MMSE Restoration Operators

• MMSE restoration operators are conditional expectation mappings that minimize the mean squared error between the true signal and its estimate.
• They can be expressed as proximity operators in various noise environments, linking classical variational formulations with modern deep learning approaches.
• Practical implementations include closed-form linear solutions, sampling methods, and neural networks, making them integral in inverse problem optimization.

Minimum Mean Square Error (MMSE) restoration operators are mappings that, given an observed degraded signal (such as a noisy, blurred, or compressed image), return an estimate minimizing the expected squared distance to the unknown ground truth under a specified probabilistic model. MMSE estimators play a foundational role across Bayesian estimation, signal processing, image restoration, sparse coding, and contemporary machine learning frameworks. Formally, for random variable pairs $(X, Y)$ on appropriate spaces, the MMSE restoration operator is defined as $y \mapsto \mathbb{E}[X \mid Y = y]$. Its unique minimizer property and connections to both classical variational principles and recent deep learning architectures make MMSE operators a cornerstone of modern inverse problem solvers.

1. Mathematical Definition and Theoretical Foundations

Let $X \in \mathbb{R}^n$ (clean image, signal, or parameter) admit a prior distribution $p_X$, and let $Y \in \mathbb{R}^m$ denote its degraded observation under a generative measurement model $p_{Y|X}$. The MMSE restoration operator $R_{\mathrm{MMSE}}$ is defined by

$$R_{\mathrm{MMSE}}(y) := \mathbb{E}[X \mid Y = y] = \int x \, p_{X|Y}(x \mid y)\,dx,$$

which is the unique solution to the minimization problem

$$\arg\min_{f:\mathbb{R}^m\to\mathbb{R}^n} \ \mathbb{E} \|X - f(Y)\|^2.$$

This definition generalizes: for any likelihood $p_{Y|X}$ (additive Gaussian noise, Poisson, or more complex corruptions and degradations), the operator structure holds, but explicit expressions may require approximations or sampling strategies depending on the tractability of the relevant integrals (Niknejad et al., 2018, Nguyen et al., 2022).

Key theoretical results characterize when the MMSE operator can be expressed as a proximity operator (i.e., the minimizer of a quadratic-plus-penalty functional) and under what noise/prior conditions this structure is convex or computationally tractable (Gribonval et al., 2018). For instance, in the additive Gaussian noise setting, for any prior $y \mapsto \mathbb{E}[X \mid Y = y]$0, the conditional mean admits a proximity-operator formulation with an (implicit) regularizer. The same applies under Poisson noise and certain exponential-family models, with generalizations to multivariate scenarios.

2. Explicit Forms in Linear and Structured Models

In the classical linear inverse problem $y \mapsto \mathbb{E}[X \mid Y = y]$1, where $y \mapsto \mathbb{E}[X \mid Y = y]$2 is a zero-mean Gaussian with covariance $y \mapsto \mathbb{E}[X \mid Y = y]$3 and $y \mapsto \mathbb{E}[X \mid Y = y]$4 is zero-mean Gaussian noise with covariance $y \mapsto \mathbb{E}[X \mid Y = y]$5, the MMSE restoration operator admits a closed-form linear solution:

$y \mapsto \mathbb{E}[X \mid Y = y]$6

This operator coincides with the solution to a regularized least-squares (Tikhonov) problem; with diagonal covariances, it reduces to the well-known ridge regression form (Buskulic et al., 12 Feb 2026). Efficient implementations leverage Cholesky factorization for moderately sized dense systems and FFTs for convolutional operators. For $y \mapsto \mathbb{E}[X \mid Y = y]$7 (underdetermined systems), the Woodbury identity enables efficient inversion with lower memory requirements.

In patch-based super-resolution or denoising using Gaussian or generalized Gaussian mixture models (GGMMs), the MMSE restoration operator takes the form of a posterior-weighted average of per-component conditional means. The direct synthesis is as follows (for input patch $y \mapsto \mathbb{E}[X \mid Y = y]$8, mixture weights $y \mapsto \mathbb{E}[X \mid Y = y]$9, and per-component means $X \in \mathbb{R}^n$0):

$X \in \mathbb{R}^n$1

where weights and means are determined from the learned GGMM, yielding an explicit patchwise MMSE estimator (Nguyen et al., 2022).

For sparse coding with possibly intractable summation over supports, MMSE can be approximated via stochastic resonance techniques: multiple sparse pursuits are run on noise-perturbed measurements, supports are aggregated, and the final estimate is a posterior-weighted or empirical mean over oracle-conditionals (Simon et al., 2018).

3. Proximal Operator Interpretations and Implicit Regularization

A landmark result by Gribonval and Nikolova establishes that, for a broad class of noise models (notably additive Gaussian noise and log-concave additive noise), MMSE restoration operators are proximity operators of (possibly non-convex) penalty functions $X \in \mathbb{R}^n$2. In detail, $X \in \mathbb{R}^n$3 satisfies

$X \in \mathbb{R}^n$4

(Gribonval et al., 2018). For Gaussian denoising, $X \in \mathbb{R}^n$5 is related to the negative log-marginal likelihood of $X \in \mathbb{R}^n$6, and Tweedie's formula connects the MMSE denoiser to the gradient of the log-marginal. The multivariate generalization holds for exponential family models, with explicit conditions characterizing when such a prox structure exists.

Such proximal operator identities justify and unify recent Plug-and-Play (PnP) algorithms and Regularization by Denoising (RED), as they allow the implicit insertion of MMSE denoisers as regularizers within broader optimization or ADMM frameworks without necessitating an explicit penalty function (Park et al., 2023).

4. Practical Approximations: Sampling and Learning-based Operators

When the underlying posterior is analytically intractable or when priors are non-parametric/non-Gaussian, MMSE restoration operators are approximated via Monte Carlo, self-normalized importance sampling (SNIS), or adaptive sampling from empirical datasets. External patch-based methods infuse datasets of clean patches, cluster them, and use adaptive mixture proposals for variance reduction in SNIS, achieving consistency to the true MMSE as sample size grows (Niknejad et al., 2018). This generalizes classical algorithms such as non-local means and applies to arbitrary likelihoods (e.g., Poisson, inpainting) with high empirical performance gains.

In sparse coding, deliberate controlled noise injection (stochastic resonance) and aggregation over the supports found by standard pursuit algorithms provide a black-box, consistent MMSE approximation—provably converging as the number of samples increases (Simon et al., 2018).

Deep neural restoration priors (either trained as denoisers or for more general degradations) can implement MMSE operators in a supervised context. Ensembles of such networks, as in the ShaRP framework, serve as effective image priors with direct links to the score of the marginal likelihood (Tweedie's formula). Combining MMSE predictors trained on various degradation models enables better suppression of structured artifacts and more robust inverse problem regularization (Hu et al., 2024).

5. Role in Optimization and Algorithmic Frameworks

The identification of MMSE operators as proximity or score operators provides a rigorous foundation for their use within modern iterative schemes, including PnP-ADMM, PnP-ISTA, and RED. For any MMSE denoiser (even mildly expansive CNNs), convergence of the PnP-ADMM iterations to stationary points can be guaranteed under mild smoothness and lower-boundedness of the implicit regularizer, without imposing nonexpansiveness conditions (Park et al., 2023). As a result, state-of-the-art restoration networks, when genuinely trained for MMSE, inherit desirable algorithmic convergence properties and can be plugged into broader composite optimization schemes for a variety of inverse problems.

Stochastic gradient schemes also exploit the linkage between MMSE restoration, score estimation, and regularization. In ShaRP, stochastic perturbations yield regularizer gradients via the restoration residual, leading to provably convergent stochastic optimization in the presence of operator bias or estimation error (Hu et al., 2024).

6. Extensions to Perceptual Criteria: Optimal Transport and Distortion–Perception Tradeoff

Traditional MMSE restoration minimizes MSE at the potential expense of perceptual fidelity. Recent theoretical developments show that the optimal estimator under a marginal distribution constraint—i.e., the minimal MSE achievable while enforcing the output distribution match the natural image distribution—can be constructed by optimal transport (OT) from the MMSE posterior mean distribution to the target data distribution. The resulting estimator, denoted $X \in \mathbb{R}^n$7, combines pixel-space or latent-space transformations with MMSE predictions (Ohayon et al., 2024, Adrai et al., 2023).

Algorithms such as Posterior-Mean Rectified Flow (PMRF) implement this principle by first performing MMSE regression, then learning an OT/flow-matching map from the MMSE estimate to the data domain via a neural ODE. The approach strictly outperforms classical posterior sampling in MSE under the perfect-perception constraint, as shown both theoretically and empirically. Similar principles underlie latent-space OT corrections of MMSE predictors in few-shot settings using pretrained VAEs (Adrai et al., 2023).

7. Empirical Performance, Stability, and Algorithmic Significance

Empirical comparisons in controlled settings demonstrate that linear MMSE estimators (LMMSE) offer robust, parameter-free restoration baselines, outperforming MAP approaches in stability and sensitivity to hyperparameters, especially in nonconvex or blind settings. LMMSE initialization substantially boosts MAP method convergence and reduces sensitivity to regularization parameter choice (Buskulic et al., 12 Feb 2026).

Patch-based and non-parametric MMSE estimators yield consistent, often state-of-the-art, performance in denoising, super-resolution, and inverse tasks, especially when leveraging structured priors or domain adaptation (Niknejad et al., 2018, Nguyen et al., 2022). Deep MMSE restoration networks, assembled as priors in ShaRP or similar frameworks, demonstrate improved artifact suppression and sample efficiency over denoiser- or diffusion-based alternatives, with convergence guarantees under standard smoothness and bounded-variance assumptions (Hu et al., 2024).

The stochastic resonance–based MMSE approximations yield close-to-optimal performance in sparse recovery, with convergence guarantees and strong PSNR gains over classical MAP inference in both synthetic and real-world data (Simon et al., 2018).

PMRF and optimal-transport-based post-processing of MMSE predictors enable navigable tradeoffs between distortion and perceptual quality, achieving near-perfect perception with bounded MSE increase, and outperforming both GAN-based and posterior-sampling approaches across multiple image restoration benchmarks (Ohayon et al., 2024, Adrai et al., 2023).

---

Key References:

• (Buskulic et al., 12 Feb 2026) A Comparative Study of MAP and LMMSE Estimators for Blind Inverse Problems
• (Park et al., 2023) Convergence of Nonconvex PnP-ADMM with MMSE Denoisers
• (Gribonval et al., 2018) On Bayesian Estimation and Proximity Operators
• (Niknejad et al., 2018) External Patch-Based Image Restoration Using Importance Sampling
• (Ohayon et al., 2024) Posterior-Mean Rectified Flow: Towards Minimum MSE Photo-Realistic Image Restoration
• (Hu et al., 2024) Stochastic Deep Restoration Priors for Imaging Inverse Problems
• (Simon et al., 2018) MMSE Approximation For Sparse Coding Algorithms Using Stochastic Resonance
• (Nguyen et al., 2022) Patch-based image Super Resolution using generalized Gaussian mixture model
• (Adrai et al., 2023) Deep Optimal Transport: A Practical Algorithm for Photo-realistic Image Restoration