalent local lattice distance changes only geometric constants. The cell-incidence distance used here and the qubit check-graph distance of Supplementary Note 3 are comparable by constants independent of L L L . Consequently a full tube of radius r r r contains the protected plaquette ribbons of Lemma 3.7 at a width proportional to r r r . We assume 1 ≤ r ≤ c α L 1 \le r \le c_\alpha L 1 ≤ r ≤ c α L , where 0 < c α < min { α / 8 , 1 / 16 } 0 < c_\alpha< \min\{\alpha/8,1/16\} 0 < c α < min { α /8 , 1/16 } . In particular,
∣ P r ∣ ≤ C ∣ C ∣ ( r + 1 ) 2 . (269) |\mathcal{P}_r| \le C|\mathcal{C}|(r+1)^2. \tag*{(269)} ∣ P r ∣ ≤ C ∣ C ∣ ( r + 1 ) 2 . (269)
The two-dimensional analogue has ∣ P r ∣ = O ( ∣ C ∣ ( r + 1 ) ) |\mathcal{P}_r| = O(|\mathcal{C}|(r+1)) ∣ P r ∣ = O ( ∣ C ∣ ( r + 1 )) .
Supplementary Lemma 4.3. (A local obstruction must approach the boundary.) Let a connected dual cluster Γ \Gamma Γ have periodic diameter s < L / 4 s < L/4 s < L /4 . Suppose it contains an Eulerian chain γ \gamma γ with γ ⋅ S = 1 \gamma\cdot S = 1 γ ⋅ S = 1 , where the dot denotes mod-two intersection. Then γ \gamma γ is null-homologous and
dist L ( Γ , C ) ≤ c f ( s + 1 ) (270) \operatorname{dist}_L(\Gamma,C) \le c_f(s+1) \tag*{(270)} dist L ( Γ , C ) ≤ c f ( s + 1 ) (270) for a geometric constant c f c_f c f . In particular, if Γ \Gamma Γ avoids T r ∗ ( C ) T_r^*(C) T r ∗ ( C ) , then s ≥ c g r s \ge c_g r s ≥ c g r for a fixed c g > 0 c_g > 0 c g > 0 . An obstruction at diameter scale s s s can be anchored among at most C ∣ C ∣ ( s + 1 ) 2 C|\mathcal{C}|(s+1)^2 C ∣ C ∣ ( s + 1 ) 2 dual edges.
Proof. Choose a vertex of Γ \Gamma Γ as origin and lift each cluster vertex to its nearest coordinate representative. Because s < L / 4 s < L/4 s < L /4 , these representatives lie in a cubical box of side 2 s + O ( 1 ) 2s + O(1) 2 s + O ( 1 ) . They preserve cluster edges: a spurious jump by a full period would require a coordinate separation at least L − 1 L-1 L − 1 , larger than the span of this box. Thus the whole cluster lifts to a contractible cubical box. Bounded small sizes can be incorporated by changing geometric constants. The lifted Eulerian chain is a cycle and bounds a dual two-chain A γ A_\gamma A γ inside the box. Each filling cell lies within c f ( s + 1 ) c_f(s+1) c f ( s + 1 ) of Γ \Gamma Γ . The boundary–intersection identity on the torus gives
1 = γ ⋅ S = ( ∂ A γ ) ⋅ S = A γ ⋅ ∂ S = A γ ⋅ C . (271) 1 = \gamma\cdot S = (\partial A_\gamma) \cdot S = A_\gamma\cdot\partial S = A_\gamma\cdot C. \tag*{(271)} 1 = γ ⋅ S = ( ∂ A γ ) ⋅ S = A γ ⋅ ∂ S = A γ ⋅ C . (271)
Hence the filling meets C C C , proving Eq. (270) . If Γ \Gamma Γ avoids the tube, then r ≤ c f ( s + 1 ) r \le c_f(s+1) r ≤ c f ( s + 1 ) . Decreasing a positive geometric constant c g c_g c g covers the bounded small-r r r cases and gives s ≥ c g r s \ge c_g r s ≥ c g r for an obstruction. A neighborhood of radius O ( s + 1 ) O(s+1) O ( s + 1 ) around the two length-L L L lines contains at most C ∣ C ∣ ( s + 1 ) 2 C|\mathcal{C}|(s+1)^2 C ∣ C ∣ ( s + 1 ) 2 dual edges. Equation (S191) guarantees that Γ \Gamma Γ has an edge in this neighborhood. Choose the first such edge under a fixed ordering as its anchor. This proves the anchor count without counting the volume of the entire filling S S S .
Supplementary Lemma 4.4. (Hard boundary enhancement restores the Wilson overlap.) Assume Eq. (265) , put β p = ∞ \beta_p = \infty β p = ∞ on P r \mathcal{P}_r P r , and put β p = β \beta_p = \beta β p = β elsewhere. There are positive constants, uniform in L L L and r r r , such that
1 − ⟨ W ( C ) ⟩ h a r d ≤ C β ∣ C ∣ ( r + 1 ) 2 e − c g r / ξ β + C β L 3 e − c β L . (272) 1-\langle W(C)\rangle_{\mathrm{hard}} \le C\beta|\mathcal{C}|(r+1)^2e^{-c_g r/\xi_\beta}+C_\beta L^3e^{-c_\beta L}. \tag*{(272)} 1 − ⟨ W ( C ) ⟩ hard ≤ C β ∣ C ∣ ( r + 1 ) 2 e − c g r / ξ β + C β L 3 e − c β L . (272) The geometric and decay constants may be adjusted from those in the preceding lemmas.
Proof. Hard protection sets every dual parameter in T r ∗ ( C ) T_r^*(C) T r ∗ ( C ) to zero, so all dual clusters avoid this tube. Exclude first the event of a cluster with periodic diameter at least L / 4 L/4 L /4 . Its probability is bounded by the second term in Eq. (272) . Suppose that C C C fails to bound a chain of open primal plaquettes. Linear duality supplies an edge cocycle a a a such that δ a ∣ A = 0 \delta a|_A = 0 δ a ∣ A = 0 and a ( C ) = 1 a(C)=1 a ( C ) = 1 . Its dual chain z = ∗ δ a z = *\delta a z = ∗ δ a is supported on D D D , satisfies ∂ z = 0 \partial z = 0 ∂ z = 0 and [ z ] = 0 [z] = 0 [ z ] = 0 , and obeys
z ⋅ S = ( δ a ) ( S ) = a ( ∂ S ) = 1. (273) z \cdot S = (\delta a)(S) = a(\partial S) = 1. \tag*{(273)} z ⋅ S = ( δ a ) ( S ) = a ( ∂ S ) = 1. (273)
Split z z z over the connected components of D D D . Each part is Eulerian, since different components share no vertices. On the event under consideration every component has diameter below L / 4 L/4 L /4 , so each part lifts to a local box and is null-homologous. Equation (S194) implies that at least one part has odd intersection with S S S . This step includes disconnected witnesses; possible nonlocal components have already been included in the excluded macroscopic event.
Let Γ \Gamma Γ contain an odd part. Lemma 4.3 implies that its diameter is at least c g r c_g r c g r . At a dyadic diameter scale [ s , 2 s ) [s,2s) [ s , 2 s ) it has an anchor among C ∣ C ∣ ( s + 1 ) 2 C|\mathcal{C}|(s+1)^2 C ∣ C ∣ ( s + 1 ) 2 edges. From an endpoint of an anchor, a cluster of diameter at least s s s contains an open connection to distance at least s / 2 s/2 s /2 , up to a fixed lattice constant. Use Eq. (266) and sum over anchors and dyadic scales. The summation is controlled by the elementary bound
∑ j ≥ 0 ( 2 j s 0 + 1 ) 2 e − a 2 j s 0 ≤ C a ( s 0 + 1 ) 2 e − a s 0 / 2 , s 0 ≥ 1 , a > 0. (274) \sum_{j\ge0}(2^j s_0+1)^2e^{-a2^j s_0}\le C_a(s_0+1)^2e^{-a s_0/2},\qquad s_0\ge1,\quad a>0. \tag*{(274)} j ≥ 0 ∑ ( 2 j s 0 + 1 ) 2 e − a 2 j s 0 ≤ C a ( s 0 + 1 ) 2 e − a s 0 /2 , s 0 ≥ 1 , a > 0. (274)
Indeed, one half of each exponential is bounded by e − a s 0 / 2 e^{-a s_0/2} e − a s 0 /2 , the remaining half absorbs the growth of 4 j 4^j 4 j in the polynomial prefactor and yields a summable series uniformly for s 0 ≥ 1 s_0 \ge1 s 0 ≥ 1 . Taking s 0 s_0 s 0 of order max { 1 , c g r } \max\{1,c_g r\} max { 1 , c g r } and adjusting constants gives the first term in Eq. (272) . Finally, Eq. (260) identifies the filling-failure probability with the Wilson deficit.
The anchor count is the reason the leading prefactor depends on the boundary length and tube cross-section. The proof does not require strengthening the plaquettes throughout the strip S S S . The second term separately controls obstructions that explore the topology of the finite torus.
Supplementary Lemma 4.5 (Finite-strength boundary enhancement). Assume Eq. (265) . If β p ≥ β R ≥ β \beta_p \ge\beta_R \ge\beta β p ≥ β R ≥ β on P r \mathcal{P}_r P r and β p = β \beta_p = \beta β p = β elsewhere, then
1 − ⟨ W ( C ) ⟩ s o f t ≤ C β ∣ C ∣ ( r + 1 ) 2 e − c g r / ξ β + C β L 3 e − c β L + 16 ∣ P r ∣ e − 2 β R . (275) 1-\langle W(C)\rangle_{\mathrm{soft}} \le C_\beta|C|(r+1)^2e^{-c_gr/\xi_\beta}+C_\beta L^3e^{-c_\beta L}+16|\mathcal{P}_r|e^{-2\beta_R}. \tag*{(275)} 1 − ⟨ W ( C ) ⟩ soft ≤ C β ∣ C ∣ ( r + 1 ) 2 e − c g r / ξ β + C β L 3 e − c β L + 16∣ P r ∣ e − 2 β R . (275) Consequently, for fixed b > 1 b>1 b > 1 and δ > 0 \delta>0 δ > 0 , the choices
r = b ξ β c g log L , β R = ( 1 2 + δ ) log L (276) r=b\frac{\xi_\beta}{c_g}\log L,\qquad\beta_R=\left(\frac{1}{2}+\delta\right)\log L \tag*{(276)} r = b c g ξ β log L , β R = ( 2 1 + δ ) log L (276) (rounding r r r upward to an integer) suffice for ⟨ W ( C ) ⟩ s o f t → 1 \langle W(C)\rangle_{\mathrm{soft}}\to1 ⟨ W ( C ) ⟩ soft → 1 as L → ∞ L\to\infty L → ∞ . The condition β R ≥ β \beta_R\ge\beta β R ≥ β holds for all sufficiently large L L L .
Proof. Let E r E_r E r be the event that at least one protected dual edge is open. For ordinary FK with cluster weight two, the conditional open probability of an edge is at most its parameter. On the tube, p e ∗ ≤ 2 e − 2 β R p_e^*\le2e^{-2\beta_R} p e ∗ ≤ 2 e − 2 β R . A union bound and Eq. (261) therefore give
ν ∗ ( E r ) ≤ 16 ∣ P r ∣ e − 2 β R . (277) \nu^*(E_r)\le16|\mathcal{P}_r|e^{-2\beta_R}. \tag*{(277)} ν ∗ ( E r ) ≤ 16∣ P r ∣ e − 2 β R . (277)
Conditional on E r c E_r^c E r c , all protected dual edges are closed. Their local factors are constant, and the remaining weight in Eq. (259) is exactly the hard-tube weight: c ( D ) c(D) c ( D ) and h ( D ) h(D) h ( D ) depend on the edge configuration, not on the values of its parameters. Thus the conditional filling-failure probability is the hard-tube deficit. Splitting according to E r E_r E r bounds the soft deficit by that hard deficit plus ν ∗ ( E r ) \nu^*(E_r) ν ∗ ( E r ) , proving Eq. (275) .
Using ∣ C ∣ = 2 L |C|=2L ∣ C ∣ = 2 L and Eq. (269) , the three contributions under Eq. (276) are respectively O ( L 1 − b log 2 L ) O(L^{1-b}\log^2 L) O ( L 1 − b log 2 L ) , O ( L 3 e − c β L ) O(L^3e^{-c_\beta L}) O ( L 3 e − c β L ) , and O ( L − 2 δ log 2 L ) O(L^{-2\delta}\log^2 L) O ( L − 2 δ log 2 L ) . They vanish, and the required geometric condition r ≤ c α L r\le c_\alpha L r ≤ c α L holds for all sufficiently large L L L .
Lemma 4.5 proves the three-dimensional Wilson-restoration statement in the main text. The same full tubes also restore AQEC for each chosen encoding under the fixed low-noise condition λ ≤ λ 0 < 1 \lambda\le\lambda_0<1 λ ≤ λ 0 < 1 of Supplementary Note 3. Indeed, the ribbons contained in these tubes impose a minimum silent distinguishing size proportional to r 2 r^2 r 2 , by Lemma 3.7. With r r r proportional to log L \log L log L , the hard-projection activity in Eq. (222) is O ( L e − c ( log L ) 2 ) O(Le^{-c(\log L)^2}) O ( L e − c ( l o g L ) 2 ) for some fixed c > 0 c>0 c > 0 . The soft–hard correction in Eq. (233) is O ( L ( r + 1 ) 2 e − β R ) = O ( L − δ log L ) O(\sqrt{L}(r+1)^2e^{-\beta_R})=O(L^{-\delta}\log L) O ( L ( r + 1 ) 2 e − β R ) = O ( L − δ log L ) . Both vanish, so the recovery and entropy bounds of Lemma 2.9 apply to each fixed logical qubit. The preparation uses O ( L log 2 L ) O(L\log^2 L) O ( L log 2 L ) enhanced plaquettes out of 3 L 3 3L^3 3 L 3 , a fraction O ( log 2 L / L 2 ) O(\log^2 L/L^2) O ( log 2 L / L 2 ) .
The radius O ( log L ) O(\log L) O ( log L ) is sufficient for this joint conclusion, whereas AQEC alone has the sufficient radius O ( log L ) O(\sqrt{\log L}) O ( log L ) established in Supplementary Note 3. Comparing these constructions does not prove that Wilson restoration requires a larger radius. Their assumptions also play different roles: low physical noise controls recovery, and subcriticality of the dual bulk controls the Wilson overlap.
Encoding distance and recovery with a common channel
The representative overlap determines both the distinguishability of two encodings and the cost of recovering without a record of which one was used. Let Q Q Q denote the entire n n n -qubit physical system, and let H L \mathcal{H}_L H L be the two-dimensional logical input space. Both representatives are enhanced in the same preparation; ∣ ψ β ⟩ |\psi_\beta\rangle ∣ ψ β ⟩ therefore includes the full, inhomogeneous measurement-strength profile. For i = 1 , 2 i=1,2 i = 1 , 2 , define
V i : H L ⟶ H Q , V i ( a ∣ 0 ⟩ L + b ∣ 1 ⟩ L ) = a ∣ ψ β ⟩ + b Z L i ∣ ψ β ⟩ . (278) V_i:\mathcal{H}_L\longrightarrow\mathcal{H}_Q,\qquad V_i(a|0\rangle_L+b|1\rangle_L)=a|\psi_\beta\rangle+bZ^{i}_{\mathrm{L}}|\psi_\beta\rangle. \tag*{(278)} V i : H L ⟶ H Q , V i ( a ∣0 ⟩ L + b ∣1 ⟩ L ) = a ∣ ψ β ⟩ + b Z L i ∣ ψ β ⟩ . (278)
The index i i i labels the representative defining the image of ∣ 1 ⟩ L |1\rangle_L ∣1 ⟩ L . A complementary physical logical X X X operator stabilizes ∣ ψ β ⟩ |\psi_\beta\rangle ∣ ψ β ⟩ and anticommutes with both Z L i Z^i_{\mathrm{L}} Z L i . The two codewords within each encoding therefore belong to opposite exact symmetry sectors. They are normalized and orthogonal, so V i † V i = I L V_i^\dagger V_i=I_L V i † V i = I L .
Write w = ⟨ ψ β ∣ Z L 1 Z L 2 ∣ ψ β ⟩ w=\langle\psi_\beta|Z^1_{\mathrm{L}}Z^2_{\mathrm{L}}|\psi_\beta\rangle w = ⟨ ψ β ∣ Z L 1 Z L 2 ∣ ψ β ⟩ . This number belongs to [ 0 , 1 ] [0,1] [ 0 , 1 ] by Lemmas 1.2 and 4.1, including the inhomogeneous and projective limits. The shared zero codeword and the vanishing cross-sector matrix elements give
V 1 † V 2 = diag ( 1 , w ) (279) V_1^\dagger V_2=\operatorname{diag}(1,w) \tag*{(279)} V 1 † V 2 = diag ( 1 , w ) (279)
in the ordered logical basis ( ∣ 0 ⟩ L , ∣ 1 ⟩ L ) (|0\rangle_L, |1\rangle_L) ( ∣0 ⟩ L , ∣1 ⟩ L ) . The channel associated with V i V_i V i is V i ( ρ L ) = V i ρ L V i † \mathcal{V}_i(\rho_L)=V_i\rho_LV_i^\dagger V i ( ρ L ) = V i ρ L V i † . All recovery channels below act on the entire physical system and output one logical qubit. They are not assumed to be spatially local or efficiently implementable.
Supplementary Lemma 4.6 (Distance between the encoding channels). For these encodings,
1 2 ∥ V 1 − V 2 ∥ ⋄ = 1 − w 2 ≤ ∥ V 1 − V 2 ∥ ∞ = 2 ( 1 − w ) . (280) \frac{1}{2}\lVert V_1-V_2\rVert_\diamond=\sqrt{1-w^2}\leq\lVert V_1-V_2\rVert_\infty=\sqrt{2(1-w)}. \tag*{(280)} 2 1 ∥ V 1 − V 2 ∥ ⋄ = 1 − w 2 ≤ ∥ V 1 − V 2 ∥ ∞ = 2 ( 1 − w ) . (280) In particular, the first expression bounds the trace distance between the two encoded outputs for every logical input, including inputs entangled with an arbitrary reference system.
Proof. For discrimination between two channels, the diamond norm can be optimized over pure input states with a reference system [21 Theorem 3.51] . Let ∣ Ω ⟩ L R |\Omega\rangle_{LR} ∣Ω ⟩ L R be any such input and let ρ L \rho_L ρ L be its reduced state on the logical qubit. The two encoded outputs are pure and have overlap
L R ⟨ Ω ∣ ( V 1 † V 2 ⊗ I R ) ∣ Ω ⟩ L R = Tr [ ρ L diag ( 1 , w ) ] = 1 − ( 1 − w ) L ⟨ 1 ∣ ρ L ∣ 1 ⟩ L . {}_{LR}\langle\Omega|(V_1^\dagger V_2\otimes I_R)|\Omega\rangle_{LR}=\operatorname{Tr}[\rho_L\operatorname{diag}(1,w)]=1-(1-w){}_L\langle1|\rho_L|1\rangle_L. L R ⟨ Ω∣ ( V 1 † V 2 ⊗ I R ) ∣Ω ⟩ L R = Tr [ ρ L diag ( 1 , w )] = 1 − ( 1 − w ) L ⟨ 1∣ ρ L ∣1 ⟩ L .
This real number lies in [ w , 1 ] [w,1] [ w , 1 ] . The trace distance between pure states with overlap z z z is 1 − ∣ z ∣ 2 \sqrt{1-|z|^2} 1 − ∣ z ∣ 2 [29 ] , so it is at most 1 − w 2 \sqrt{1-w^2} 1 − w 2 . The input ∣ 1 ⟩ L |1\rangle_L ∣1 ⟩ L attains this bound, without a reference system. The operator-norm equality follows from
( V 1 − V 2 ) † ( V 1 − V 2 ) = diag ( 0 , 2 ( 1 − w ) ) . (V_1-V_2)^\dagger(V_1-V_2)=\operatorname{diag}(0,2(1-w)). ( V 1 − V 2 ) † ( V 1 − V 2 ) = diag ( 0 , 2 ( 1 − w )) .
Supplementary Lemma 4.7 (Transfer of a fixed recovery channel). Let N \mathcal{N} N be the same subsequent physical noise channel for both encodings, and let R \mathcal{R} R be a fixed recovery channel. Then
1 2 ∥ R N V 1 − R N V 2 ∥ ⋄ ≤ 1 − w 2 . (281) \frac{1}{2}\lVert\mathcal{R}\mathcal{N}V_1-\mathcal{R}\mathcal{N}V_2\rVert_\diamond\leq\sqrt{1-w^2}. \tag*{(281)} 2 1 ∥ RN V 1 − RN V 2 ∥ ⋄ ≤ 1 − w 2 . (281) Using the same maximally entangled input as in the main text, now with K = 2 K=2 K = 2 , write
F e ( T ) = ⟨ Φ 2 ∣ ( id R ⊗ T ) ( ∣ Φ 2 ⟩ ⟨ Φ 2 ∣ ) ∣ Φ 2 ⟩ . F_{\mathrm{e}}(\mathcal{T})=\langle\Phi_2|(\operatorname{id}_R\otimes\mathcal{T})(|\Phi_2\rangle\langle\Phi_2|)|\Phi_2\rangle. F e ( T ) = ⟨ Φ 2 ∣ ( id R ⊗ T ) ( ∣ Φ 2 ⟩ ⟨ Φ 2 ∣ ) ∣ Φ 2 ⟩ . The entanglement fidelities of the two recovered channels satisfy
∣ F e ( R N V 1 ) − F e ( R N V 2 ) ∣ ≤ 1 − w 2 . (282) |F_{\mathrm{e}}(\mathcal{R}\mathcal{N}V_1)-F_{\mathrm{e}}(\mathcal{R}\mathcal{N}V_2)|\leq\sqrt{1-w^2}. \tag*{(282)} ∣ F e ( RN V 1 ) − F e ( RN V 2 ) ∣ ≤ 1 − w 2 . (282) Their errors relative to the identity also obey
1 2 ∥ R N V 2 − id L ∥ ⋄ ≤ 1 2 ∥ R N V 1 − id L ∥ ⋄ + 1 − w 2 . \frac{1}{2}\lVert\mathcal{R}\mathcal{N}V_2-\operatorname{id}_L\rVert_\diamond\leq\frac{1}{2}\lVert\mathcal{R}\mathcal{N}V_1-\operatorname{id}_L\rVert_\diamond+\sqrt{1-w^2}. 2 1 ∥ RN V 2 − id L ∥ ⋄ ≤ 2 1 ∥ RN V 1 − id L ∥ ⋄ + 1 − w 2 .
Proof. Postcomposition by a channel contracts the diamond norm [21 Proposition 3.48] , giving (S202). Apply this bound to the maximally entangled input. Measuring the projector ∣ Φ 2 ⟩ ⟨ Φ 2 ∣ |\Phi_2\rangle\langle\Phi_2| ∣ Φ 2 ⟩ ⟨ Φ 2 ∣ changes its outcome probability by at most the trace distance, proving (S203). The final statement follows from the triangle inequality for diamond norm. No conversion from entanglement fidelity to worst-case fidelity is used.
The preceding lemma is sufficient for the three-dimensional application: when the chosen encoding is recoverable and w → 1 w\to1 w → 1 , a recovery selected for that encoding also works for the second encoding. This implication uses a common physical noise channel and does not give the decoder a record of the representative choice.
Nonvanishing distance between two general encoding maps would not, by itself, exclude a common recovery: two orthogonal code subspaces, for example, can carry distinguishable encoding flags. The present family has the additional structure of a shared physical codeword ∣ ψ β ⟩ |\psi_\beta\rangle ∣ ψ β ⟩ for logical input ∣ 0 ⟩ L |0\rangle_L ∣0 ⟩ L . The common-recovery optimum can then be evaluated exactly.
Supplementary Lemma 4.8 (Common recovery without a representative record). In the absence of subsequent noise, suppose that either V 1 V_1 V 1 or V 2 V_2 V 2 is used, and that its identity is not supplied to the recovery. The optimal worst-choice entanglement fidelity of a common recovery is
max R min i = 1 , 2 F e ( R V i ) = 1 + ( 1 + w ) / 2 2 . (283) \max_{\mathcal{R}}\min_{i=1,2}F_{\mathrm{e}}(\mathcal{R}V_i)=\frac{1+\sqrt{(1+w)/2}}{2}. \tag*{(283)} R max i = 1 , 2 min F e ( R V i ) = 2 1 + ( 1 + w ) /2 . (283) Consequently, after any common subsequent noise channel N \mathcal{N} N ,
inf R max i = 1 , 2 [ 1 − F e ( R N V i ) ] ≥ 1 − ( 1 + w ) / 2 2 . (284) \inf_{\mathcal{R}}\max_{i=1,2}[1-F_{\mathrm{e}}(\mathcal{R}\mathcal{N}V_i)]\geq\frac{1-\sqrt{(1+w)/2}}{2}. \tag*{(284)} R inf i = 1 , 2 max [ 1 − F e ( RN V i )] ≥ 2 1 − ( 1 + w ) /2 . (284)
Proof. Within this proof, write ∣ a ⟩ = V 1 ∣ 1 ⟩ L |a\rangle= V_1|1\rangle_L ∣ a ⟩ = V 1 ∣1 ⟩ L and ∣ b ⟩ = V 2 ∣ 1 ⟩ L |b\rangle= V_2|1\rangle_L ∣ b ⟩ = V 2 ∣1 ⟩ L . Both are orthogonal to ∣ ψ β ⟩ |\psi_\beta\rangle ∣ ψ β ⟩ , and ⟨ a ∣ b ⟩ = w \langle a|b\rangle= w ⟨ a ∣ b ⟩ = w . First let w < 1 w < 1 w < 1 , and define the orthonormal physical states
∣ s + ⟩ = ∣ a ⟩ + ∣ b ⟩ 2 ( 1 + w ) , ∣ s − ⟩ = ∣ a ⟩ − ∣ b ⟩ 2 ( 1 − w ) , c = ( 1 + w ) / 2 . |s_+\rangle= \frac{|a\rangle+ |b\rangle}{\sqrt{2(1+w)}}, \qquad|s_-\rangle= \frac{|a\rangle- |b\rangle}{\sqrt{2(1-w)}}, \qquad c = \sqrt{(1+w)/2}. ∣ s + ⟩ = 2 ( 1 + w ) ∣ a ⟩ + ∣ b ⟩ , ∣ s − ⟩ = 2 ( 1 − w ) ∣ a ⟩ − ∣ b ⟩ , c = ( 1 + w ) /2 .
They are orthogonal to ∣ ψ β ⟩ |\psi_\beta\rangle ∣ ψ β ⟩ , and ∣ a ⟩ = c ∣ s + ⟩ + 1 − c 2 ∣ s − ⟩ |a\rangle= c|s_+\rangle+ \sqrt{1-c^2}|s_-\rangle ∣ a ⟩ = c ∣ s + ⟩ + 1 − c 2 ∣ s − ⟩ , ∣ b ⟩ = c ∣ s + ⟩ − 1 − c 2 ∣ s − ⟩ |b\rangle= c|s_+\rangle- \sqrt{1-c^2}|s_-\rangle ∣ b ⟩ = c ∣ s + ⟩ − 1 − c 2 ∣ s − ⟩ . Consider the equally weighted mixture M = ( V 1 + V 2 ) / 2 \mathcal{M} = (\mathcal{V}_1+\mathcal{V}_2)/2 M = ( V 1 + V 2 ) /2 . Its action is
M ( ∣ 0 ⟩ L ⟨ 0 ∣ L ) = ∣ ψ β ⟩ ⟨ ψ β ∣ , M ( ∣ 1 ⟩ L ⟨ 1 ∣ L ) = c 2 ∣ s + ⟩ ⟨ s + ∣ + ( 1 − c 2 ) ∣ s − ⟩ ⟨ s − ∣ , M ( ∣ 0 ⟩ L ⟨ 1 ∣ L ) = c ∣ ψ β ⟩ ⟨ s + ∣ . \begin{aligned}
\mathcal{M}(|0\rangle_L\langle0|_L) &= |\psi_\beta\rangle\langle\psi_\beta|,\\
\mathcal{M}(|1\rangle_L\langle1|_L) &= c^2|s_+\rangle\langle s_+| + (1-c^2)|s_-\rangle\langle s_-|,\\
\mathcal{M}(|0\rangle_L\langle1|_L) &= c|\psi_\beta\rangle\langle s_+|.
\end{aligned} M ( ∣0 ⟩ L ⟨ 0 ∣ L ) M ( ∣1 ⟩ L ⟨ 1 ∣ L ) M ( ∣0 ⟩ L ⟨ 1 ∣ L ) = ∣ ψ β ⟩ ⟨ ψ β ∣ , = c 2 ∣ s + ⟩ ⟨ s + ∣ + ( 1 − c 2 ) ∣ s − ⟩ ⟨ s − ∣ , = c ∣ ψ β ⟩ ⟨ s + ∣.
For an arbitrary recovery R \mathcal{R} R with Kraus operators R μ R_\mu R μ , entanglement fidelity is one quarter of the sum of two correct-output population terms and two conjugate coherence terms. Each population term is at most one. The relevant coherence obeys
∣ ⟨ 0 ∣ L R ( ∣ ψ β ⟩ ⟨ s + ∣ ) ∣ 1 ⟩ L ∣ = ∣ ∑ μ ⟨ 0 ∣ L R μ ∣ ψ β ⟩ ⟨ s + ∣ R μ † ∣ 1 ⟩ L ∣ ≤ ( ∑ μ ∣ ⟨ 0 ∣ L R μ ∣ ψ β ⟩ ∣ 2 ) 1 / 2 ( ∑ μ ∣ ⟨ 1 ∣ L R μ ∣ s + ⟩ ∣ 2 ) 1 / 2 ≤ 1. \begin{aligned}
\left|\langle0|_L\mathcal{R}(|\psi_\beta\rangle\langle s_+|)|1\rangle_L\right| = \left|\sum_\mu\langle0|_L R_\mu|\psi_\beta\rangle\langle s_+|R_\mu^\dagger|1\rangle_L\right| \\
&\le\left(\sum_\mu|\langle0|_L R_\mu|\psi_\beta\rangle|^2\right)^{1/2}\left(\sum_\mu|\langle1|_L R_\mu|s_+\rangle|^2\right)^{1/2} \le1.
\end{aligned} ∣ ⟨ 0 ∣ L R ( ∣ ψ β ⟩ ⟨ s + ∣ ) ∣1 ⟩ L ∣ = μ ∑ ⟨ 0 ∣ L R μ ∣ ψ β ⟩ ⟨ s + ∣ R μ † ∣1 ⟩ L ≤ ( μ ∑ ∣ ⟨ 0 ∣ L R μ ∣ ψ β ⟩ ∣ 2 ) 1/2 ( μ ∑ ∣ ⟨ 1 ∣ L R μ ∣ s + ⟩ ∣ 2 ) 1/2 ≤ 1.
Hence F e ( R M ) ≤ ( 2 + 2 c ) / 4 F_e(\mathcal{R}\mathcal{M}) \le(2+2c)/4 F e ( R M ) ≤ ( 2 + 2 c ) /4 . By linearity of entanglement fidelity in the channel,
min i F e ( R V i ) ≤ 1 2 ∑ i F e ( R V i ) = F e ( R M ) ≤ 1 + c 2 . \min_i F_e(\mathcal{R}\mathcal{V}_i) \le\frac{1}{2}\sum_i F_e(\mathcal{R}\mathcal{V}_i) = F_e(\mathcal{R}\mathcal{M}) \le\frac{1+c}{2}. i min F e ( R V i ) ≤ 2 1 i ∑ F e ( R V i ) = F e ( R M ) ≤ 2 1 + c .
To attain the bound, take recovery Kraus operators
R + = ∣ 0 ⟩ L ⟨ ψ β ∣ + ∣ 1 ⟩ L ⟨ s + ∣ , R − = ∣ 1 ⟩ L ⟨ s − ∣ . R_+ = |0\rangle_L\langle\psi_\beta| + |1\rangle_L\langle s_+|,\qquad R_- = |1\rangle_L\langle s_-|. R + = ∣0 ⟩ L ⟨ ψ β ∣ + ∣1 ⟩ L ⟨ s + ∣ , R − = ∣1 ⟩ L ⟨ s − ∣.
Complete these to a trace-preserving channel on the remaining physical subspace, which the encoded inputs never occupy. For both i i i , the recovered channel preserves the logical populations and multiplies the off-diagonal matrix elements by c c c . It therefore has entanglement fidelity ( 1 + c ) / 2 (1+c)/2 ( 1 + c ) /2 for each encoder, proving (S204). At w = 1 w=1 w = 1 the encoders coincide and their isometric inverse recovers perfectly. Finally, every R N \mathcal{R}_N R N is itself a physical-to-logical channel; optimizing over this restricted class cannot improve the noiseless optimum. This proves (S205).
For the two-dimensional boundary-enhancement regime in which w → 0 w\to0 w → 0 , the last bound tends to ( 1 − 1 / 2 ) / 2 ≃ 0.1464 (1-1/\sqrt{2})/2 \simeq0.1464 ( 1 − 1/ 2 ) /2 ≃ 0.1464 . This is a lower bound on the worst-choice entanglement infidelity when the representative choice is unknown to the recovery. It is compatible with vanishing recovery error for each separately specified encoding. In three dimensions, w → 1 w\to1 w → 1 removes this particular obstruction, while Lemma 4.7 combines with the AQEC upper bound to establish a common recovery. Neither statement asserts that two disjoint physical regions each contain an independently recoverable copy of the logical qubit.
Changing the boundary used for a controlled logical operation
Changing the physical boundary used for a logical operation provides another way to test representative equivalence. Here the target encoding is held fixed, and only the controlled string is changed. Use V = V 1 V=V_1 V = V 1 , with V ∣ 0 ⟩ L = ∣ ψ β ⟩ V|0\rangle_L=|\psi_\beta\rangle V ∣0 ⟩ L = ∣ ψ β ⟩ and V ∣ 1 ⟩ L = Z ℓ 1 ∣ ψ β ⟩ V|1\rangle_L=Z^{\ell_1}|\psi_\beta\rangle V ∣1 ⟩ L = Z ℓ 1 ∣ ψ β ⟩ . These are the same representatives as above: Z ℓ 1 Z^{\ell_1} Z ℓ 1 and Z ℓ 2 Z^{\ell_2} Z ℓ 2 commute, square to the identity, and anticommute with an exact X X X logical operator that stabilizes ∣ ψ β ⟩ |\psi_\beta\rangle ∣ ψ β ⟩ . We keep the logical basis fixed, so the physical string Z ℓ 1 Z^{\ell_1} Z ℓ 1 implements logical X L = ∣ 0 ⟩ L ⟨ 1 ∣ L + ∣ 1 ⟩ L ⟨ 0 ∣ L X_L=|0\rangle_L\langle1|_L+|1\rangle_L\langle0|_L X L = ∣0 ⟩ L ⟨ 1 ∣ L + ∣1 ⟩ L ⟨ 0 ∣ L .
Supplementary Lemma 4.9 (Controlled-operation error from the Wilson overlap). Let w = ⟨ ψ β ∣ Z ℓ 1 Z ℓ 2 ∣ ψ β ⟩ ∈ [ 0 , 1 ] w=\langle\psi_\beta|Z^{\ell_1}Z^{\ell_2}|\psi_\beta\rangle\in[0,1] w = ⟨ ψ β ∣ Z ℓ 1 Z ℓ 2 ∣ ψ β ⟩ ∈ [ 0 , 1 ] . For an ideal control qubit c c c , define operators on H c ⊗ H Q \mathcal{H}_c\otimes\mathcal{H}_Q H c ⊗ H Q , with I Q I_Q I Q the physical-target identity,
G j = ∣ 0 ⟩ c ⟨ 0 ∣ c ⊗ I Q + ∣ 1 ⟩ c ⟨ 1 ∣ c ⊗ Z ℓ j , j = 1 , 2. (285) G_j = |0\rangle_c\langle0|_c\otimes I_Q + |1\rangle_c\langle1|_c\otimes Z^{\ell_j},\qquad j=1,2. \tag*{(285)} G j = ∣0 ⟩ c ⟨ 0 ∣ c ⊗ I Q + ∣1 ⟩ c ⟨ 1 ∣ c ⊗ Z ℓ j , j = 1 , 2. (285)
The ideal gate CNOTc → L _{c\to L} c → L acts on H c ⊗ H L \mathcal{H}_c \otimes\mathcal{H}_L H c ⊗ H L . With I c I_c I c and I L I_L I L the identities on the control and logical input, respectively,
G 1 ( I c ⊗ V ) = ( I c ⊗ V ) CNOT c → L , (286) G_1(I_c \otimes V) = (I_c \otimes V)\operatorname{CNOT}_{c\to L}, \tag*{(286)} G 1 ( I c ⊗ V ) = ( I c ⊗ V ) CNOT c → L , (286)
V † Z ℓ 2 V = w X L , V † Z ℓ 1 Z ℓ 2 V = w I L , (287) V^\dagger Z^{\ell_2} V = wX_L,\qquad V^\dagger Z^{\ell_1}Z^{\ell_2}V = wI_L, \tag*{(287)} V † Z ℓ 2 V = w X L , V † Z ℓ 1 Z ℓ 2 V = w I L , (287)
∥ G 2 ( I c ⊗ V ) − ( I c ⊗ V ) CNOT c → L ∥ ∞ = 2 ( 1 − w ) . (288) \left\|G_2(I_c \otimes V) - (I_c \otimes V)\operatorname{CNOT}_{c\to L}\right\|_\infty= \sqrt{2(1-w)}. \tag*{(288)} ∥ G 2 ( I c ⊗ V ) − ( I c ⊗ V ) CNOT c → L ∥ ∞ = 2 ( 1 − w ) . (288)
For control input ∣ 1 ⟩ c |1\rangle_c ∣1 ⟩ c , the probability that G 2 G_2 G 2 takes the target outside the fixed code space is 1 − w 2 1-w^2 1 − w 2 for every target input. These statements remain valid for logical inputs entangled with an arbitrary reference system. Writing G j G_j G j for the channel from the control and logical target to the control and physical target induced by G j ( I c ⊗ V ) G_j(I_c \otimes V) G j ( I c ⊗ V ) , one also has
1 2 ∥ G 1 − G 2 ∥ ⋄ = 1 − w 2 . (289) \frac{1}{2}\left\|G_1-G_2\right\|_\diamond=\sqrt{1-w^2}. \tag*{(289)} 2 1 ∥ G 1 − G 2 ∥ ⋄ = 1 − w 2 . (289)
Proof. The identities Z ℓ 1 ∣ ψ β ⟩ = V ∣ 1 ⟩ L Z^{\ell_1}|\psi_\beta\rangle=V|1\rangle_L Z ℓ 1 ∣ ψ β ⟩ = V ∣1 ⟩ L and ( Z ℓ 1 ) 2 ∣ ψ β ⟩ = V ∣ 0 ⟩ L (Z^{\ell_1})^2|\psi_\beta\rangle=V|0\rangle_L ( Z ℓ 1 ) 2 ∣ ψ β ⟩ = V ∣0 ⟩ L give Z ℓ 1 V = V X L Z^{\ell_1}V=VX_L Z ℓ 1 V = V X L , proving Eq. (286) . The exact X X X logical symmetry makes the expectations of Z ℓ 1 Z^{\ell_1} Z ℓ 1 and Z ℓ 2 Z^{\ell_2} Z ℓ 2 vanish. Since Z ℓ 1 Z^{\ell_1} Z ℓ 1 and Z ℓ 2 Z^{\ell_2} Z ℓ 2 commute, the diagonal entries of V † Z ℓ 2 V V^\dagger Z^{\ell_2}V V † Z ℓ 2 V vanish and its off-diagonal entries both equal w w w . Similarly, V † Z ℓ 1 Z ℓ 2 V V^\dagger Z^{\ell_1}Z^{\ell_2}V V † Z ℓ 1 Z ℓ 2 V has diagonal entries w w w and vanishing off-diagonal entries. This proves Eq. (287) .
Subtracting the first controlled implementation from the second leaves only the control-1 block. Its squared norm is determined by
[ ( Z ℓ 2 − Z ℓ 1 ) V ] † [ ( Z ℓ 2 − Z ℓ 1 ) V ] = V † ( 2 I Q − 2 Z ℓ 1 Z ℓ 2 ) V = 2 ( 1 − w ) I L . \left[(Z^{\ell_2}-Z^{\ell_1})V\right]^\dagger\left[(Z^{\ell_2}-Z^{\ell_1})V\right]=V^\dagger(2I_Q-2Z^{\ell_1}Z^{\ell_2})V=2(1-w)I_L. [ ( Z ℓ 2 − Z ℓ 1 ) V ] † [ ( Z ℓ 2 − Z ℓ 1 ) V ] = V † ( 2 I Q − 2 Z ℓ 1 Z ℓ 2 ) V = 2 ( 1 − w ) I L .
The largest singular value is therefore 2 ( 1 − w ) \sqrt{2(1-w)} 2 ( 1 − w ) , giving Eq. (288) . Projection onto the fixed code is given by V V † VV^\dagger V V † . Equation (S208) gives
V † Z ℓ 2 V V † Z ℓ 2 V = ( V † Z ℓ 2 V ) † ( V † Z ℓ 2 V ) = w 2 I L . V^\dagger Z^{\ell_2}VV^\dagger Z^{\ell_2}V=(V^\dagger Z^{\ell_2}V)^\dagger(V^\dagger Z^{\ell_2}V)=w^2I_L. V † Z ℓ 2 V V † Z ℓ 2 V = ( V † Z ℓ 2 V ) † ( V † Z ℓ 2 V ) = w 2 I L .
For any logical state ρ L \rho_L ρ L , the probability to remain in the code after the remote string is consequently Tr ( ρ L w 2 I L ) = w 2 \operatorname{Tr}(\rho_L w^2I_L)=w^2 Tr ( ρ L w 2 I L ) = w 2 . Tensoring these operator identities with a reference identity proves the corresponding statements for entangled inputs. Finally, the overlap of the two control–target isometries is
( I c ⊗ V † ) G 1 † G 2 ( I c ⊗ V ) = ∣ 0 ⟩ c ⟨ 0 ∣ c ⊗ I L + w ∣ 1 ⟩ c ⟨ 1 ∣ c ⊗ I L . (I_c\otimes V^\dagger)G_1^\dagger G_2(I_c\otimes V)=|0\rangle_c\langle0|_c\otimes I_L+w|1\rangle_c\langle1|_c\otimes I_L. ( I c ⊗ V † ) G 1 † G 2 ( I c ⊗ V ) = ∣0 ⟩ c ⟨ 0 ∣ c ⊗ I L + w ∣1 ⟩ c ⟨ 1 ∣ c ⊗ I L .
For an arbitrary pure input including a reference, the output overlap is 1 − ( 1 − w ) Pr ( c = 1 ) 1-(1-w)\Pr(c=1) 1 − ( 1 − w ) Pr ( c = 1 ) and belongs to [ w , 1 ] [w,1] [ w , 1 ] . The pure-state trace distance is at most 1 − w 2 \sqrt{1-w^2} 1 − w 2 and attains that value for control ∣ 1 ⟩ c |1\rangle_c ∣1 ⟩ c . Convexity handles mixed inputs, proving Eq. (289) . □ \square □
Combining this lemma with the main-text Wilson-restoration theorem proves the encoded-CNOT corollary. In the enhanced three-dimensional regime, the two distant boundaries implement the same controlled logical operation with vanishing error. In the two-dimensional regime, using the remote representative in the fixed target encoding instead gives a leakage probability tending to one when the control is ∣ 1 ⟩ c |1\rangle_c ∣1 ⟩ c . Each representative still acts exactly as a logical X X X for its own encoding. The distinction is whether it can replace the chosen representative while the target encoding is held fixed.
The control and controlled strings in this comparison are ideal. A controlled string can be realized algebraically as a product of controlled-Z Z Z gates on its support; geometric locality, depth, and fault-tolerant implementation of those gates require a separate circuit analysis. The result is an operational consequence of representative equivalence, and places no general prohibition on other logical-CNOT constructions in two dimensions.