Inference, Certification, and Evidence

An information claim is the end of an inference chain, not a synonym for a measured correlator or a computed entropy. The chain begins with a calibrated instrument or simulator, passes through estimators and model assumptions, propagates finite-sample and continuum errors, challenges the result with alternative states and independent methods, and ends with the strongest statement actually identified by the data. This chapter supplies that chain for entropies, randomized measurements, replica protocols, Gaussian reconstruction, detector channels, modular response, witnesses, and reproducible evidence.

Helpful background. Information-Measure Domain and Comparison Atlas matches formulas to domains. Detector and instrument validation supplies calibration and causal measurement models. Complete Lattice Error Budgets develops general numerical inference machinery, while From Measured Intensity to Many-Body Claim develops platform-specific evidence. This chapter focuses on what those results license as QFT information statements.

The claim is bounded by the weakest identified link

Represent the inference as

$$D_{\rm raw} \xrightarrow{\ C\ } D_{\rm calibrated} \xrightarrow{\ \widehat\theta,M\ } \widehat\theta \xrightarrow{\ U,\Lambda\to\infty\ } I \xrightarrow{\ A,H_0\ } \text{bounded claim}.$$

$C$ is calibration and data reduction; $\widehat\theta$ is an estimator under model $M$; $U$ is its joint uncertainty; $\Lambda$ denotes volume, cutoff, truncation, and continuum limits; $A$ is the set of alternatives tested; and $H_0$ is a null model. Each arrow has a validation test. A precise final number cannot repair an unidentifiable model or an untested continuum extrapolation.

Certification is claim shaped. Randomized Rényi experiments demonstrate one route from controlled measurements to an entropy functional Brydges et al. 2019, pp. 260–263, while classical shadows show how observable-dependent sample bounds replace full tomography Huang, Kueng, and Preskill 2020, Theorems 1–2. A negative entanglement witness certifies entanglement for states within its measurement model but does not reconstruct entropy. A covariance matrix fixes a Gaussian state but does not exclude a non-Gaussian state with identical second moments. An OTOC indicates operator influence under its contour and normalization but does not certify recovery without a decoder test. Durable data and metadata should also follow reusable provenance principles Wilkinson et al. 2016, Principles F1–R1.

Choose a route

Goal	Route	Stop when you can…
Build the inference chain	From Field Data to Information Claims	name every calibration, estimator, assumption, limit, alternative, and claim ceiling
Estimate entropy	Estimating Entropy and Rényi Quantities → Randomized Measurements and Classical Shadows	distinguish direct observables, property estimators, and model-based reconstruction
Use multiple copies	entropy estimation → Replica, Swap, and Multicopy Protocols	verify copy identity, region matching, and permutation calibration
Reconstruct Gaussian data	Gaussian State Reconstruction from Correlators	certify physical covariance and state the Gaussian assumption
Learn an instrument	Detector and Channel Parameter Tomography	establish identifiability rather than reporting a narrow ridge as precision
Avoid full tomography	Estimating Modular Response and Relative Entropy	bound neglected modular terms on the stated perturbative domain
Certify a resource	Witnesses for Entanglement, Scrambling, and Recovery	state the exact conclusion and adversarial states still compatible with the data
Propagate all errors	Continuum Bias and Correlated Uncertainty → Cross-Method Checks and Adversarial Null Tests	separate shared bias from method-specific failure
Maintain and reproduce evidence	Claims, Replication, Correction, and Retraction → Reproducible Reporting Standard	preserve provenance while updating every dependent conclusion

From observations to licensed statements

Entropy, tomography, witness, and recovery methods enter at the estimator stage, but all share calibration, uncertainty, continuum, and alternative-model tests. The final statement is no stronger than the least validated arrow. The diagram is schematic and not to scale.

Compare the methods

Assumptions, failure scales, and reproducibility requirements for representative inference methods

Method	Target	Measurement design	Principal assumption	Dominant error	Continuum ceiling	Held-out test	Reproducibility minimum
Covariance reconstruction	Gaussian state and entropy	Calibrated quadratures or two-point functions	Gaussianity and mode completeness	Physicality projection and covariance noise	Fixed smeared modes or controlled mode growth	Higher connected moments	Raw moments, calibration, and covariance bootstrap
Randomized measurements	Selected observables, purity, or shadows	Known random ensemble and readout map	Invertible ensemble channel and bounded shadow norm	Inversion variance and readout bias	Declared energy or mode truncation	Unseen observables and ensemble condition number	Random seeds or unitaries, counts, and inversion code
Swap or cyclic protocol	nth power trace of a regional state	n matched copies and a regional permutation	Copy identity and permutation fidelity	Drift, mismatch, and nonlinear logarithm bias	Matched physical region and copy regulator	Calibration product states and copy swaps	Per-copy records and permutation calibration
Detector or channel tomography	Channel parameters or response	Informationally complete calibrated probes	Identifiable channel family	Parameter degeneracy and model discrepancy	Energy-constrained input domain	Held-out preparations and residual structure	Probe states, transfer function, and likelihood
Modular response	Perturbative relative entropy	Stress response and a known modular kernel	Correct region and state expansion	Truncation of higher orders	Fixed region geometry and regulator matching	Direct regulated calculation in a solvable case	Kernel, correlators, and perturbative residual
Finite witness	Entanglement or another task-specific property	Chosen observable family	Witness hypotheses and trusted calibration	False negatives and adversarial moment matching	Only the tested observable class	Known positive and negative states	Witness coefficients, uncertainty, and decision rule

Failure channels and controls

Different estimators can share the same calibration or normalization bias, so numerical agreement is not automatically independent replication. Adversarial nulls, held-out observables, regulator variation, and genuinely independent implementations set the claim ceiling and trigger correction when needed. The diagram is schematic.

The twelve pages

1. From Field Data to Information Claims defines the calibrated inference chain.
2. Entropy and Rényi Estimation Protocols compares direct, indirect, and model-based estimators.
3. Randomized Measurements and Classical Shadows controls ensemble inversion and shadow variance.
4. Replica, Swap, and Multi-Copy Measurement Protocols measures permutation invariants across matched copies.
5. Correlation-Matrix and Gaussian-State Reconstruction certifies covariance physicality and propagates error.
6. Detector, Channel, and Local-Parameter Tomography distinguishes identifiability from precision.
7. Modular-Response and Relative-Entropy Estimation uses accessible response without claiming full-state reconstruction.
8. Witnesses for Entanglement, Scrambling, and Recoverability matches finite observations to bounded conclusions.
9. Continuum Extrapolation, Bias, and Uncertainty combines statistical and systematic errors without double counting.
10. Cross-Method Benchmarks and Adversarial Null Tests designs independent comparisons.
11. Claim–Evidence Records, Replication, and Retraction Handling maintains dated scientific conclusions when evidence changes.
12. Reproducible Reporting Standard for QFT Information gives the minimum record for independent reproduction.

Conventions and evidence boundaries

Every uncertainty states whether it is a standard deviation, confidence interval, credible interval, or deterministic bound and how coverage was checked. Nonlinear transformations are applied within resampling or posterior propagation, not only to central values. Continuum fits state their correction ansatz, fit window, covariance, and regulator family. Witness decisions state the null, decision threshold, multiplicity treatment, and calibration uncertainty.

General algorithms, Monte Carlo, and continuum-fit machinery are developed in Volume 8; instrument platforms and material-specific claims are treated in Volume 12. This chapter concerns the boundary between QFT information evidence and the statements it licenses. A reproducible verification workflow specifies synthetic cross-method tests; its existence is not evidence that any estimator has passed them.

Review the chapter

Moment matching. A measured covariance violates a separability criterion. What is certified, and what is not?

Verification criteria

After calibration and uncertainty are included, the witness can certify entanglement under its measurement assumptions. The covariance determines the full state and entropy only if Gaussianity is justified; a non-Gaussian state can share the same covariance.

Agreement. A replica estimator and covariance estimator agree. Why might this not be independent confirmation?

Verification criteria

They may share state preparation, region definition, normalization, detector calibration, regulator, or analysis code. Independence must be mapped at each stage, and at least one held-out or adversarial test should target the shared failure channel.

Correction. A finer regulator reveals a bias outside the published interval. What changes?

Verification criteria

Preserve the original result and inputs, add the new evidence and corrected analysis, update every conclusion that depended on the old extrapolation, explain which assumptions failed, and distinguish a corrected numerical value from a fully withdrawn claim.

References

• Brydges, Tiff, Andreas Elben, Petar Jurcevic, Benoît Vermersch, Christine Maier, Ben P. Lanyon, Peter Zoller, Rainer Blatt, and Christian F. Roos. “Probing Rényi Entanglement Entropy via Randomized Measurements.” Science 364 (2019): 260–263. DOI. Open PDF.
• Huang, Hsin-Yuan, Richard Kueng, and John Preskill. “Predicting Many Properties of a Quantum System from Very Few Measurements.” Nature Physics 16 (2020): 1050–1057. DOI. Open PDF.
• Wilkinson, Mark D., Michel Dumontier, IJsbrand Jan Aalbersberg, et al. “The FAIR Guiding Principles for Scientific Data Management and Stewardship.” Scientific Data 3 (2016): 160018. DOI.