Diagonal, Switching, and Locally Optimal Unraveling for efficient quantum trajectories in metastable open quantum systems

AI summary

Overview

Research area: Quantum physics — open quantum systems, numerical methods for quantum trajectories, and dissipative phase transitions.

Technical level: Advanced. The paper assumes familiarity with Lindblad master equations, quantum jump / Monte Carlo wavefunction methods, non-Hermitian effective Hamiltonians, and phase-space semiclassics.

Scope (one sentence): The paper introduces DiSLOU, a three-layer numerical scheme that makes quantum-trajectory simulations of metastable open quantum systems — where rare switching events between long-lived attractors must be sampled over very long times — dramatically cheaper without loss of accuracy, benchmarked on a driven-dissipative Kerr resonator and on a two-mode dissipative cat system.

What This Paper Is About

In many driven open quantum systems, the state lingers for a very long time in one of several long-lived "metastable" configurations before a rare event switches it to another one — exactly the physics behind dissipative phase transitions and the bit-flip errors of cat qubits. Simulating these rare switches with quantum trajectories is expensive because the simulator must simultaneously resolve fast dynamics within each attractor and integrate over timescales long enough to collect switching statistics. The paper's goal is to cut that cost by changing how trajectories are represented and propagated, without changing the physics they describe.

Key Contributions

1. A three-layer composite method (DiSLOU) — "Diagonal, Switching, and Locally Optimal Unraveling" — that combines gauge switching, diagonal propagation of the no-jump evolution, and local reduced eigenspaces into a single composable algorithm for metastable systems.

2. Layer I (locally optimal unraveling): A finite set of N_g gauges is chosen in advance, roughly one per metastable region, with shifts selected to cancel the corresponding jump amplitudes. The gauge is selected at trajectory start by minimizing the activity, and re-selected after a jump only if a candidate's activity beats the current one by a hysteresis factor η (0 < η ≤ 1), which prevents repeated switching near basin boundaries.

3. Layer II (diagonal propagation and first passage): The non-Hermitian effective Hamiltonian is diagonalized once per gauge at cost O(N³), after which the no-jump state is advanced to any later time at cost O(N) instead of O(N²) per step. Jump times are located by directly solving for the first passage of the survival probability through the random threshold u, rather than by step-by-step ODE integration plus root finding.

4. Layer III (local reduced eigenspaces): Because long no-jump intervals suppress fast-decaying eigenmodes, the state is propagated in the span of the m^(g) ≪ N least-decaying eigenmodes, at leading cost O(m^(g)²) rather than O(N²), with a fallback to the full space when switching events redistribute weight across many modes. An open-source Julia implementation built on the quantum-trajectory solver of QuantumToolbox.jl is released.

Main Findings

• Benchmark on the driven-dissipative Kerr resonator: DiSLOU trajectories agree with master-equation evolution within statistical uncertainty and show no measurable loss of accuracy against standard quantum trajectories. The largest performance improvement occurs in the bistable regime, where standard trajectories are most expensive.

• Jump suppression in the optimal gauge: For the bright branch at Δ/κ = 13, K/κ = 1.33, ε/κ = 8.616, the measurement record in the original gauge carries 223 jump counts, while the count drops to 35 in the locally optimal gauge over the same time interval.

• Longer no-jump intervals: In the selected gauge, the survival probability takes 13.7 times longer to cross the survival threshold than in the original ζ = 0 gauge.

• Spectral separation in shifted gauges: Most eigenmodes of the shifted effective Hamiltonians decay quickly and separate from a small number of low-activity modes. The least-decaying eigenmodes of the ζ = ζ^(±) gauges have Wigner functions resembling the metastable states, with amplitudes ⟨â⟩ roughly corresponding to −ζ^(g).

• Mode weights stay local between jumps: During propagation in the optimal gauge, only the first modes retain sizable weights, and higher modes are repopulated after each quantum jump — the empirical basis for the reduced-space truncation.

• Switching requires the full space: Close to a branch switch, the event activity rises sharply and weight spreads across many modes (shown at κt ≃ 134706, immediately after the last jump before the photon-number drop), motivating the fallback procedure.

• Rare bit flips in the cat manifold: Applied to a two-mode dissipatively stabilized parametric oscillator, including the effects of a strongly coupled two-level system, DiSLOU gives a performance improvement of over one order of magnitude while preserving the resulting statistics, including estimates of the decay rate of stored quantum information.

• Documented numerical caveat: Because the eigenvectors of the non-Hermitian Hamiltonian are generally non-orthogonal, coordinates lose up to log₁₀ κ(V) significant digits, with κ(V) the condition number of the eigenvector matrix. The implementation warns when κ(V) ≥ ε_mach^(−1/2) and stops when κ(V) ≥ ε_mach^(−1), i.e., where the Hamiltonian is numerically not diagonalizable, as at an exceptional point.

Methodology in Plain English

A quantum trajectory is one randomly sampled run of a dissipative quantum system: the state evolves smoothly under a non-Hermitian Hamiltonian and is occasionally kicked by a discrete "jump." Averaging many trajectories recovers the density matrix. The trouble in metastable systems is that jumps are frequent inside each attractor, so each trajectory is long and jump-heavy, and the switch of interest is rare, so many long trajectories are needed.

The authors attack this on three fronts.

First, the representation of the jumps is not unique: shifting each jump operator by a complex scalar (a "gauge") and compensating the Hamiltonian leaves the averaged physics identical but changes individual trajectories. Physically this corresponds to mixing the output field with a local oscillator before detection. The authors pre-compute a small set of shifts, roughly one per metastable region, pick the one that minimizes the jump rate at the start of a trajectory, and only swap gauges after a jump, with a hysteresis factor to avoid dithering.

Second, because the Hamiltonian is time independent within a gauge, they diagonalize it once. In the eigenbasis the no-jump evolution is a set of independent phases, so reaching any later time costs O(N). The wavefunction is rebuilt in the original basis when observables or jump operators are needed, and the jump time is found as the first passage of the survival probability through a random threshold — no time stepping. A QR decomposition of the eigenvector matrix replaces explicit inversion when converting states back into eigen-coordinates.

Third, since low-activity modes dominate between jumps, they truncate to the m^(g) slowest-decaying eigenmodes and use a thin QR of that eigenvector block so that norms come for free and observables become small matrices. Whenever the reduced description fails — typically at a switching event — the simulation falls back to the full space. The three layers are validated against independent density-matrix evolution and against standard trajectories.

Why This Matters

Metastability is where much of the interesting physics of driven open systems lives — dissipative phase transitions, bistable resonators, and the engineered noise bias of cat qubits — but it is precisely where brute-force trajectory simulation becomes impractical, since the cost of resolving every jump scales with the same activity that makes the attractors long-lived. DiSLOU makes the per-trajectory cost largely independent of the jump rate and of the metastable manifold's fast internal dynamics, so that simulation time is spent on the rare events that matter. It also demonstrates that gauge freedom, usually treated as a variance-reduction or entanglement-reduction trick, can be exploited as a computational accelerator when combined with spectral methods.

Real-world applications:

• Simulating and calibrating cat-qubit quantum computers, where the quantity of interest is the rare bit-flip rate of the encoded logical state.
• Modeling driven-dissipative nonlinear resonators in superconducting-circuit experiments, including bistability and dissipative phase transitions.
• Estimating the decay rate of quantum information stored in a dissipatively stabilized bosonic mode, including in the presence of a strongly coupled two-level system such as a transmon.
• Providing efficient reference simulations for verifying approximate or low-rank solvers of Lindblad dynamics in the large-system regime.

Industry relevance: The work is directly tied to hardware efforts on dissipatively stabilized cat qubits, which rely on a noise bias between bit flips and phase flips; accurately and cheaply predicting the rare bit-flip rate is a design requirement for such processors. The released Julia package lowers the barrier to adopting the method on existing quantum-trajectory toolchains.

Future Directions

• Robustness of the diagonal layer: The method assumes a diagonalizable effective Hamiltonian. Strongly non-normal cases with ill-conditioned eigenvector matrices require a different treatment, such as a Schur decomposition, and behavior at or near exceptional points remains a limitation.
• Systematic gauge selection: The paper offers two practical recipes — semiclassical basins of attraction and clustering of short trial trajectories — but does not establish how many gauges N_g are needed or how suboptimal choices degrade performance beyond raising jump counts.
• Truncation control: How to choose the reduced dimension m^(g) (fixed number versus a cutoff) and how to make the fallback criterion for switching events principled rather than reactive are open design questions.
• Broader validation: The benchmarks cover the Kerr resonator and a two-mode dissipative cat system; whether the same spectral separation and mode locality hold for other metastable platforms, larger system sizes, or time-dependent drives is not established in the presented content.

Target Audience

Quantum-optics and open-quantum-systems researchers who run large-scale trajectory simulations; superconducting-qubit and cat-qubit experimentalists who need bit-flip and information-decay rates; numerical-methods developers working on Lindblad solvers, low-rank and corner-space approximations, and Krylov or spectral approaches; and graduate students with a background in quantum trajectories who want a concrete, implementable acceleration strategy. Readers without familiarity with unravelings and non-Hermitian effective Hamiltonians will find the middle sections demanding, though the introduction and the layer descriptions are accessible at an intermediate level.

Authors’ abstract

Open quantum systems can display metastability, with multiple long-lived states coexisting. Examples include the bistability of bright and dim manifolds in nonlinear driven resonators, or the long-lived coherent states of opposite phases in parametrically-driven systems, the building block of cat-state-based quantum computers. Metastability often emerges at system sizes where density-matrix evolution is impractical. Quantum trajectories mitigate this issue by evolving wavefunctions rather than density matrices. Metastable states then emerge as attractors towards which the system rapidly relaxes until a rare switching event brings it to another attractor. Simulating them remains costly, as one must resolve fast intra-attractor dynamics over timescales long enough to gather switching statistics. To tackle this problem, we introduce the Diagonal, Switching, and Locally Optimal Unraveling (DiSLOU) method, which combines three optimization layers: (i) Switching among locally optimal unravelings that minimize the number of quantum jumps in each metastable region. (ii) Diagonal propagation: the no-jump evolution is accelerated by diagonalizing the non-Hermitian effective Hamiltonian once per attractor. (iii) Local reduced eigenspaces: the no-jump state can be faithfully propagated in a truncated subspace of only slowly decaying eigenmodes. We benchmark DiSLOU on the driven-dissipative Kerr resonator, finding no measurable loss of accuracy and the largest speedup in the bistable regime, where standard trajectories are most expensive. We then apply the method to the more challenging problem of accurately simulating rare bit-flip events in the cat manifold emerging in a two-mode dissipatively stabilized parametric oscillator, obtaining a performance improvement of over one order of magnitude. An implementation of DiSLOU in Julia is available as an open-source package.