Randomized Trotter Formula in Quantum Simulation

• RTF is a method that randomizes the ordering of Hamiltonian fragments to mitigate fixed coherent bias in product-formula simulations.
• It leverages BCH error cancellation by replacing a deterministic bias with an averaged effective error, while introducing minimal spectral broadening.
• Embedded in frameworks like SPRINT and GRADE, RTF enables near-integrable splitting and significant reductions in quantum gate counts.

Searching arXiv for the cited papers to ground the article in current sources. Randomized Trotter Formula (RTF) denotes a class of product-formula methods for Hamiltonian simulation in which the ordering of Hamiltonian fragments is randomized across Trotter steps rather than held fixed. In the chemistry-oriented formulation developed in "Theory and practice of Trotter product formulas for quantum chemistry" (Casares et al., 29 Jun 2026), the defining randomized ingredient is explicit: “the ordering of the fragments in these products is changed from Trotter step to Trotter step,” with the captioned clarification that “the term ‘randomized’ refers to the ordering of the fragments in the product formula.” In that setting, RTFs are embedded in a broader design framework, SPRINT, that combines randomized fragment ordering with near-integrable splitting, symmetry protection, processing, and QROM-based compilation for electronic-structure Hamiltonians (Casares et al., 29 Jun 2026).

1. Definition and formal scope

In the algorithmic setting, the Hamiltonian is decomposed into fragments,

$H=\sum_j H_j,$

and a product formula approximates

$$e^{-i\tau H}\approx U(\tau)=\prod_k \exp(-i\tau a_k H_k).$$

The basic deterministic instances are the first-order formula

$U_{1}(\tau) = \prod_j \exp(-i H_j \tau)$

and the symmetric second-order Strang formula

$$U_{2}(\tau) = \underbrace{\prod_{j=1}^L \exp(-i H_j \tau/2)}_{U_{1}(\tau/2)} \times \underbrace{\prod_{j=L}^1 \exp(-i H_j \tau/2)}_{U^\dagger_{1}(-\tau/2)}.$$

Randomized Trotterization modifies this deterministic template by changing the order of fragments from step to step. The stepwise composition is written as

$$U(J\tau) = \prod_{j=1}^J U_j(\tau)\approx \exp\left(-i\tau\sum_j H_{\text{eff},j}\right),$$

so that, to leading BCH order, the effective Hamiltonians of different steps add. The specific randomized mechanism emphasized in SPRINT is permutation-based: the method randomizes fragment orderings, typically within fragment groups $A$ and $B$, rather than randomizing the Hamiltonian decomposition itself, the symmetry sector, or the low-rank factorization (Casares et al., 29 Jun 2026).

This places RTF squarely in the family of randomized product formulas. The source material also distinguishes it from fully stochastic simulation primitives: SPRINT is described as a “stepwise randomized ordering scheme,” not “a fully stochastic simulation primitive like qDRIFT,” although qDRIFT and other randomized compilation methods can optionally be used for a residual group $H_C$ (Casares et al., 29 Jun 2026).

2. Deterministic product formulas and the algebra of error terms

The analysis of RTFs in the cited chemistry setting is based on BCH expansions. For two operators $X$ and $Y$, the BCH series is presented as

$$e^{-i\tau H}\approx U(\tau)=\prod_k \exp(-i\tau a_k H_k).$$0

and the symmetric BCH expansion as

$$e^{-i\tau H}\approx U(\tau)=\prod_k \exp(-i\tau a_k H_k).$$1

For first order,

$$e^{-i\tau H}\approx U(\tau)=\prod_k \exp(-i\tau a_k H_k).$$2

with

$$e^{-i\tau H}\approx U(\tau)=\prod_k \exp(-i\tau a_k H_k).$$3

For second order,

$$e^{-i\tau H}\approx U(\tau)=\prod_k \exp(-i\tau a_k H_k).$$4

where

$$e^{-i\tau H}\approx U(\tau)=\prod_k \exp(-i\tau a_k H_k).$$5

Higher even orders are generated through the Suzuki recursion

$$e^{-i\tau H}\approx U(\tau)=\prod_k \exp(-i\tau a_k H_k).$$6

with

$$e^{-i\tau H}\approx U(\tau)=\prod_k \exp(-i\tau a_k H_k).$$7

The randomized component exploits the antisymmetry of commutators under order swaps. If an error term contains a nested commutator with no repeated innermost fragments, then exchanging $$e^{-i\tau H}\approx U(\tau)=\prod_k \exp(-i\tau a_k H_k).$$8 and $$e^{-i\tau H}\approx U(\tau)=\prod_k \exp(-i\tau a_k H_k).$$9 flips its sign:

$U_{1}(\tau) = \prod_j \exp(-i H_j \tau)$0

This sign flip is the algebraic basis for expectation-level cancellation under random reordering. Deterministic orderings therefore generate a fixed coherent BCH bias, whereas randomized orderings replace that fixed bias by an averaged effective error Hamiltonian plus fluctuations around it (Casares et al., 29 Jun 2026).

3. Chemistry-structured realizations: SPRINT, GRADE, and near-integrability

The most developed modern RTF realization in the supplied sources is SPRINT, short for “Symmetry-Protected Randomized near-Integrable Trotter.” It is presented not as a single universal algebraic product formula but as a framework whose workflow is: factorize the Hamiltonian into fast-forwardable fragments; group fragments by norm into $U_{1}(\tau) = \prod_j \exp(-i H_j \tau)$1, $U_{1}(\tau) = \prod_j \exp(-i H_j \tau)$2, and possibly $U_{1}(\tau) = \prod_j \exp(-i H_j \tau)$3; choose a near-integrable product formula; randomize the ordering within groups $U_{1}(\tau) = \prod_j \exp(-i H_j \tau)$4 and $U_{1}(\tau) = \prod_j \exp(-i H_j \tau)$5 at each Trotter step; apply symmetry protection when auxiliary orbitals are present; optionally add processing; use QROM to batch commuting diagonal rotations; and optionally simulate $U_{1}(\tau) = \prod_j \exp(-i H_j \tau)$6 with qDRIFT, RTE, or STAR (Casares et al., 29 Jun 2026).

A major structural ingredient is the Generalized Rank Decomposition (GRADE), defined through

$U_{1}(\tau) = \prod_j \exp(-i H_j \tau)$7

with the residual approximated by Pauli strings:

$U_{1}(\tau) = \prod_j \exp(-i H_j \tau)$8

The source states that this decomposition unifies direct qubit mappings ($U_{1}(\tau) = \prod_j \exp(-i H_j \tau)$9), compressed double factorization when $$U_{2}(\tau) = \underbrace{\prod_{j=1}^L \exp(-i H_j \tau/2)}_{U_{1}(\tau/2)} \times \underbrace{\prod_{j=L}^1 \exp(-i H_j \tau/2)}_{U^\dagger_{1}(-\tau/2)}.$$0, and isometric THC when $$U_{2}(\tau) = \underbrace{\prod_{j=1}^L \exp(-i H_j \tau/2)}_{U_{1}(\tau/2)} \times \underbrace{\prod_{j=L}^1 \exp(-i H_j \tau/2)}_{U^\dagger_{1}(-\tau/2)}.$$1 and $$U_{2}(\tau) = \underbrace{\prod_{j=1}^L \exp(-i H_j \tau/2)}_{U_{1}(\tau/2)} \times \underbrace{\prod_{j=L}^1 \exp(-i H_j \tau/2)}_{U^\dagger_{1}(-\tau/2)}.$$2 (Casares et al., 29 Jun 2026).

The near-integrable organization is central. The paper writes

$$U_{2}(\tau) = \underbrace{\prod_{j=1}^L \exp(-i H_j \tau/2)}_{U_{1}(\tau/2)} \times \underbrace{\prod_{j=L}^1 \exp(-i H_j \tau/2)}_{U^\dagger_{1}(-\tau/2)}.$$3

with $$U_{2}(\tau) = \underbrace{\prod_{j=1}^L \exp(-i H_j \tau/2)}_{U_{1}(\tau/2)} \times \underbrace{\prod_{j=L}^1 \exp(-i H_j \tau/2)}_{U^\dagger_{1}(-\tau/2)}.$$4 comprising the dominant one-body fragment and leading two-body fragment, and $$U_{2}(\tau) = \underbrace{\prod_{j=1}^L \exp(-i H_j \tau/2)}_{U_{1}(\tau/2)} \times \underbrace{\prod_{j=L}^1 \exp(-i H_j \tau/2)}_{U^\dagger_{1}(-\tau/2)}.$$5 the many smaller fragments. The purpose is to avoid applying a uniformly high-order Suzuki formula to a fragment tail whose norms are much smaller. In this design, RTF is not a standalone trick; it is combined with a norm hierarchy created by factorization.

The two flagship near-integrable kernels are summarized below.

Kernel	Formula	Role
$$U_{2}(\tau) = \underbrace{\prod_{j=1}^L \exp(-i H_j \tau/2)}_{U_{1}(\tau/2)} \times \underbrace{\prod_{j=L}^1 \exp(-i H_j \tau/2)}_{U^\dagger_{1}(-\tau/2)}.$$6	$$U_{2}(\tau) = \underbrace{\prod_{j=1}^L \exp(-i H_j \tau/2)}_{U_{1}(\tau/2)} \times \underbrace{\prod_{j=L}^1 \exp(-i H_j \tau/2)}_{U^\dagger_{1}(-\tau/2)}.$$7	$$U_{2}(\tau) = \underbrace{\prod_{j=1}^L \exp(-i H_j \tau/2)}_{U_{1}(\tau/2)} \times \underbrace{\prod_{j=L}^1 \exp(-i H_j \tau/2)}_{U^\dagger_{1}(-\tau/2)}.$$8 treated effectively at second order and $$U_{2}(\tau) = \underbrace{\prod_{j=1}^L \exp(-i H_j \tau/2)}_{U_{1}(\tau/2)} \times \underbrace{\prod_{j=L}^1 \exp(-i H_j \tau/2)}_{U^\dagger_{1}(-\tau/2)}.$$9 at first order
$$U(J\tau) = \prod_{j=1}^J U_j(\tau)\approx \exp\left(-i\tau\sum_j H_{\text{eff},j}\right),$$0	$$U(J\tau) = \prod_{j=1}^J U_j(\tau)\approx \exp\left(-i\tau\sum_j H_{\text{eff},j}\right),$$1	higher-order near-integrable kernel

For the second-order near-integrable kernel,

$$U(J\tau) = \prod_{j=1}^J U_j(\tau)\approx \exp\left(-i\tau\sum_j H_{\text{eff},j}\right),$$2

while for the fourth-order kernel,

$$U(J\tau) = \prod_{j=1}^J U_j(\tau)\approx \exp\left(-i\tau\sum_j H_{\text{eff},j}\right),$$3

This suggests that the principal value of randomized ordering emerges when it is layered onto a structured near-integrable split rather than applied to an unstructured fragment list (Casares et al., 29 Jun 2026).

4. Randomization, coherent bias cancellation, and variance-induced broadening

The error model used in the source is not centered on worst-case diamond-norm analysis. Instead, it is described as a chemistry-specific spectral error metric based on peak shifts in spectroscopy, estimated through perturbation theory on the effective Hamiltonian. For a deterministic $$U(J\tau) = \prod_{j=1}^J U_j(\tau)\approx \exp\left(-i\tau\sum_j H_{\text{eff},j}\right),$$4-th order formula with

$$U(J\tau) = \prod_{j=1}^J U_j(\tau)\approx \exp\left(-i\tau\sum_j H_{\text{eff},j}\right),$$5

the perturbed eigenvalues and eigenvectors satisfy

$$U(J\tau) = \prod_{j=1}^J U_j(\tau)\approx \exp\left(-i\tau\sum_j H_{\text{eff},j}\right),$$6

and

$$U(J\tau) = \prod_{j=1}^J U_j(\tau)\approx \exp\left(-i\tau\sum_j H_{\text{eff},j}\right),$$7

For randomized product formulas, the source defines an average effective Hamiltonian

$$U(J\tau) = \prod_{j=1}^J U_j(\tau)\approx \exp\left(-i\tau\sum_j H_{\text{eff},j}\right),$$8

and a shot-dependent one

$$U(J\tau) = \prod_{j=1}^J U_j(\tau)\approx \exp\left(-i\tau\sum_j H_{\text{eff},j}\right),$$9

where $A$0 is the random ordering at step $A$1. The resulting interpretation is explicit: randomized ordering reduces coherent BCH bias by replacing a fixed ordering-dependent $A$2 with an averaged $A$3, but fluctuations $A$4 introduce dephasing and spectral broadening. The broadening rate is

$A$5

This tradeoff is one of the defining technical features of RTF in the chemistry setting. The coherent contribution appears as an average peak shift, while the stochastic contribution appears as line broadening. The paper states that, in the studied regime, the broadening is “usually small enough to be worth it,” and also states that randomization is “most effective at low Trotter orders and incurs no additional computational overhead” (Casares et al., 29 Jun 2026).

Randomization is only one controlled cancellation mechanism in SPRINT. Symmetry protection is introduced when auxiliary orbitals create leakage between physical and auxiliary subspaces. The symmetry-protected step is written as

$A$6

and is used so that leakage is reduced to

$A$7

Processing supplies another cancellation device. For the fourth-order near-integrable kernel, the processor is

$A$8

and conjugation yields

$A$9

A plausible implication is that the most effective RTF architectures are composite: randomized ordering handles sign-sensitive commutator cancellation, while symmetry protection and processing suppress structurally distinct error channels.

5. Compilation, resource estimates, and the Li$B$0Mn$B$1O case study

The cited source argues that product formulas remain attractive because of low qubit requirements, and it supplements randomized ordering with QROM-based compilation. Within a fragment, commuting diagonal $B$2 rotations can be implemented by precomputing the phase

$B$3

The baseline Toffoli cost is

$B$4

whereas the QROM implementation cost is

$B$5

The stated consequence is about $B$6–$B$7 savings on diagonal blocks and around $B$8–$B$9 overall step savings depending on parameters (Casares et al., 29 Jun 2026).

The principal numerical application is X-ray absorption spectroscopy for a CAS(22e,18o) Li$H_C$0Mn$H_C$1O cluster. The spectroscopy signal is written as

$H_C$2

with

$H_C$3

The resource-estimation assumptions are listed as target peak-shift precision $H_C$4 eV $H_C$5 Ha, $H_C$6 Ha, $H_C$7 Ha, $H_C$8, $H_C$9, shots $X$0, and CDF factorization with $X$1 fragments.

Under those assumptions, the source reports the following central findings for the $X$2 active space. Overall SPRINT gives a $X$3 Toffoli reduction over the prior Trotter state of the art, is only $X$4 more Toffolis than qubitization, and uses $X$5 fewer logical qubits. The approximate attribution of savings is: tighter Trotter error estimation $X$6, near-integrability $X$7, randomization about $X$8–$X$9, and QROM/compilation $Y$0. For near-integrability alone, $Y$1 gives about $Y$2 fewer Toffolis than the standard second-order formula at $Y$3, and $Y$4 gives about $Y$5 fewer Toffolis than standard fourth-order Suzuki at $Y$6, asymptotically approaching $Y$7 and $Y$8, respectively (Casares et al., 29 Jun 2026).

The updated XAS resource table entry for SPRINT/CDF at $Y$9 is 100 logical qubits, $$e^{-i\tau H}\approx U(\tau)=\prod_k \exp(-i\tau a_k H_k).$$00 total Toffolis, and a largest circuit of $$e^{-i\tau H}\approx U(\tau)=\prod_k \exp(-i\tau a_k H_k).$$01 Toffolis. The qubitization comparison at $$e^{-i\tau H}\approx U(\tau)=\prod_k \exp(-i\tau a_k H_k).$$02, $$e^{-i\tau H}\approx U(\tau)=\prod_k \exp(-i\tau a_k H_k).$$03, gives 336 qubits, 2495 Toffolis per walk step, $$e^{-i\tau H}\approx U(\tau)=\prod_k \exp(-i\tau a_k H_k).$$04, and $$e^{-i\tau H}\approx U(\tau)=\prod_k \exp(-i\tau a_k H_k).$$05 for unit-time cost scaling. The same source states that GRADE improves per-step gate cost and can beat CDF on raw step complexity, but that in the tested spectroscopy problems it suffers from larger Trotter and especially leakage errors; even after symmetry protection, leakage remained dominant, and the best full pipeline used CDF + SPRINT rather than GRADE + SPRINT (Casares et al., 29 Jun 2026).

6. Terminological boundaries: RTF versus quantum stochastic Trotter formulas

The label “randomized Trotter formula” should not be conflated with quantum stochastic Lie–Trotter formulas in Hudson–Parthasarathy theory. Two supplied sources are especially explicit on this point. "A homomorphism theorem and a Trotter product formula for quantum stochastic flows with unbounded coefficients" proves a deterministic strong Trotter product formula for quantum stochastic flows with unbounded coefficients, where the stochasticity is intrinsic quantum stochastic noise on Bosonic Fock space and the objects of interest are operator-valued cocycles solving QSDEs; it is not about random ordering of factors, Monte Carlo term selection, or probabilistic averaging over product formulas (Das et al., 2010). Likewise, J. Martin Lindsay and Kalyan B. Sinha’s "A quantum stochastic Lie-Trotter product formula" establishes a Lie–Trotter product formula for unitary quantum stochastic cocycles with constant bounded coefficients, with fixed dyadic partitions and deterministic concatenation; again, the stochastic element is the quantum noise process, not randomized scheduling in the algorithmic Hamiltonian-simulation sense (Lindsay et al., 2010).

This distinction is mathematically substantive. In quantum stochastic Trotter theory, one combines QS generators or structure matrices and studies convergence of cocycles on Fock space. In RTF for Hamiltonian simulation, one deliberately randomizes fragment orderings inside a digital product formula to cancel certain low-order commutator contributions in expectation. The former concerns stochastic dynamics of the modeled evolution; the latter concerns stochasticity injected into the simulation protocol itself. A plausible implication is that the two literatures share the general language of product formulas and generator splitting but address different notions of randomness, different error metrics, and different application domains.

Within the algorithmic meaning of RTF, the strongest message of the supplied material is therefore specific rather than universal: randomized ordering is most effective when embedded in a structured simulation stack with factorization, norm hierarchy, near-integrability, and compilation-aware implementation. The same source also records the main caveats. SPRINT may not help if no norm hierarchy exists; randomization helps mainly at low order; practical gains rely on estimated matrix elements of BCH commutators rather than worst-case guarantees; the benefit estimates from sampled random orderings were not fully converged for larger systems; spectral broadening may matter in stricter spectroscopy settings; and asymptotically qubitization still wins in gate count (Casares et al., 29 Jun 2026).