Common Inference Errors in Quantum Research
Many weak research conclusions arise from a correct calculation being used to answer a broader question than it actually addresses. The most useful audit is therefore to look for a hidden change of task, model, limit, or denominator between the evidence and the conclusion.
A small numerical residual is not a physical error bar
Section titled “A small numerical residual is not a physical error bar”An eigensolver can solve a truncated matrix accurately while the basis, boundary, or model remains inadequate. Independent convergence studies must vary the approximations that affect the physical observable. Refining one parameter does not certify all the others.
A related error is fitting a thermodynamic transition from a small range of sizes without checking crossing drift or alternative fits. The localization assessment illustrates why accessible finite-scale behavior and asymptotic phase claims require separate language.
Conditional success is not unconditional performance
Section titled “Conditional success is not unconditional performance”A protocol can improve accepted-output fidelity by rejecting more trials. That can be a legitimate strategy, but report acceptance probability, delivered rate, and whether success is known before the output is used. Do not compare one protocol’s selected outcomes with another’s full ensemble.
The same issue appears when failed optimization runs, unstable devices, or inconvenient disorder samples are omitted. A reproducible selection rule belongs in the method and the resource ledger, not only in a supplementary explanation after the headline number.
A quantum resource is not automatically a quantum advantage
Section titled “A quantum resource is not automatically a quantum advantage”Entanglement, squeezing, or coherence can be present without improving the intended task. Specify the comparator and equalize relevant resources: total time, samples, energy, loss, input access, precision, training, and verification. A favorable asymptotic exponent can coexist with an unfavorable finite-instance cost.
Likewise, a faster optimizer is not a faster physical protocol if it produces pulses that exceed bandwidth or require more calibration. Resource estimation and metrology provide task-specific comparisons.
A signature is not a unique mechanism
Section titled “A signature is not a unique mechanism”Several models can reproduce a peak, plateau, oscillation, or transport anomaly. Ask which control or observable distinguishes the proposed mechanism from its strongest alternative. More precise measurement of a nondiscriminating observable may leave the central uncertainty unchanged.
A model trained on the same data used to score its success can hide this problem. Reserve meaningful held-out interventions, parameter regimes, or independent datasets. A random split of nearly identical samples may test interpolation while the scientific claim concerns generalization.
A theorem does not survive silently changed hypotheses
Section titled “A theorem does not survive silently changed hypotheses”Finite-dimensional results may require new work for unbounded operators. Independent-noise guarantees may not cover long-range correlations. A proof about average behavior may not bound rare failures. Read the quantifiers and conditions before using a theorem as an architectural guarantee.
Similarly, failure to prove a claim is not evidence that it is false, and failure of one algorithm is not proof that the physical task is impossible. The conjecture guide turns these distinctions into useful restricted projects.
A practical audit exercise
Section titled “A practical audit exercise”Choose one topic assessment and identify the strongest conclusion its evidence supports. Then deliberately write a broader, tempting version and name the extra assumption it requires. Design one calculation or control that tests that assumption.
The deliverable is a claim comparison, an explicit missing step, and a proposed test with a failure criterion. This exercise often produces a better research question than adding another unconstrained parameter sweep.