Quantum ARMA Models: SchWARMA Noise Simulation

• Quantum ARMA models extend classical ARMA processes by mapping scalar correlations to sequences of CPTP maps, enabling simulation of correlated quantum noise.
• The SchWARMA framework employs a geometric formulation on the complex Stiefel manifold to enforce physical channel constraints while integrating ARMA-driven dynamics.
• Efficient simulation and parameter estimation are achieved by using ARMA-driven tangent-space perturbations, reducing computational cost compared to fine-grained noise integration.

Searching arXiv for the cited papers and adjacent work on quantum ARMA / autoregressive quantum models. {"query":"all:SchWARMA model-based approach time-correlated noise quantum circuits (Schultz et al., 2020) OR all:\"quantum autoregressive\" ARMA", "max_results": 10} {"query":"id:(Schultz et al., 2020) OR id:(Luo et al., 2020) OR id:(Bortone et al., 2023) OR id:(Mohanty et al., 9 Apr 2026)", "max_results": 10} Quantum autoregressive moving average models are constructions that transfer the logic of classical autoregressive moving average (ARMA) processes from scalar time series to quantum objects. In the most explicit formulation in the present literature, the target objects are sequences of completely positive trace-preserving (CPTP) maps acting on density matrices, with temporal correlations inherited from classical ARMA processes and physical validity enforced by a manifold representation of quantum channels. This formulation, denoted Schrödinger Wave ARMA (SchWARMA), is designed for parameterizing and simulating time-correlated noise in quantum circuits and for importing techniques from signal processing, control theory, and system identification into quantum noise modeling (Schultz et al., 2020). Related literature uses autoregressive ideas in broader quantum settings—such as autoregressive probabilistic representations of open quantum systems or autoregressive quantum states with filter-based correlations—but these works either do not explicitly define quantum ARMA or use the terminology in a quantum-inspired sense rather than as a channel-valued ARMA model (Luo et al., 2020, Bortone et al., 2023, Mohanty et al., 9 Apr 2026).

1. Classical ARMA origin and the quantum generalization

In classical time series analysis, a scalar discrete-time process $\{y_k\}$ is modeled by an ARMA$(p,q)$ process,

$$y_k=\sum_{i=1}^p a_i\,y_{k-i}+\sum_{j=0}^q b_j\,x_{k-j},$$

where $x_k$ are i.i.d. innovations, $\{a_i\}$ are autoregressive coefficients, and $\{b_j\}$ are moving-average coefficients. For i.i.d. Gaussian driving noise, the power spectral density is

$$S_y(\omega)=\frac{\left|\sum_{k=0}^q b_k e^{-ik\omega}\right|^2}{\left|1+\sum_{k=1}^p a_k e^{-ik\omega}\right|^2}.$$

A central motivation for the quantum extension is that ARMA models are dense in the set of discrete-time spectra, so any discrete-time PSD can be approximated to arbitrary accuracy by a suitable ARMA model (Schultz et al., 2020).

SchWARMA transfers this idea from real-valued outputs to sequences of quantum channels. Instead of scalar outputs $y_k$, one constructs temporally correlated error channels $\mathcal{S}_k$ that are CPTP, physically interpretable as noisy gates, and driven by a parametric classical ARMA model. The conceptual bridge has three parts: choose a smooth manifold representation of the space of CPTP maps; drive tangent-space elements by classical ARMA processes $y_k^{(\ell)}$; and map those tangent vectors back to the CPTP manifold via a matrix exponential on a Stiefel manifold. In this construction, the ARMA outputs remain classical scalars, but they weight fixed tangent directions in the space of CPTP maps, and the exponentiated result is a valid quantum channel at each discrete time step (Schultz et al., 2020).

A recurrent misconception is that “quantum ARMA” must mean an intrinsically quantum analogue of scalar recurrence relations acting directly on amplitudes or density matrices. In the SchWARMA framework, the ARMA structure is classical at the level of innovations and state variables, while the quantum part enters through the manifold-valued map that converts those correlated scalar processes into physically valid CPTP operations. This preserves the classical interpretability of AR and MA coefficients while enforcing quantum channel constraints exactly (Schultz et al., 2020).

2. Geometric formulation on the manifold of CPTP maps

The SchWARMA construction identifies a quantum channel $(p,q)$0 on an $(p,q)$1-dimensional system with a point on a complex Stiefel manifold. A CPTP map in Kraus form,

$(p,q)$2

is encoded by stacking Kraus operators into

$(p,q)$3

The trace-preserving constraint becomes $(p,q)$4, so $(p,q)$5 has orthonormal columns and therefore lies on the complex Stiefel manifold

$(p,q)$6

For parameterizing noisy maps, one fixes a Kraus representation, even though different Kraus representations of the same channel correspond to equivalent points under right-unitary transformations of the Kraus set (Schultz et al., 2020).

The tangent space at a point $(p,q)$7 has the form

$(p,q)$8

where $(p,q)$9 is skew-Hermitian, $$y_k=\sum_{i=1}^p a_i\,y_{k-i}+\sum_{j=0}^q b_j\,x_{k-j},$$0 is arbitrary, and $$y_k=\sum_{i=1}^p a_i\,y_{k-i}+\sum_{j=0}^q b_j\,x_{k-j},$$1 is unitary. The decomposition is physically suggestive: $$y_k=\sum_{i=1}^p a_i\,y_{k-i}+\sum_{j=0}^q b_j\,x_{k-j},$$2 generates unitary rotations in the Kraus set, while $$y_k=\sum_{i=1}^p a_i\,y_{k-i}+\sum_{j=0}^q b_j\,x_{k-j},$$3 generates non-unitary or dissipative directions. The exponential map

$$y_k=\sum_{i=1}^p a_i\,y_{k-i}+\sum_{j=0}^q b_j\,x_{k-j},$$4

takes a tangent vector, exponentiates a block skew-Hermitian matrix in a larger unitary group, and projects back to the first $$y_k=\sum_{i=1}^p a_i\,y_{k-i}+\sum_{j=0}^q b_j\,x_{k-j},$$5 columns, yielding a new point on the Stiefel manifold and hence a new CPTP map (Schultz et al., 2020).

For a target gate $$y_k=\sum_{i=1}^p a_i\,y_{k-i}+\sum_{j=0}^q b_j\,x_{k-j},$$6 in a circuit, the reference point is the corresponding ideal channel. With $$y_k=\sum_{i=1}^p a_i\,y_{k-i}+\sum_{j=0}^q b_j\,x_{k-j},$$7, one takes the perfect channel to be represented by $$y_k=\sum_{i=1}^p a_i\,y_{k-i}+\sum_{j=0}^q b_j\,x_{k-j},$$8, and perturbations around this point describe noisy implementations of the intended gate. This is the geometric setting in which ARMA-driven tangent perturbations are inserted (Schultz et al., 2020).

3. Formal SchWARMA model and physical interpretation

The formal model couples $$y_k=\sum_{i=1}^p a_i\,y_{k-i}+\sum_{j=0}^q b_j\,x_{k-j},$$9 classical ARMA processes to fixed tangent directions. For each mode $x_k$0,

$x_k$1

where $x_k$2 are zero-mean unit-variance i.i.d. Gaussian, or other chosen, noise innovations. One then chooses fixed tangent matrices

$x_k$3

with $x_k$4 skew-Hermitian and $x_k$5 arbitrary, forms the total tangent perturbation

$x_k$6

and defines the noisy channel at time $x_k$7 by

$x_k$8

A quantum circuit with SchWARMA noise then acts as

$x_k$9

with each step CPTP by construction (Schultz et al., 2020).

The state variables of the ARMA model are the classical scalars $\{a_i\}$0, and the innovations are the i.i.d. classical noises $\{a_i\}$1. The quantum noise process is the sequence $\{a_i\}$2 obtained by mapping ARMA outputs through fixed tangent directions and the Stiefel exponential. In this sense, the model is linear in the noise at the tangent level and nonlinear, through the matrix exponential, in the space of CPTP maps (Schultz et al., 2020).

This division between classical and quantum structure also clarifies the model’s notion of memory. The temporal correlations are classical correlations in the coefficients that determine successive channels, not a recurrence relation over density operators. The authors stress that at the bath level the dynamics are assumed CP-divisible and Markovian in the quantum sense, while the errors are temporally correlated at the gate level because the random channel at time $\{a_i\}$3 depends probabilistically on previous channels through the AR recursion. SchWARMA therefore captures classical non-Markovian temporal structure in gate noise while preserving CPTP maps at every step; it does not rely on non-CP intermediate maps and is not a model of true quantum non-Markovianity in that stronger sense (Schultz et al., 2020).

4. Spectral modeling, Lindbladian connection, and channel classes

All temporal correlations in SchWARMA are inherited from the underlying classical ARMA processes. Each ARMA component has a known autocovariance and PSD, and the total tangent perturbation is a linear combination of these outputs. For high-fidelity gates, the exponentials can be linearized around the identity, and the correlation functions of the error maps can be related directly to the ARMA covariance structure. Because ARMA models are dense in the space of discrete-time spectra, the framework can be used to design gate-time noise with effective spectra matching band-limited noise, multipole or “peaky” spectra, and $\{a_i\}$4 noise. The paper reports explicit numerical quantum noise spectroscopy examples in which reconstructing the spectrum from SchWARMA-simulated data reproduces the intended spectra (Schultz et al., 2020).

The tangent-space parameterization also connects directly to Lindblad generators. For a tangent element with skew-Hermitian $\{a_i\}$5 and arbitrary blocks $\{a_i\}$6, the infinitesimal limit of a SchWARMA step reproduces a Lindblad master equation,

$\{a_i\}$7

with $\{a_i\}$8. In this infinitesimal correspondence, the $\{a_i\}$9 block is the Hamiltonian part and the $\{b_j\}$0 blocks are Lindblad operators. A plausible implication is that SchWARMA can be understood as a finite-time analogue of Lindbladian evolution whose generator amplitudes are themselves driven by correlated ARMA processes (Schultz et al., 2020).

Several special cases illustrate the range of channel classes encompassed by the construction. For single-qubit $\{b_j\}$1 dephasing, one sets $\{b_j\}$2, $\{b_j\}$3, $\{b_j\}$4, and $\{b_j\}$5, obtaining

$\{b_j\}$6

a random $\{b_j\}$7 rotation with angle $\{b_j\}$8. If the ARMA process is i.i.d., the temporal correlations disappear and the model reduces to an i.i.d. per-gate dephasing channel. With three tangent directions $\{b_j\}$9, $$S_y(\omega)=\frac{\left|\sum_{k=0}^q b_k e^{-ik\omega}\right|^2}{\left|1+\sum_{k=1}^p a_k e^{-ik\omega}\right|^2}.$$0, and $$S_y(\omega)=\frac{\left|\sum_{k=0}^q b_k e^{-ik\omega}\right|^2}{\left|1+\sum_{k=1}^p a_k e^{-ik\omega}\right|^2}.$$1, and three independent ARMA processes, averaging over realizations yields a depolarizing channel with error probabilities determined by the variances of the driving processes. For amplitude damping, a non-unitary tangent direction with $$S_y(\omega)=\frac{\left|\sum_{k=0}^q b_k e^{-ik\omega}\right|^2}{\left|1+\sum_{k=1}^p a_k e^{-ik\omega}\right|^2}.$$2 yields Kraus operators

$$S_y(\omega)=\frac{\left|\sum_{k=0}^q b_k e^{-ik\omega}\right|^2}{\left|1+\sum_{k=1}^p a_k e^{-ik\omega}\right|^2}.$$3

which is exactly an amplitude damping channel with fluctuating damping angle $$S_y(\omega)=\frac{\left|\sum_{k=0}^q b_k e^{-ik\omega}\right|^2}{\left|1+\sum_{k=1}^p a_k e^{-ik\omega}\right|^2}.$$4 (Schultz et al., 2020).

5. Simulation, estimation, and circuit-level use

The simulation workflow is gate-scale rather than fine-time integration. For a circuit with $$S_y(\omega)=\frac{\left|\sum_{k=0}^q b_k e^{-ik\omega}\right|^2}{\left|1+\sum_{k=1}^p a_k e^{-ik\omega}\right|^2}.$$5 gates and system dimension $$S_y(\omega)=\frac{\left|\sum_{k=0}^q b_k e^{-ik\omega}\right|^2}{\left|1+\sum_{k=1}^p a_k e^{-ik\omega}\right|^2}.$$6, one precomputes tangent directions for the desired noise types, chooses ARMA orders and coefficients to match a target spectrum, initializes the ARMA states, and then at each time step updates the classical ARMA variables, forms $$S_y(\omega)=\frac{\left|\sum_{k=0}^q b_k e^{-ik\omega}\right|^2}{\left|1+\sum_{k=1}^p a_k e^{-ik\omega}\right|^2}.$$7, computes $$S_y(\omega)=\frac{\left|\sum_{k=0}^q b_k e^{-ik\omega}\right|^2}{\left|1+\sum_{k=1}^p a_k e^{-ik\omega}\right|^2}.$$8 through the Stiefel exponential, and applies it to the current state. In important unitary-noise cases with $$S_y(\omega)=\frac{\left|\sum_{k=0}^q b_k e^{-ik\omega}\right|^2}{\left|1+\sum_{k=1}^p a_k e^{-ik\omega}\right|^2}.$$9, the noisy channel reduces to a single unitary matrix exponential, so the simulation is as cheap as a standard state-vector simulator. The paper states a computational advantage from operating at the gate timescale rather than integrating a stochastic Liouville equation at small $y_k$0 noise correlation time, and reports speedups of $y_k$1, for example a factor $y_k$2 compared to a fine Trotterization (Schultz et al., 2020).

The same framework supports parameter identification in a forward-modeling setting. The sketched identification strategy is to run noisy circuits multiple times, measure a quantity such as survival probability, relate the resulting decay to the underlying PSD through known filter functions,

$y_k$3

reconstruct $y_k$4 by regression or nonnegative least squares, and then fit a classical ARMA model to the reconstructed spectrum using standard identification tools. This suggests a pipeline in which experimentally inferred spectra are converted into ARMA coefficients and then lifted into channel-valued SchWARMA noise for circuit simulation (Schultz et al., 2020).

Representative validation studies are at both circuit and continuously driven levels. The paper reports 5-qubit surface-code $y_k$5- and $y_k$6-stabilizer simulations comparing full Trotterized stochastic Liouville dynamics against a SchWARMA model at the gate timescale; process infidelities from both methods agree to within Monte Carlo sampling error, with dominant differences of $y_k$7 for 1000 trajectories. It also reports Landau–Zener systems, 2- and 3-level, and adiabatic Grover search with collective dephasing, where SchWARMA noise is applied at a coarser step $y_k$8. With $y_k$9, the paper finds excellent agreement with full stochastic simulations and analytical results, together with about 100$\mathcal{S}_k$0 reduction in noise-simulation cost (Schultz et al., 2020).

6. Broader usage of autoregressive and ARMA language in quantum research

The term “quantum autoregressive moving average models” is used more broadly in adjacent literature than in the strict SchWARMA sense. One direction is probabilistic open-system simulation with autoregressive neural networks. In “Autoregressive Transformer Neural Network for Simulating Open Quantum Systems via a Probabilistic Formulation,” the density matrix is mapped exactly to a classical probability distribution via an informationally complete POVM, and that distribution is represented as

$\mathcal{S}_k$1

using an autoregressive Transformer. The work explicitly states that it does not formalize an explicit quantum MA or ARMA structure, but it interprets Lindbladian dissipation and the projection step onto the variational manifold as suggestive of a moving-average-like response to environmental noise (Luo et al., 2020).

A related line of work studies autoregressive quantum states in variational many-body settings. “Impact of conditional modelling for a universal autoregressive quantum state” defines normalized conditional wavefunctions,

$\mathcal{S}_k$2

and introduces filters as analogues to convolutional layers that encode translationally symmetrized correlations. The paper presents these filters as MA-like in the sense that they act as finite-range convolutional correlation structures inside an autoregressive factorization, but it does not claim a channel-valued quantum ARMA model and emphasizes that autoregressive normalization materially constrains expressivity in many systems (Bortone et al., 2023).

A third usage is quantum-inspired ARIMA for classical time series. “QARIMA: A Quantum Approach To Classical Time Series Analysis” keeps the classical ARIMA$\mathcal{S}_k$3 equation but replaces differencing selection, ACF/PACF lag discovery, and AR/MA parameter estimation with swap-test-based diagnostics and fixed-configuration VQCs. Its seven stated quantum contributions are differencing selection, QACF, QPACF, swap-test primitives with delayed-matrix construction, VQC-AR, VQC weak-lag refinement, and VQC-MA. This is a quantum approach to classical time-series analysis, not a model of quantum channel noise or open-system dynamics, but it shows that ARMA language is also being imported into quantum-computing methodology from the opposite direction (Mohanty et al., 9 Apr 2026).

These neighboring usages motivate a careful terminological distinction. In the strict sense established by SchWARMA, a quantum ARMA model is a sequence of CPTP maps driven by classical ARMA processes and constrained geometrically to remain physical. In adjacent neural-state and quantum-inspired forecasting literature, “autoregressive” often refers to conditional factorizations over spatial indices or to quantum-assisted estimation of classical ARIMA components, while the “moving-average” role is either implicit, filter-like, or purely classical (Schultz et al., 2020, Luo et al., 2020, Bortone et al., 2023, Mohanty et al., 9 Apr 2026).

7. Limitations, misconceptions, and extension paths

The main limitations of SchWARMA are geometric and inferential. The Stiefel-manifold representation grows with Kraus rank and system dimension, so matrix exponentials become expensive for large $\mathcal{S}_k$4 or large $\mathcal{S}_k$5. Accuracy analyses based on Magnus expansions assume small noise, such that higher-order terms and non-commutativity effects within a gate are negligible. The model is also a gate-timescale coarse-graining: it captures noise integrated over a gate duration, but it does not resolve fine intra-gate pulse-shaping effects unless one refines the time step or incorporates gate-dependent filtering into the ARMA design. Finally, while the paper sketches parameter-identification methods, it does not fully develop a system-identification theory for estimating both ARMA coefficients and tangent directions directly from quantum data (Schultz et al., 2020).

A common misunderstanding is to equate temporally correlated quantum noise with quantum non-Markovianity in the sense of non-CP intermediate maps. The SchWARMA construction explicitly avoids that move: it assumes CP-divisible dynamics and preserves CPTP maps at each step. Another misunderstanding is to view the ARMA part as acting directly on superoperators. The paper states that there is no ARMA structure directly at the level of superoperators; instead there is a linear ARMA structure in tangent-space coordinates, which is sufficient to generate correlated channel sequences while hard-coding CPTP constraints (Schultz et al., 2020).

The extensions suggested in the literature remain largely classical on the driving-process side. Proposed families include ARIMA for nonstationary noise, SARMA or PARMA for periodic or seasonal structure, and GARCH for volatility clustering. Multi-qubit or multivariate SchWARMA would use coupled multivariate ARMA processes and tangent directions acting across multiple qubits. Time-varying ARMA coefficients are suggested for drifting environments, and integration with optimal control or feedback design is presented as a natural continuation of the signal-processing viewpoint. This suggests an increasingly broad program in which quantum noise is modeled by physically constrained channel geometry while the temporal statistics are supplied by progressively richer classical time-series families (Schultz et al., 2020).