Symmetry Verification: Quantum Error Mitigation

• Symmetry Verification is a technique that leverages invariant properties to project noisy quantum states onto desired subspaces, reducing error rates.
• It employs group-theoretic and circuit-level methods, including post-selection and spatio-temporal stabilizers, to ensure fidelity in both VQE and QAOA algorithms.
• Experimental validations show improvements in energy error and approximation ratios, with symmetry checks effectively filtering out detrimental noise.

Symmetry verification (SV) denotes a family of procedures that exploit known invariances to test, enforce, or certify membership in a symmetry sector. In quantum computing, the dominant usage is a quantum error mitigation technique that projects a noisy state onto the eigenspace of a symmetry operator and discards symmetry-violating runs; later work extends the idea from states to circuits and to channels. In few-particle physics, a distinct earlier usage denotes symmetrization verification of two-particle states by one-particle detection and tomography. Related verification literatures use symmetry to reduce proof obligations, infer quantified invariants, or certify symmetry-breaking constraints, but these are methodologically distinct from post-selection-based quantum SV (Sagastizabal et al., 2019, Kakkar et al., 2022, Tsubouchi et al., 17 Mar 2025, Sancho, 2011).

1. State-level SV as projection onto a symmetry sector

In the conventional, state-oriented formulation, one assumes that the ideal output state $|\psi\rangle$ is stabilized by some unitary $\hat S\in\mathcal P^N$, so that $\hat S|\psi\rangle=|\psi\rangle$ and $[\hat S,H]=0$ for the relevant Hamiltonian. The Hilbert space then splits as $\mathcal H=\mathcal H_+\oplus\mathcal H_-$, with projectors

$$P_+ = \frac{I+\hat S}{2}, \qquad P_-=\frac{I-\hat S}{2}.$$

If a noisy preparation yields $\rho$, symmetry verification implements

$$\rho \mapsto \rho_{\mathrm{SV}}=\frac{P_+\rho P_+}{\mathrm{Tr}[P_+\rho]},$$

that is, post-selection onto the desired symmetry subspace (Sagastizabal et al., 2019).

A more general group-theoretic formulation replaces a single Pauli symmetry by a finite or compact group $G$ with unitary representation $\{U_g\}$. The Hilbert space decomposes into irreducible-representation sectors labeled by $\hat S\in\mathcal P^N$0, with projectors

$\hat S\in\mathcal P^N$1

Conventional SV then prepares a noisy output $\hat S\in\mathcal P^N$2, measures the POVM $\hat S\in\mathcal P^N$3, and post-selects the desired sector $\hat S\in\mathcal P^N$4, yielding

$\hat S\in\mathcal P^N$5

This formulation makes explicit that conventional SV is state-level: it requires the input to lie in a symmetry sector and uses symmetry only at the end (Tsubouchi et al., 17 Mar 2025).

The operational meaning of SV is therefore selective rejection of symmetry-breaking noise. If the noise channel can be written as $\hat S\in\mathcal P^N$6, and if each error Kraus $\hat S\in\mathcal P^N$7 satisfies $\hat S\in\mathcal P^N$8, then after SV the surviving state is proportional to the ideal output, with success rate $\hat S\in\mathcal P^N$9. This identifies the central limitation of state-level SV: it filters only errors that move weight outside the target sector, while in-sector errors remain invisible (Tsubouchi et al., 17 Mar 2025).

2. Embedding SV in VQE and QAOA

Within VQE, SV can be implemented entirely in post-processing. For an $\hat S|\psi\rangle=|\psi\rangle$0-qubit Hamiltonian $\hat S|\psi\rangle=|\psi\rangle$1 commuting with a Pauli-string symmetry $\hat S|\psi\rangle=|\psi\rangle$2, the corrected expectation value of any Pauli observable $\hat S|\psi\rangle=|\psi\rangle$3 in target sector $\hat S|\psi\rangle=|\psi\rangle$4 is

$\hat S|\psi\rangle=|\psi\rangle$5

The symmetry-verified density matrix is then reconstructed as

$\hat S|\psi\rangle=|\psi\rangle$6

For the $\hat S|\psi\rangle=|\psi\rangle$7 molecule in a minimal STO-3G basis, Bravyi-Kitaev mapped and with two qubits projected out, the effective two-qubit Hamiltonian commutes with $\hat S|\psi\rangle=|\psi\rangle$8, so total two-qubit parity is conserved (Sagastizabal et al., 2019).

The same work couples SV to positivity restoration. Direct linear-inversion tomography may yield $\hat S|\psi\rangle=|\psi\rangle$9 with small negative eigenvalues, so one solves the semidefinite program

$[\hat S,H]=0$0

SV is then applied to $[\hat S,H]=0$1 or to its Pauli coefficients. In this setting, relaxation and residual excitation change total excitation number and therefore violate parity, whereas dephasing commutes with $[\hat S,H]=0$2 and is not removed by the projection (Sagastizabal et al., 2019).

For QAOA, the standard example is MaxCut on $[\hat S,H]=0$3 qubits, where

$[\hat S,H]=0$4

is invariant under the global bit-flip $[\hat S,H]=0$5 symmetry

$[\hat S,H]=0$6

If $[\hat S,H]=0$7 is the ideal QAOA state, then $[\hat S,H]=0$8, and measuring $[\hat S,H]=0$9 while retaining only runs with outcome $\mathcal H=\mathcal H_+\oplus\mathcal H_-$0 implements

$\mathcal H=\mathcal H_+\oplus\mathcal H_-$1

Under local noise after each QAOA layer, with $\mathcal H=\mathcal H_+\oplus\mathcal H_-$2, the ideal fidelity is $\mathcal H=\mathcal H_+\oplus\mathcal H_-$3, while after SV

$\mathcal H=\mathcal H_+\oplus\mathcal H_-$4

Thus the fidelity improvement ratio is

$\mathcal H=\mathcal H_+\oplus\mathcal H_-$5

This makes explicit that SV improves fidelity by conditioning on the weight remaining in the target parity sector (Kakkar et al., 2022).

3. Noise models, acceptance probability, and finite-shot performance

Under depolarizing single-qubit noise,

$\mathcal H=\mathcal H_+\oplus\mathcal H_-$6

the QAOA analysis derives for depth $\mathcal H=\mathcal H_+\oplus\mathcal H_-$7

$\mathcal H=\mathcal H_+\oplus\mathcal H_-$8

with

$\mathcal H=\mathcal H_+\oplus\mathcal H_-$9

Writing

$$P_+ = \frac{I+\hat S}{2}, \qquad P_-=\frac{I-\hat S}{2}.$$0

one obtains

$$P_+ = \frac{I+\hat S}{2}, \qquad P_-=\frac{I-\hat S}{2}.$$1

and in practice replaces $$P_+ = \frac{I+\hat S}{2}, \qquad P_-=\frac{I-\hat S}{2}.$$2 when errors accumulate over $$P_+ = \frac{I+\hat S}{2}, \qquad P_-=\frac{I-\hat S}{2}.$$3 layers. Under dephasing,

$$P_+ = \frac{I+\hat S}{2}, \qquad P_-=\frac{I-\hat S}{2}.$$4

only even-weight errors commute with $$P_+ = \frac{I+\hat S}{2}, \qquad P_-=\frac{I-\hat S}{2}.$$5, giving the closed form

$$P_+ = \frac{I+\hat S}{2}, \qquad P_-=\frac{I-\hat S}{2}.$$6

These formulas identify a basic trade-off: fidelity improves only by discarding runs, so acceptance probability controls sampling overhead (Kakkar et al., 2022).

A finite-shot theory makes this trade-off explicit. Let $$P_+ = \frac{I+\hat S}{2}, \qquad P_-=\frac{I-\hat S}{2}.$$7 be the projector onto the symmetry sector, let $$P_+ = \frac{I+\hat S}{2}, \qquad P_-=\frac{I-\hat S}{2}.$$8 be a symmetry-compatible observable, let $$P_+ = \frac{I+\hat S}{2}, \qquad P_-=\frac{I-\hat S}{2}.$$9 be the acceptance probability, let $\rho$0 be the post-selection flag, let $\rho$1 be the measurement result of $\rho$2 when $\rho$3, and set $\rho$4. The clipped estimator is

$\rho$5

For every finite $\rho$6,

$\rho$7

with remainder bounded by

$\rho$8

Because Bernoulli postselection yields

$\rho$9

there is no ratio-bias cross term $$\rho \mapsto \rho_{\mathrm{SV}}=\frac{P_+\rho P_+}{\mathrm{Tr}[P_+\rho]},$$0 (Alfaro, 13 Jun 2026).

The same analysis isolates two irreducible effects. First, if $$\rho \mapsto \rho_{\mathrm{SV}}=\frac{P_+\rho P_+}{\mathrm{Tr}[P_+\rho]},$$1 and $$\rho \mapsto \rho_{\mathrm{SV}}=\frac{P_+\rho P_+}{\mathrm{Tr}[P_+\rho]},$$2, then

$$\rho \mapsto \rho_{\mathrm{SV}}=\frac{P_+\rho P_+}{\mathrm{Tr}[P_+\rho]},$$3

with

$$\rho \mapsto \rho_{\mathrm{SV}}=\frac{P_+\rho P_+}{\mathrm{Tr}[P_+\rho]},$$4

so the bias floor scales as

$$\rho \mapsto \rho_{\mathrm{SV}}=\frac{P_+\rho P_+}{\mathrm{Tr}[P_+\rho]},$$5

Second, postselection inflates variance by $$\rho \mapsto \rho_{\mathrm{SV}}=\frac{P_+\rho P_+}{\mathrm{Tr}[P_+\rho]},$$6: $$\rho \mapsto \rho_{\mathrm{SV}}=\frac{P_+\rho P_+}{\mathrm{Tr}[P_+\rho]},$$7 When comparing SV against an unmitigated estimator, the critical sample budget scales as $$\rho \mapsto \rho_{\mathrm{SV}}=\frac{P_+\rho P_+}{\mathrm{Tr}[P_+\rho]},$$8 or $$\rho \mapsto \rho_{\mathrm{SV}}=\frac{P_+\rho P_+}{\mathrm{Tr}[P_+\rho]},$$9, depending on whether the leading variance terms cancel. This result replaces an asymptotic “SV always helps” narrative by an operating-window picture in which benefit depends jointly on noise rate, acceptance probability, and finite budget (Alfaro, 13 Jun 2026).

4. From state symmetries to circuit and channel symmetries

State-based SV fails when the output state does not share a fixed stabilizer independent of the input, even though the circuit itself obeys useful commutation relations. Circuit-oriented SV addresses this by using spatio-temporal stabilizers (STS). For a circuit $G$0, an STS specifies partial operators $G$1 at times $G$2 so that

$G$3

or equivalently

$G$4

Hence,

$G$5

An ancilla-assisted check inserts controlled-$G$6 operations at the designated times and measures the ancilla in the $G$7-basis; a $G$8 outcome signals a symmetry violation. If a noisy Kraus component anti-commutes with one of the $G$9, it is filtered by the check. In a circuit with $\{U_g\}$0 faulty gates, any residual error must involve an even number of anti-commuting faults, so the error probability scales $\{U_g\}$1 rather than $\{U_g\}$2. The sampling overhead factor is

$\{U_g\}$3

A related commutativity test uses the quantum switch, whose controlled superposition of orders $\{U_g\}$4 and $\{U_g\}$5 filters anti-commutator-violating noise without full tomography (Xiong et al., 2021).

Channel-level generalization is provided by symmetric channel verification (SCV). A channel $\{U_g\}$6 is $\{U_g\}$7-symmetric if

$\{U_g\}$8

If the ideal channel commutes with an operator $\{U_g\}$9 having $\hat S\in\mathcal P^N$00 distinct eigenvalues and projectors $\hat S\in\mathcal P^N$01, define

$\hat S\in\mathcal P^N$02

A quantum-phase-estimation-like circuit applies controlled powers of $\hat S\in\mathcal P^N$03, the noisy channel $\hat S\in\mathcal P^N$04, controlled inverse powers of $\hat S\in\mathcal P^N$05, an inverse QFT on the ancilla, and post-selection on ancilla outcome $\hat S\in\mathcal P^N$06. The resulting trace-nonincreasing map is

$\hat S\in\mathcal P^N$07

If the noise satisfies $\hat S\in\mathcal P^N$08, then $\hat S\in\mathcal P^N$09; in particular, for $\hat S\in\mathcal P^N$10 with $\hat S\in\mathcal P^N$11, SCV yields pure $\hat S\in\mathcal P^N$12 with success rate $\hat S\in\mathcal P^N$13 (Tsubouchi et al., 17 Mar 2025).

A hardware-efficient variant, virtual SCV (VSCV), reconstructs the same purification only at the level of expectation-value estimation. It uses one ancilla and controlled-Pauli Clifford gadgets to realize virtual supermaps

$\hat S\in\mathcal P^N$14

from which $\hat S\in\mathcal P^N$15 is reconstructed by Pauli expansion. A depolarizing error $\hat S\in\mathcal P^N$16 on the ancilla introduces only a constant factor $\hat S\in\mathcal P^N$17, which cancels in normalization. In the Clifford-only regime, SCV under Pauli symmetry is stated to be the optimal purification method (Tsubouchi et al., 17 Mar 2025).

5. Experimental realizations and observed regimes

On a two-qubit circuit QED processor for $\hat S\in\mathcal P^N$18, SV was implemented with an exchange-interaction ansatz gate

$\hat S\in\mathcal P^N$19

which preserves total excitation number and hence parity $\hat S\in\mathcal P^N$20. Full tomography used 36 tensor-product single-qubit pre-rotations, with $\hat S\in\mathcal P^N$21 shots per generation during optimization and $\hat S\in\mathcal P^N$22 shots for final tomography. Over the $\hat S\in\mathcal P^N$23 dissociation curve, the reported averages were

$\hat S\in\mathcal P^N$24

with raw VQE energy errors typically $\hat S\in\mathcal P^N$25 Hartree and symmetry-verified errors $\hat S\in\mathcal P^N$26 Hartree (Sagastizabal et al., 2019).

For QAOA on an 11-qubit IonQ trapped-ion device with all-to-all connectivity, experiments covered all non-isomorphic graphs of size $\hat S\in\mathcal P^N$27 at depths $\hat S\in\mathcal P^N$28, using a global parity-check circuit. Across 3- and 4-node graphs and $\hat S\in\mathcal P^N$29, SV changed the approximation ratio from $\hat S\in\mathcal P^N$30 up to $\hat S\in\mathcal P^N$31, with mean $\hat S\in\mathcal P^N$32. The post-selection success probability $\hat S\in\mathcal P^N$33 was $\hat S\in\mathcal P^N$34, and the reported two-qubit error rates of $\hat S\in\mathcal P^N$35 matched the simulation-predicted regime in which SV improves the QAOA objective (Kakkar et al., 2022).

Circuit-level STS checks have been demonstrated on QFT and QAOA instances. For QFT under asymmetric Pauli channels with $\hat S\in\mathcal P^N$36 error-rate ratio $\hat S\in\mathcal P^N$37, purity increased by $\hat S\in\mathcal P^N$38 at $\hat S\in\mathcal P^N$39; under depolarizing noise, net benefit appeared when two-qubit gate error was $\hat S\in\mathcal P^N$40. For single-stage QAOA, single-ancilla STS produced a purity boost of $\hat S\in\mathcal P^N$41 at $\hat S\in\mathcal P^N$42, while a cat-state two-ancilla STS added $\hat S\in\mathcal P^N$43 purity. On IBMQ_Lima, a four-gate XX-rotation circuit protected by a reduced STS on one qubit showed a total error probability drop from $\hat S\in\mathcal P^N$44 to $\hat S\in\mathcal P^N$45 (Xiong et al., 2021).

SCV and VSCV have likewise been evaluated on Hamiltonian simulation and phase estimation workloads. For an $\hat S\in\mathcal P^N$46 Heisenberg chain under local depolarizing noise, SCV on the entire circuit reduced the trace distance $\hat S\in\mathcal P^N$47 from $\hat S\in\mathcal P^N$48 to $\hat S\in\mathcal P^N$49. In a SELECT operation for 2D Fermi-Hubbard qubitization, applying virtual SCV on all idling system qubits reduced the system-noise contribution to $\hat S\in\mathcal P^N$50 per qubit, so the total noise became $\hat S\in\mathcal P^N$51, yielding a quadratic improvement in $\hat S\in\mathcal P^N$52 (Tsubouchi et al., 17 Mar 2025).

Finite-shot comparisons against virtual distillation further sharpen the empirical picture. Gate-level simulation and archived runs on ibm_marrakesh and ibm_kingston reported that, at $\hat S\in\mathcal P^N$53, SV had $\hat S\in\mathcal P^N$54, significantly below the unmitigated $\hat S\in\mathcal P^N$55, while virtual distillation remained above SV and could suffer denominator instability. In the tested QAOA instances, calibrated SV was therefore the practical winner, but the same study explicitly rejects a universal-winner interpretation and instead emphasizes regime structure (Alfaro, 13 Jun 2026).

6. Distinct usages and related symmetry-based verification paradigms

A separate meaning of the phrase appears in two-particle quantum physics. The 2011 proposal for symmetrization verification considers two identical particles in a state

$\hat S\in\mathcal P^N$56

with the upper sign for bosons and the lower sign for fermions. Detection of one particle at position $\hat S\in\mathcal P^N$57 is modeled by

$\hat S\in\mathcal P^N$58

and the surviving one-particle state is

$\hat S\in\mathcal P^N$59

Equivalently,

$\hat S\in\mathcal P^N$60

The proposal does not follow a Hong-Ou-Mandel-type approach; instead it uses one-particle destructive detection plus state tomography of the survivor, and it can also generate single-particle superposition states (Sancho, 2011).

Outside quantum information, symmetry-based verification has additional meanings. In combinatorial optimisation, “Certified Symmetry and Dominance Breaking for Combinatorial Optimisation” realizes symmetry verification through machine-checkable cutting-planes certificates, lex-leader pseudo-Boolean constraints, and dominance-based inference rules, enabling certification of symmetry breaking in SAT, maximum clique, and constraint programming (Bogaerts et al., 2022). In parameterized distributed protocols, SymIC3 and IC3PO use symmetry groups, clause orbits, and quantifier inference from orbit structure to derive quantified inductive invariants without a priori quantifier-template search (Goel et al., 2021). In symbolic model checking, dynamic symmetry reduction avoids constructing the orbit relation BDD and instead interleaves reachability with on-the-fly representative selection and state symmetries (Appold, 2010). In Coq proofs for triangulation algorithms, explicit rotation, reflection, and orbit lemmas collapse “without-loss-of-generality” case splits into canonical representatives (Bertot, 2018). In hybrid-systems scenario verification, SceneChecker uses symmetry abstractions and refinement to build quotient automata over geometrically similar plan segments and reports substantial speedups while using existing reachability tools as black-box subroutines (Sibai et al., 2020).

These literatures share a common structural idea—explicit exploitation of symmetry to reduce ambiguity, redundancy, or noise—but they do not instantiate a single formalism. In quantum error mitigation, SV is fundamentally a projection-and-postselection procedure; in symmetrization verification it is a measurement-induced test of exchange symmetry; in formal verification and optimisation it is a certification or reduction methodology. A plausible implication is that “symmetry verification” is best treated as a family resemblance term whose precise meaning is fixed by the ambient theory of states, channels, clauses, or transition systems.