Higher-Order Trotter

First-order Trotter (chapter 131) is a beautiful idea — split the Hamiltonian into its local terms, exponentiate each one for a small time slice, interleave. Error O(t^2/n), gate count O(J t^2 / \epsilon). For many problems that is enough. But run a real quantum-chemistry calculation — say, the electronic structure of a 10-atom molecule — and you want chemical accuracy, which means \epsilon \sim 10^{-3} Hartree. Plug that into first-order Trotter's gate count and you are looking at 10^{10} to 10^{12} elementary gates for even modest time evolutions. No near-term hardware can run that circuit before noise swamps the signal. The question is whether a cleverer splitting can get the same precision for vastly fewer gates.

It can. The trick is symmetry. Order your exponentials so that the leading error term — the thing proportional to \Delta t^2 — cancels itself out. You are left with a smaller leading error proportional to \Delta t^3, which is hugely smaller when \Delta t is already small. Masuo Suzuki worked out the recipe in 1991 [1], building on an earlier observation by Gilbert Strang in the context of classical differential equations [2]. This chapter walks through the Strang formula (second-order), the Suzuki recursion (general order-2k), and the practical cost-precision trade-off that tells you which order to pick.

The picture: asymmetry wastes precision

Before any algebra, see why symmetry matters. Look again at the first-order Trotter step on two terms:

This is asymmetric: you run A first, then B. If you had run B first, then A, you would get

Both approximate e^{-i(A+B)\Delta t}, but with opposite-sign leading errors (the \tfrac{1}{2}[A, B]\Delta t^2 correction flips sign when you swap order, since [A, B] = -[B, A]). The two first-order approximations are equally wrong in opposite directions.

What if you averaged them? Or — better, since you need a unitary at the end — sandwiched one term between half-doses of the other?

That sandwich construction is exactly the Strang splitting.

The second-order (Strang) formula

Define

Claim: S_2(\Delta t) \;=\; e^{-i(A+B)\Delta t \;+\; O(\Delta t^3)}. The quadratic-in-\Delta t error term has vanished.

Why the quadratic error cancels

Expand each factor to second order in \Delta t:

Multiply out S_2 = (\text{left half-dose of A}) \cdot (\text{full dose of B}) \cdot (\text{right half-dose of A}) and collect by powers of \Delta t.

Order \Delta t^0: I. Fine.

Order \Delta t^1: -i(A/2 + B + A/2)\Delta t = -i(A + B)\Delta t. Exactly what you want — this is the linear term in e^{-i(A+B)\Delta t}.

Order \Delta t^2: here is the magic. The contributions are

• (-iA/2)^2 \cdot 2 — two ways of picking the half-dose-A twice (left-left impossible; so left\cdotright): gives -\tfrac{1}{4}A^2 \Delta t^2 \cdot 2 / 2 = -\tfrac{A^2}{4}\Delta t^2. Wait — include both self-contractions: the "left A/2" with itself (-A^2 \Delta t^2 / 8) plus "right A/2" with itself (-A^2 \Delta t^2 / 8) plus the cross-term "left A/2 times right A/2" (-\Delta t^2 / 4 \cdot A \cdot A). Sum: -A^2\Delta t^2/2.
• (-iB)^2/2! = -B^2 \Delta t^2 / 2.
• Cross terms A/2 \cdot B (from left-A with full-B) giving -\tfrac{1}{2}AB\Delta t^2; B \cdot A/2 (from full-B with right-A) giving -\tfrac{1}{2}BA\Delta t^2. Sum: -\tfrac{1}{2}(AB + BA)\Delta t^2.

Total order-\Delta t^2 contribution:

Why this is the good news: the (A+B)^2 is exactly the quadratic term in the true expansion e^{-i(A+B)\Delta t} = I - i(A+B)\Delta t - \tfrac{1}{2}(A+B)^2\Delta t^2 + \ldots. The leftover — anything that does not factor as (A+B)^2 — is the error. And here there is none at order \Delta t^2. The AB and BA cross terms combined symmetrically into the same (A+B)^2 expansion.

The leading error is third-order. Work out order \Delta t^3 with the same bookkeeping, and the commutators [A, [A, B]] and [B, [A, B]] enter. They do not cancel. But they come with \Delta t^3, which is much smaller than \Delta t^2 when \Delta t \ll 1. Over n Trotter steps, total error is n \cdot O(\Delta t^3) = O(t^3 / n^2).

Reading the improvement. For first-order, halving \epsilon requires doubling n. For second-order, halving \epsilon requires only \sqrt{2} \approx 1.41 times more steps — dramatically cheaper at fixed precision.

The Suzuki recursion — fourth order and beyond

Second-order kills the \Delta t^2 error. What kills \Delta t^3? A more elaborate symmetric combination. Suzuki 1991 [1] gave a general recursive construction: given an order-2k symmetric formula S_{2k}, build an order-(2k+2) formula by

with

Why this specific coefficient works: the combination 2 S_{2k}(p_k \Delta t) + S_{2k}((1-4p_k)\Delta t) + 2 S_{2k}(p_k \Delta t) uses an "evaluation at two scales" trick. The two scales p_k and (1-4p_k) are chosen so that the leading \Delta t^{2k+1} error terms cancel when summed with the right weights. The specific formula for p_k is what makes this cancellation work — it is the only choice that kills the next-order error while preserving the symmetric structure.

The fourth-order formula, unwrapped, is

with p_1 = 1 / (4 - 4^{1/3}) \approx 0.4145. The (1 - 4p_1) = 1 - 4/(4 - 4^{1/3}) \approx -0.6579 — note the negative time coefficient! Evolving "backwards" in time is how the formula cancels the leading error.

Suzuki's recursion continues indefinitely. Sixth order: apply the recursion to S_4 to get S_6 with error O(\Delta t^7) per step, at cost of 25 fundamental substeps. Eighth order: error O(\Delta t^9), cost 125 substeps. Every two extra orders multiplies the cost per step by 5.

The trade-off — which order to use

You now have a family of Trotter formulas indexed by order 2k. Each has per-step error O(\Delta t^{2k+1}) and per-step gate cost c_{2k} \cdot J where c_2 \approx 2, c_4 \approx 10, c_6 \approx 50, c_8 \approx 250, etc. To hit total error \epsilon over time t:

Total gate count:

You have two knobs: k (order) and the prefactors. Increasing k helps through the exponents but hurts through the prefactor 5^{k-1}. There is an optimum, which depends on (t, \epsilon).

Rule of thumb. For loose precision (\epsilon \geq 10^{-2}) and short time (t \leq 10), first- or second-order Trotter is fine. For chemistry-level precision (\epsilon \sim 10^{-3}–10^{-4}) and modest time, fourth-order is the sweet spot. For the tightest targets (\epsilon \sim 10^{-8}) or very long time, the optimal Trotter order creeps up to 6 or 8 — but at that point you should be using LCU or qubitisation instead (see chapter 133 and 134), which have fundamentally better scaling.

Examples

Common confusions

• "Higher order is always better." No. Each order doubles the per-step cost roughly, and the asymptotic scaling improvement diminishes. At some point — usually around 4th to 6th order — the constant factors overwhelm the scaling advantage, and raising the order increases the total gate count. For very loose precision targets, even second-order may be worse than first-order after accounting for the 2× per-step cost.

• "The optimal Trotter order is infinity." No. In the limit of large order, n_{2k} \to t (you need at least one step per unit of time), so gate count is roughly 5^{k-1} \cdot J \cdot t — exponentially growing in k. The optimum is finite and depends on (t, \epsilon).

• "Suzuki formulas are unique." No. Different arrangements with different coefficients can all achieve the same order of error. Yoshida 1990 [3] independently derived fourth-order splitting formulas (his motivation was symplectic integrators for classical Hamiltonian systems). The Suzuki family is one particular construction; Blanes-Casas and others have derived more compact alternatives. All produce order-2k error with different prefactors.

• "Negative time coefficients are unphysical." They look strange but cause no trouble. The fourth-order Suzuki formula contains the substep S_2((1 - 4 p_1) \Delta t) with a negative coefficient — you are running the evolution "backwards" for a short time. Each substep is still a valid unitary (e^{iH\tau} is unitary for any real \tau, positive or negative). The forward-and-backward dance is exactly what makes the error cancel.

• "Higher-order Trotter replaces the lower orders." For ultimate precision, yes. For NISQ-era hardware, often no — because higher-order Trotter has more gates per step, and each gate introduces noise. A noisy fourth-order circuit can have worse effective error than a cleaner second-order one despite the better scaling. The interaction of Trotter error and gate noise is a lively area of research.

• "All quantum simulation uses Trotter." Modern algorithms go beyond. Linear combinations of unitaries (Berry-Childs-Kothari 2015 [4]) and qubitisation (Low-Chuang 2017) achieve O(t + \log(1/\epsilon)) scaling — optimal in both t and \epsilon, beating any finite-order Trotter asymptotically. For near-term hardware, though, Trotter is often simpler and has smaller constants.

Going deeper

The full Suzuki recursion

Suzuki's 1991 paper [1] gave the general recursion in closed form. Starting from any symmetric order-2k formula S_{2k},

This produces a sequence of formulas of order 2, 4, 6, 8, \ldots with per-step costs c_2 = 2, c_4 = 10, c_6 = 50, c_8 = 250, \ldots — factor-5 growth per two orders. The asymptotic gate count to reach precision \epsilon is

The exponent 1 + 1/(2k) \to 1 as k \to \infty, approaching linear scaling in t. But the prefactor 5^{k-1} rules out letting k run to infinity — there is a crossover order at which it becomes cheaper to use a different algorithm entirely.

Yoshida 1990 — independent fourth-order construction

Haruo Yoshida independently derived fourth-order splitting formulas in 1990 [3], a year before Suzuki. Yoshida's motivation was classical mechanics: symplectic integrators for Hamiltonian ODE systems need to preserve both energy and phase-space volume, which requires splitting the Hamiltonian into kinetic and potential parts and applying a symmetric Strang-like formula. Yoshida's fourth-order formula has slightly different coefficients than Suzuki's but the same order of error. In modern Hamiltonian-simulation papers you will see "Suzuki-Yoshida product formulas" as a catch-all term for this whole family.

BCH and the algebraic structure

The deeper reason these formulas exist is algebraic. Consider the Lie algebra generated by A and B — the vector space of all nested commutators. A general product e^{a_1 A} e^{b_1 B} e^{a_2 A} e^{b_2 B} \ldots has a generator (via BCH) that is an element of this Lie algebra, with coefficients that are polynomials in the a_i and b_i. The Suzuki recursion amounts to a clever choice of the a_i and b_i such that all Lie-algebra-valued polynomials up to a given order match the target e^{(A+B)\Delta t}. The technical tool is the Hall basis of a free Lie algebra, and constructing optimal splitting formulas is equivalent to solving a set of polynomial equations in the Hall basis coefficients. Hatano-Suzuki 2005 and later Blanes-Casas-Murua work through this in detail.

Post-Trotter methods and the end of Trotter

Modern quantum Hamiltonian simulation has moved beyond Trotter for high-precision applications. Three milestones:

• Taylor series LCU (Berry-Childs-Cleve-Kothari-Somma 2014, 2015 [4]). Expand e^{-iHt} as a Taylor series \sum_k (-iHt)^k/k!, truncate, and implement the sum via linear combinations of unitaries. Gate count scales as O(\tau \log(\tau/\epsilon)) where \tau = t\|H\|_1. Asymptotic beat over any Trotter order.

• Qubitisation (Low-Chuang 2017). Block-encode H in an ancilla-assisted unitary and use Quantum Signal Processing to implement e^{-iHt} via polynomial approximation. Gate count O(t \|H\| + \log(1/\epsilon)) — optimal in both parameters.

• qDRIFT (Campbell 2019). Randomly sample terms from H. Expected gate count O(t^2 \|H\|_1^2/\epsilon) — matches first-order Trotter scaling but independent of the number of terms, which is a huge win for Hamiltonians with many small pieces.

For long-time and high-precision simulations (quantum chemistry, lattice field theory), qubitisation is the algorithm of choice in modern fault-tolerant compilers. For short-time NISQ applications, Trotter — often fourth-order — still dominates because the constant factors are small and the ancilla overhead is zero.

Application to quantum chemistry

Quantum-chemistry simulation has been a proving ground for Trotter optimisation. The typical workflow: (1) compute one- and two-electron integrals classically; (2) apply Jordan-Wigner or Bravyi-Kitaev to get a Pauli-sum Hamiltonian with O(N^4) terms where N is the number of spin-orbitals; (3) group commuting Paulis together to reduce effective term count; (4) apply fourth-order Trotter with carefully optimised term ordering; (5) measure the energy via quantum phase estimation.

Babbush et al. 2018 estimated that simulating the active site of nitrogenase (FeMoCo, implicated in industrial nitrogen fixation) needs around 10^{12} T-gates with fourth-order Trotter. With qubitisation, that drops to 10^{10}. Both are beyond current hardware — but within the realm where engineering progress matters, not complexity-class barriers. The target is often quoted as "100 logical qubits, 10^9 T-gates" for the first molecule where quantum computing provides useful chemistry answers that classical methods cannot. That crossover depends on Trotter-and-beyond improvements as much as on hardware.

Indian research on quantum simulation

IISc Bangalore's quantum-chemistry group has published comparative studies of Trotter orders for small molecules on NISQ hardware. TIFR theoretical condensed-matter has analysed Trotter errors for Hubbard-model simulations relevant to high-T_c physics. India's National Quantum Mission funds a dedicated track on quantum algorithms, including Hamiltonian-simulation benchmarking; published results from Indian collaborators have appeared in multiple arXiv preprints on fourth-order Trotter compilations in 2024–2025. As fault-tolerant hardware matures over the next decade, Indian groups are poised to contribute both to theory (tighter error bounds, new recursions) and to application (pharmacological molecules, materials science).

Where this leads next

• Trotter-Suzuki Decomposition — chapter 131, the first-order foundation.
• Linear Combinations of Unitaries — chapter 133, the first post-Trotter algorithm with better asymptotic scaling.
• Qubitisation and Block Encoding — chapter 134, the optimal-scaling method for modern Hamiltonian simulation.
• Local Hamiltonians — chapter 130, the structural assumption every Trotter method exploits.
• Hamiltonian Simulation — the overall landscape of simulation algorithms.

References

1. Masuo Suzuki, General theory of fractal path integrals with applications to many-body theories and statistical physics, J. Math. Phys. 32, 400 (1991) — DOI:10.1063/1.529425.
2. Wikipedia, Lie product formula and Trotter–Suzuki decomposition.
3. Haruo Yoshida, Construction of higher order symplectic integrators, Phys. Lett. A 150, 262 (1990) — DOI:10.1016/0375-9601(90)90092-3.
4. Berry, Childs, Kothari, Hamiltonian simulation with nearly optimal dependence on all parameters (2015) — arXiv:1501.01715.
5. Seth Lloyd, Universal Quantum Simulators, Science 273, 1073 (1996) — DOI:10.1126/science.273.5278.1073.
6. John Preskill, Lecture Notes on Quantum Computation, Chapter 7 — theory.caltech.edu/~preskill/ph229.