Quantum Enhanced Pauli Propagation (QuEPP)

• Quantum Enhanced Pauli Propagation (QuEPP) is a collection of Pauli-basis techniques used to simulate quantum circuits, estimate observables, and design error-mitigation workflows.
• It leverages methods such as Monte Carlo propagation, symmetry merging, and inverse noise mapping to control computational complexity and manage noise.
• Recent formulations incorporate quantum memory and PTM learning to enable efficient hybrid simulation and error-cancellation in large-scale quantum computations.

Searching arXiv for the cited QuEPP and Pauli propagation papers to ground the article in current literature. Quantum Enhanced Pauli Propagation (QuEPP) denotes a family of Pauli-basis methods for simulating quantum circuits and dynamics, estimating observables, learning channel representations, and constructing hybrid error-mitigation workflows. Its common substrate is the propagation or backpropagation of Pauli operators through gates or channels, usually in the Heisenberg picture, so that expectation values can be expressed in terms of Pauli coefficients, Pauli paths, or Pauli transfer matrices. The label is not standardized across the literature: foundational works introduce Pauli propagation without using the term, while later papers use “QuEPP” for several related, but distinct, enhancements built on the same Pauli-operator backbone (Rall et al., 2019, Caro, 2022, Martinez et al., 22 Jan 2025, Majumder et al., 15 Mar 2026, Eddins et al., 18 Jun 2026).

1. Terminology and historical scope

A recurrent source of confusion is the assumption that QuEPP names a single canonical algorithm. Current usage is broader. The original paper “Simulation of Qubit Quantum Circuits via Pauli Propagation” introduces Schrödinger propagation and Heisenberg propagation and explicitly notes that it does not use the term “Quantum Enhanced Pauli Propagation” (Rall et al., 2019). The symmetry-merging work likewise states that it does not define “QuEPP” explicitly, even though its symmetry-adapted framework naturally fits that label in later syntheses (Teng et al., 12 Dec 2025).

Usage	Representative paper	Core mechanism
Pauli propagation techniques	(Rall et al., 2019)	Monte Carlo propagation in the Pauli basis
PTM-learning with quantum memory	(Caro, 2022)	Learn a channel via Choi-state access
Non-unital-noise QuEPP	(Martinez et al., 22 Jan 2025)	Pauli backpropagation with LOWESA-AD / MC-LOWESA-AD
Hybrid CPT-based QuEPP	(Majumder et al., 15 Mar 2026)	Clifford ensemble plus global rescaling
Noise-canceling-observable QuEPP	(Eddins et al., 18 Jun 2026)	Propagate observables through inverse noise maps

The broader lineage of the subject predates the recent QuEPP terminology. In fault-tolerant computation, “Pauli propagation” or “Pauli tracking” refers to the classical tracking of byproduct operators in a Pauli frame, with correction statuses $s_k \in \{I,X,Z,XZ\}$ updated through Clifford+T teleportation gadgets in $O(m)$ time for $m$ gates (Paler et al., 2014). In qudit settings, the “error probability tensor” extends generalized Pauli propagation to $D$-dimensional Clifford circuits and stabilizer codes, propagating generalized Pauli error statistics by permutations and convolutions on the index space $\mathbb{Z}_D^{2n}$ (Miller et al., 2018). These are not identical to contemporary QuEPP formulations, but they establish the same central motif: Pauli-structured evolution as the primitive computational object.

2. Pauli-basis formalism

For $n$ qubits, the Pauli basis $\mathcal{P}_n = \{I,X,Y,Z\}^{\otimes n}$ gives the decomposition

$$A = \frac{1}{2^n}\sum_{\sigma \in \mathcal{P}_n} \mathrm{Tr}(\sigma A)\,\sigma,$$

so that expectation values reduce to contractions of Pauli coefficients (Rall et al., 2019). In Heisenberg form, an operator evolves as

$P(t)=U^\dagger P(0)U,\qquad \frac{dP}{dt}= i[H,P],$

and layered circuits induce repeated Pauli-basis updates through the Pauli transfer matrix (PTM) (Teng et al., 12 Dec 2025). The PTM of a channel $\mathcal{E}$ is

$O(m)$0

or equivalently $O(m)$1, and expectation values become bilinear forms in PTM entries, state coefficients, and observable coefficients (Caro, 2022).

Pauli propagation is attractive because Clifford gates are non-branching in this basis: each Pauli maps to another Pauli up to sign. Non-Clifford Pauli rotations branch. In the symbolic formulation, if $O(m)$2 acts on a Pauli $O(m)$3, then

$O(m)$4

and in the anticommuting case this becomes a two-term branch (Monaco et al., 18 Dec 2025). The layered expansion can therefore be viewed as a Pauli-path sum. In the symmetry-merging paper, a single string can branch to as many as $O(m)$5 strings across a layer with $O(m)$6 Pauli gates, and worst-case support can reach $O(m)$7 strings (Teng et al., 12 Dec 2025).

The 2019 Monte Carlo formulation associates this propagation with a stabilizer norm

$O(m)$8

Schrödinger propagation yields sample complexity

$O(m)$9

whereas Heisenberg propagation yields

$m$0

For Clifford channels, $m$1, so each sample is produced in linear time with no variance blow-up from the Clifford layers (Rall et al., 2019). This state-independence of Heisenberg variance is one of the main reasons Pauli backpropagation became central in later QuEPP formulations.

3. Compression, symmetry, and operator-structure exploitation

A major branch of QuEPP research is devoted to reducing the combinatorial growth of Pauli paths without sacrificing exactness under relevant structure. The clearest recent example is symmetry-merging Pauli propagation. If a circuit or Hamiltonian has a finite symmetry group $m$2 with layerwise covariance $m$3, Pauli strings can be partitioned into symmetry orbits $m$4, and only a representative $m$5 from each orbit need be propagated (Teng et al., 12 Dec 2025). The merged evolution is exact whenever both the dynamics and initial state commute with the symmetry action, so that $m$6. Burnside’s lemma yields the representative count

$m$7

which becomes the number of $m$8-ary necklaces for translation symmetry and $m$9 for full permutation symmetry. Consequently, translation gives asymptotic $D$0 savings, while full permutation symmetry gives $D$1 effective space ratio and therefore exponential reduction (Teng et al., 12 Dec 2025). The same paper reports that, in $D$2-qubit all-to-all XXZ dynamics, symmetry merging remained stable at truncation thresholds where standard Pauli propagation diverged earlier, and that symmetry PP with $D$3 matched standard PP at $D$4 in the reported runs (Teng et al., 12 Dec 2025).

A second compression line treats Pauli propagation symbolically. “Symbolic Pauli Propagation for Gradient-Enabled Pre-Training of Quantum Circuits” represents propagated observables as explicit trigonometric polynomials in the circuit parameters, then truncates by Pauli weight $D$5 and frequency $D$6 (Monaco et al., 18 Dec 2025). Under a locality-induced coefficient decay assumption $D$7, the joint truncation error obeys

$D$8

with a corresponding gradient bound (Monaco et al., 18 Dec 2025). In the reported $D$9-qubit ANNNI VQE experiment, propagating only three observables once and reusing their symbolic forms was sufficient to map the $\mathbb{Z}_D^{2n}$0 phase diagram accurately with $\mathbb{Z}_D^{2n}$1 and $\mathbb{Z}_D^{2n}$2 (Monaco et al., 18 Dec 2025).

A third line emphasizes operator complexity rather than symmetry. “Pauli Propagation: Simulating Quantum Spin Dynamics via Operator Complexity” introduces the Operator Stabilizer Rényi entropy

$\mathbb{Z}_D^{2n}$3

for the Pauli-coefficient vector $\mathbb{Z}_D^{2n}$4, and proves a Top-$\mathbb{Z}_D^{2n}$5 truncation bound

$\mathbb{Z}_D^{2n}$6

with $\mathbb{Z}_D^{2n}$7 prescribed by $\mathbb{Z}_D^{2n}$8 and a target $\mathbb{Z}_D^{2n}$9 (Shao et al., 25 Oct 2025). For the $n$0 1D Heisenberg model, the paper proves that a local observable has only $n$1 nonzero coefficients after $n$2 Trotter steps, establishing an explicit compressibility theorem in the XX limit (Shao et al., 25 Oct 2025). This suggests that a substantial part of QuEPP can be read as the search for structural regimes—symmetry, locality, low frequency, low operator entropy—in which Pauli growth remains computationally sparse.

4. Noisy dynamics and controlled truncation

A second major component of the QuEPP literature addresses realistic noise. The 2025 non-unital-noise paper extends Pauli backpropagation from depolarizing noise to amplitude damping and more general single-qubit non-unital channels in normal form (Martinez et al., 22 Jan 2025). For amplitude damping with parameter $n$3,

$n$4

so the identity component must be tracked explicitly. The paper decomposes noisy $n$5 layers into a small set of processes and introduces two truncation algorithms: deterministic LOWESA-AD and Monte Carlo MC-LOWESA-AD. Their runtime is

$n$6

respectively, with $n$7 the split cutoff and $n$8 the number of sampled trees. The error guarantees are

$n$9

for LOWESA-AD, where $\mathcal{P}_n = \{I,X,Y,Z\}^{\otimes n}$0 is a lower bound on the minimum number of damping splits in discarded branches, and

$\mathcal{P}_n = \{I,X,Y,Z\}^{\otimes n}$1

for MC-LOWESA-AD with probability at least $\mathcal{P}_n = \{I,X,Y,Z\}^{\otimes n}$2 (Martinez et al., 22 Jan 2025). The same paper also proves an expected-damping improvement

$\mathcal{P}_n = \{I,X,Y,Z\}^{\otimes n}$3

when each noisy rotation is preceded by a uniformly random single-qubit Clifford (Martinez et al., 22 Jan 2025).

The companion paper on arbitrary local incoherent noise shifts the viewpoint from worst-case circuits to average-case ensembles (Angrisani et al., 22 Jan 2025). Under layer distributions invariant under independent single-qubit random gates, and for local noise with normal-form contraction $\mathcal{P}_n = \{I,X,Y,Z\}^{\otimes n}$4, high-weight Pauli contributions decay exponentially on average, enabling path-weight truncation and an effective depth

$\mathcal{P}_n = \{I,X,Y,Z\}^{\otimes n}$5

The paper proves polynomial-time classical estimation of expectation values with inverse-polynomial error for arbitrary local noise, including non-unital channels, and numerically validates the approach on a $\mathcal{P}_n = \{I,X,Y,Z\}^{\otimes n}$6 lattice under amplitude damping and dephasing, as well as on an $\mathcal{P}_n = \{I,X,Y,Z\}^{\otimes n}$7 lattice with amplitude damping (Angrisani et al., 22 Jan 2025). The contrast between this average-case result and the worst-case impossibility constructions under non-unital noise is one of the clearest examples of QuEPP’s dependence on ensemble assumptions rather than merely on gate syntax.

5. Hybrid quantum-enhanced formulations

The most literal uses of the name QuEPP arise when classical Pauli propagation is explicitly augmented by quantum resources. One such formulation is PTM learning with quantum memory. “Learning Quantum Processes and Hamiltonians via the Pauli Transfer Matrix” shows that copies of a channel’s normalized Choi state $\mathcal{P}_n = \{I,X,Y,Z\}^{\otimes n}$8 allow the PTM entries

$\mathcal{P}_n = \{I,X,Y,Z\}^{\otimes n}$9

to be learned with

$$A = \frac{1}{2^n}\sum_{\sigma \in \mathcal{P}_n} \mathrm{Tr}(\sigma A)\,\sigma,$$0

copies, while any learner without quantum memory needs $$A = \frac{1}{2^n}\sum_{\sigma \in \mathcal{P}_n} \mathrm{Tr}(\sigma A)\,\sigma,$$1 channel queries even under strong promises (Caro, 2022). The same framework supports efficient prediction of $$A = \frac{1}{2^n}\sum_{\sigma \in \mathcal{P}_n} \mathrm{Tr}(\sigma A)\,\sigma,$$2 when the state, observable, or PTM is Pauli-sparse, and it extends to Hamiltonian learning via short-time dynamics and polynomial interpolation (Caro, 2022). Here the “quantum enhancement” is neither noise mitigation nor truncation, but coherent access to channel information that collapses an otherwise exponential query barrier.

A second formulation is the paper explicitly titled “Quantum Enhanced Pauli Propagation” (Majumder et al., 15 Mar 2026). It begins from Clifford perturbation theory (CPT), in which a target circuit is expanded into a weighted ensemble of classically simulable Clifford circuits,

$$A = \frac{1}{2^n}\sum_{\sigma \in \mathcal{P}_n} \mathrm{Tr}(\sigma A)\,\sigma,$$3

The low-order CPT terms provide a truncated classical approximation, while noisy quantum expectation values of the same Clifford ensemble are used to infer a global rescaling factor $$A = \frac{1}{2^n}\sum_{\sigma \in \mathcal{P}_n} \mathrm{Tr}(\sigma A)\,\sigma,$$4. The QuEPP estimator is

$$A = \frac{1}{2^n}\sum_{\sigma \in \mathcal{P}_n} \mathrm{Tr}(\sigma A)\,\sigma,$$5

where $$A = \frac{1}{2^n}\sum_{\sigma \in \mathcal{P}_n} \mathrm{Tr}(\sigma A)\,\sigma,$$6 is the portion of the noisy target expectation not accounted for by the low-order simulated ensemble (Majumder et al., 15 Mar 2026). The paper emphasizes that no noise characterization is required, proves an asymptotically unbiased route as the ensemble is enlarged, and reports experiments on IBM Heron hardware for random mirror circuits up to $$A = \frac{1}{2^n}\sum_{\sigma \in \mathcal{P}_n} \mathrm{Tr}(\sigma A)\,\sigma,$$7 qubits and depth $$A = \frac{1}{2^n}\sum_{\sigma \in \mathcal{P}_n} \mathrm{Tr}(\sigma A)\,\sigma,$$8, as well as Trotterized Hamiltonian evolution (Majumder et al., 15 Mar 2026).

A third formulation uses inverse noise maps to construct modified observables. In “Computing noise-canceling observables via Pauli propagation,” the target observable is propagated through inverse channels so that

$$A = \frac{1}{2^n}\sum_{\sigma \in \mathcal{P}_n} \mathrm{Tr}(\sigma A)\,\sigma,$$9

The paper implements this idea through Propagated Noise Absorption (PNA) and Euclid, both of which truncate the Pauli expansion of $P(t)=U^\dagger P(0)U,\qquad \frac{dP}{dt}= i[H,P],$0 rather than inserting physical inverse channels into the hardware circuit (Eddins et al., 18 Jun 2026). Numerical benchmarks cover $P(t)=U^\dagger P(0)U,\qquad \frac{dP}{dt}= i[H,P],$1- and $P(t)=U^\dagger P(0)U,\qquad \frac{dP}{dt}= i[H,P],$2-qubit mirror TFI circuits, and experiments are reported on a $P(t)=U^\dagger P(0)U,\qquad \frac{dP}{dt}= i[H,P],$3-qubit superconducting processor. For the non-Clifford regime, the paper states that PNA’s squared sampling cost tracks roughly the square root of optimized shaded-lightcone PEC cost, indicating a quadratic sampling speedup, whereas at Clifford points that separation vanishes under commuting-lightcone pruning (Eddins et al., 18 Jun 2026).

Measurement and hardware layers have also been brought under the QuEPP umbrella. “Enhanced measurements on quantum computers via the simultaneous probing of non-commuting Pauli operators” defines QuEPP as a two-copy “double” measurement scheme: because $P(t)=U^\dagger P(0)U,\qquad \frac{dP}{dt}= i[H,P],$4 for all Pauli strings, one Bell-basis measurement per qubit pair provides Bernoulli samples for all $P(t)=U^\dagger P(0)U,\qquad \frac{dP}{dt}= i[H,P],$5 simultaneously, giving access to $P(t)=U^\dagger P(0)U,\qquad \frac{dP}{dt}= i[H,P],$6 for every Pauli term in an observable (Simon et al., 1 Sep 2025). Bayesian fusion with occasional single-copy sign-resolving measurements and adaptive shot allocation then minimizes the estimator variance. Separately, “Quantum simulation of Pauli channels and dynamical maps” shows how ancilla-assisted circuits can implement arbitrary Pauli channels and certain one-parameter Pauli dynamical maps on hardware, with a theorem characterizing when a full $P(t)=U^\dagger P(0)U,\qquad \frac{dP}{dt}= i[H,P],$7-qubit family can be realized with only one parameter-dependent single-qubit rotation (Basile et al., 2023). These papers broaden QuEPP from propagation-as-simulation to propagation-as-measurement design and propagation-as-noise synthesis.

6. Applications, limitations, and research directions

The application domain of QuEPP is correspondingly broad. PauliPropagation.jl systematizes bit-level Pauli-string encodings, tree-search propagation, PTM-defined gates, and overlap interfaces, and positions Pauli propagation as a general classical framework for simulating digital quantum systems (Rudolph et al., 27 May 2025). Symmetry-aware versions target all-to-all Heisenberg and XXZ dynamics (Teng et al., 12 Dec 2025). Symbolic variants target classical pre-training and gradient evaluation for VQE-like workloads (Monaco et al., 18 Dec 2025). Operator-complexity variants target nonequilibrium spin dynamics in regimes where tensor-network state entanglement is the wrong bottleneck (Shao et al., 25 Oct 2025). Error-mitigation variants target Clifford subcircuits and measurement-based workflows: propagated probabilistic error cancellation fuses inverse Pauli channels on Clifford segments and proves

$P(t)=U^\dagger P(0)U,\qquad \frac{dP}{dt}= i[H,P],$8

with additional savings from $P(t)=U^\dagger P(0)U,\qquad \frac{dP}{dt}= i[H,P],$9 and $\mathcal{E}$0 boundary reductions (Scheiber et al., 2024). Compilation variants use Pauli error propagation paths to reschedule commuting gates according to the weighted estimated success probability (WESP) on hardware with spatially varying gate errors (Saravanan et al., 2022). In a different register, qudit generalizations based on the error probability tensor enable exact analytical propagation through generalized Pauli channels, stabilizer reductions, and repeater architectures (Miller et al., 2018).

The limits are equally varied. The Monte Carlo Pauli-propagation formalism remains exponentially sensitive to large $\mathcal{E}$1, adaptivity, and highly local observables in Schrödinger mode (Rall et al., 2019). Symmetry merging loses force when orbits are small or symmetry is broken by disorder, boundary conditions, or non-covariant noise (Teng et al., 12 Dec 2025). Non-unital-noise truncation guarantees are presently strongest for single-qubit normal-form channels and alternating Clifford-plus-rotation architectures rather than arbitrary multi-qubit non-unital noise (Martinez et al., 22 Jan 2025). Average-case polynomial simulability under arbitrary local noise depends on layer invariance under single-qubit designs and does not cover adversarial circuits that violate those assumptions (Angrisani et al., 22 Jan 2025). Hybrid inverse-channel observable methods require invertible or regularizable noise models and are sensitive to calibration mismatch, crosstalk, and PTM ill-conditioning (Eddins et al., 18 Jun 2026). PTM-learning variants assume quantum memory and Choi-state access; two-copy measurement variants assume the ability to prepare and jointly measure identical copies (Caro, 2022, Simon et al., 1 Sep 2025).

A plausible synthesis is that QuEPP now names not one algorithm but a design pattern: represent dynamics, noise, or measurements in the Pauli basis; exploit algebraic structure to keep that representation tractable; and, when classical propagation alone is insufficient, use quantum resources to learn PTMs, probe non-commuting Pauli data, correct truncation bias, or absorb inverse noise into observables. Under that reading, QuEPP is best understood as the convergence of Pauli propagation, PTM methods, structure-aware compression, and hybrid quantum-classical inference into a single research program rather than a single fixed protocol (Majumder et al., 15 Mar 2026, Eddins et al., 18 Jun 2026).