Rotating local axes trims the memory cost of quantum simulations

When the representation becomes the bottleneck

Simulating quantum dynamics does not always mean tracking the quantum state. In the Heisenberg picture, a calculation instead follows how an observable—the mathematical operator associated with a measurement—evolves through a circuit.

One approach, called Pauli propagation, represents that observable as a weighted list of Pauli strings. Each string places either an identity operator or one of three Pauli operators, associated with the x, y and z axes, on every qubit. Quantum gates can turn one string into several, making the list grow. To stay within memory limits, simulations discard smaller coefficients, introducing truncation error.

But a long list need not indicate irreducible complexity. An operator pointing between two coordinate axes may require several terms in one basis and fewer in another. In their arXiv preprint, submitted on November 8, 2026, Pineda-Jiménez and Collura ask how much computational cost comes from this choice of axes. The supplied record provides no journal publication or evidence of peer review.

What rotating the axes can change

The researchers separate the representation's complexity into two contributions using the squared coefficients of the Pauli strings as a probability distribution.

The first concerns occupation: which qubits carry a non-identity operator in each string. Rotating the three Pauli axes on individual qubits cannot change the total probability associated with any occupation pattern. Consequently, the distributions of operator mass—the number of occupied sites—and operator length—the position of the rightmost occupied site—also remain unchanged.

The second contribution concerns flavour: whether an occupied site carries an x, y or z component. Local rotations can concentrate those components into fewer strings.

The authors prove that occupation statistics impose a lower bound on complexity in every local Pauli frame. This is not a universal limit on all possible representations: more general changes of basis, including ones that mix the identity with other components, can go below it. Nor is all flavour complexity necessarily removable by rotating individual sites, because correlations between sites also matter.

A basis that follows the calculation

Their method, frame-adaptive Pauli propagation, or FAPP, updates the axes at the two qubits touched by each gate. It computes a small local polarisation tensor describing how the operator's weight is distributed among the three flavours, then aligns the axes with that tensor's principal directions.

The frame change is merged with the gate operation, avoiding a separate expansion of the whole stored list. The rotation itself is exact for the stored operator; keeping only the largest coefficients remains an approximation.

This rule optimises the flavour distribution at each site individually, not necessarily the full operator. Comparisons with numerically optimised local frames on a ten-qubit chain suggest it captures much of the available improvement, but it is not always optimal. The authors also find that local polarisation helps indicate when rotations are likely to help, without providing an exact universal predictor.

Savings across several tests

The researchers benchmarked FAPP against ordinary Pauli propagation using the same string budgets and gate sequences. For Ising-model dynamics, both methods were compared with the same discretised, or Trotter, circuit, isolating truncation error rather than error from approximating continuous evolution.

In ten-qubit chain examples, FAPP reduced the squared operator error by factors of 28 and 23 at selected times and a budget of 16,384 strings. Reaching a specified squared error generally required two to four times fewer strings. A separate chain-size study found roughly threefold savings for an initially local operator, persisting across the sizes studied. The implementation took 1.2–1.5 times longer at a fixed string budget.

On an 11-by-11 Ising lattice, where boundary effects were negligible over the relevant times, savings reached about sevenfold at fixed certified accuracy. In random Clifford circuits supplemented with T gates, the reported memory reduction was twofold to eightfold. Clifford gates normally permute Pauli strings; T gates can split them, creating an opportunity for adaptive frames.

Where the gains stop

FAPP did not improve every task. Propagating total magnetisation as one operator required 10–20% more strings than standard propagation in the true-error comparison. Its constituent terms favoured different axes. Evolving those terms separately recovered savings, although certifying their combined error requires additional care.

Better accuracy for the whole operator also did not guarantee comparably better individual expectation values. For large lattices without exact references, agreement between high-budget calculations was a consistency check, not an accuracy certificate.

Finally, required memory still grew exponentially with evolution time. Even substantial compression therefore bought only a modest extension of reachable time in two dimensions. The result is a practical improvement to a representation, not a removal of quantum simulation's underlying difficulty.