When two quantum models fit, which control can we trust?
Two different quantum models can explain the same measurements. A small one-qubit study asks what we can still promise about a control action.
You have a set of measurements, a model that fits them and a control action that works nicely in simulation. It is a reassuring chain of reasoning until another model turns out to fit the same data. Under that model, the same control performs differently. The measurements have not changed; what they allow you to promise has.
That is the question I explored in a September 2026 technical study. Instead of choosing one model too early, I keep the models the experiment cannot yet distinguish. Then I ask what performance is possible across all of them, and which extra measurement would help with the task.
The study offers a defined research question and a small reproducible example, not a general certification method or a hardware result. The equations below make the example precise. The main idea is simpler: more measurements of the same kind do not always settle the doubt that matters.
Keep every explanation the data still allows
System identification depends on what the experiment can actually prepare, control and observe. Two models can be operationally indistinguishable under one experiment library even if their Hamiltonians differ. Burgarth and Yuasa’s identification framework makes the attainable experiment setting central to the question of what can be known.
I therefore start with a declared physical model domain and retain every model compatible with the data and assumptions. This set may have disconnected branches. A confidence interval around one fitted parameter vector can miss another branch entirely. Repeating the same non-discriminating measurement reduces sampling noise without resolving that structural ambiguity.
It helps to distinguish two kinds of uncertainty here:
- Exact observational equivalence means two models predict the same outcome probabilities for every protocol in a specified identification library. That is an equivalence relation.
- A finite-data consistency set contains models whose predictions satisfy stated statistical constraints. Being close to measured frequencies does not itself define an equivalence relation.
The set need not collapse to one model. Classical data-informativity work already shows that useful analysis or control can sometimes proceed without unique identification. Keeping all compatible models is therefore a starting point from prior research, not a novelty claim of this study.
For a control (u), task loss ℓ and additional perturbation budget Δ(r), the certificate I want has the form:
B_D(u, r) ≥ sup{ℓ(u, m, d) : m ∈ M(D), d ∈ Δ(r)}. Issue a certificate only if the data-compatible set M(D) is nonempty and B_D(u, r) ≤ ε, where ε is the declared loss tolerance.
The physical assumptions, allowed controls, target, measurement frame, loss and units of the perturbation radius must all be specified. A sampled parameter sweep can find failures; it cannot certify a continuum of models that it has not searched.
A single qubit makes the problem visible
Take a deliberately restricted, calibrated qubit model:
H = (ω + u)Z/2
The unknown drift is in two physical branches around ω = ±1. Identification prepares the X-positive state, evolves freely for one unit of time and reads out X. Its positive-outcome probability is (1 + cos ω)/2. Because cosine is even, ω = +1 and ω = −1 give exactly the same X statistics. More shots of that experiment cannot tell them apart.
Now fix the control duration at T = π/2, allow only commuting Z control u ∈ [−2, 2], and ask for the calibrated target gate Rz(π/2). The process fidelity is Fₚ(u,ω) = cos²[(ω + u − 1)π/4]. Process fidelity here is the gate-overlap measure for the stated unitary target; it is not average gate fidelity.
For the compatible pair ω = ±1, Fₚ(u,+1) + Fₚ(u,−1) = 1 for every allowed control. So the smaller fidelity cannot exceed 0.50. A controller fitted only to the positive branch chooses u = 0: its fidelity is 1 on that branch and 0 on the other. The blue curve in the analytic figure below is what survives when both models are retained.
The 0.50 limit belongs to this particular example: its fixed duration, calibrated frame and allowed controls. Change those choices and the result can change. It tells us what the ambiguity costs in this task, rather than setting a limit for quantum control in general.
A calibrated Y readout produces (1 + sin ω)/2, which separates the signs. At ω = ±1, the X and Y probabilities are about 0.770 in both cases for X, but about 0.921 and 0.079 for Y. The useful feature of Y is that it removes a global alias that matters to the task. Local parameter precision alone can miss that distinction.
The same number of shots can teach different things
I also compared two predeclared synthetic measurement designs, each using 8,192 reset shots and a separately constructed 95% model-set coverage statement. Design A takes two X batches. Design B takes one X and one Y batch. Both use the same physical prior and true simulated value ω = 1. The study fixes the NumPy seed at 20260923 and records the batch counts: X gives 3,091/4,096 and 3,146/4,096 successes; Y gives 3,790/4,096.
| Evaluation | Process-fidelity floor |
|---|---|
| X-only: positive branch retained, negative branch wrongly discarded | 0.999235 apparent |
| Same controller, both X-compatible branches retained | 0.000019 actual worst case |
| X-only: a robust controller over both branches | 0.488360 |
| X plus Y: controller over the retained single interval | 0.998215 |
The X-plus-Y compatible interval is approximately [0.9817915, 1.0893959]. It is wider than one X-only branch. The improvement comes from removing the competing sign, not from a narrower local interval. In this analytic model, the midpoint control is approximately u = −0.035594 and the worst-case process fidelity is cos²[(b − a)π/8] ≈ 0.998215. For a separately defined constant drift |d| ≤ r, the same stated model retains Fₚ ≥ 0.99 through r ≈ 0.073735 normalized angular-frequency units.
These are rounded evaluations of closed-form extrema on the declared model. The implementation used ordinary floating-point arithmetic, not an outward-rounded interval proof engine. The comparison is one synthetic realization, not evidence that X-plus-Y or a certificate-oriented adaptive policy wins generally.
Confidence in the model set is not yet control performance
A finite-shot model set needs a statistical contract. For fixed binary protocols with independent reset shots, the study uses Hoeffding radii and a union bound to retain the true stationary model with probability at least 1 − α, conditional on the declared physical domain and sampling assumptions. An adaptive experiment sequence would need a valid sequential error budget or confidence sequence; repeatedly reusing an ordinary fixed-time 95% interval after optional stopping would not suffice.
The performance calculation then needs a deterministic upper bound over the entire retained set. If the set contains the true model and the bound is sound, an issued certificate covers that model. Under the stated assumptions, the probability of issuing an invalid certificate is at most the model-set noncoverage probability. This is not automatically the error rate among certificates that happen to be issued.
If no model remains compatible, something in the data and assumptions does not fit. That is a reason to investigate. Drift, uncertain calibration and preparation or readout errors need a place in the model or a separate test. Repeating a measurement cannot repair a missing physical explanation.
Choose the measurement that helps the task
The research direction is to score a feasible new experiment by its effect on a certified task loss or robustness radius, accounting for shots, interrogation time and computation. That is a different objective from minimizing a fitted Hamiltonian’s covariance. Active Hamiltonian learning already selects informative queries, so an adaptive measurement rule by itself would not be new. A contribution would need a defensible task-based objective, a sound verifier and a measured gain under matched resources.
A related two-parameter sensing example makes the same issue visible. If a target phase a and nuisance phase b only enter as a + b, one sensitive readout cannot distinguish them. A second calibrated protocol that responds to −a + b can separate them under a stated principal phase domain and stable nuisance assumptions. Nuisance-aware quantum estimation is established literature; this reduced calculation specifies what an eventual application experiment would have to resolve. It is not a simulation or result from an atom interferometer.
What I would try next
The next technical step is a verifier for low-dimensional, possibly disconnected compatible-model sets. It must keep every unresolved branch, produce an upper loss bound over a continuum and expose how conservative that bound is. A controller optimiser may propose candidates; an independent verification step must certify them. Published differential sensitivity bounds provide a comparison baseline, but a local sensitivity result does not automatically cover a remote compatible branch.
Only after that can fixed, estimation-oriented and certificate-oriented experiment designs be compared fairly on the same physical prior, control class, shot budget, interrogation time and computation. Failures to certify and model-mismatch cases belong in the results. A later optical-lattice application would require a specified apparatus model, available measurements and actual experimental access.
My earlier QANTIS research supplies experience with quantum experiments, uncertainty and classical baselines. QFlow Studio could keep the eventual experiment record reproducible. Neither is the scientific contribution claimed here. The present output is a research formulation, analytic examples and a bounded synthetic calculation. A general method, an adaptive-policy benchmark, an open-system extension and laboratory validation remain open work.
Sources
- Burgarth and Yuasa, Quantum System Identification (2012) journals.aps.org
- van Waarde et al., Data informativity (2020) arxiv.org
- O'Neil et al., Differential sensitivity bounds for dynamic quantum control (2024) arxiv.org
- Dutt et al., Active learning of quantum system Hamiltonians (2023) journals.aps.org
- Suzuki, Yang and Hayashi, Quantum state estimation with nuisance parameters (2020) arxiv.org