Rounded Clocks and Finite Precision: A Transfer Principle Between Continuous and Discrete Quantum Dynamics
Michael Jarret
quant-ph
Oct 5, 2026 · v2
math-ph
TL;DR
The definitions, theorems, and examples, including the comparison identity and the O(1/T) transfer theorem, are tagged with links to Lean formalizations.
Abstract
I present a short approach to discretizing continuous unitary evolution in finite dimensions. For a time-dependent Hamiltonian $H(s)$ with $s\in[0,1]$ and an integer $T\ge1$, let $U(T)$ be the evolution of $H(t/T)$ over time $T$, and let $W_T=e^{-iH((T-1)/T)}\cdots e^{-iH(0)}$ be the product of the unitaries generated by its samples. With a half-step phase at each end, $V(T)=e^{-iH(1)/2}W_Te^{iH(0)/2}$ is the evolution generated by $H(\lfloor t+1/2\rfloor/T)$, where $\lfloor t+1/2\rfloor$ is a rounded clock, time at finite precision. Whereas Duhamel's formula only bounds $\|V(T)-U(T)\|$ by a constant, I show that $\|V(T)-U(T)\|=O(1/T)$ in operator norm, with explicit constants, whenever $H'$ is Lipschitz and every sample has eigenvalue differences at most a fixed width below $2π$. The gain comes from cancellation within and between time steps. This general transfer theorem assumes neither adiabaticity nor a spectral gap, and any operator-norm bound for $U(T)$ transfers to $V(T)$ with an added $O(1/T)$. Nonetheless, discrete adiabatic theorems for $W_T$ follow from continuous ones, because the end phases commute with the spectral projectors of $H(0)$ and $H(1)$. For a product of samples of a $C^2$ unitary family, a continuously moving arc of fixed length $\vartheta$ that avoids the spectrum determines a Hermitian logarithm meeting the width condition, and the $O(1/T)$ bound has coefficients $O(\vartheta^{-3})$. In the easier direction, Duhamel's formula shows that rounding the clock to midpoints and truncating the entries of a regulated time-dependent Hamiltonian at a fixed binary place gives a product within $T$ times the error of the rounded, truncated Hamiltonian.
Problem
Products of unitaries generated by samples of a time-dependent Hamiltonian at fixed unit steps are only bounded by a constant distance from the continuous evolution under Duhamel's formula. Existing fixed-step bounds that do not grow with T cover only a gapped eigenstate and each requires its own analysis.
Approach
The sampled product, with half-step phases at both ends, is viewed as the evolution of the Hamiltonian run on a rounded clock. A comparison identity integrates Duhamel's formula by parts against a correction that may jump, which captures cancellation within and between time steps. A converse direction combines midpoint clock rounding with bit truncation of the matrix entries. The definitions, theorems, and examples are marked as formalized in Lean.
Results
For Lipschitz H' and eigenvalue differences of every sample below a fixed width under 2π, the error is O(1/T) in operator norm with explicit constants, and no gap or adiabaticity is assumed. Discrete adiabatic theorems follow from continuous ones, and an application shows that circuits keep the inverse-linear gap dependence of gap-adapted schedules.