Spectral Signal-to-Noise Ratio (SSNR)

• SSNR is a frequency-domain measure that compares recoverable signal to noise, adapting its definition to the specific estimator or operator used.
• It is applied across domains—such as Gaussian spectral estimation, numerical differentiation, and Brillouin microspectroscopy—providing task-specific fidelity insights.
• SSNR highlights that increased raw signal or observation time does not guarantee improved inference without selecting the appropriate processing method.

Spectral signal-to-noise ratio (SSNR) denotes a family of frequency-domain or spectrum-domain measures that quantify recoverable signal relative to stochastic variability, additive noise, or noise-dominated spectral components. In the literature considered here, the term is not uniform: for a single-trajectory power spectral density of a centered Gaussian process it is naturally identified with the inverse squared coefficient of variation of the estimator (Squarcini et al., 2022); for numerical differentiation it becomes an operator-shaped SNR defined after differentiation, equivalently by frequency-weighted spectra (Verma et al., 24 Jan 2025); in Brillouin microspectroscopy it is the ratio of spectral intensity to the standard deviation of additive noise, either per pixel or averaged over detector samples (Xiang et al., 2019); in STEM phase retrieval it is a dose-aware Fourier-space ratio of mean recovered signal to sample standard deviation across repeated reconstructions (Varnavides et al., 25 Jul 2025); and in graph-Laplacian analysis it emerges as eigenvalue separation between informative spectral modes and a noise bulk or plateau (Ding et al., 2020).

1. Conceptual scope and definitional variability

Across these domains, SSNR is always spectrum-resolved in the broad sense, but the object being compared changes with the inferential task. In one setting the central object is a finite-time PSD estimator, in another it is the output of a differentiation operator, in another it is a pixelated spectrum, and in others it is a reconstructed Fourier component or an eigenvalue of a kernel matrix. This variability is explicit in the surveyed literature (Squarcini et al., 2022, Verma et al., 24 Jan 2025, Xiang et al., 2019, Varnavides et al., 25 Jul 2025, Ding et al., 2020).

Context	Quantity serving as SSNR	Primary object
Single-trajectory Gaussian spectra	$$\mu^2(\omega,T)/\operatorname{Var}S(\omega,T)=1/\gamma^2(\omega,T)$$	Finite-time PSD estimator
Numerical differentiation	$\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$	Operator output
Brillouin microspectroscopy	Spectral intensity divided by noise standard deviation; average per-pixel SNR $I_\infty\Delta/(X\sigma)$	Detector-sampled spectrum
STEM phase retrieval	Mean Fourier component divided by sample standard deviation across reconstructions	Reconstructed spatial frequency
Graph Laplacian spectra	Separation of informative eigenvalues from a noise bulk or plateau	Kernel/GL spectrum

A persistent theme is that SSNR is task-dependent rather than merely data-dependent. The relevant denominator may be estimator variance, additive detector noise, Poisson-limited reconstruction variability, or the noise-dominated part of a random-matrix spectrum. A plausible implication is that any SSNR claim is meaningful only after specifying the operator, estimator, or reconstruction pipeline to which it refers.

2. Single-trajectory spectra of centered Gaussian processes

For an arbitrary real-valued centered Gaussian process $X(t)$ with covariance $C(t_1,t_2)$ that is symmetric, continuous, and non-negative on a finite interval, the single-trajectory PSD is defined by

$$S(\omega,T)=\frac{1}{T}\left|\int_0^T e^{i\omega t}X(t)\,dt\right|^2,$$

with ensemble mean $\mu(\omega,T)=\overline{S(\omega,T)}$. The paper does not require stationarity for the finite-time analysis; stationarity is only needed if one wants to relate to a “true” infinite-time PSD. The central fluctuation metric is the noise-to-signal ratio

$$\gamma(\omega,T)=\frac{\sqrt{\operatorname{Var}S(\omega,T)}}{\mu(\omega,T)},$$

which is exactly the relative standard deviation of the spectral estimate, and whose inverse is a natural intrinsic SSNR of the estimator (Squarcini et al., 2022).

The key result is the universal inequality

$1\le \gamma(\omega,T)\le \sqrt{2},$

equivalently

$$\mu^2(\omega,T)\le \operatorname{Var}S(\omega,T)\le 2\mu^2(\omega,T).$$

Hence a natural estimator SSNR is

$\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$0

with the sharp bound

$\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$1

This establishes that a single-trajectory PSD of a centered Gaussian process is intrinsically noisy even in the absence of additive measurement noise. The lower- and upper-fluctuation limits are attained in concrete regimes: for any centered Gaussian process, $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$2 for all $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$3; for subdiffusive fractional Brownian motion and for the Brownian gyrator, $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$4 as $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$5; and for superdiffusive fractional Brownian motion, $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$6 as $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$7.

The paper also derives the full universal PDF of $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$8 in terms of $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$9 and $I_\infty\Delta/(X\sigma)$0, with exponential tails modulated by a modified Bessel factor. Its maximum is at $I_\infty\Delta/(X\sigma)$1, so the most probable value is the smallest allowed spectral value, while the mean is maintained by rare large fluctuations. The corresponding “typical” spectrum, defined by the geometric mean,

$I_\infty\Delta/(X\sigma)$2

is

$I_\infty\Delta/(X\sigma)$3

with $I_\infty\Delta/(X\sigma)$4. Therefore $I_\infty\Delta/(X\sigma)$5 always, ranging from $I_\infty\Delta/(X\sigma)$6 at $I_\infty\Delta/(X\sigma)$7 down to $I_\infty\Delta/(X\sigma)$8 as $I_\infty\Delta/(X\sigma)$9. The practical consequence is that the ensemble mean is not representative of what is typically observed in a single run.

Averaging over independent spectra changes the scaling. If $X(t)$0 statistically independent trajectories are averaged, the coefficient of variation becomes $X(t)$1 and the effective estimator SSNR becomes $X(t)$2. This suggests that ensemble averaging, rather than merely increasing $X(t)$3, is the direct route to high SSNR for this class of spectral estimators.

3. Operator-shaped SSNR in numerical differentiation

For numerical differentiation, the traditional RMS-based SNR is ineffective because differentiation is a high-pass operator and the damage done by noise depends strongly on its frequency content. In the harmonic example

$X(t)$4

traditional SNR depends only on amplitude ratios, whereas the derivative errors scale with $X(t)$5 or $X(t)$6. The paper therefore defines relevant SNRs after the differentiation operator, not before it (Verma et al., 24 Jan 2025).

For harmonic noise, the derivative errors are exactly the derivatives of the noise:

$X(t)$7

with

$X(t)$8

The paper then sets

$X(t)$9

so that

$C(t_1,t_2)$0

A more symmetric derivative-based form is

$C(t_1,t_2)$1

and analogously for the second derivative.

For white noise, continuous-time differentiation is replaced by discrete-time backward differences. If $C(t_1,t_2)$2 are uncorrelated samples, then the first-derivative noise term has RMS $C(t_1,t_2)$3 and the second-derivative noise term has RMS $C(t_1,t_2)$4. The paper defines

$C(t_1,t_2)$5

so that

$C(t_1,t_2)$6

The ratio $C(t_1,t_2)$7 quantifies the stronger amplification of noise in higher-order differentiation.

The general statement is operator-theoretic. For a linear operator $C(t_1,t_2)$8 with frequency response $C(t_1,t_2)$9, a relevant SNR is

$$S(\omega,T)=\frac{1}{T}\left|\int_0^T e^{i\omega t}X(t)\,dt\right|^2,$$0

or, in the spectral domain,

$$S(\omega,T)=\frac{1}{T}\left|\int_0^T e^{i\omega t}X(t)\,dt\right|^2,$$1

Differentiation is the case $$S(\omega,T)=\frac{1}{T}\left|\int_0^T e^{i\omega t}X(t)\,dt\right|^2,$$2, so the effective SSNR weights signal and noise by $$S(\omega,T)=\frac{1}{T}\left|\int_0^T e^{i\omega t}X(t)\,dt\right|^2,$$3. This is why a naive RMS SNR on raw data can remain almost constant while derivative RMSE changes by orders of magnitude.

4. Spectral SNR in Brillouin microspectroscopy

In Brillouin microspectroscopy, the paper defines SNR as “the ratio of the spectral intensity to the standard deviation of additive noise sources, such as detector noise.” At the pixel level, if the noise-free intensity is $$S(\omega,T)=\frac{1}{T}\left|\int_0^T e^{i\omega t}X(t)\,dt\right|^2,$$4 and the additive noise has standard deviation $$S(\omega,T)=\frac{1}{T}\left|\int_0^T e^{i\omega t}X(t)\,dt\right|^2,$$5, the per-pixel SNR is essentially $$S(\omega,T)=\frac{1}{T}\left|\int_0^T e^{i\omega t}X(t)\,dt\right|^2,$$6. For the detector as a whole, the average per-pixel spectral SNR is

$$S(\omega,T)=\frac{1}{T}\left|\int_0^T e^{i\omega t}X(t)\,dt\right|^2,$$7

where $$S(\omega,T)=\frac{1}{T}\left|\int_0^T e^{i\omega t}X(t)\,dt\right|^2,$$8 is the integrated spectral intensity, $$S(\omega,T)=\frac{1}{T}\left|\int_0^T e^{i\omega t}X(t)\,dt\right|^2,$$9 is the pixel size, $\mu(\omega,T)=\overline{S(\omega,T)}$0 is the detector width, and $\mu(\omega,T)=\overline{S(\omega,T)}$1 is the standard deviation of additive Gaussian white noise (Xiang et al., 2019).

The measurement model is

$\mu(\omega,T)=\overline{S(\omega,T)}$2

and denoising is performed by maximum entropy reconstruction (MER) or wavelet analysis (WA). MER maximizes

$\mu(\omega,T)=\overline{S(\omega,T)}$3

with entropy functional $\mu(\omega,T)=\overline{S(\omega,T)}$4 and Gaussian-noise fidelity

$\mu(\omega,T)=\overline{S(\omega,T)}$5

WA uses soft-thresholding of wavelet coefficients with the universal threshold

$\mu(\omega,T)=\overline{S(\omega,T)}$6

The central estimation targets are the Brillouin shift $\mu(\omega,T)=\overline{S(\omega,T)}$7 and linewidth $\mu(\omega,T)=\overline{S(\omega,T)}$8, obtained by Lorentzian fitting after denoising.

The Cramér–Rao analysis makes the link between spectral SNR and parameter precision explicit: the variance of the shift estimator scales as $\mu(\omega,T)=\overline{S(\omega,T)}$9, equivalently $$\gamma(\omega,T)=\frac{\sqrt{\operatorname{Var}S(\omega,T)}}{\mu(\omega,T)},$$0. In simulations, raw fitting collapses at $$\gamma(\omega,T)=\frac{\sqrt{\operatorname{Var}S(\omega,T)}}{\mu(\omega,T)},$$1, WA improves bias and variance for $$\gamma(\omega,T)=\frac{\sqrt{\operatorname{Var}S(\omega,T)}}{\mu(\omega,T)},$$2, and MER yields the lowest variance and bias across $$\gamma(\omega,T)=\frac{\sqrt{\operatorname{Var}S(\omega,T)}}{\mu(\omega,T)},$$3 to $$\gamma(\omega,T)=\frac{\sqrt{\operatorname{Var}S(\omega,T)}}{\mu(\omega,T)},$$4. At $$\gamma(\omega,T)=\frac{\sqrt{\operatorname{Var}S(\omega,T)}}{\mu(\omega,T)},$$5, MER achieves $$\gamma(\omega,T)=\frac{\sqrt{\operatorname{Var}S(\omega,T)}}{\mu(\omega,T)},$$6 bias and $$\gamma(\omega,T)=\frac{\sqrt{\operatorname{Var}S(\omega,T)}}{\mu(\omega,T)},$$7 relative uncertainty in $$\gamma(\omega,T)=\frac{\sqrt{\operatorname{Var}S(\omega,T)}}{\mu(\omega,T)},$$8 in simulations. For experimental Brillouin spectra of distilled water, the measured speed of sound at $$\gamma(\omega,T)=\frac{\sqrt{\operatorname{Var}S(\omega,T)}}{\mu(\omega,T)},$$9 is $1\le \gamma(\omega,T)\le \sqrt{2},$0 m/s with MER, $1\le \gamma(\omega,T)\le \sqrt{2},$1 m/s with WA, and $1\le \gamma(\omega,T)\le \sqrt{2},$2 m/s with no denoising; the expected value at $1\le \gamma(\omega,T)\le \sqrt{2},$3 is about $1\le \gamma(\omega,T)\le \sqrt{2},$4 m/s. The corresponding RMS fit errors are $1\le \gamma(\omega,T)\le \sqrt{2},$5, $1\le \gamma(\omega,T)\le \sqrt{2},$6, and $1\le \gamma(\omega,T)\le \sqrt{2},$7.

A notable feature of this literature is that effective SSNR is interpreted not only as a pointwise ratio of counts to noise, but also through the information content of the spectrum for downstream parameter estimation. This suggests that a cleaner-looking spectrum is not sufficient by itself; the relevant criterion is whether denoising improves $1\le \gamma(\omega,T)\le \sqrt{2},$8 and $1\le \gamma(\omega,T)\le \sqrt{2},$9 while remaining consistent with the noise model.

5. Dose-aware SSNR in STEM phase retrieval

In STEM phase retrieval, SSNR is defined directly in Fourier space from an ensemble of reconstructions of the same object:

$$\mu^2(\omega,T)\le \operatorname{Var}S(\omega,T)\le 2\mu^2(\omega,T).$$0

where $$\mu^2(\omega,T)\le \operatorname{Var}S(\omega,T)\le 2\mu^2(\omega,T).$$1 is the $$\mu^2(\omega,T)\le \operatorname{Var}S(\omega,T)\le 2\mu^2(\omega,T).$$2-th reconstructed image and $$\mu^2(\omega,T)\le \operatorname{Var}S(\omega,T)\le 2\mu^2(\omega,T).$$3 its Fourier transform. The numerator is the mean recovered signal at spatial frequency $$\mu^2(\omega,T)\le \operatorname{Var}S(\omega,T)\le 2\mu^2(\omega,T).$$4, while the denominator is the sample standard deviation across repeated reconstructions. The paper further relates SSNR to detective quantum efficiency through

$$\mu^2(\omega,T)\le \operatorname{Var}S(\omega,T)\le 2\mu^2(\omega,T).$$5

For white-noise objects, $$\mu^2(\omega,T)\le \operatorname{Var}S(\omega,T)\le 2\mu^2(\omega,T).$$6, where $$\mu^2(\omega,T)\le \operatorname{Var}S(\omega,T)\le 2\mu^2(\omega,T).$$7 is the total electron dose (Varnavides et al., 25 Jul 2025).

The motivation is the inadequacy of the contrast transfer function (CTF) as a practical performance metric. CTF describes the maximum usable signal in the infinite-dose, noise-free limit, but it neglects finite fluence and Poisson counting statistics. SSNR corrects this by incorporating dose, Poisson-limited variability, and algorithmic behavior. The paper shows closed-form noise terms for three direct methods:

$$\mu^2(\omega,T)\le \operatorname{Var}S(\omega,T)\le 2\mu^2(\omega,T).$$8

Accordingly, COM has SSNR that vanishes at low $$\mu^2(\omega,T)\le \operatorname{Var}S(\omega,T)\le 2\mu^2(\omega,T).$$9 because the noise diverges like $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$00, even though its CTF tends to unity there; parallax has SSNR that mirrors its oscillatory CTF; and direct ptychography has SSNR proportional to $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$01.

The same paper distinguishes sharply between direct and iterative ptychography. For white-noise objects, center-of-mass, parallax, and direct ptychography exhibit dose-independent SSNRs in the sense that their frequency dependence and relative magnitude are fixed by intrinsic noise propagation. Iterative ptychography behaves differently: at low fluence its SSNR converges to that of direct ptychography, while at high fluence it saturates at

$\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$02

corresponding to a DQE limit of about $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$03 relative to the ideal Zernike-phase TEM reference. The paper connects this saturation to recent quantum Fisher information bounds. A central consequence is that CTF can significantly overestimate practical performance, especially in low-dose regimes.

6. SSNR-like spectral separation in graph Laplacian spectra

For kernel-based graph Laplacians built from high-dimensional noisy point clouds, the paper does not explicitly define SSNR, but it systematically characterizes how informative spectral modes separate from noise-dominated modes. The model is

$\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$04

where the signal has low-rank covariance with spikes $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$05 and the noise is isotropic with covariance $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$06. The component-wise SNR is

$\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$07

The affinity matrix is

$\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$08

with bandwidth $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$09, and the analysis proceeds in the high-dimensional regime $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$10 (Ding et al., 2020).

With $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$11 and $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$12, the spectral behavior changes by regime. For low SNR, $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$13, the bulk eigenvalues are governed by a shifted Marchenko–Pastur law and only a finite number of outliers carry signal information. For $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$14 with the classic bandwidth $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$15, the noisy affinity matrix is close in operator norm to

$\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$16

and only a small set of leading eigenvalues are nontrivial: $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$17 when $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$18, and $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$19 when $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$20. The remainder form a flat baseline near $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$21. This is naturally interpreted as a spectral signal band above a noise plateau.

For very high SNR, $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$22, the same bandwidth choice becomes destructive: off-diagonal affinities vanish exponentially and $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$23, so the spectrum is uninformative. The paper therefore introduces the signal-adaptive scale $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$24. With this bandwidth, $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$25 approximates a scaled and shifted clean-signal kernel for all $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$26, and for $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$27 it approaches $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$28 itself. A quantile-based adaptive bandwidth recovers the same scaling without prior knowledge of $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$29. This suggests an SSNR viewpoint in which informative eigenvalue separation depends jointly on signal strength and kernel bandwidth; strong signal alone does not guarantee an informative spectrum.

7. Cross-domain regularities and interpretive cautions

A recurring misconception is that “spectral SNR” is a single universal quantity. The surveyed work shows instead that SSNR is estimator-specific and operator-specific. In Gaussian single-trajectory spectroscopy it is an intrinsic property of a finite-time PSD estimator; in numerical differentiation it is the SNR after a high-pass operator; in Brillouin microspectroscopy it is tied to additive detector noise and to the Fisher information for peak parameters; in STEM it is a dose-aware Fourier-space reproducibility metric; and in graph Laplacian analysis it is effectively the spectral separation of informative modes from noise-dominated bulk behavior (Squarcini et al., 2022, Verma et al., 24 Jan 2025, Xiang et al., 2019, Varnavides et al., 25 Jul 2025, Ding et al., 2020).

Another recurring point is that increasing raw signal strength, observation time, or nominal transfer does not automatically improve practical spectral fidelity. For single-trajectory Gaussian spectra, $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$30 does not vanish with $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$31 and may converge only to $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$32 or $\mathrm{RMS}(H*y)/\mathrm{RMS}(H*\eta)$33, so single-run SSNR remains order one (Squarcini et al., 2022). For differentiation, traditional time-domain SNR can remain almost unchanged while derivative error worsens dramatically with noise frequency (Verma et al., 24 Jan 2025). For STEM, a CTF near unity can coexist with poor low-dose recoverability, and iterative ptychography only at high fluence approaches its quantum-limited plateau (Varnavides et al., 25 Jul 2025). For graph Laplacians, excessively strong signal combined with a mis-specified bandwidth can collapse the kernel matrix toward the identity (Ding et al., 2020).

A final regularity is the distinction between raw spectral appearance and inferential utility. In Brillouin microspectroscopy, the decisive question is not only how noisy a spectrum looks, but whether denoising moves shift and linewidth estimates toward the Cramér–Rao limit (Xiang et al., 2019). In Gaussian single-trajectory spectra, the average spectrum can be a poor descriptor because the typical spectrum is substantially smaller than the mean (Squarcini et al., 2022). In differentiation and STEM, the relevant SSNR is defined after the transform or reconstruction, not before it (Verma et al., 24 Jan 2025, Varnavides et al., 25 Jul 2025). This suggests that SSNR is best understood as a task-conditioned measure of recoverable spectral information rather than as a generic scalar property of data alone.