When a quantum processor produces a result that seems impossible to reproduce on a laptop, the natural question is: has anyone actually tried? Increasingly, the answer comes from a family of classical techniques that skip the brute-force approach to simulation entirely. Pauli propagation is among the most influential, and it has reshaped how researchers judge claims of quantum utility.
The Core Idea: Evolve the Question, Not the State
Simulating a quantum circuit the obvious way means tracking the full quantum state, whose memory cost doubles with every added qubit. At 127 qubits, that is hopeless.
Pauli-based methods take a different route. They estimate specific measurement results without storing the full quantum state, evolving the measured observable backward through the circuit (the Heisenberg picture) and representing it as a weighted sum of Pauli strings. In other words, rather than asking “what is the whole state at the end?”, they ask “what does my one measurement of interest look like, traced back to the start?”
How the Method Works
The procedure has a clear structure:
- Start with an observable, such as a Pauli-Z on a single qubit.
- Propagate it backward through each gate. Clifford gates like CNOT map Pauli words to Pauli words, with possible sign or phase changes, while other gates can split a term into several Pauli strings.
- Truncate the expansion at each step to keep it manageable.
- Evaluate the result against the initial state.
Truncation is where the craft lies. Coefficient-based truncation drops terms whose coefficients fall below a threshold, so lower thresholds keep more terms and cost more, while weight-based truncation drops strings with more non-identity operators than a set maximum. The result is an approximation whose quality depends heavily on the choices made.
Why It Matters
The appeal is easy to see. The method avoids the 2^n memory cost of a full statevector, and it suits cases where only a few observables are needed. It can also be efficient for structured, locally scrambling circuits, and circuit depth is not a reliable measure of difficulty, since some deeper circuits are easier to approximate than shallow ones. That last point surprises many people: a deep circuit is not automatically a hard one.
A Case Study: The 127-Qubit Experiment
The most discussed test case is IBM’s 2023 Nature experiment. It used a kicked Ising model on the 127-qubit Eagle processor. At the time it was presented as beyond easy classical reach. Follow-up classical work then reproduced the results with Sparse Pauli Dynamics, tensor networks and belief propagation between 2023 and 2025.
For practitioners who want to experiment with these techniques directly, hosted pauli path simulators offer a way to run such calculations without building custom infrastructure. BlueQubit’s implementation, for example, lets users submit circuits, define observables, set thresholds and retrieve estimates, with a Python SDK for programmatic runs. Its benchmark estimated ⟨Z₆₂⟩ after 20 Trotter steps across rotation angles, and its results matched the converged classical reference within about 0.01 for the angles shown. On runtime, the article reports seconds to minutes per angle, versus hours for matrix product states in the original comparison, with full statevector simulation impractical. Those figures are described as representative, not guaranteed.
Scale has been pushed further elsewhere. A related method called ORQA, run on Fugaku, handled the same 127-qubit model while tracking up to one trillion Pauli strings.
Where the Method Struggles
No approximation is free, and honest use means knowing the limits. The article lists several:
- Growth in significant Pauli terms can make memory and runtime climb quickly, and truncation can remove high-weight terms that matter for the final result.
- Lower thresholds do not always improve accuracy, because convergence can be non-monotonic.
- Theoretical efficiency guarantees apply only to specific circuit families, noise models or observables.
- Stable estimates across thresholds are useful evidence but not a rigorous error bound, and the method can’t efficiently simulate every quantum circuit.
The cautionary note is concrete: in a follow-up study, some angles that looked like matches with IBM’s published results were far from converged when tested across multiple thresholds. A single run can look convincing and still mislead.
A Crowded, Productive Toolbox
Pauli propagation isn’t a single product but an ecosystem. Libraries include PauliPropagation.jl in Julia, Qiskit’s pauli-prop, NVIDIA’s cuQuantum Pauli propagation for GPUs, and a PennyLane tutorial. Related methods such as Sparse Pauli Dynamics, LOWESA, OBPPP, ORQA and Clifford Perturbation Theory each target different circuit types and assumptions, alongside matrix product states and tensor networks. Choosing among them is part of the work.
What Researchers Actually Use It For
Three use cases stand out:
- Benchmarking hardware. Classical estimates serve as reference values for comparing quantum processor results.
- Algorithm development. Researchers can explore parameters and check whether classical approximation suffices before using hardware.
- Mapping simulation limits. Studying which circuit properties make classical simulation tractable helps design harder quantum benchmarks.
Best Practices for Using Truncated Simulation
Because truncation introduces approximation error, discipline matters. The article recommends repeating runs across thresholds and comparing against independent references or hardware data. In practice, that means sweeping thresholds, watching for convergence, cross-checking with a different method, and never trusting a single configuration.
The Bigger Lesson
The takeaway is not that quantum computers are unimpressive. It is that claims need testing against the strongest classical competition. The IBM 127-qubit case shows that large, complex quantum computations don’t automatically establish advantage over every classical method. And rigorous, multi-threshold and cross-method comparisons remain necessary.
For the field, that is healthy. Every credible classical attack raises the bar for the next quantum experiment, and every careful benchmark makes eventual real advantage easier to believe.




