Fixed operator order beats adaptive screening

Yu-Hao Chen of National Taiwan University and Renata Wong of Chang Gung University and colleagues demonstrate a new approach to variational imaginary-time evolution (VITE) that achieves a relative error of 10^-4 using 1. 6-5. The team’s technique fixes the order of circuit layers in advance, building them from nested commutators of the Hamiltonian, and appending new layers only when a calculated residual exceeds a threshold.

This pool-free method proved particularly effective on the mixed-field Ising chain and tori, with the advantage persisting as the chain scaled from 8 to 16 sites. The work reveals that this efficiency relies on “broken integrability,” meaning the technique isn’t universally superior to adaptive methods.

Dynamic VITE: Residual Trigger for Circuit Growth

6 to 5. The trigger activates when the relative McLachlan residual, a measure of circuit progress, exceeds a set threshold, offering a streamlined alternative to methods requiring constant screening and re-optimization of operator pools.

This reduction in computational demand is particularly notable given that screening accounts for 87-91% of the circuits used by adaptive methods on spin models. Researchers discovered that the advantage of dynamic VITE is not universal, and is dependent on the system being modeled; tests on the mixed-field Ising chain, tori, and stretched H₄ systems demonstrated consistent benefits, while transverse-field Ising models did not see the same performance gains.

The team replayed recorded runs with a single circuit per orbit of the dihedral group, reducing the number of circuits needed by factors of 8-18 at a relative error of 10^-4 and 13-27 at 10^-6, though this optimization does not apply to single-string circuits. This suggests that the method’s efficiency is linked to the ability to exploit symmetries within the system being simulated, and that the reduction in circuits is not simply a matter of faster convergence.

The performance of dynamic VITE is governed by a principle the researchers term “broken integrability,” which dictates that the technique is most effective when the system lacks certain simplifying symmetries. “The question we address is when this pool-free strategy is cheaper than the adaptive methods, and why,” the paper states, highlighting the focus on identifying the conditions under which this new approach outperforms established techniques.

On the H₄ system at 2. 0 Å, adaptive VITE built a shorter circuit with 14 layers compared to dynamic VITE’s 19. But this circuit required 157 imaginary-time steps to converge, versus 63 for the dynamic VITE method. This suggests a trade-off between circuit length and the number of steps needed to reach a solution, and highlights the importance of considering the overall computational cost rather than solely focusing on circuit size.

Nested Commutators Define Fixed-Order Layer Schedule

Building operator schedules from nested commutators offers a distinct approach to variational imaginary-time evolution, circumventing the need for operator selection during circuit construction. This technique constructs a fixed sequence of layers derived from the Hamiltonian, grouping them by symmetry and particle number, a process reliant on operator algebra alone. The resulting schedule eliminates the computational demand of choosing which operators to add at each step, a key difference from adaptive methods like ADAPT-VQE.

The method’s efficiency is determined by a relative residual, denoted as ρ, which quantifies how closely the circuit approximates the true imaginary-time evolution direction. Defined as the squared norm of the exact direction missed divided by the variance of the Hamiltonian, ρ ranges from zero to one and is a trigger for appending the next layer.

Tests on systems ranging from transverse-field Ising models to the Hubbard chain demonstrate the method’s adaptability across different physical systems and dimensions. Specific performance metrics reveal the advantage of this approach. These results highlight the method’s capacity to generate effective circuits for both spin and fermionic systems, with varying degrees of complexity and accuracy. Nested commutators have previously been employed as counterdiabatic terms and within operator pools for ADAPT-VQE, but this work uniquely arranges them into a fixed schedule.

This fixed order contrasts with adaptive methods that dynamically select operators, and the study aims to determine when this pool-free strategy proves more efficient. The method’s reliance on pre-calculated commutators and a residual-based trigger offers a potentially scalable alternative to the screening and re-optimization inherent in many variational algorithms.

McLachlan Residual Measures Imaginary-Time Trajectory Deficiency

A relative error of 10^-4 was achieved with 1. This performance was observed across five non-integrable systems, the mixed-field Ising chain and tori, alongside stretched H₄ molecules, suggesting broad applicability beyond simplified models. The method’s success hinges on a dynamic trigger based on the relative McLachlan residual, a dimensionless metric measuring the deficiency of the circuit’s trajectory. This trigger activates when the residual exceeds a predetermined threshold, signaling the addition of a new layer to the quantum circuit.

Unlike adaptive methods that rely on screening an operator pool at each growth step and often re-optimizing all parameters, this approach avoids those computationally intensive processes. The researchers found that screening and re-optimization constitute the bulk of the computational demand for adaptive methods, and bypassing them offers a significant advantage.

The study reveals that the technique requires “broken integrability” to outperform adaptive methods, meaning it thrives when the system’s inherent symmetries are disrupted. The researchers also detail the underlying principle of the method, rooted in McLachlan’s variational principle, which minimizes a distance metric measuring the circuit’s deviation from the exact imaginary-time trajectory. This metric, is a key indicator of the circuit’s accuracy and guides the layer-growth process.

Relative Residual (ρ) Quantifies Circuit Convergence

This new quantification underpins a dynamic VITE method that appends layers when ρ exceeds a predetermined threshold, a so-called “residual trigger”. Tests on five non-integrable systems (the mixed-field Ising chain and tori, and stretched H₄) reveal that this pool-free approach can achieve a relative error of 10⁻⁴ with fewer circuits than adaptive methods. Screening and re-optimization account for most of the adaptive methods’ cost. The researchers found that its success hinges on the underlying physics of the system.

On spin models, a longitudinal-field scan shows that the advantage requires broken integrability, which makes adaptive screening expensive. The advantage is lost on transverse-field Ising models, at tight targets (relative error 10⁻⁶) and where the schedule front-loads weak layers. The team’s work builds upon existing methods that use nested commutators of the Hamiltonian, arranging them into a fixed schedule that eliminates the need for operator selection, a computationally intensive step in adaptive algorithms.

Pool-Free Approach Reduces Adaptive Screening Costs

The technique appends new layers when a specific residual threshold is exceeded, a process the authors term the “residual trigger”. On the spin models, the method reached a relative error of ${10}^{–4}$ with 1.6-5.4 times fewer circuits than the cheapest of six to eight adaptive variants. The researchers tested systems with 6 to 16 qubits, revealing that the advantage requires broken integrability.

Performance Advantage on Broken Integrability Systems

Specifically, the advantage vanishes on transverse-field Ising models, and at stringent error targets of 10⁻⁶ when the computational schedule prioritizes weaker layers early on. Dynamic VITE, however, only incurs the cost of the McLachlan system necessary for evolution, a key factor in its efficiency. The advantage observed is not universal; it requires both a computational schedule that reaches the target accuracy in few layers and the presence of “broken integrability,” which increases the expense of adaptive screening.

The study found that on the five systems tested, the mixed-field Ising chain and tori, and H₄, dynamic VITE reached a relative error of 10⁻⁴ using 1. Even with stronger baseline methods, dynamic VITE consistently required fewer circuits on all five systems at every target error rate, needing 1. 6-2. 3-6. “A symmetry-class layer moves many strings with one parameter, in the direction the commutators single out,” the paper explains, highlighting why single-string pools demand more operators to achieve the same accuracy.

More like this

Quantum Error CorrectionGottesman Kitaev Preskill codes gain a dynamical formalism

Quantum PhysicsLocal feedback controls correlations between distant quantum spins

Quantum AlgorithmsGenerative quantum eigensolver learns from hardware at 70 qubits

Quantum AlgorithmsRydberg atoms tested on 12-vertex graphs