TL;DR
This paper introduces the prior-force family, a mathematical framework that unifies implicit priors from Gaussian denoisers (used in PnP and RED) with classical Maximum A Posteriori (MAP) estimation. The authors reveal that the noise level in these methods is not just a tuning parameter or step size, but a fundamental prior parameter that shifts reconstruction preferences from sharp peaks to broader regions of high probability mass as noise increases.
What Problem Does This Paper Address?
In modern imaging, we often solve "inverse problems"—like deblurring or inpainting—by using deep-learning-based denoisers as implicit priors. Popular frameworks like Plug-and-Play (PnP) and Regularization by Denoising (RED) replace explicit mathematical priors with these denoisers.
However, a major mystery remains: What exactly do these denoisers "prefer" when the data is ambiguous? For example, if an image is missing a large chunk (inpainting), how does the denoiser choose between multiple plausible completions? While we know these methods work, we lack a clear understanding of how the "noise level" parameter affects the final decision, especially in the "null space" where the observed data provides no information.
Methods
The authors bridge the gap between implicit denoisers and explicit priors by defining the prior-force family $W_t$: $$W_t(x) = t^{-1}(D_t^{-1}(x) - I)$$ where $D_t$ is a Gaussian denoiser at noise variance $t$. This "force" is essentially the normalized denoising residual. By using this definition, the authors show that PnP, RED, and MAP estimation all solve equilibrium equations that differ only in their prior terms ($W_t$, the score force, and the MAP force, respectively).
Key theoretical contributions include:
- Induced Energy ($R_t$): For Bayes-optimal denoisers, the prior force is the gradient of an "induced energy."
- The Variance-Ordering Law: A new law stating that as the noise level $t$ increases, the energy of branches with higher "retained posterior variance" decreases faster.
- Proximal Hulls: The authors prove that, at leading order, PnP minimizes the proximal hull of the MAP energy. This explains why PnP can select "bridges" between modes rather than just the modes themselves.
Datasets
The study uses two main types of data for evaluation:
- Synthetic Gaussian Mixtures (GMMs): Used to provide "exact" priors where the induced energy and posterior covariance can be calculated analytically. This allows for rigorous verification of the proposed structural identities.
- MNIST Digits: A GMM prior with 20 components was fitted to 20,000 MNIST digits. The authors then performed inpainting experiments on 40 test images to see if their theory could predict which "digit component" the PnP method would select at different noise scales.
Key Insights and Ideas
- Noise Level is a Prior Parameter: The most significant "aha moment" is that changing the noise level $t$ in a PnP algorithm actually changes the underlying prior. It is not merely a step size for optimization.
- From Height to Mass: At small noise levels, the implicit prior behaves like MAP, selecting the highest peak (mode). As noise increases, the preference shifts toward modes with more "mass" (broader regions of high probability), even if their peaks are lower.
- Bridges vs. Cores: Because PnP relates to the proximal hull, it can select a "blend" of signals between two modes (a bridge) if the noise level is high enough to merge their influence. This is visualized in Figure 1.
Key Results
- Predictive Accuracy: On 100 unseen MNIST-like instances, the authors' "component scores" correctly predicted the PnP decision in 96.3% of cases, significantly outperforming the trivial rule that the decision simply stays in the MAP component (77.7%).
- Structural Verification: Numerical tests (Table 4) confirmed that identities like Tweedie’s formula and the new variance-ordering law hold with relative errors as low as $10^{-10}$ to $10^{-13}$.
- Inpainting Behavior: Figure 3 demonstrates that for narrow MAP peaks with little mass, PnP will switch to a different, more "massive" component as the noise level $t$ grows.
Limitations and Future Work
- Non-Symmetric Jacobians: The theory relies on Bayes-optimal denoisers. Many trained neural networks have non-symmetric Jacobians, meaning an "induced energy" might not strictly exist. Extending this to general black-box networks is an open challenge.
- Manifold Priors: The current analysis assumes the prior has "full affine span." Real image data often lies on low-dimensional manifolds, which might cause the MAP energy to degenerate while PnP scores remain well-behaved.
- Annealing: The paper studies fixed noise levels. How these decisions evolve during "annealing" (gradually reducing noise during iteration) remains an open question.
Summary updated