Analog quantum simulators have been able to show dynamics that look right. They have not, as a rule, been able to say how wrong they might be. A collaboration centered at the Technical University of Munich and the University of Innsbruck / IQOQI has now attached quantitative error bars to a trapped-ion simulator with up to 51 ions. The paper is Tristan Kraft, Manoj K. Joshi and colleagues, “Bounded-Error Quantum Simulation via Hamiltonian and Lindbladian Learning,” Physical Review X 16, 031037 (2026), DOI 10.1103/s96t-n8tx. APS published it on 13 August 2026. IQOQI’s news note is 18 August. The PRX PDF was fetched for this article.
The object is a one-dimensional chain of 40Ca+ ions running a long-range Ising interaction. The method learns the coherent Hamiltonian and the dissipative Lindbladian from data, then pushes those uncertainties into observables. On 10 ions, classical simulation can still check the answer. On 51 ions, the same protocol still returns a bound. That is a calibration result. It is not a logical qubit, and it is not a product.
What happened
Kraft and Barbara Kraus are at TUM’s School of Natural Sciences and the Munich Center for Quantum Science and Technology. Peter Zoller is at Innsbruck’s Institute for Theoretical Physics and IQOQI. The hardware run is Joshi, Florian Kranzl, Johannes Franke, Rainer Blatt and Christian F. Roos at IQOQI and Innsbruck’s Institute for Experimental Physics. Grenoble (CNRS/LPMMC), Jülich PGI-8, SISSA, INO-CNR Florence, Copenhagen and Quobly are on the author list.
IQOQI’s 18 August note states the claim in one sentence: the simulator should return a result with error margins, not a single number.
The PRX experiment section is more specific. Linear strings of up to 51 40Ca ions sit in a radio-frequency trap. Qubits are the S1/2, m = +1/2 and D5/2, m = +5/2 Zeeman states, linked by a 729 nm quadrupole transition. A bichromatic beam, waist 105 μm (10 ions) or 310 μm (51 ions), couples all ions off-resonantly to transverse motional modes and engineers a long-range spin-spin Hamiltonian. Crystal lengths are 71 μm (10 ions) and 263 μm (51 ions). Fitted couplings: J0 = 576 rad/s, α = 1.19 (10 ions) and J0 = 185 rad/s, α = 0.97 (51 ions), both at detuning 2π × 25 kHz. Trap frequencies sit in the table below.
Two learning routes. On 10 ions they use an integral method. Two product states evolve at tr ∈ {0, 0.5, 1, 1.5} ms. Almost all up-to-three-qubit Pauli correlations are estimated from randomized measurements. Budget: 3.2 × 10^5 shots. A long-range Hamiltonian plus a dephasing Lindbladian beats truncated nearest-neighbor models. Collective dephasing dominates. Learned X-X and Y-Y couplings are nearly equal, so total magnetization is approximately conserved. A simulation that flips measurement outcomes “0” → “1” with probability 2% reproduces the experimental residual-norm scaling; an ideal simulation recovers shot-noise scaling.
On 51 ions they switch to a differential (derivative) method. The Hamiltonian ansatz allows up to two-qubit Paulis; decoherence is again a dephasing Lindbladian. Fig. 7 caption states the parameter count: 3N + 9N(N − 1)/2 + N² = 14,229 parameters. Acquisition: NU = 200 random settings, NM = 200 shots, Nt = 11 equally spaced times between 0 and 1 ms. Sample complexity of the underlying protocol scales as O(3^{w1+w2} log(N) ε^{-2}) — logarithmic in system size if the weights stay fixed.
Short-time bounds on the 10-ion data use η = 0.05 (95% confidence) and a Gaussian error δ ≈ 13 rad/s. A Hanson-Wright envelope contains Ns = 100 simulated trajectories. For 51 ions, Fig. 8 shows predicted error bounds on quench dynamics; the learned model is checked against a tensor-network representation in TeNPy. That classical check is still available for this 1D chain. The paper’s second, harder claim is that short-time bounds can also be written from measurements alone, without a full master-equation simulation — the step you would need if the system outran classical methods. Demonstrating that the 51-ion learned model sits inside its own bounds is not the same as having left the classically simulable regime.
Chart-ready numbers
| Quantity | 10 ions | 51 ions |
|---|---|---|
| Ion | 40Ca+ | 40Ca+ |
| Crystal length | 71 μm | 263 μm |
| ωz | 2π × 0.217 MHz | 2π × 0.115 MHz |
| Beam waist wz | 105 μm | 310 μm |
| J0 | 576 rad/s | 185 rad/s |
| α | 1.19 | 0.97 |
| Detuning Δ | 2π × 25 kHz | 2π × 25 kHz |
| Time grid | 0, 0.5, 1, 1.5 ms | 11 points, 0–1 ms |
| Shots / settings | 3.2 × 10^5 shots | 200 settings × 200 shots × 11 times |
| Parameters learned | long-range XY + fields + dephasing | 14,229 |
| Classical check | full master equation | TeNPy tensor network |
| Confidence example | η = 0.05 (95%) | protocol as in Fig. 8 |
IQOQI’s note, quoting Zoller, already names the sequel: two-dimensional simulators, where classical verification gets harder faster, and a contest scored on verifiable margin of error, not only speed.
Why it matters
Analog simulation is the branch of quantum hardware that already has tens of particles and a native many-body Hamiltonian. What it has lacked is a way to treat the device as a measuring instrument. Spectroscopy quotes a line with an uncertainty. A quantum quench has usually quoted a curve. Bounded-error quantum simulation is an attempt to make the curve look like spectroscopy: a prediction plus a data-derived interval.
The 10-ion half of the paper is the credibility check. Dynamics at N = 10 can still be integrated classically, so a wrong error bar would show. The 51-ion half is the scaling check: 14,229 parameters, logarithmic sample-complexity in N, a 1-millisecond window. Together they argue that “we learned the generator from the first part of the quench and predicted the rest, with a bound” is a workflow, not a slogan.
If analog simulators are going to be used as instruments, they need error bars that do not secretly assume the target Hamiltonian. Learning the actual Hamiltonian and Lindbladian is how you stop reporting the textbook model by mistake. That is closer to metrology than to a quantum-advantage press release. It is also a different stack from Innsbruck’s measurement-free Grover run: analog many-body spins here, encoded logical qubits there.
Limits
Read the geometry. The chain is 1D. Interactions fall as a power law, not as a 2D lattice. TeNPy still represents the 51-ion dynamics; the authors use that as a check. They have not shown a regime where the bound is the only certificate because the classical simulation has died.
The 14,229-parameter ansatz is large. Regularization and a physically motivated dephasing model keep it identifiable at the stated shot budget. Change the noise, and the bound moves. SPAM errors were not folded into the 51-ion M matrix; the authors say statistical uncertainty in the derivatives dominated at this budget and that SPAM can be added later following the cited protocol.
On 10 ions the measured nearest-neighbor couplings ran slightly larger than the theoretical J0 ≈ 288 rad/s; the authors blame a few-degree tilt of the trap axes. That residual is why you learn the Hamiltonian instead of trusting the design document.
IQOQI’s “up to 51 qubits” lede is the paper’s 51 ions: analog spins, not logical qubits.
What to watch next
The authors already named the sequel: 2D analog simulators, where classical tensors strain sooner. A published bound on a 2D quench, without a TeNPy crutch, would be the first time this method is doing work that a laptop cannot. Second: a digital-circuit version. The abstract says the techniques “directly extend to digital quantum simulation”; that extension is not in the figures. Third: whether the short-time, measurement-only bound — the one that does not integrate a master equation — becomes the default certificate on the Innsbruck 51-ion machine, or remains a theoretical appendix to a classically checked curve.
Until then the sourced claim is: Physical Review X 16, 031037 (13 Aug 2026), 10- and 51-ion 40Ca+ analog Ising simulators, 14,229 parameters at N = 51, error bars from Hamiltonian and Lindbladian learning, not a fault-tolerant processor.
