Logical Pauli Channels in Quantum Error Correction

• Logical Pauli Channels are encoded noise maps defined as classical mixtures of logical Pauli operators applied to the logical subspace in stabilizer codes.
• They arise naturally in syndrome-conditioned error correction processes and can be reconstructed solely from syndrome data under correctable Pauli noise conditions.
• Twirling and Pauli conjugation techniques enforce or restore the Pauli channel structure, while deviations under non-Pauli noise highlight the limits of the approximation.

Logical Pauli channels are effective encoded-noise maps in which the action on the logical degrees of freedom is expressed as a classical mixture of logical Pauli operators. In stabilizer quantum error correction, they arise after syndrome measurement and recovery, as syndrome-conditioned maps attached to detector regions, and as the output of twirling procedures that eliminate off-diagonal Pauli-transfer-matrix elements. Recent work also clarifies the boundaries of this description: under correctable Pauli noise, logical Pauli channels can be reconstructed from syndrome data alone, whereas in approximate GKP error correction and under pure or heralded loss the exact logical map can be genuinely non-Pauli, so Pauli-channel models require explicit justification rather than assumption (Wagner et al., 2022, Girling et al., 11 Aug 2025, Jafarzadeh et al., 18 Apr 2025, Harris et al., 18 Apr 2025).

1. Definition and algebraic structure

For an $[[n,k]]$ stabilizer code with stabilizer group $S\subset P^n$, a phenomenological Pauli noise channel on the data qubits is written as

$$\Lambda(\rho)\;=\;\sum_{e\in P^n}P(e)\;e\,\rho\,e^\dagger.$$

The associated logical distribution is the stabilizer-coset average

$P_L(e)\;=\;\frac1{|S|}\sum_{s\in S}P(e\,s),$

and the logical channel is

$$\Lambda_L(\rho)\;=\;\sum_{\ell\in P^n}P_L(\ell)\,\ell\,\rho\,\ell^\dagger.$$

Only $\ell\in S^\perp=L$ occur with nonzero weight, so in Kraus form one may equivalently write

$$\Lambda_L(\rho)\;=\;\sum_{\bar\ell\in L/S}p_{\bar\ell}\,\bigl(U_{\bar\ell}\,\rho\,U_{\bar\ell}^\dagger\bigr),$$

where $U_{\bar\ell}$ is any representative of the logical coset $\bar\ell\in L/S$. The Choi matrix of $\Lambda_L$ is supported entirely on the logical subspace and inherits the block-diagonal structure of the underlying Pauli channel when written in the logical Pauli basis (Wagner et al., 2022).

A syndrome-conditioned version of the same object appears in detector-region analysis. If a detector region is formed from two or more syndrome extraction gadgets and the observed detector outcome is $S\subset P^n$0, then after twirling the physical errors into a Pauli channel the conditioned action is

$S\subset P^n$1

with $S\subset P^n$2. A Pauli error on the full block can be uniquely factored as

$S\subset P^n$3

which induces logical-coset probabilities

$S\subset P^n$4

In particular, $S\subset P^n$5 is the chance of no logical error given detector $S\subset P^n$6 (Girling et al., 11 Aug 2025).

These definitions distinguish two related uses of the term. In one use, the logical Pauli channel is the logical reduction of a physical Pauli channel. In the other, it is a syndrome-resolved object attached to a specific detector outcome. This suggests that the logical Pauli channel is not a single universal abstraction, but a family of encoded-noise descriptions indexed by the operational information retained during error correction.

2. Syndrome-conditioned logical channels in detector regions

In detector-region tomography, a “detector region” consists of two or more repeated syndrome extraction gadgets whose parity outcome is labeled by $S\subset P^n$7. If two successive rounds with outcomes $S\subset P^n$8 and $S\subset P^n$9 are grouped into one detector $$\Lambda(\rho)\;=\;\sum_{e\in P^n}P(e)\;e\,\rho\,e^\dagger.$$0, then normalizing the corresponding post-measurement map by $$\Lambda(\rho)\;=\;\sum_{e\in P^n}P(e)\;e\,\rho\,e^\dagger.$$1 gives

$$\Lambda(\rho)\;=\;\sum_{e\in P^n}P(e)\;e\,\rho\,e^\dagger.$$2

The resulting channel is therefore a logical Pauli channel conditioned on the observed detector parity (Girling et al., 11 Aug 2025).

The protocol introduced for this setting is SPAM-robust flag-based “detector region tomography” (LSD-DRT). Its detector design uses $$\Lambda(\rho)\;=\;\sum_{e\in P^n}P(e)\;e\,\rho\,e^\dagger.$$3 sequential syndrome extractions and forms $$\Lambda(\rho)\;=\;\sum_{e\in P^n}P(e)\;e\,\rho\,e^\dagger.$$4 disjoint detectors

$$\Lambda(\rho)\;=\;\sum_{e\in P^n}P(e)\;e\,\rho\,e^\dagger.$$5

This enforces that noise in each detector region is i.i.d. The experiment prepares a $$\Lambda(\rho)\;=\;\sum_{e\in P^n}P(e)\;e\,\rho\,e^\dagger.$$6 eigenstate of each logical Pauli $$\Lambda(\rho)\;=\;\sum_{e\in P^n}P(e)\;e\,\rho\,e^\dagger.$$7, runs the $$\Lambda(\rho)\;=\;\sum_{e\in P^n}P(e)\;e\,\rho\,e^\dagger.$$8 gadgets, and then destructively measures $$\Lambda(\rho)\;=\;\sum_{e\in P^n}P(e)\;e\,\rho\,e^\dagger.$$9 together with the full stabilizer set, so that one simultaneously obtains the logical outcome and the final syndrome.

After Pauli twirl, each conditioned channel $P_L(e)\;=\;\frac1{|S|}\sum_{s\in S}P(e\,s),$0 diagonalizes on the normalizer: $P_L(e)\;=\;\frac1{|S|}\sum_{s\in S}P(e\,s),$1 For a multiset of detector outcomes $P_L(e)\;=\;\frac1{|S|}\sum_{s\in S}P(e\,s),$2, the expectation value of measuring $P_L(e)\;=\;\frac1{|S|}\sum_{s\in S}P(e\,s),$3 depends only on the detector counts $P_L(e)\;=\;\frac1{|S|}\sum_{s\in S}P(e\,s),$4: $P_L(e)\;=\;\frac1{|S|}\sum_{s\in S}P(e\,s),$5 The term $P_L(e)\;=\;\frac1{|S|}\sum_{s\in S}P(e\,s),$6 is the SPAM offset. Fitting $P_L(e)\;=\;\frac1{|S|}\sum_{s\in S}P(e\,s),$7, $P_L(e)\;=\;\frac1{|S|}\sum_{s\in S}P(e\,s),$8, and $P_L(e)\;=\;\frac1{|S|}\sum_{s\in S}P(e\,s),$9 to repeated-length data decouples SPAM from the detector-region eigenvalues.

The hardware assumptions are explicit. Between gadgets one applies a random Pauli on the data, termed “Pauli frame randomization,” to enforce a stochastic Pauli channel on average. Every few gadgets one swaps data and ancilla qubits and measures out all physical qubits in order to flush leakage back into the computational space. The method is stated to be most suitable for flag-based syndrome measurement schemes (Girling et al., 11 Aug 2025).

3. Identifiability and statistical reconstruction

For arbitrary stabilizer codes, subsystem codes, and data syndrome codes, the logical error channel induced by Pauli noise can be estimated from syndrome data under minimal conditions. In the stabilizer-code setting, the relevant correctness assumption is that the physical Pauli channel factorizes into independent local parts,

$$\Lambda_L(\rho)\;=\;\sum_{\ell\in P^n}P_L(\ell)\,\ell\,\rho\,\ell^\dagger.$$0

with each $$\Lambda_L(\rho)\;=\;\sum_{\ell\in P^n}P_L(\ell)\,\ell\,\rho\,\ell^\dagger.$$1 acting only on a support region $$\Lambda_L(\rho)\;=\;\sum_{\ell\in P^n}P_L(\ell)\,\ell\,\rho\,\ell^\dagger.$$2, such that every union $$\Lambda_L(\rho)\;=\;\sum_{\ell\in P^n}P_L(\ell)\,\ell\,\rho\,\ell^\dagger.$$3 is a correctable region and each local channel satisfies $$\Lambda_L(\rho)\;=\;\sum_{\ell\in P^n}P_L(\ell)\,\ell\,\rho\,\ell^\dagger.$$4. Under these conditions, Theorem 1 states that the logical channel $$\Lambda_L(\rho)\;=\;\sum_{\ell\in P^n}P_L(\ell)\,\ell\,\rho\,\ell^\dagger.$$5 is uniquely determined by the syndrome measurement statistics (Wagner et al., 2022).

The reconstruction is expressed in terms of Walsh–Hadamard moments

$$\Lambda_L(\rho)\;=\;\sum_{\ell\in P^n}P_L(\ell)\,\ell\,\rho\,\ell^\dagger.$$6

and canonical moments $$\Lambda_L(\rho)\;=\;\sum_{\ell\in P^n}P_L(\ell)\,\ell\,\rho\,\ell^\dagger.$$7 defined by Möbius inversion. For each measured stabilizer $$\Lambda_L(\rho)\;=\;\sum_{\ell\in P^n}P_L(\ell)\,\ell\,\rho\,\ell^\dagger.$$8,

$$\Lambda_L(\rho)\;=\;\sum_{\ell\in P^n}P_L(\ell)\,\ell\,\rho\,\ell^\dagger.$$9

and after taking logarithms,

$\ell\in S^\perp=L$0

One solves the sparse linear system for $\ell\in S^\perp=L$1, reconstructs the logical moments

$\ell\in S^\perp=L$2

and finally inverts the Walsh–Hadamard transform on $\ell\in S^\perp=L$3: $\ell\in S^\perp=L$4 To estimate each $\ell\in S^\perp=L$5 within $\ell\in S^\perp=L$6 requires $\ell\in S^\perp=L$7 syndrome rounds. The linear solve is $\ell\in S^\perp=L$8 in the worst case, or quasi-linear time if one exploits locality (Wagner et al., 2022).

A complementary statistical route is used in LSD-DRT. For each logical $\ell\in S^\perp=L$9 and detector-count vector $$\Lambda_L(\rho)\;=\;\sum_{\bar\ell\in L/S}p_{\bar\ell}\,\bigl(U_{\bar\ell}\,\rho\,U_{\bar\ell}^\dagger\bigr),$$0, one gathers binary outcomes $$\Lambda_L(\rho)\;=\;\sum_{\bar\ell\in L/S}p_{\bar\ell}\,\bigl(U_{\bar\ell}\,\rho\,U_{\bar\ell}^\dagger\bigr),$$1 and models

$$\Lambda_L(\rho)\;=\;\sum_{\bar\ell\in L/S}p_{\bar\ell}\,\bigl(U_{\bar\ell}\,\rho\,U_{\bar\ell}^\dagger\bigr),$$2

The classical post-processing employs a conjugate Beta–Binomial hierarchy: the likelihood is $$\Lambda_L(\rho)\;=\;\sum_{\bar\ell\in L/S}p_{\bar\ell}\,\bigl(U_{\bar\ell}\,\rho\,U_{\bar\ell}^\dagger\bigr),$$3; the prior on $$\Lambda_L(\rho)\;=\;\sum_{\bar\ell\in L/S}p_{\bar\ell}\,\bigl(U_{\bar\ell}\,\rho\,U_{\bar\ell}^\dagger\bigr),$$4 is $$\Lambda_L(\rho)\;=\;\sum_{\bar\ell\in L/S}p_{\bar\ell}\,\bigl(U_{\bar\ell}\,\rho\,U_{\bar\ell}^\dagger\bigr),$$5; priors on $$\Lambda_L(\rho)\;=\;\sum_{\bar\ell\in L/S}p_{\bar\ell}\,\bigl(U_{\bar\ell}\,\rho\,U_{\bar\ell}^\dagger\bigr),$$6 are rescaled $$\Lambda_L(\rho)\;=\;\sum_{\bar\ell\in L/S}p_{\bar\ell}\,\bigl(U_{\bar\ell}\,\rho\,U_{\bar\ell}^\dagger\bigr),$$7; and weakly-informative or flat priors are placed on $$\Lambda_L(\rho)\;=\;\sum_{\bar\ell\in L/S}p_{\bar\ell}\,\bigl(U_{\bar\ell}\,\rho\,U_{\bar\ell}^\dagger\bigr),$$8, $$\Lambda_L(\rho)\;=\;\sum_{\bar\ell\in L/S}p_{\bar\ell}\,\bigl(U_{\bar\ell}\,\rho\,U_{\bar\ell}^\dagger\bigr),$$9, and all hyperparameters. Posterior sampling, for example via PyMC, yields joint samples of $U_{\bar\ell}$0, $U_{\bar\ell}$1, and $U_{\bar\ell}$2, from which one computes credible intervals for $U_{\bar\ell}$3 and, via the Walsh–Hadamard transform, for $U_{\bar\ell}$4 (Girling et al., 11 Aug 2025).

These two frameworks answer different identifiability questions. The moment method shows when syndrome data alone determine a logical Pauli channel. LSD-DRT shows how to estimate syndrome-dependent logical Pauli channels while separating SPAM from detector-region noise.

4. Twirling, locality restoration, and Pauli-diagonal logical maps

For a generic physical noise channel $U_{\bar\ell}$5 acting on an $U_{\bar\ell}$6 stabilizer code, Pauli twirling over the $U_{\bar\ell}$7-qubit Pauli group maps $U_{\bar\ell}$8 to

$U_{\bar\ell}$9

which has no off-diagonal Pauli-transfer-matrix elements. Because the recovery super-operator $\bar\ell\in L/S$0 is itself a Pauli channel, the logically twirled channel

$\bar\ell\in L/S$1

remains a logical Pauli channel,

$\bar\ell\in L/S$2

This is the standard route by which coherent physical noise is converted into an incoherent encoded Pauli model (Cai et al., 2019).

In approximate GKP error correction with finite-energy ancillae, the situation is subtler. Inserting non-unitary damping operators

$\bar\ell\in L/S$3

into teleportation-based stabilizer-measurement circuits causes the naïve identification “one round of error correction $\bar\ell\in L/S$4 one qubit-CPTP map” to fail. The stated reasons are that finite-energy GKP states leak outside the ideal code space and that adjacent rounds become classically correlated through the outcome-dependent shift of the envelope. Two twirls are introduced to restore locality and trace-preservation at the logical level: a stabilizer twirl over the GKP stabilizer group and a Pauli-shift twirl over the four logical displacements $\bar\ell\in L/S$5. Once the noise has been twirled in either way, the decoded logical map on the qubit is a Pauli channel,

$\bar\ell\in L/S$6

Under the stabilizer-twirled approximation with standard-binning decoding, one finds $\bar\ell\in L/S$7, $\bar\ell\in L/S$8, and in leading order

$\bar\ell\in L/S$9

Under the Pauli-twirled approach with optimized lookup decoding, all three nontrivial error probabilities can be nonzero. For $\Lambda_L$0, corresponding to $\Lambda_L$1 dB effective squeezing, the GRN–standard-binning, Pauli-twirled + standard-binning, and Pauli-twirled + optimal curves are all virtually indistinguishable, with differences $\Lambda_L$2. In the high-energy limit $\Lambda_L$3, both twirling procedures collapse to the same stabilizer-twirled channel (Jafarzadeh et al., 18 Apr 2025).

A common misconception is that twirling is merely a calculational convenience. In the finite-energy GKP setting, the results instead identify it as the mechanism that restores a bona fide logical Pauli-channel description.

5. Exact non-Pauli logical channels and the limits of Pauli modeling

The strongest qualification to the logical-Pauli-channel framework comes from pure loss and heralded loss acting on approximate GKP qubits. A single-mode pure-loss channel of transmissivity $\Lambda_L$4 admits both a photon-counting Kraus form,

$\Lambda_L$5

and a heterodyne Kraus form,

$\Lambda_L$6

Composed with finite-energy damping and ideal GKP error correction, this induces a qubit-level map whose process matrix $\Lambda_L$7 in the GKP-Pauli basis $\Lambda_L$8 is obtained from conditional Kraus operators $\Lambda_L$9 and Bloch-vector components

$S\subset P^n$00

The exact unconditional channel follows after averaging over $S\subset P^n$01 and then over the syndrome $S\subset P^n$02 (Harris et al., 18 Apr 2025).

A stochastic Pauli channel has a process matrix diagonal in $S\subset P^n$03. By contrast, the exact pure-loss logical channel has generically nonzero $S\subset P^n$04. The data state that for moderate loss, for example $S\subset P^n$05, the magnitude of the off-diagonal $S\subset P^n$06 elements becomes appreciable, signaling coherent non-Pauli distortions. If one Pauli-twirls the logical map, all of these off-diagonals vanish and the result becomes a classical mixture of $S\subset P^n$07. In the small-loss limit $S\subset P^n$08, $S\subset P^n$09, the dominant errors are small-displacement-like, with $S\subset P^n$10 and $S\subset P^n$11. Hence for $S\subset P^n$12 the Pauli-twirl is a decent approximation, but for larger loss or higher GKP squeezing the exact non-Pauli corrections are essential (Harris et al., 18 Apr 2025).

Heralded loss is even more strongly non-Pauli. If one retains the ancilla outcome instead of tracing it out, then photon-subtraction and loss-plus-heterodyne measurements generate conditional logical maps with process matrices

$S\subset P^n$13

For photon-subtraction with $S\subset P^n$14, the Bloch components are obtained by discrete derivatives of the heterodyne expression, and numerically one finds coherent rotations outside the Pauli tetrahedron for $S\subset P^n$15. Averaging over $S\subset P^n$16 but not the ancilla outcome yields highly non-Pauli logical operations that, when applied to GKP Pauli eigenstates, trace out magic-rich trajectories (Harris et al., 18 Apr 2025).

These results correct another common misconception: the logical channel induced by physically relevant noise is not always a Pauli channel, even when the code and recovery are Pauli structured. In such cases, the Pauli description is an approximation obtained by twirling or coarse-graining, not the exact encoded dynamics.

6. Noise tailoring, Pauli conjugation, and mitigation strategies

One mitigation strategy is to tailor the noise so that the assumptions underlying logical Pauli-channel models become empirically valid. In detector-region tomography, Pauli frame randomization is inserted between gadgets to enforce a stochastic Pauli channel on average, and leakage-flushing SWAP operations are used to return leaked population to the computational subspace. The stated effect is that several noise diagnostic tests for fault tolerance improve significantly when using such tailoring and mitigation strategies (Girling et al., 11 Aug 2025).

A distinct strategy, developed for coherent physical noise, is Pauli conjugation. Instead of averaging over all Pauli gates, one deterministically sandwiches the noise between a chosen pair $S\subset P^n$17: $S\subset P^n$18 After recovery,

$S\subset P^n$19

and the corresponding logical channel is again a Pauli channel on the logical space,

$S\subset P^n$20

The search for an optimal conjugation removes stabilizer and logical generators, removes any Pauli that commutes with every term in $S\subset P^n$21, and groups the remainder into equivalence classes under the code-plus-noise symmetry group $S\subset P^n$22 (Cai et al., 2019).

Under global coherent $S\subset P^n$23 noise,

$S\subset P^n$24

the optimal conjugation scheme exceeds both do-nothing and full twirling in logical fidelity for the benchmark codes studied. For the Steane $S\subset P^n$25 code, the thresholds are $S\subset P^n$26 with no mitigation, $S\subset P^n$27 with twirling, and $S\subset P^n$28 with conjugation. For the 9-qubit Shor $S\subset P^n$29 code, the corresponding values are $S\subset P^n$30, $S\subset P^n$31, and $S\subset P^n$32. For the distance-3 surface $S\subset P^n$33 code, they are $S\subset P^n$34, $S\subset P^n$35, and $S\subset P^n$36. The simulations further state that the scheme is robust to gate errors and that, with logical twirling interleaved between rounds, the favorable single-round comparison persists over multiple rounds (Cai et al., 2019).

Taken together, these mitigation results show that logical Pauli channels can be either enforced by randomized tailoring or exploited through deterministic Pauli conjugation, depending on whether the objective is faithful stochastic modeling