Structural Signal-to-Noise Ratio (SSNR)

• SSNR is a dose-aware metric that quantifies the ratio of recoverable structural signal to random noise using statistical measures in spatial and Fourier domains.
• It is applied in STEM phase retrieval and astrophotometry to set resolution limits and guide experimental design by assessing the reliability of structural parameter estimation.
• SSNR improves upon traditional metrics like the contrast transfer function by directly incorporating physical noise sources and sample statistics to yield accurate reconstructions.

The Structural Signal-to-Noise Ratio (SSNR) quantifies the ratio of recoverable structural signal to random noise in imaging, inference, and model-fitting tasks where random fluctuations—arising from physical measurement noise or limited experimental dose—compete with features of scientific interest. SSNR serves as a dose-aware metric that can be defined in both spatial and Fourier domains, effectively capturing the frequency-dependent or region-specific fidelity of reconstruction methods. It is employed to inform the reliability of structural parameter estimation in fields such as scanning transmission electron microscopy (STEM) and stellar photometry, superseding traditional “contrast-transfer” or “total SNR” metrics by directly integrating physical noise sources and sample statistics (Varnavides et al., 25 Jul 2025, Narbutis et al., 2015).

1. Mathematical Foundations of SSNR

Let $v(\mathbf r)$ denote the reconstructed or observed signal as a function of position, resulting from a physical process (e.g., STEM imaging, simulated star cluster profile) subject to additive noise. For a linear imaging model,

$v(\mathbf r) = (h * O)(\mathbf r) + n(\mathbf r)$

where $h$ is the instrument response (point-spread function), $O(\mathbf r)$ is the true object, and $n(\mathbf r)$ is a zero-mean noise process. In the Fourier domain, where $H(\mathbf q)$ is the transfer function and $O(\mathbf q)$ the object, this is

$$\tilde{v}(\mathbf q) = H(\mathbf q) O(\mathbf q) + n(\mathbf q)$$

The SSNR for spatial frequency $\mathbf q$ is defined via ensemble statistics over $M$ realizations: $$\mathrm{SSNR}(\mathbf q) = \frac{\left| \overline{\tilde v}(\mathbf q) \right|^2}{\sigma^2(\mathbf q)}$$ with mean $$\overline{\tilde v}(\mathbf q) = \frac{1}{M}\sum_i \tilde v_i(\mathbf q)$$ and sample variance $\sigma^2(\mathbf q)$; for white-noise objects, this approximates

$$\mathrm{SSNR}(\mathbf q) \approx \frac{\left| H(\mathbf q) O(\mathbf q) \right|^2}{\langle |n(\mathbf q)|^2 \rangle}$$

This explicit dose-aware estimate is critical for comparing methods or experimental regimes in which the noise is Poisson-limited and inversely proportional to incident dose or flux (Varnavides et al., 25 Jul 2025).

2. SSNR in Model-Based Structural Inference

In unresolved star cluster photometry, SSNR is operationalized via the total flux enclosed within a photometric aperture (e.g., $r_{\rm phot}=10$ px for 80% cluster flux), with Poisson noise from both object and background. The signal-to-noise is defined by (Narbutis et al., 2015): $$\mathrm{S/N} = \frac{f_{\rm phot}}{\sqrt{f_{\rm phot} + \pi r_{\rm phot}^2 \mu_{\rm sky}}}$$ where $f_{\rm phot}$ is the aperture flux, $\mu_{\rm sky}$ the mean sky background. Multiple Poisson realizations are used to assess parameter posteriors via MCMC, enabling quantification of SSNR-driven uncertainties in derived quantities such as core radius, tidal radius, or half-light radius.

A key quantitative finding is that, for all realistic backgrounds, structural parameters (e.g., half-light radius) of King-profile star clusters can be measured to $\lesssim 20\%$ accuracy only for $\mathrm{S/N} \gtrsim 50$ within the 80%-flux aperture. Below this threshold, parameter uncertainties (especially for outskirt-sensitive quantities like tidal radius) become unacceptably large (Narbutis et al., 2015).

3. SSNR in STEM Phase Retrieval

In STEM, the conventional contrast transfer function (CTF) only predicts the maximum transferable signal and ignores the impact of finite electron dose. SSNR corrects this by relating mean squared signal to the dose-dependent noise variance at every spatial frequency. For major phase retrieval methods, dose-aware closed-form SSNRs are available (Varnavides et al., 25 Jul 2025):

Method	SSNR Functional Form (at $\mathbf q$)	Dose Scaling
Center-of-mass	$|\mathbf q|[v*v^*](\mathbf q)\sqrt{N_e}$	$\propto \sqrt{N_e}$
Parallax	$\sin[\chi(\mathbf q)][A*A](\mathbf q)\sqrt{N_e}$	$\propto \sqrt{N_e}$
Direct ptychography	$\sqrt{2C_{\rm pty}(\mathbf q)}\sqrt{N_e}$	$\propto \sqrt{N_e}$
Iterative ptychography	Plateaus at $v_0^2/2$ (high dose); $\approx$ direct ptychography for low dose	$\propto \sqrt{N_e}$ at low dose, saturates at high dose

At low dose ($N_e \ll N_0$), iterative ptychography does not surpass direct ptychography in SSNR, but at high dose it converges towards the quantum Fisher limit $v_0^2 / 2$, at which detector noise no longer improves with increasing dose (Varnavides et al., 25 Jul 2025).

4. Comparison with Traditional Metrics and Implications

Traditional metrics such as the CTF systematically overestimate recovery fidelity by failing to penalize for finite, Poisson-limited statistics. For example, the CTF may predict ideal contrast even as actual fluctuations dominate the measurement. SSNR directly incorporates the noise floor, capturing effects like Thon-ring oscillations and quantifying the frequency cutoff where genuine signal exceeds fluctuations (Varnavides et al., 25 Jul 2025).

A commonly used criterion is that $\mathrm{SSNR}(\mathbf q)\gtrsim 1$ indicates that signal at that frequency is reliably detectable; more stringent cutoffs (3–5) are used to define effective resolution. This enables rigorous dose budgeting and method benchmarking in experiment and simulation.

5. Experimental and Simulation Workflow for SSNR

To estimate SSNR, practitioners typically:

1. Acquire or simulate $M \gtrsim 50$ independent realizations at fixed dose.
2. For each realization, compute the reconstructed signal (e.g., $v_i(\mathbf r)$).
3. Transform to Fourier domain and calculate mean and variance at each frequency.
4. Compute $\mathrm{SSNR}(\mathbf q)$ via the above ratio.
5. Optionally radially average SSNR over frequency shells for visualization and summary.

Such workflows are applicable to both experimental data (with repeated measurements) and Monte Carlo simulations with injected synthetic noise (Varnavides et al., 25 Jul 2025, Narbutis et al., 2015).

6. Parameter Degeneracies and SSNR Dependence

In MCMC-based inference (e.g., fit of King profiles to star clusters), SSNR critically determines the geometry of posterior uncertainty. At low SNRs (e.g., S/N $\sim$ 20), parameter posteriors for core and tidal radius are highly degenerate and non-elliptical, with “banana-shaped” distributions indicating that distinct parameter changes can yield nearly indistinguishable profiles after convolution with PSF and noise. Higher SSNR shrinks posterior volumes but can leave residual degeneracies due to the intrinsic model structure; even at S/N = 100, the primary axis of $r_c$–$r_t$ correlation may persist (Narbutis et al., 2015). Parameter covariances such as between $r_t$ and $f_{\rm cl}$ (profile extent and total flux) or $r_t$ and $\mu_{\rm sky}$ (extent and background) are also SSNRSensitive.

7. Guidelines for Experimental Design and Method Selection

In unresolved astrophysical imaging, S/N $\gtrsim$ 50 within an aperture enclosing 80% of the total flux is required for $\lesssim$ 20% precision in half-light radius, independent of sky background over a factor of 100 in intensity (Narbutis et al., 2015).

For STEM phase retrieval, the SSNR framework enables direct comparison of acquisition and reconstruction strategies. If working at doses below the iterative-single-crossover (few hundred e⁻ Å⁻²), simpler methods (parallax, direct ptychography) may match the SSNR of iterative approaches at much lower computational cost. At very high dose, iterative methods approach the quantum limit, beyond which further dose yields only marginal SSNR improvement (Varnavides et al., 25 Jul 2025).

---

References

• "Structural Parameters of Star Clusters: Signal to Noise Effects" (Narbutis et al., 2015)
• "Beyond Contrast Transfer: Spectral SNR as a Dose-Aware Metric for STEM Phase Retrieval" (Varnavides et al., 25 Jul 2025)