Comparing Simulation Methods

We now have a full toolbox: first-order Trotter, higher-order Suzuki formulas, LCU, qubitization, and QSP. None is universally best. This lesson lays the methods side by side so you can pick the right one for a given Hamiltonian, accuracy target, and hardware.

The cost dictionary

Let $H$ act on $n$ qubits, with evolution time $t$ , target error $\varepsilon$ , and (for the algebraic methods) coefficient 1-norm $\alpha = \sum_\ell |\alpha_\ell|$ from the Pauli decomposition.

| Method | Gate / query cost | Ancillas | Error in $\varepsilon$ | | --- | --- | --- | --- | | First-order Trotter | $O\!\left(\dfrac{(\,\Lambda t)^2}{\varepsilon}\right)$ | $0$ | $1/\varepsilon$ | | Order- $2k$ Suzuki | $O\!\left((\Lambda t)^{1+1/2k}\,(1/\varepsilon)^{1/2k}\right)$ | $0$ | $(1/\varepsilon)^{1/2k}$ | | LCU (Taylor series) | $O\!\left(\alpha t \,\dfrac{\log(\alpha t/\varepsilon)}{\log\log(\cdots)}\right)$ | $O(\log L)$ | $\log(1/\varepsilon)$ | | Qubitization / QSP | $O\!\left(\alpha t + \log(1/\varepsilon)\right)$ | $O(\log L)$ | $\log(1/\varepsilon)$ |

Here $\Lambda$ is a norm of the (commutators of the) Hamiltonian terms — for product formulas the relevant quantity is the sum of commutator norms, not $\alpha$ (Childs et al.). The headline contrasts are:

Accuracy. Product formulas pay polynomially in $1/\varepsilon$ ; the algebraic methods pay only logarithmically. For high-precision chemistry or many-body physics this is decisive.
Time. Qubitization and QSP achieve the additive form $O(\alpha t + \log 1/\varepsilon)$ , which is provably optimal. Trotter's $t$ -dependence is worse (e.g. $t^2$ at first order).
Ancillas. Product formulas use none. The algebraic methods need an index register and block-encoding workspace.

Where each method wins

First-order / low-order Trotter. Best when you have no spare qubits, the time $t$ is short, and modest accuracy suffices. The constant factors are tiny and the circuits are shallow — ideal for near-term noisy hardware where every ancilla and every gate costs fidelity.

Higher-order Suzuki. Best when you still want zero ancillas but need better accuracy, and $t$ is moderate. The 4th-order formula often hits the sweet spot: its $(1/\varepsilon)^{1/4}$ scaling beats first order long before the constant-factor overhead (a 4th-order step is several times longer than a 1st-order step) catches up.

LCU / qubitization / QSP. Best in the asymptotic, fault-tolerant regime: large $t$ , very small $\varepsilon$ , and a Hamiltonian with a favourable (small) 1-norm $\alpha$ . The logarithmic accuracy dependence and optimal $\alpha t$ scaling dominate once the problem is big enough to amortise the block-encoding overhead.

Commutator scaling: a subtle Trotter advantage

The table's worst-case Trotter bound is often pessimistic. Childs et al. proved that the true error scales with the sum of nested commutator norms of the Hamiltonian terms. When the terms nearly commute — as in geometrically local lattice Hamiltonians — these commutators are small, and product formulas can outperform their generic bounds by large factors. This is one reason Trotter remains competitive for physically structured problems despite its inferior $\varepsilon$ -scaling.

Choosing a method: a checklist

Do you have ancilla qubits to spare? No → product formula. Yes → consider LCU/qubitization.
How accurate must the answer be? Low/medium → product formula. Very high → algebraic method.
Is the evolution time large? Large $t$ favours the additive $O(\alpha t)$ scaling of qubitization/QSP over Trotter's super-linear $t$ .
Is the Hamiltonian geometrically local? Then exploit commutator scaling — a high-order product formula may beat its worst-case bound substantially.
What is the coefficient 1-norm $\alpha$ ? A large $\alpha$ inflates every algebraic-method cost; pre-processing to reduce $\alpha$ (e.g. by re-grouping Pauli terms) is often worthwhile.

There is no single winner — the art of Hamiltonian simulation is matching the method to the structure of the problem and the constraints of the machine.