Observable-targeted variational quantum simulation of Hamiltonian dynamics

Technical Summary: Observable-targeted variational quantum simulation of Hamiltonian dynamics

Problem Statement
Standard Variational Quantum Simulation (VQS) aims to reproduce the evolution of the full quantum state by updating circuit parameters to minimize the distance between the variational state derivative and the exact Hamiltonian derivative (typically using McLachlan's variational principle). However, many practical applications require only the expectation values of a few specific observables rather than the complete state. In standard VQS, the variational update distributes accuracy across the entire tangent space of the state, treating all directions equally. Under a finite measurement budget, this approach may waste resources estimating variational matrices in directions that do not significantly impact the target observable, leading to suboptimal accuracy for the specific quantity of interest.

Methodology
The authors propose an Observable-targeted VQS method that updates circuit parameters specifically to reproduce the evolution of a selected set of observables (a "dictionary") rather than the full state.

1. Error Identity and Observable Selection:
   The method is grounded in an exact error identity (Theorem 1) which relates the final error in a target expectation value $\langle A \rangle$ to the accumulated mismatch (residual) between the variational and Hamiltonian dynamics, weighted by the target observable transported backward in time. This identity motivates selecting a dictionary of observables whose linear span approximates the "target orbit" (the set of operators $\{e^{iHs}Ae^{-iHs}\}$). The authors propose constructing this dictionary systematically using repeated commutators of the target observable $A$ with the Hamiltonian $H$ (Heisenberg Krylov spaces).

2. Variational Update:
   Instead of minimizing the Hilbert-Schmidt norm of the state residual (as in standard McLachlan VQS), the targeted method minimizes the mismatch in the equations of motion for the selected dictionary observables. This is formulated as a least-squares problem:
   $$\min_u \| G(\theta)u - b(\theta) \|^2_2$$
   where $G$ encodes the derivatives of the dictionary observables with respect to circuit parameters, and $b$ encodes their Hamiltonian derivatives (expectation values of commutators $i[H, O_a]$).

3. Measurement Efficiency:
   For Pauli observables and circuits where each parameter corresponds to a single Pauli rotation, the required coefficients ($G$ and $b$) can be estimated using only Pauli measurements on the original and parameter-shifted circuits. This eliminates the need for ancillary qubits and controlled operations (Hadamard tests) typically required for overlap estimation in standard VQS implementations.

Key Contributions

• Theoretical Framework: Derivation of an exact error identity linking the target observable error to the residual dynamics, providing a rigorous justification for focusing the variational update on specific observables.
• Algorithmic Construction: A systematic procedure for building observable dictionaries using nested commutators to approximate the target orbit.
• Implementation Efficiency: Demonstration that for Pauli-based circuits, the targeted update avoids ancillary qubits and controlled operations, reducing circuit depth and measurement overhead compared to standard Hadamard-test implementations.
• Numerical Benchmarking: Comprehensive comparison across six-qubit spin, fermionic (Fermi-Hubbard), and molecular (LiH) models under finite measurement statistics.

Results
The authors compared the observable-targeted method against standard McLachlan VQS using identical ansätze, measurement budgets (shots per velocity evaluation), and time-step sizes. The performance metric was the median time $T_\epsilon$ until the target error exceeded a prescribed tolerance.

• Performance Gain: Across all tested models (Mixed-field Ising, XXZ, Disordered Heisenberg, XY, Fermi-Hubbard, and LiH), the targeted method extended the median reachable time within the error tolerance.
• Quantitative Improvements: The ratio of reachable times (Targeted / McLachlan) ranged from 1.06 to 4.22.
  • The largest improvement (factor of 4.22) was observed in the XY chain at a shot budget of $10^7$.
  • In the Fermi-Hubbard model at $10^6$ shots, the median time increased by a factor of 3.81.
  • For the LiH molecular model, the targeted method showed significant gains, with 85% of trajectories remaining within tolerance up to the simulation limit ($T=3$) at high shot budgets, compared to a McLachlan median of 1.23.
• Measurement Trade-offs: While the targeted method requires estimating more coefficients if the dictionary is large, the authors note that for $K \ll p$ (where $K$ is dictionary size and $p$ is parameter count), it can require fewer coefficients than the full Gram matrix of standard VQS. In the LiH case, despite a larger number of measurement settings, the targeted approach yielded superior accuracy.

Significance and Claims
The paper claims that directing the variational update toward the target observable allows for extended accurate simulation without increasing the measurement cost per time step. By concentrating the available accuracy on the quantity to be predicted rather than the entire state, the method mitigates the impact of finite measurement noise on the specific observable of interest.

The authors emphasize that this approach is particularly relevant for near-term quantum processors where circuit depth is constrained and measurement resources are limited. They conclude that the observable intended for measurement is "information available before the simulation starts" and that variational methods should leverage this information rather than discarding it. The work does not claim to solve the general problem of state simulation but offers a practical optimization for observable-specific dynamics.