Gaussian Surrogates for Poisson Imaging: Some Theoretical and Empirical Results

arXiv:2602.17274 · eess.IV, stat.ML · Submitted 2026-02-19 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Gaussian Surrogates for Poisson Imaging".

Jane: In imaging inverse problems where measurements follow a Poisson distribution, this paper investigates whether using Gaussian surrogate objectives can yield Mean Squared Error (MSE) comparable to Poisson-based methods,

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So, we've seen how the paper, "Gaussian Surrogates for Poisson Imaging: Some Theoretical and Empirical Results," investigates using Gaussian objectives to match the performance of Poisson-based methods when dealing with Poisson noise, especially in those tricky low-dose situations. We’ve covered the theoretical analysis of why unregularized methods struggle and how these Gaussian surrogates offer a better MSE bound than their counterparts.

Jane: That's right, Tom; the authors demonstrate that proper regularization can effectively mitigate the variance spikes inherent in low-count measurements, showing that simple quadratic objectives are surprisingly competitive with Poisson Maximum A Posteriori estimators. The empirical results confirmed this across different count levels in CT imaging, validating the theoretical claims made in "Gaussian Surrogates for Poisson Imaging: Some Theoretical and Empirical Results."

Lu: The implications here stretch beyond just image quality; it suggests a new way to approach ill-posed inverse problems where the noise structure is non-Gaussian, offering a more flexible mathematical toolkit for reconstruction.

Meng: From an engineering standpoint, this means we can start designing reconstruction algorithms around these Gaussian surrogates as a baseline for low-dose data processing, rather than sticking strictly to the Poisson likelihood formulation which proved unstable in those regimes.

Lalam: I think the broader impact is in how we build AI models that interpret noisy or sparse data; if our underlying reconstruction methods are more stable and accurate, any downstream AI interpretation will benefit from a higher quality input signal.

Tom: Exactly, Lalam; it’s about building a foundation that handles uncertainty better. So, to wrap up on "Gaussian Surrogates for Poisson Imaging: Some Theoretical and Empirical Results," the authors suggest that these surrogate objectives provide provably smaller MSE in the low-dose regime when regularized compared to standard Poisson methods.

Jane: That’s the central message; they prove that even with Poisson noise, a carefully chosen Gaussian surrogate objective can lead to reconstructions with comparable error metrics, particularly when regularization is applied. We need to keep an eye on how this methodology can be integrated into real-world imaging systems.

Lu: I think the future work should explore how these heteroscedastic objectives could be generalized to handle even more complex noise structures found in real-world sensor data, expanding the applicability of this framework significantly.

Meng: I'm curious if they have suggestions on how to optimize the regularization parameter gamma i based on the specific characteristics of the measurement operator A for different physical imaging modalities. That would be a key practical step for implementation.

Lalam: It really opens up avenues for developing more resilient AI systems, allowing them to operate effectively in environments where data is inherently sparse or low-count, which is a huge step forward in practical application.

Conclusion: Tom: So, we've just been deep into the technical details of "Gaussian Surrogates for Poisson Imaging: Some Theoretical and Empirical Results," and now it's time to talk about what this whole thing actually means for us on air today.

Jane: That’s right, Tom; we’re shifting gears from the math to the big picture implications of this research.

Lu: From my perspective at Tsinghua, this paper opens up a really interesting avenue for how we model noise in inverse problems that don't follow standard Gaussian assumptions.

Meng: I'm curious about what these theoretical bounds translate to in terms of actual system design when we’re trying to build something practical.

Lalam: I think the real cultural impact here is how it helps us build trust in AI-driven reconstruction because the underlying math becomes more robust under real-world, noisy conditions.

Tom: Exactly, Lalam; that robustness is what makes this work so compelling when we consider how much of our critical infrastructure relies on accurate imaging.

Jane: The authors are presenting a way to bridge the gap between theoretically ideal Poisson models and the more manageable Gaussian frameworks in practical scenarios.

Lu: They prove that you can use these simpler quadratic objectives, like those based on Weighted Least Squares, and get results that are comparable to the standard Poisson Maximum A Posteriori estimators.

Meng: That comparison is interesting because it suggests we might not always need to implement a complex Poisson likelihood model if we can find a regularized Gaussian surrogate that performs similarly.

Lalam: It means we can design AI models for reconstruction that are less sensitive to the exact nature of the noise process, which should improve their generalizability across different data sets.

Tom: And the empirical validation in 2D parallel-beam CT really backs up those theoretical claims by showing these methods hold up across different count levels <ref:2602.17274#pg1>.

Jane: It’s a very reassuring finding that we can achieve better error metrics with these Gaussian surrogates, especially when dealing with the low-dose situations where traditional approximations often fail.

Lu: The paper suggests that proper regularization is the essential ingredient here for dampening those variance spikes that plague low-count measurements.

Meng: So, the main point seems to be that for many practical imaging problems, a carefully chosen surrogate objective can deliver competitive performance without needing the full complexity of a Poisson likelihood setup.

Lalam: It’s about making AI systems smarter by giving them more stable mathematical foundations when they're facing messy data.

Tom: Right; so, the title and authors point to this work as a significant step in finding more efficient and reliable ways to process noisy imaging data without sacrificing accuracy.

Alexandra Spitzer, Lorenzo Baldassari, Valentin Derbanot, Ivan Dokmanic

Department of Mathematics and Computer Science, University of Basel · INSA-Lyon, Universite Claude Bernard Lyon 1, CNRS, Inserm

eess.IV, stat.ML

Submitted: 2026-02-19

Updated: 2026-10-05

Code: https://github.com/AlexandraSpitzer/tomoreconstruction-gaussian-poisson

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 83/100

The gist: In imaging inverse problems where measurements follow a Poisson distribution, this paper investigates whether using Gaussian surrogate objectives can yield Mean Squared Error (MSE) comparable to

Key concepts

Poisson MLE: This is the standard method used to estimate image data when measurements are expected to follow a Poisson distribution, which is common in low-dose imaging. It minimizes a function derived from the Poisson log-likelihood, but it suffers from large variance spikes when counts are very low.
Gaussian Surrogate Objectives: These are alternative mathematical objectives based on Gaussian distributions used to approximate the true Poisson likelihood. The paper tests two types: one that assumes varying noise levels (heteroscedastic) and one assuming uniform noise (homoscedastic), to see if they perform better than the standard Poisson method.
Regularization: Regularization is a technique used to stabilize solutions in ill-posed problems. In this context, it acts by shrinking the influence of modes (features) that would otherwise show excessive variance growth due to low counts, effectively damping these unstable components.

Terminology

Regularization — 47× in this paper · explained in 3 episodes
Inverse Problem — 14× in this paper · explained in 2 episodes
Weighted Least Squares — 6× in this paper
Poisson Maximum A Posteriori — 5× in this paper
Mean Squared Error — 3× in this paper
Benchmark — 2× in this paper · explained in 2 episodes
Robustness — 2× in this paper · explained in 5 episodes
Accuracy — 1× in this paper · explained in 1 episode
Core — 1× in this paper · explained in 2 episodes
Generalizability — 1× in this paper · explained in 6 episodes
Poisson Maximum Likelihood Estimator — 1× in this paper
Reconstruction Error — 1× in this paper · explained in 2 episodes

Summary

In imaging inverse problems where measurements follow a Poisson distribution, this paper investigates whether using Gaussian surrogate objectives can yield Mean Squared Error (MSE) comparable to Poisson-based methods, particularly in low-dose regimes where standard approximations might fail. The core finding is that proper regularization can mitigate the variance spikes inherent in low-count measurements, suggesting that simple quadratic objectives may be surprisingly competitive with Poisson Maximum A Posteriori (MAP) estimators.

Theoretical Analysis of Poisson MLE

The analysis begins by considering the Poisson Maximum Likelihood Estimator (MLE), which minimizes an extended-valued function derived from the Poisson log-likelihood: xMLE,P ∈ arg min x∈X+ LP (x; y), LP (x; y):= Xm j=1 (s(Ax)j − yj log(s(Ax)j) in the low-dose regime where µj ≪ 1. In a stylized diagonal model, the unregularized Poisson MLE incurs large MSE due to variance spikes: small diagonal entries push individual modes into an effective low-count regime where single-photon events create variance spikes (Section 2.1). The analysis shows that for a mode with expected count µj, the per-mode MSE ratio relative to the Poisson MLE satisfies E(ˆxTik,P,i(yi) − x ⋆ i) squared E(ˆxMLE,P,i − x ⋆ i) squared = (sai) 2h 2 i + O(µj), µj → 0. This ratio is explicitly related to the regularization level γ i:= τ (sai) squared, where γ i can be viewed as an effective regularization level: the larger γ i, the stronger the shrinkage induced by the penalty relative to the Poisson sensitivity of the forward map.

Surrogate Objectives and MSE Comparison

The paper studies two primary Gaussian surrogate objectives designed to replace or approximate Poisson likelihoods. First, a heteroscedastic objective motivated by the normal approximation of Poisson data, which is shown to achieve smaller MSE than the Poisson MLE (Proposition A.1). Second, a homoscedastic Gaussian surrogate derived from the simplest Weighted Least Squares (WLS) model: yj ∼ Ns(Ax⋆)j, 1. The analysis of this homoscedastic MAP estimator with Tikhonov regularization shows that in the low-dose regime, it yield[s] provably smaller MSE in the same low-dose regime, and we compare it with Poisson maximum a posteriori (MAP) with Tikhonov regularization (Section 2.2.1). Specifically, for the homoscedastic surrogate, the per-mode MSE ratio is shown to be bounded by E(ˆxTik,G,i(yi) − x ⋆ i) squared E(ˆxMLE,P,i − x ⋆ i) squared = c(ϵ) squared + O(µj), µj → 0, where the constant c(ϵ) squared is bounded between "1/4 and √5 - 1/2 for all ε > 0."

Experimental Validation in Computed Tomography

The theoretical findings are validated through numerical experiments on a stylized diagonal model and, more broadly, on 2D parallel-beam CT. The experiments compare three objectives: (1) Poisson MAP, (2) regularized HG MAP (with Tikhonov prior), and (3) penalized weighted least squares (PWLS). The results indicate that for low count levels (Average expected number of counts c=10), the reconstruction obtained with homoscedastic LS or regularized HG MAP is better than Poisson MAP. As the count level increases to c=1000, anatomical structure becomes more visually coherent for all methods, and there is little or no perceptual difference. The conclusion drawn from these experiments is that in the low-count regime, Poisson MAP is not consistently better in MSE, as simple regularized quadratic solvers match Poisson MAP in MSE across tested count levels.

Key Findings on Regularization and Low-Dose Behavior

The paper establishes a critical role for regularization when dealing with ill-posed problems under Poisson noise. The analysis demonstrates that regularization acts primarily by damping these modes that would otherwise exhibit large variance growth.

Improvements for AI systems

As a fastidious and diligent researcher, I have analyzed the core findings of this paper concerning Gaussian Surrogates for Poisson Imaging. The key insight is that in low-count regimes, simple quadratic objectives (like homoscedastic Gaussian surrogates) can yield MSE comparable to or better than Poisson Maximum A Posteriori (MAP) estimators, provided proper regularization is applied.

Here are the specific improvements and capabilities this research enables for AI systems:

The core improvement lies in developing more robust and computationally efficient inverse problems solvers, particularly for low-dose medical imaging and scientific data where photon/count statistics are Poisson distributed.

Enhance Robustness of Low-Dose Reconstruction Algorithms:
Develop Computationally Tractable, High-Performance Solvers:
Improve Performance in Noisy or Sparse Data Environments:

Specific Improvements and Capabilities:

The paper demonstrates that the unregularized Poisson Maximum Likelihood Estimator (MLE) suffers from large MSE spikes at low dose because single-count events create variance spikes. The proposed Gaussian surrogates (especially the homoscedastic one) can mitigate this instability by yielding a linear estimator, which is both computationally simple and analytically tractable.
This suggests that replacing complex Poisson likelihood maximization (like Richardson-Lucy or ML-EM) with simpler, quadratic objectives derived from Gaussian approximations can lead to equivalent MSE performance in the low-dose regime.
The research provides explicit analytical bounds on the MSE ratio for both Tikhonov regularization applied to Poisson objectives and Gaussian surrogates, showing that the performance gain depends critically on the effective regularization level (the parameter related to dose and noise variance).
In practical applications (Computed Tomography experiments), this translates to developing reconstruction pipelines where standard quadratic solvers (like Ordinary Least Squares or simple Tikhonov regularization) can achieve MSE comparable to sophisticated Poisson MAP methods, even when data is severely photon-limited.

Specific Capabilities of the Improved AI System:

An AI system designed for medical image reconstruction (e.g., CT, PET) that handles low-dose scans will use a surrogate objective function instead of the exact Poisson likelihood to optimize its reconstruction weights or parameters.
The system can be trained using simplified Gaussian priors (like a homoscedastic model) which are computationally cheaper than full Poisson solvers, leading to faster inference times during clinical use without sacrificing image quality in low-count scenarios.
The system's regularization strength can be dynamically tuned based on the estimated dose level and noise characteristics, utilizing the analytical insights derived from the paper (e.g., identifying the balanced regime where regularization starts dominating) to prevent variance explosion at low counts.
The resulting AI model will exhibit superior stability in noisy environments, specifically by effectively damping noise spikes generated by rare events without requiring complex, iterative Poisson solvers that are prone to instability when signal-to-noise ratios are low.