Statistical Efficiency as Model-Dependent Uncertainty

The highlighted sentence—“Statistical efficiency concerns uncertainty under a specified observation model”—is a compact way of separating one part of the paper’s argument from two nearby but different questions. In this Rydberg receiver problem, one may ask whether the signal is uniquely recoverable in an ideal noiseless world, whether the recovered estimate has small variance in a noisy world, and whether the mathematical model corresponds to the physical apparatus. The sentence says that statistical efficiency belongs to the second question: once an observation model has been chosen, how much uncertainty remains in estimating the signal parameters?

This distinction matters because the paper studies a self-calibrating receiver. The receiver is not simply measuring a known sinusoid with known reference amplitude. It must infer the signal while also eliminating or estimating nuisance quantities such as local-oscillator amplitudes, detuning, and phase increments. The same measurement design can therefore look successful from one viewpoint and weak from another. It may be identifiable, meaning the noiseless equations have a unique solution, but still statistically fragile if small noise causes large changes in the inferred signal. Conversely, an estimator may be statistically efficient relative to an assumed noise model while still being physically biased if that assumed model omits a real systematic effect.

The observation model is the probabilistic contract

An observation model specifies how the measured data are generated from the unknown parameters. In the parent document, the ideal noiseless dressed-state separation model has the form, in essence,

$$ \Omega_{jk}=L_j e^{i\phi_{jk}}+s, $$

where $s=a e^{i\theta}$ is the complex signal Rabi frequency, $L_j$ is the local-oscillator amplitude at setting $j$, and $\phi_{jk}$ is the phase used for the $k$th spectrum at that setting. The measured squared separation is modeled through quantities of the form

$$ y_{jk}=D+L_j^2+P+2L_j\sqrt{P}\cos(\phi_{jk}-\theta)+\varepsilon_{jk}, $$

where $P=a^2$ is signal power, $D=\delta^2$ is squared detuning, and $\varepsilon_{jk}$ is measurement noise. This equation is representative of the paper’s squared-separation model: a background term plus a sinusoidal interference term plus noise.

The phrase “under a specified observation model” means that statistical statements are conditional on assumptions such as the following: the mean of $y_{jk}$ is exactly the expression above; the phase values are the intended or modeled values; the noise variables have a particular variance structure; and the noise is independent, Gaussian, or at least sufficiently regular for Fisher-information calculations. If the paper assumes homoscedastic squared-separation noise, for example, it is assuming that the variance of $\varepsilon_{jk}$ is the same across the relevant measurements, or has a stated setting-dependent form if generalized.

Once that probabilistic contract is fixed, one can ask a precise statistical question: how much information do the observations contain about $P$ and $\theta$, after accounting for nuisance parameters such as $L_1,L_2$ and $D$? That is the domain of statistical efficiency.

Efficiency is about variance, not uniqueness

Identifiability is a deterministic property of the mapping from parameters to noiseless observations. If two different parameter vectors produce exactly the same noiseless data, the parameters are not identifiable. If the mapping is one-to-one on the relevant domain, then the parameters are identifiable. This is why the parent document says identifiability “concerns uniqueness from noiseless observations.”

Statistical efficiency enters only after noise is introduced. Suppose the parameters are identifiable, so the noiseless equations have a unique solution. With noisy measurements, an estimator $\hat{\theta}$ of the signal phase will fluctuate from experiment to experiment. Its variance might be small or large. The Cramér–Rao bound formalizes a lower bound on this variance for unbiased estimators under regularity conditions. If $\eta$ denotes the vector of unknown parameters and $I(\eta)$ is the Fisher information matrix, then a standard form of the bound is

$$ \operatorname{Cov}(\hat{\eta}) \succeq I(\eta)^{-1}, $$

when $I(\eta)$ is nonsingular and the estimator is unbiased in the usual local sense [Kay 1993; Van Trees et al. 2013]. The symbol $\succeq$ means that the difference between the left and right sides is positive semidefinite. In words, no unbiased estimator can have covariance smaller than the inverse Fisher information matrix in all directions.

An estimator is often called statistically efficient if it attains, or asymptotically attains, the relevant information bound. But the highlighted sentence is using “statistical efficiency” more broadly to mean the uncertainty analysis itself: how much precision is possible under the chosen noise model, and how much precision is lost by estimating nuisance parameters rather than knowing them.

This is why the paper can discuss “information loss caused by self-calibration.” Self-calibration may preserve identifiability: the signal can still be recovered in principle. But because some data must also determine $L_j$, $D$, or phase increments, less information may remain for $P$ than in an externally calibrated benchmark. The relevant question is not merely “is there a unique solution?” but “how sharply is that solution determined when the observations are noisy?”

Nuisance parameters and information loss

The parent document’s local context emphasizes that the protocol separates an incident signal from imperfectly known local-oscillator quantities. In statistical terms, the local-oscillator amplitudes, detuning, and phase increments are often nuisance parameters: they are not the final quantities of interest, but they affect the observations and must be handled.

A useful way to see the issue is to imagine estimating a signal amplitude from a sinusoid whose offset and scale are unknown. If the offset and scale were known, variation in the data could be attributed more directly to the signal. If they are unknown, some of the same data must be used to infer them. The Fisher information for the signal alone is then reduced by the directions in data space that are also explainable by changes in nuisance parameters.

Mathematically, if the full parameter vector is divided into a parameter of interest $\psi$ and nuisance parameters $\lambda$, the Fisher information matrix can be partitioned as

$$ I(\psi,\lambda)= \begin{pmatrix} I_{\psi\psi} & I_{\psi\lambda}\\ I_{\lambda\psi} & I_{\lambda\lambda} \end{pmatrix}. $$

When $\lambda$ is unknown, the effective information for $\psi$ is commonly represented, under regular conditions, by the Schur complement

$$ I_{\text{eff}}(\psi) = I_{\psi\psi} - I_{\psi\lambda}I_{\lambda\lambda}^{-1}I_{\lambda\psi}. $$

The second term is the information lost because changes in $\psi$ can be partially mimicked by changes in $\lambda$. If $I_{\psi\lambda}=0$, then the parameter of interest is locally orthogonal to the nuisance parameters, and estimating the nuisance parameters does not reduce the information for $\psi$ in that local model. This is the statistical meaning behind the parent document’s later claim that, under uniform phase sampling with a homoscedastic squared-separation noise model, the signal phase is orthogonal to the LO-amplitude and detuning nuisance parameters. That statement is not a universal physical truth; it is a consequence of the specific observation model and sampling symmetry.

The parent document’s formula for the phase Cramér–Rao bound illustrates this model-dependent conclusion. It gives, for two independent settings,

$$ \operatorname{CRB}(\theta) = \left( \sum_{j=1}^{2} \frac{2K L_j^2 P}{\sigma_j^2} \right)^{-1}, $$

where $K$ is the number of uniformly sampled phases per setting, $L_j$ is the LO amplitude at setting $j$, $P$ is signal power, and $\sigma_j^2$ is the squared-separation noise variance at that setting. The structure is intuitive: stronger LO amplitude, stronger signal power, more phase samples, or lower noise all increase information and reduce phase uncertainty. But the conclusion that nuisance elimination leaves this phase information unchanged depends on the assumed sinusoidal model, uniform sampling, and noise structure.

Efficiency does not guarantee physical correctness

The final phrase of the highlighted sentence, “under a specified observation model,” is the safeguard against overclaiming. Statistical efficiency is internal to a model. It does not by itself prove that the model is physically correct.

For example, suppose the true apparatus has an LO-power-dependent detuning contribution, but the observation model assumes a common squared detuning $D$ independent of LO setting. The resulting estimator might be efficient for the wrong model: it could have small variance around a biased value. Similarly, if the commanded phase increments differ systematically from the actual phase increments at the atoms, the fitted sinusoid may look internally consistent while returning a wrong signal phase or amplitude. The parent document explicitly treats such ambiguities as “limits of internal calibration.”

This is a standard distinction in statistical estimation. The Cramér–Rao bound is a statement about variance under specified regularity and distributional assumptions. It does not bound systematic error caused by model misspecification. If the mean function is wrong, or if the noise model is wrong in a way that affects weighting or identifiability, then the computed CRB describes an idealized experiment rather than the actual one. In engineering terms, it is a precision calculation, not a complete accuracy guarantee.

Why the sentence is central to the paper’s logic

The highlighted sentence helps organize the paper’s three layers of reasoning. First, the paper asks whether the unknowns can be uniquely determined from ideal spectra: this is identifiability. Second, it asks how uncertainty propagates through the self-calibrating inversion and how much Fisher information is lost relative to calibrated benchmarks: this is statistical efficiency. Third, it asks whether the ideal dressed-state separation model and its calibration assumptions match the physical Rydberg apparatus: this is physical calibration.

The important inference is that these layers are related but not interchangeable. Identifiability is necessary for meaningful local estimation, but it does not ensure low variance. Statistical efficiency can be derived cleanly only after fixing a noise model, but it does not ensure absence of systematic bias. Physical calibration can validate or invalidate the absolute scale and interpretation of the model, but it is not the same as Fisher-information optimality.

Thus the highlighted passage is not merely definitional. It is a warning about the scope of the paper’s CRB and information-loss results. Those results can show, for example, that a phase estimate suffers no nuisance-parameter information penalty under a particular uniform-sampling, homoscedastic squared-separation model. They cannot, by themselves, show that the laboratory system obeys that model, that the extracted separations are unbiased, or that all phase commands correspond to the actual fields at the atoms. To verify the paper’s reasoning, one should therefore check both the internal statistical derivations and the physical assumptions connecting the mathematical observations to real spectra.

References

[Kay 1993] S. M. Kay, Fundamentals of Statistical Signal Processing, Volume I: Estimation Theory. Prentice Hall, 1993.

[Van Trees et al. 2013] H. L. Van Trees, K. L. Bell, and Z. Tian, Detection, Estimation, and Modulation Theory, Part I, 2nd ed. Wiley, 2013.

[Stoica and Marzetta 2001] P. Stoica and T. L. Marzetta, “Parameter estimation problems with singular information matrices,” IEEE Transactions on Signal Processing, vol. 49, no. 1, pp. 87–90, 2001.

[Holloway et al. 2017] C. L. Holloway, M. T. Simons, J. A. Gordon, A. Dienstfrey, D. A. Anderson, and G. Raithel, “Electric field metrology for SI traceability: Systematic measurement uncertainties in electromagnetically induced transparency in atomic vapor,” Journal of Applied Physics, vol. 121, no. 23, Article 233106, 2017.

[Bruning et al. 1974] J. H. Bruning, D. R. Herriott, J. E. Gallagher, D. P. Rosenfeld, A. D. White, and D. J. Brangaccio, “Digital wavefront measuring interferometer for testing optical surfaces and lenses,” Applied Optics, vol. 13, no. 11, pp. 2693–2703, 1974.