Subsystem-Balanced Pauli Twirling (SB-PT)

• Subsystem-Balanced Pauli Twirling (SB-PT) is a method that applies balanced random Pauli or Clifford operations to selected quantum subsystems to transform structured noise into a symmetric effective channel while preserving invariant features.
• It is used in measurement error mitigation and quantum key distribution to reduce independent error components with efficient O(4^τ) circuit scaling and achieve isotropic depolarization.
• SB-PT leverages algebraic cancellation and reduced twirling sets to optimize resource allocation and rigorously preserve subsystem relationships critical for interference and error symmetry.

Subsystem-Balanced Pauli Twirling (SB-PT) denotes a class of twirling constructions in which random Pauli or Clifford operations are balanced across selected subsystems so that structured noise is converted into a more symmetric effective channel while preserving the task-relevant invariant structure. In model-free measurement error mitigation (MF-MEM), SB-PT enforces Pauli operator balance on the measuring subsystem of a Pauli observable and, for a weight-$\tau$ Pauli observable, removes all independent error components using only $O(4^\tau)$ random circuits (Xu et al., 22 Sep 2025). In measurement-device-independent quantum key distribution (MDI-QKD), the same balancing idea appears as correlated twirling, where identical unitaries act as $U\otimes U$ and the relative polarization operator is converted into an isotropic depolarizing channel; the paper explicitly identifies its correlated twirling protocol as a concrete realization of SB-PT (Pewkhom et al., 8 May 2026). Related formulations connect subsystem-balanced twirling to reduced Pauli twirling sets for arbitrary channels and to local symmetric Clifford twirling around non-Clifford layers (Cai et al., 2018, Tsubouchi et al., 2024).

1. Formal definitions and scope

At the level of channels, twirling is the supermap

$$\mathcal{T}_G(\Lambda)(\rho)=\frac{1}{|G|}\sum_{U\in G}U^\dagger\,\Lambda(U\rho U^\dagger)\,U,$$

with the single-qubit Pauli-twirling specialization

$$T(\Lambda)(\rho)=\frac{1}{4}\sum_{P\in\{I,X,Y,Z\}}P^\dagger \Lambda(P\rho P^\dagger)P.$$

For subsystem-balanced constructions, the twirling set is not treated as an undifferentiated global group. Instead, the balancing is imposed on selected subsystems: in MF-MEM the relevant subsystem is the support of the measured Pauli observable; in MDI-QKD it is the two-party polarization subspace acted on by correlated $U\otimes U$ conjugation; in reduced Pauli twirling it is a partition of the $n$-qubit register into local factors $G_1,\dots,G_k$; and in symmetric Clifford twirling it is a collection of local Clifford groups constrained to commute with specified Pauli subgroups (Xu et al., 22 Sep 2025, Pewkhom et al., 8 May 2026, Cai et al., 2018, Tsubouchi et al., 2024).

Setting	Balanced object	Effective outcome
MF-MEM	Pauli cover on $\mathrm{supp}(Z_r)$	State-independent scaling for the target observable
MDI-QKD	Correlated $U\otimes U$ twirl on both arms	Isotropic depolarizing channel in the relative polarization subspace
Reduced channel twirling	Product set $O(4^\tau)$0	Pauli-diagonal channel with reduced twirling-set size
Symmetric Clifford twirling	Local Cliffords commuting with subsystem symmetries	Randomization of non-invariant Pauli components

A central distinction from standard uncorrelated Pauli twirling is that SB-PT preserves a structured relation between subsystems. In the MDI-QKD setting, this relation is the relative indistinguishability needed for Hong-Ou-Mandel interference; in MF-MEM it is the support pattern of the measured observable; in symmetric Clifford twirling it is the commutation relation with the subgroup $O(4^\tau)$1 associated with the non-Clifford layer (Pewkhom et al., 8 May 2026, Xu et al., 22 Sep 2025, Tsubouchi et al., 2024).

2. Algebraic mechanism of subsystem balance

The algebraic criterion underlying reduced Pauli twirling is the off-diagonal cancellation condition

$O(4^\tau)$2

where $O(4^\tau)$3 is the Pauli support of the error channel and $O(4^\tau)$4 is the commutator function defined by $O(4^\tau)$5. When this condition holds, all off-diagonal $O(4^\tau)$6 terms with $O(4^\tau)$7 vanish in the Pauli basis, and the twirled channel becomes a Pauli channel with diagonal coefficients $O(4^\tau)$8 (Cai et al., 2018).

For subsystem factorizations, the same cancellation can be enforced locally. If $O(4^\tau)$9 and the projected factors satisfy local zero-sum conditions, then

$U\otimes U$0

This is the formal basis for subsystem-balanced constructions derived from reduced twirling sets. The paper further gives size relations

$U\otimes U$1

showing that the required twirl set can be much smaller than the full $U\otimes U$2-element Pauli set when the channel’s Pauli support is structured (Cai et al., 2018).

In MF-MEM, the cancellation mechanism is specialized to classical measurement noise in the Pauli-transfer-matrix representation. For a twirling set $U\otimes U$3,

$U\otimes U$4

SB-PT constructs $U\otimes U$5 so that each $U\otimes U$6 appears equally often on each qubit of $U\otimes U$7. Theorem 2 states that for the trigger set $U\otimes U$8, one has

$U\otimes U$9

so all independent error components supported on the measuring subsystem are exactly canceled (Xu et al., 22 Sep 2025).

In correlated twirling for MDI-QKD, the same algebraic theme is realized through unitary 2-design averaging and Schur’s Lemma. The relative operator $$\mathcal{T}_G(\Lambda)(\rho)=\frac{1}{|G|}\sum_{U\in G}U^\dagger\,\Lambda(U\rho U^\dagger)\,U,$$0 is averaged by conjugation with a 12-element set $$\mathcal{T}_G(\Lambda)(\rho)=\frac{1}{|G|}\sum_{U\in G}U^\dagger\,\Lambda(U\rho U^\dagger)\,U,$$1, yielding an isotropic depolarizing channel rather than an axis-dependent rotation (Pewkhom et al., 8 May 2026).

3. Correlated twirling as SB-PT in MDI-QKD

For MDI-QKD, fiber-induced polarization drift is modeled by the relative rotation

$$\mathcal{T}_G(\Lambda)(\rho)=\frac{1}{|G|}\sum_{U\in G}U^\dagger\,\Lambda(U\rho U^\dagger)\,U,$$2

The correlated twirling supermap acts as

$$\mathcal{T}_G(\Lambda)(\rho)=\frac{1}{|G|}\sum_{U\in G}U^\dagger\,\Lambda(U\rho U^\dagger)\,U,$$3

and, on the relative single-qubit state,

$$\mathcal{T}_G(\Lambda)(\rho)=\frac{1}{|G|}\sum_{U\in G}U^\dagger\,\Lambda(U\rho U^\dagger)\,U,$$4

The depolarizing parameter is fixed by

$$\mathcal{T}_G(\Lambda)(\rho)=\frac{1}{|G|}\sum_{U\in G}U^\dagger\,\Lambda(U\rho U^\dagger)\,U,$$5

so that for $$\mathcal{T}_G(\Lambda)(\rho)=\frac{1}{|G|}\sum_{U\in G}U^\dagger\,\Lambda(U\rho U^\dagger)\,U,$$6,

$$\mathcal{T}_G(\Lambda)(\rho)=\frac{1}{|G|}\sum_{U\in G}U^\dagger\,\Lambda(U\rho U^\dagger)\,U,$$7

The paper’s analytical “suppression by a factor of $$\mathcal{T}_G(\Lambda)(\rho)=\frac{1}{|G|}\sum_{U\in G}U^\dagger\,\Lambda(U\rho U^\dagger)\,U,$$8” appears in the protected basis guessing probability

$$\mathcal{T}_G(\Lambda)(\rho)=\frac{1}{|G|}\sum_{U\in G}U^\dagger\,\Lambda(U\rho U^\dagger)\,U,$$9

in contrast to worst-case unprotected scaling $$T(\Lambda)(\rho)=\frac{1}{4}\sum_{P\in\{I,X,Y,Z\}}P^\dagger \Lambda(P\rho P^\dagger)P.$$0 along a bad axis (Pewkhom et al., 8 May 2026).

The protocol is executed as a virtual post-processing step. Alice and Bob use a public random beacon to select the same $$T(\Lambda)(\rho)=\frac{1}{4}\sum_{P\in\{I,X,Y,Z\}}P^\dagger \Lambda(P\rho P^\dagger)P.$$1 for each transmission window, conceptually scramble with $$T(\Lambda)(\rho)=\frac{1}{4}\sum_{P\in\{I,X,Y,Z\}}P^\dagger \Lambda(P\rho P^\dagger)P.$$2, transmit through independent channels $$T(\Lambda)(\rho)=\frac{1}{4}\sum_{P\in\{I,X,Y,Z\}}P^\dagger \Lambda(P\rho P^\dagger)P.$$3 and $$T(\Lambda)(\rho)=\frac{1}{4}\sum_{P\in\{I,X,Y,Z\}}P^\dagger \Lambda(P\rho P^\dagger)P.$$4, and then compute $$T(\Lambda)(\rho)=\frac{1}{4}\sum_{P\in\{I,X,Y,Z\}}P^\dagger \Lambda(P\rho P^\dagger)P.$$5 during sifting by a deterministic look-up table. The procedure requires no quantum hardware changes, uses roughly 4 bits of public randomness per choice from $$T(\Lambda)(\rho)=\frac{1}{4}\sum_{P\in\{I,X,Y,Z\}}P^\dagger \Lambda(P\rho P^\dagger)P.$$6, and adds only constant-time classical processing per successful detection event (Pewkhom et al., 8 May 2026).

The reported performance gains are explicitly quantitative. Simulations show that the induced symmetry neutralizes catastrophic axis-dependent failures, extending the Y-bias tolerance from $$T(\Lambda)(\rho)=\frac{1}{4}\sum_{P\in\{I,X,Y,Z\}}P^\dagger \Lambda(P\rho P^\dagger)P.$$7 to $$T(\Lambda)(\rho)=\frac{1}{4}\sum_{P\in\{I,X,Y,Z\}}P^\dagger \Lambda(P\rho P^\dagger)P.$$8 radians and increasing the absolute angular misalignment tolerance at the $$T(\Lambda)(\rho)=\frac{1}{4}\sum_{P\in\{I,X,Y,Z\}}P^\dagger \Lambda(P\rho P^\dagger)P.$$9 QBER threshold from $U\otimes U$0 to $U\otimes U$1. The Bell-projector definition

$U\otimes U$2

is used together with attenuation and detector-dark-count parameters $U\otimes U$3, $U\otimes U$4, and $U\otimes U$5 to quantify the extended secure-distance regime. The same paper states that detector-independence is unchanged, composability is preserved because the twirl is public pre-processing, and “twirling operators commute with intensity modulation,” ensuring compatibility with weak-coherent-pulse decoy protocols (Pewkhom et al., 8 May 2026).

4. SB-PT for model-free measurement error mitigation

In MF-MEM, measurement is modeled as a classical stochastic map with transfer matrix $U\otimes U$6 after complete dephasing, and the reduced Pauli-transfer matrix $U\otimes U$7 acts only on the Pauli-$U\otimes U$8 sector. Under the tensor-product-noise model $U\otimes U$9, one has $n$0 with single-qubit parameters $n$1 and $n$2, where $n$3 and $n$4. For a Pauli observable $n$5, the measuring subsystem is $n$6 with weight $n$7 (Xu et al., 22 Sep 2025).

SB-PT constructs a stratified twirling set $n$8 of size $n$9, $G_1,\dots,G_k$0, such that each Pauli $G_1,\dots,G_k$1 appears equally often on every qubit in $G_1,\dots,G_k$2, while Paulis on the complement are sampled uniformly at random. For a weight-$G_1,\dots,G_k$3 observable, this yields the stated $G_1,\dots,G_k$4 circuit scaling. The associated mitigated estimator is

$G_1,\dots,G_k$5

with $G_1,\dots,G_k$6, and under full Pauli twirling the measured expectation reduces to

$G_1,\dots,G_k$7

Theorem 3 gives the SB-PT error bound

$G_1,\dots,G_k$8

where

$G_1,\dots,G_k$9

The paper states that this bound is strictly tighter than the random-twirling bound at equal twirling-set size (Xu et al., 22 Sep 2025).

To extend the method to arbitrary observables, the same work introduces a hardware-efficient measurement transformation $\mathrm{supp}(Z_r)$0 such that

$\mathrm{supp}(Z_r)$1

The basic primitives are a CNOT ladder

$\mathrm{supp}(Z_r)$2

a weight-reduction map

$\mathrm{supp}(Z_r)$3

and a location-shift gadget

$\mathrm{supp}(Z_r)$4

The measurement-transformation circuit uses $\mathrm{supp}(Z_r)$5 two-qubit gates and has linear depth, while unified SB-PT suppresses off-diagonals of the combined effective channel $\mathrm{supp}(Z_r)$6 on the effective support (Xu et al., 22 Sep 2025).

The reported numerical results are also explicit. For sparse observables with $\mathrm{supp}(Z_r)$7, a marked improvement occurs once the number of random circuits exceeds $\mathrm{supp}(Z_r)$8. For global observables, MT(sub) reduces global $\mathrm{supp}(Z_r)$9 to weight 1 and achieves near-optimal mitigation with $U\otimes U$0, matching or exceeding alternatives at $U\otimes U$1, corresponding to a greater than 16-fold improvement in sampling efficiency. The favorable scaling persists up to at least 10 qubits, and robustness is reported under moderate correlated measurement errors and coherent gate errors $U\otimes U$2 (Xu et al., 22 Sep 2025).

5. Reduced twirling sets and symmetric Clifford realizations

Reduced Pauli twirling provides a constructive route to subsystem-balanced sets tailored to the Pauli support $U\otimes U$3 of the error channel. For an error channel $U\otimes U$4, the construction begins with $U\otimes U$5, where $U\otimes U$6, and builds a generating set $U\otimes U$7 through quotient-table machinery so that the commutator table of $U\otimes U$8 matches the image of $U\otimes U$9 inside an auxiliary group $O(4^\tau)$00. The paper states that full Pauli twirling on $O(4^\tau)$01 qubits uses $O(4^\tau)$02, but the reduced construction can be exponentially smaller in symmetric cases; for a global $O(4^\tau)$03-field on $O(4^\tau)$04 spins with $O(4^\tau)$05, the construction yields $O(4^\tau)$06 and $O(4^\tau)$07, compared with generating-set size $O(4^\tau)$08 for full Pauli twirling and brute-force set size $O(4^\tau)$09 (Cai et al., 2018).

The same paper proves that one-gate twirling is equivalent to a stabilizer measurement with discarded outcome. If $O(4^\tau)$10 is the decomposition into parts commuting and anticommuting with a Pauli $O(4^\tau)$11, then

$O(4^\tau)$12

This permits replacement of some twirl gates by existing stabilizer checks, thereby reducing depth and randomization overhead (Cai et al., 2018).

Symmetric Clifford twirling supplies a second realization of SB-PT. For a non-Clifford unitary $O(4^\tau)$13, define

$O(4^\tau)$14

Twirling by $O(4^\tau)$15 randomizes the Pauli components that do not commute with the symmetry subgroup while leaving invariant components unchanged. In the canonical case $O(4^\tau)$16 on qubit 1, $O(4^\tau)$17; $O(4^\tau)$18 or $O(4^\tau)$19 noise on qubit 1 is uniformly spread over $O(4^\tau)$20 and $O(4^\tau)$21, whereas $O(4^\tau)$22 noise is invariant (Tsubouchi et al., 2024).

The quantitative scrambling metric is the distance

$O(4^\tau)$23

For single-qubit $O(4^\tau)$24 noise under full symmetric Clifford twirling, $O(4^\tau)$25 is reduced from $O(4^\tau)$26 to $O(4^\tau)$27; for single-qubit depolarizing noise, $O(4^\tau)$28 is reduced from $O(4^\tau)$29 to $O(4^\tau)$30; and for invariant $O(4^\tau)$31 noise, $O(4^\tau)$32 remains $O(4^\tau)$33. The $O(4^\tau)$34-sparse local version gives

$O(4^\tau)$35

and is described as hardware-light, with $O(4^\tau)$36 requiring at most one CNOT plus single-qubit Cliffords. For Pauli noise with total error $O(4^\tau)$37, cost-optimal rescaling gives sampling overhead $O(4^\tau)$38, compared with probabilistic error cancellation scaling $O(4^\tau)$39 (Tsubouchi et al., 2024).

6. Distinctions, assumptions, and limitations

SB-PT is not identical to standard uncorrelated Pauli twirling. In the MDI-QKD use case, standard independent twirling on each subsystem can destroy the relative alignment needed for two-photon interference, whereas subsystem-balanced correlated twirling applies the same unitary to each subsystem and preserves the relative structure needed for Hong-Ou-Mandel interference. The same section of the MDI-QKD paper also distinguishes Pauli-only averaging from the 12-element unitary 2-design actually used there: the protocol is stronger than Pauli-only twirling and guarantees the desired isotropy with minimal set size (Pewkhom et al., 8 May 2026).

The assumptions are similarly context-dependent. The MDI-QKD analysis assumes that polarization noise is well modeled by unitary $O(4^\tau)$40 rotations with negligible non-unitary effects such as polarization-dependent loss, mode coupling, or dephasing not captured by the $O(4^\tau)$41 model. If drift varies faster than the synchronization window, averaging may be imperfect, although the protocol still symmetrizes statistics over time. In MF-MEM, the strongest guarantees target the tensor-product trigger structure of classical measurement noise; strongly correlated noise spanning across subsystems, non-Pauli measurement errors, and coherent errors not fully randomized by SB-PT can leave residual bias and may require larger $O(4^\tau)$42 or complementary methods. In reduced Pauli twirling, strongly nonlocal Pauli support can break subsystem factorization and force the use of nonlocal generators or enlarged local sets. In symmetric Clifford twirling, components commuting with $O(4^\tau)$43, notably $O(4^\tau)$44-type components under $O(4^\tau)$45-rotations, are invariants and therefore set the performance floor (Xu et al., 22 Sep 2025, Cai et al., 2018, Tsubouchi et al., 2024, Pewkhom et al., 8 May 2026).

Several comparisons clarify the method’s scope. Hardware automatic polarization control in QKD attempts to enforce $O(4^\tau)$46 physically, using reference pulses and downtime; correlated twirling is hardware-free and instead replaces axis-dependent failures with isotropic depolarization. In MF-MEM, SB-PT avoids full noise-map reconstruction and exponential calibration cost, but it is not a matrix-inversion method and does not reconstruct the complete measurement channel. A plausible implication is that SB-PT is best viewed as a symmetry-engineering and resource-allocation strategy rather than a universal replacement for full characterization: it trades exact channel inversion for structured cancellation, reduced overhead, and compatibility with the native symmetries of the task (Pewkhom et al., 8 May 2026, Xu et al., 22 Sep 2025).