Optimal Convergence Rate of Lie-Trotter Approximation for Quantum Thermal Averages

Technical Summary: Optimal Convergence Rate of Lie–Trotter Approximation for Quantum Thermal Averages

Problem Statement
The calculation of the quantum partition function $Z = \text{tr}[e^{-\beta \hat{H}}]$ and associated thermal averages is central to quantum statistical mechanics. For systems with unbounded Hamiltonians, such as those describing particles in potentials on $\mathbb{R}$, obtaining rigorous, non-asymptotic error bounds for the Lie–Trotter product formula remains a significant mathematical challenge. While the Lie–Trotter approximation $Z_N = \text{tr}[(e^{-\beta \hat{H}_0/N} e^{-\beta \hat{V}/N})^N]$ is the foundation of widely used computational methods like Path Integral Monte Carlo (PIMC) and Path Integral Molecular Dynamics (PIMD), formal analyses often suggest a second-order convergence rate of $O(1/N^2)$ without providing rigorous proofs for unbounded operators. Previous rigorous results have largely focused on operator norm errors, which do not directly translate to the accuracy of thermal averages (trace errors), or have established suboptimal rates such as $O(1/N)$.

Methodology
The authors develop a rigorous error analysis within a path integral framework, specifically targeting the trace error $Z_N - Z$ and the error in thermal averages $|\langle \hat{O} \rangle_N - \langle \hat{O} \rangle|$. The core of the methodology involves:

1. Path Integral Representation: The trace quantities are expressed as integrals over discrete Brownian bridges. The Lie–Trotter approximation is formulated as an expectation over these discrete paths, weighted by a discrete Boltzmann factor.
2. Successive Difference Analysis: Instead of bounding the total error $Z_N - Z$ directly, the authors analyze the successive difference $Z_N - Z_{2N}$. By leveraging the Araki–Lieb–Thirring inequality, they establish that $Z_N \geq Z_{2N} \geq Z$, allowing the total error to be bounded by a telescoping sum of successive differences.
3. Symmetric Splitting and Cancellation: The difference $Z_N - Z_{2N}$ is expressed as a path integral involving the difference between "even-step" and "odd-step" evaluations of the potential and observables. The authors demonstrate that the leading-order terms in these differences cancel, leaving a remainder that scales as $O(1/N^2)$.
4. Stochastic Estimates: The analysis relies on bounding the second moments of the differences between path variables at adjacent time steps. This requires detailed estimates of the covariance structure of discrete Brownian bridges.
5. Domain Truncation (for Confining Potentials): For potentials on the infinite domain $\mathbb{R}$, the authors employ a domain truncation strategy. They approximate the trace by integrating over a finite interval $[-R, R]$ and rigorously bound the truncation error (which decays exponentially with $R$) against the algebraic approximation error. The parameter $R$ is optimized as a function of $N$ to balance these errors.

Key Contributions and Results

• Periodic Potentials (Optimal Rate): For a particle in a smooth, periodic potential $V(x)$ on a finite interval, the paper establishes an optimal convergence rate of $O(1/N^2)$ for both the partition function and thermal averages.

  • Result: $|Z_N - Z| \leq C/N^2$ and $|\langle \hat{O} \rangle_N - \langle \hat{O} \rangle| \leq C/N^2$.
  • The constant $C$ is explicitly derived in closed form, depending on the mass $m$, inverse temperature $\beta$, and bounds on the derivatives of the potential and observable.

• Confining Potentials (Nearly Optimal Rate): For a particle in a smooth, confining potential on $\mathbb{R}$ (with at least quadratic growth), the paper proves a nearly optimal convergence rate of $O((\log N + 1)^{3/2} / N^2)$.

  • Result: $|Z_N - Z| \leq C(\log N + 1)^{3/2} / N^2$.
  • The logarithmic factor arises from the domain truncation technique used to handle the unbounded integration domain. The authors note that this factor is likely an artifact of the proof technique, as an alternative operator splitting (Appendix C) can recover the optimal $O(1/N^2)$ rate under stronger regularity assumptions (uniformly bounded derivatives of a modified potential).

• Thermal Averages: The bounds on the partition function and the observable trace $W_N$ are combined to derive rigorous bounds for the thermal average $\langle \hat{O} \rangle_N$, confirming the second-order accuracy for practical simulations.

Significance
The paper provides a firm mathematical foundation for the second-order accuracy of path integral simulations in quantum statistical mechanics. By moving beyond operator norm bounds to directly analyze trace errors, the results validate the efficiency and accuracy of PIMC and PIMD methods for systems with unbounded Hamiltonians. The explicit error constants allow for quantitative predictions of the number of imaginary time steps $N$ required to achieve a desired precision. The work resolves a long-standing challenge in mathematical physics by rigorously confirming the $O(1/N^2)$ convergence rate for the Lie–Trotter formula in physically relevant settings, distinguishing the analysis of trace errors from previous norm-based analyses. Numerical verification in one dimension supports the theoretical findings, showing that the rescaled error converges to a constant, suggesting the logarithmic factor in the confining case may be removable.