Norm-Aware Product Formulas for Hamiltonian Simulations

Technical Summary: Norm-Aware Product Formulas for Hamiltonian Simulations

Problem Statement
The efficient simulation of quantum-mechanical time evolution, governed by the unitary operator $U(t) = e^{-iHt}$, is a central task in quantum computing, particularly for quantum chemistry. While advanced techniques like Quantum Singular Value Transformation offer optimal scaling, they are currently infeasible for Noisy Intermediate-Scale Quantum (NISQ) devices due to resource constraints. Consequently, product formulas (Trotter-Suzuki decompositions) remain the standard digital approach. However, standard Lie–Trotter–Suzuki (LTS) formulas treat all Hamiltonian terms equally, ignoring potential hierarchies in the operator norms of the Hamiltonian's components. Recent work, specifically the THRIFT product formula, demonstrated that exploiting a hierarchy where a dominant term $H_0$ is separated from a perturbation $\alpha H^{(1)}$ (with $\alpha \ll 1$) can suppress the leading error term by a factor of $\alpha$. However, the THRIFT approach requires implementing grouped exponentials of the form $e^{-i(H_0 + \alpha H^{(1)}_k)t}$, which is efficient only for specific systems (e.g., one- and two-qubit interactions) but lacks a general, resource-efficient synthesis for arbitrary $k$-qubit operators ($k > 2$).

Methodology
The authors propose a new family of first- and second-order product formulas designed to exploit norm hierarchies while retaining the general product-formula structure (Eq. 3), where the Hamiltonian is decomposed into a sum of individually exponentiable terms.

1. Theoretical Framework:

   • The Hamiltonian is structured as $H = H_0 + \alpha H^{(1)}$, where $H_0$ is the dominant term and $H^{(1)} = \sum H^{(1)}_k$ represents the perturbation scaled by a small parameter $\alpha$.
   • First-Order Formulas: The authors derive a "Partial-Strang" (PS) decomposition. By applying a symmetric Strang splitting to the dominant term $H_0$ and approximating the perturbation $e^{-i\alpha H^{(1)}t}$ using a standard first-order Trotter step, they achieve an error scaling of $O(\alpha^2 t^2)$. This improves upon the standard first-order LTS error of $O(\alpha t^2)$.
   • Second-Order Formulas: Extending this logic, the authors construct second-order formulas by substituting the perturbation blocks with second-order symmetric LTS decompositions. By optimizing a parameter $\lambda$ (specifically $\tilde{\lambda} \approx 0.211325$), they construct a decomposition that cancels the leading error term linear in $\alpha$, resulting in an error scaling of $O(\alpha^2 t^3)$. This outperforms the standard second-order LTS scaling of $O(\alpha t^3)$.
   • Generality: Unlike THRIFT, these formulas do not require the implementation of grouped exponentials $e^{-i(H_0 + \alpha H^{(1)}_k)t}$. Instead, they use the standard form $e^{-iH_0 t}$ and $e^{-i\alpha H^{(1)}_k t}$, making them applicable to Hamiltonians with terms acting on an arbitrary number of qubits.

2. Numerical Implementation:

   • The study employs exact classical emulation of the product formulas using Python, NumPy, and SciPy.
   • Two distinct test cases are used:
     • Spin Systems: A 2D spin-$1/2$ XY model on a $3 \times 3$ lattice with open boundary conditions.
     • Molecular Systems: Real-time propagation of electronic Hamiltonians for five molecules ($H_2$, $LiH$, $BeH_2$, $H_2O$, $NH_3$) using the Double-Factorization (DF) framework. In this framework, the Hamiltonian is partitioned into a dominant one-body term ($H_0$) and perturbative two-body terms ($H^{(1)}_k$).

Key Results

• Spin-XY Model:
  • In the perturbative regime (small $\alpha$), the proposed first-order PS formula ($U^{[1]}_{PS}$) and the THRIFT formula ($U^{[1]}_{THRIFT}$) both exhibit error scaling proportional to $\alpha^2$, significantly outperforming the standard LTS formula ($U^{[1]}_{LTS}$), which scales linearly with $\alpha$.
  • Among second-order methods, the proposed optimized formulas ($U^{[2]}_{\tilde{\lambda}}$ and $U^{[2]}_{OG}$) and THRIFT ($U^{[2]}_{THRIFT}$) achieve the lowest absolute errors, systematically outperforming standard second-order LTS and the unoptimized PS formula in the perturbative regime.
  • As $\alpha \to 1$, the theoretical advantage diminishes as the norm hierarchy assumption breaks down, and the errors of different methods converge.
• Molecular Hamiltonians:
  • For the molecules studied, the norm ratio $\alpha_k = \|H^{(1)}_k\| / \|H_0\|$ is consistently small ($< 0.5$), validating the applicability of the norm-aware approach.
  • The proposed first-order formulas ($U^{[1]}_{PS}$ and $U^{[1]}_{\tilde{\lambda}}$) reduce the state-propagation error by a factor of approximately 4.5 to 8 compared to the standard first-order LTS formula.
  • Second-order formulas ($U^{[2]}_{PS}$, $U^{[2]}_{\tilde{\lambda}}$, $U^{[2]}_{OG}$) further reduce errors compared to first-order methods. Notably, the $U^{[2]}_{PS}$ formula proves highly competitive with more complex optimized schemes despite its simpler construction.
  • The results confirm that the error reduction is driven by the structure of the commutators, specifically the suppression of the error contribution from the commutator between the dominant one-body term and the two-body terms.

Significance and Claims
The paper claims that these norm-aware product formulas offer a practical and widely applicable alternative to standard LTS decompositions for Hamiltonians exhibiting a norm hierarchy.

• General Applicability: The primary advantage over the THRIFT method is that these formulas retain the general product-formula structure, allowing them to be applied to Hamiltonians with terms acting on arbitrary numbers of qubits without requiring complex, non-standard gate syntheses.
• Resource Efficiency: The proposed formulas achieve error suppression comparable to or better than THRIFT and standard high-order LTS methods but with a gate count that lies between standard second- and fourth-order LTS methods.
• Algorithmic Impact: The authors posit that reducing the approximation error of the propagator directly improves the accuracy of downstream quantum algorithms, such as Quantum Phase Estimation (QPE), and can potentially reduce the overall circuit depth required for a given target accuracy.
• Modesty: The authors acknowledge that the error bounds are derived for short time steps and that the performance gains depend on the specific structure of the Hamiltonian (specifically the ratio of norms and commutator structures). They note that for molecular systems, the gain is an upper estimate based on spectral bounds, though the observed gains in simulations align with the order of magnitude predicted.

In conclusion, the work demonstrates that explicitly accounting for operator norm hierarchies within the general product-formula framework can yield significant accuracy improvements for both spin models and molecular electronic structure simulations, making it a viable strategy for near-term quantum hardware.