Model Order Reduction for Parametric Dissipative Quantum Systems

1. Introduction

• Exponential growth of the density-matrix space makes repeated high-fidelity steady-state solves costly for parameter-dependent many-body systems.For L sites with local dimension s, the Hermitian matrix space has dimension s^(2L).Source: §1 · p. 1
• The paper treats steady states both as zero-eigenvalue vectors of the Lindbladian and through trace-constrained residual minimization.The reduced space is formed from selected steady-state snapshots.Source: §1 · p. 2

1.1. Related work.

• Prior open-system reduction methods primarily target a fixed system, while [10] applies residual-driven reduced-basis ideas using the smallest eigenvalue of L(µ)†L(µ).This work instead seeks a reusable surrogate for an entire parametric family.Source: §1.1 · p. 2

1.2. Main contributions.

• A unified framework treats the eigenvalue and new trace-constrained formulations, derives residual-based state-error estimates, and establishes greedy convergence under analytic parameter dependence.The convergence analysis accounts for finite training sets; the paper identifies its rate result as the first for greedy reduced-basis constructions in a parametric eigenvalue setting.Source: §1.2 · p. 3

1.3. General notation.

• The paper views N × N Hermitian matrices as a real Hilbert space with the Frobenius inner product and induced norm.It denotes this space by H.Source: §1.3 · p. 3

2.1. Quantum states and Lindbladian operators.

• A finite-dimensional Lindbladian maps Hermitian matrices to zero-trace Hermitian matrices, and its steady-state problem has a solution in the quantum-state set.That solution is unique if and only if the kernel has dimension one.Source: §2.1 · p. 3
• The zero-trace Hermitian space is the orthogonal complement of the identity, and unit-trace Hermitian matrices form its identity-shifted affine space.The quantum-state set is the intersection of that affine space with the positive semidefinite matrices; it is compact and convex.Source: §2.1 · p. 4

2.2. Parametric formulation and approximability of the steady state solution set.

• The parametric formulation assumes a compact parameter domain and an affine Lindbladian decomposition with fixed operators and scalar coefficient functions.Nonnegativity of the dissipative coefficient functions is required for the Lindbladian structure.Source: §2.2 · p. 4
• The steady-state solution set is not necessarily a topological manifold, and its Kolmogorov widths quantify how well it can be approximated by low-dimensional linear spaces.Linear and affine widths are equivalent up to an index shift.Source: §2.2 · pp. 4–5

• Kolmogorov widths quantify how well the steady-state solution set can be approximated by low-dimensional spaces.Source: §3 · p. 5
• The section aims to turn that approximability into a constructive reduced model.Source: §3 · p. 5
• Its roadmap covers reduced-basis methods, offline reduced-operator assembly, two reduced formulations, and optional positivity post-processing.Source: §3 · p. 5

3.1. Reduced basis method.

• Uniqueness is unnecessary for constructing snapshot spaces or residual indicators, but is imposed for state-error estimates and convergence of a single-valued solution map.Source: §3.1 · p. 5
• Offline, the method builds a reduced space from snapshots and stores parameter-independent quantities; online, it assembles and solves a reduced formulation for each parameter.The affine formulation has unit trace, while the eigenvalue formulation is trace-normalized when that normalization is well-defined.Source: §3.1 · p. 5
• Reduced observable evaluation uses a once-computed reduced observable representation, avoiding state reconstruction and other full-order computations for subsequent parameter queries.This holds unless positivity post-processing is applied.Source: §3.1 · p. 5

3.1.1. Subspace assembly strategies.

• POD retains Gram-matrix eigenvectors above a cutoff and minimizes mean-square snapshot projection error, but requires all snapshot solves in advance.Greedy construction instead selects parameters by maximizing an error indicator over a finite training set and enriches the space with the selected state.Source: §3.1.1 · p. 6

3.2. Offline computation of reduced operators.

• The reduced operator is the Galerkin projection of the self-adjoint positive semidefinite operator onto V, represented by an r × r matrix.Its representation uses the Frobenius-orthonormal reduced basis.Source: §3.2 · p. 6
• Online assembly uses precomputed component-operator actions on the reduced basis and requires only coefficient evaluations and reduced-matrix operations, not the full-order Lindbladian.The affine decomposition supplies the parameter-dependent coefficients for assembling the reduced operator.Source: §3.2 · pp. 6–7
• Symmetry reduces storage to (P + Q)(P + Q + 1)/2 parameter-independent r × r matrices.The reduced matrix for a new parameter is then assembled from those stored matrices.Source: §3.2 · pp. 6–7
• Offline quantities can be updated incrementally when a basis matrix is added, without recomputing previously assembled entries.Only the new component actions and corresponding reduced-matrix rows and columns are needed.Source: §3.2 · p. 7

3.3. Reduced formulations and greedy selection rules.

• Both reduced formulations use the compressed operator and a parameter-independent vector representing the trace on the reduced space.The two formulations support both greedy selection and online approximation.Source: §3.3 · p. 7

3.3.1. Reduced eigenvalue formulation.

• EigGreedy uses the squared residual of a normalized eigenvector for the smallest eigenvalue of A_r(µ), then normalizes its trace online when nonzero.The eigenvalue formulation's error indicator is the squared residual norm.Source: §3.3.1 · p. 7

3.3.2. Affine residual formulation.

• The affine residual formulation enforces unit trace in the reduced solve and obtains its minimizer through a saddle-point system.Feasibility follows for the greedy spaces and for the stated nontrivial POD-space construction.Source: §3.3.2 · p. 7
• Under a unique full steady state, A(µ) is positive definite on A0, making the reduced objective strictly convex on the feasible affine space and the saddle-point system nonsingular.Source: §3.3.2 · p. 8
• The affine approximation corresponds to solving the steady-state problem in V ∩ A0, and its squared residual indicator defines AffGreedy.Source: §3.3.2 · p. 8

3.3.3. Algorithmic formulation and computational comparison.

• Each greedy step selects the training parameter maximizing its formulation's indicator, computes that steady state, and enriches the basis with it.Source: §3.3.2 · p. 8, §3.3.3 · p. 8
• The construction starts from span{I}; as the nested spaces grow, parameters below tolerance can be removed from subsequent sweeps.The remaining parameters form an active training set.Source: §3.3.3 · p. 8
• At each active parameter, the eigenvalue method solves for the smallest eigenpair, while the affine method solves an (r + 1)-dimensional saddle-point system; both dense solves have cubic worst-case complexity in r.Source: §3.3.3 · p. 8

3.4.1. Projection onto the quantum-state set.

• The exact correction is the unique Frobenius-metric projection onto the quantum-state set, and it preserves the input matrix's eigenbasis.This reduces the projection to projecting the eigenvalue vector onto the unit simplex.Source: §3.4.1 · p. 8
• The exact projection requires a complete spectral decomposition of an N × N Hermitian matrix and may become prohibitively expensive for large N.Source: §3.4.1 · p. 9

3.4.2. Partial spectral clipping.

• Partial spectral clipping corrects the computed negative spectral components; it is certified to yield a quantum state when all negative eigenvalues are included, in particular if λN−k+1 ≥ 0.The procedure computes only the k algebraically smallest eigenpairs.Source: §3.4.2 · pp. 9–10
• When k ≪ N, clipping may be substantially cheaper, but both corrections require full-matrix reconstruction and spectral computations, so the paper makes them optional online steps.Source: §3.4.2 · p. 10

4.1. Effect of positivity post-processing.

• Exact projection onto the quantum-state set does not increase approximation error, while partial spectral clipping scales it by a k-dependent prefactor.The stated prefactor is 1 for exact projection and at most 2 + k for partial clipping.Source: §4.1 · pp. 10–11

4.2. Residual-based state-error estimates.

• For a unique steady state, the state error is bounded by the Lindbladian residual divided by the trace-zero inf-sup constant and the absolute trace.The inf-sup constant is positive exactly when the steady state is unique; the spectral gap and minimal angle provide a lower bound for it.Source: §4.2 · pp. 11–12

4.3.1. Generic weak greedy convergence framework for RBM.

• Weak-greedy convergence is exponential or sub-exponential when the solution set is compact and its Kolmogorov widths decay at the corresponding rate.Exponential convergence follows for width decay exponent α2 = 1, and sub-exponential convergence for 0 < α2 < 1.Source: §4.3.1 · pp. 12–13

4.3.2. Compact solution set and valid weak greedy selection rules.

• With continuous coefficients, a unique steady state throughout the compact parameter domain, and a reduced space containing I, both reduced-formulation indicators are efficient and reliable.These assumptions also yield a continuous steady-state map and a compact solution set.Source: §4.3.2 · pp. 13–14

4.4. Analytic regularity, width decay and convergence of the greedy algorithm.

• The convergence analysis next imposes assumptions on parameter dependence and domain geometry, then uses steady-state analyticity and tensorized Chebyshev interpolation to derive width decay.It also relates the idealized full-domain estimate to the finite training set used in Algorithm 1.Source: §4.4 · p. 15
• Real-analytic parameter dependence makes the steady-state map real-analytic on a neighborhood of the parameter set.This follows under Assumptions 4.9 and 4.12.Source: §4.4 · p. 15
• For a hyperrectangle parameter domain, analytic regularity yields exponential or sub-exponential decay of the approximation widths.The result also applies when the domain is covered by a surjective analytic parametrization from a compact hyperrectangle, provided the composed steady-state map extends real-analytically.Source: §4.4 · p. 16
• The idealized eigenvalue and affine greedy procedures inherit exponential or sub-exponential convergence from the width decay.The result assumes both procedures start from the span of the identity and maximize their indicators over the full parameter set.Source: §4.4 · p. 17
• For either formulation, finite-training-set greedy convergence bounds have constants independent of iteration and training set, while discretization error is bounded by the steady-state Lipschitz constant times the fill distance.The training-set state-space error is also bounded by √C⋆ εtol.Source: §4.4 · pp. 17–18

5. Numerical experiments

• The experiments assess solution-set compressibility, greedy convergence, approximation accuracy, residual indicators, offline costs, online speedups, and observable evaluation.The numerical study is organized around the two test models and these evaluation aspects.Source: §5 · p. 18

5.1.1. Boundary-driven Fermi–Hubbard chain.

• The first test model is a boundary-driven spinless Fermi–Hubbard chain with one fermionic mode per site.Its parameter is µ = (g, δ) ∈ [0, 2] × [−0.25, 0.25].Source: §5.1.1 · p. 18

5.1.2. XYZ chain.

• The second test model is a dissipative spin-1/2 XYZ chain with parameters µ = (Jx, Jy) ∈ [−1, 1]².The Lindbladian has an affine form with three parameter coefficients.Source: §5.1.2 · pp. 18–19

5.1.3. High-fidelity solver, parameter grids, and error measures.

• The reference solver uses sparse vectorization and a trace-lifted normal equation solved by conjugate gradients with relative tolerance 10^-14.Computed solutions are symmetrized and normalized to unit trace; spectral gap and trace-zero stability constants are evaluated by Lanczos methods at sampled parameters.Source: §5.1.2 · p. 19, §5.1.3 · p. 19
• Training grids contain 492 FH and 612 XYZ parameters, while errors are evaluated on 132 uniform validation points.Grid refinement was used because coarser grids produced visibly faster singular-value decay, indicating an insufficiently resolved snapshot ensemble.Source: §5.1.3 · p. 19
• At every training parameter, the eigensolver detected a single zero mode separated from the remaining spectrum, confirming sampled-parameter uniqueness.The authors say this indicates uniqueness throughout the parameter domain.Source: §5.1.3 · p. 19

5.2. Empirical compressibility and system-size dependence.

• Using the singular-value cutoff ε = 3.2 × 10^-7, retained POD dimension rε(L) increases with L but grows much more slowly than ambient dimension N² = 4L; XYZ consistently requires more modes than FH.Source: §5.2 · p. 19, Table 1 · p. 20
• For L=5 to 8, fitted slopes change from −3.39 to −1.97 for FH and −1.22 to −0.84 for XYZ, consistent with sub-exponential decay; decay slows with size and is faster for FH.The normalized singular values decay over several orders of magnitude, indicating empirical compressibility of both sampled solution sets.Source: §5.2 · p. 19, Figure 1 · p. 20

5.3.1. Offline cost.

• Both greedy procedures substantially reduce offline construction time relative to POD, and AffGreedy is faster than EigGreedy.Indicator evaluation takes 15.5 s versus 27.2 s for FH, and 728 s versus 1736 s for XYZ, for AffGreedy and EigGreedy respectively.Source: §5.3.1 · p. 20, Table 2 · p. 20

5.3.2. Greedy convergence.

• For L=8, the POD singular values and greedy indicators decay over several orders of magnitude at comparable rates, with faster decay for FH than XYZ.This is qualitatively consistent with the sub-exponential convergence predicted for both greedy constructions.Source: §5.3.2 · p. 20, Figure 2 · p. 21
• At common dimensions, POD has the smallest projection errors; among greedy spaces, AffGreedy generally has lower absolute error, while EigGreedy has lower relative error at larger dimensions.The difference is consistent with EigGreedy's unit-Frobenius normalization and AffGreedy's unit-trace scaling.Source: Figure 3 · p. 21, §5.3.2 · p. 21

5.3.3. Accuracy at fixed RB dimension.

• Residual-driven greedy accuracy is limited in finite precision because squared indicators must resolve values roughly at the square of the target projection-error scale.Stability factors impose a further reduction for EigGreedy and AffGreedy, respectively, so reliable grid comparisons can fail before vector errors reach numerical precision.Source: §5.3.3 · p. 22
• At the largest common RB dimensions, all methods stay below 4.2 × 10^-5 maximal absolute and 2.7 × 10^-5 maximal relative error for both models.POD is most accurate at fixed dimension, while AffGreedy generally has smaller approximation errors than EigGreedy; no significant loss of positivity is observed at the final greedy dimensions.Source: §5.3.3 · pp. 21–22, Figure 4 · p. 22

5.4.1. Indicator effectivity.

• For both models, both greedy indicators and squared projection errors decay at comparable rates; AffGreedy has narrower effectivity ranges and larger intrinsic weakness parameters.Median effectivity is of order 10 for FH EigGreedy and 1 for FH AffGreedy; for XYZ the corresponding orders are 100 and 10.Source: §5.4.1 · p. 22, Figure 5 · p. 23

5.4.2. Validation of the residual-based state-error bound.

• Sampled trace-zero stability constants have minima of 2.35 × 10^-3 for FH and 7.33 × 10^-2 for XYZ, providing numerical evidence for uniqueness across both parameter domains.The positive values vary smoothly over the validation grids.Source: §5.4.2 · p. 23
• The β-based state-error bound holds at all 169 validation parameters for both greedy methods and is markedly tighter for XYZ.At final dimensions, median bound-to-error ratios are 97–102 for FH and about 13.6 for XYZ; using β directly reduces median bounds by factors of approximately 1.9 and 4.8, respectively, relative to the gap–angle bound.Source: §5.4.2 · p. 23

5.5.1. Online cost.

• Relative to high-fidelity solves, the reported speedups are of order 10^3 and 10^2, respectively, for the two final reduced models at L = 8.Reported online times include reconstruction of the RB approximation.Source: §5.5.1 · p. 24, Table 3 · p. 24

5.5.2. Reduced evaluation of observables.

• Reduced observable values are computed directly from reduced coefficient vectors, without reconstructing the full-order RB state at validation points.For both greedy constructions, one-time reduced-observable construction takes less than 0.07 s and each subsequent evaluation takes less than one microsecond.Source: §5.5.2 · pp. 24–25
• Observable errors decrease by several orders of magnitude with RB dimension; at the largest common dimensions, current and magnetization errors are below 4 × 10^-8 and 2 × 10^-7, respectively.The current error falls below one percent of its maximal high-fidelity magnitude at r=30, and the magnetization error does so at r=60.Source: Figure 7 · p. 25, §5.5.2 · p. 25

• The paper develops a reduced-basis framework for parametric Lindbladian steady states, unifying the eigenvalue and trace-constrained residual formulations in a common greedy construction.Source: §6 · p. 25
• Both formulations admit offline–online decomposition, and positive semidefiniteness can optionally be imposed through post-processing.Source: §6 · p. 25
• Under uniqueness, residual-based estimates control state error, and analytic parameter dependence yields generalized-exponential width and greedy convergence for both formulations.The analysis also quantifies the effect of a finite training set.Source: §6 · p. 25
• POD is generally most accurate at fixed reduced-basis dimension, while AffGreedy avoids the full snapshot ensemble and offers the more favorable offline-cost and accuracy balance among the two greedy procedures.Source: §6 · p. 25