--- title: Adaptive Posterior Diffusion Sampling (AdaPS) url: https://www.emergentmind.com/topics/adaptive-posterior-diffusion-sampling-adaps type: topic --- # Adaptive Posterior Diffusion Sampling (AdaPS) Adaptive Posterior Diffusion Sampling (AdaPS) encompasses a family of frameworks and algorithms that leverage adaptive or data-dependent procedures to sample from complex posterior distributions using diffusion models. AdaPS finds critical application in compressed sensing, Bayesian inference, and general inverse problems, systematically updating the sampling or measurement process based on posterior uncertainty or fidelity to observed data. While differing instantiations of the term appear across the literature—including adaptive measurement selection in compressed sensing, adaptive guidance scaling in inverse-problem diffusion sampling, and adaptive-metric Markov Chain Monte Carlo (MCMC) for Bayesian posterior sampling—the unifying theme is iterative adaptation driven by posterior information, typically without additional training or parameter tuning. ## 1. Foundational Formulation and Principles Across its variants, AdaPS is predicated on the use of diffusion models as generative or prior distributions, often for ill-posed problems where direct inversion is intractable. In the classical compressed sensing (CS) setting [2407.08256], the goal is to reconstruct $x \in \mathbb{R}^D$ from $d \ll D$ measurements: \[ y = Hx + \varepsilon,\qquad \varepsilon \sim \mathcal{N}(0, \sigma^2 I) \] with $H \in \mathbb{R}^{d \times D}$ possibly determined adaptively. The Bayesian posterior is then \[ p(x \mid y) \propto p_\theta(x)\,p(y \mid x) \] where $p_\theta(x)$ is the diffusion-based prior, typically realized by a pre-trained Denoising Diffusion Probabilistic Model (DDPM). In inverse problems, the general observation model similarly takes $y = A x + \varepsilon$ for a known (possibly ill-conditioned) operator $A$ [2511.18471], and the diffusion prior guides the sampling trajectory toward plausible solutions. Crucially, AdaPS modifies the standard (prior) reverse diffusion process by conditioning each sampling step on observed data, and—unlike traditional approaches—adapts key aspects of the sampling or measurement loop based on current posterior estimates or guidance statistics. ## 2. Posterior Sampling and Adaptive Measurement Selection In compressed sensing, AdaPS enables adaptive acquisition by quantifying posterior uncertainty and greedily selecting measurements that maximize expected information gain. From a current set of partial observations $y_{0:nr}$, AdaPS generates $s$ samples from the posterior $p_\theta(x \mid y_{0:nr})$ using a zero-shot diffusion sampler (e.g., DDRM). The empirical covariance of these samples serves as an estimate of the posterior uncertainty: \[ \widehat{\mathrm{Cov}}[x \mid y_{0:nr}] = \frac{1}{s} \sum_{i=1}^s \bar{x}_i \bar{x}_i^\top, \quad \bar{x}_i = x_i - \frac{1}{s} \sum_j x_j \] Measurement selection in the unconstrained case reduces to choosing the top-$r$ eigenvectors of this covariance, while in constrained scenarios (e.g., selecting k-space lines in MRI), the optimal measurement is found by maximizing the expected posterior variance in feasible directions [2407.08256]. AdaPS thus alternates: 1. Sampling the current posterior, 2. Estimating uncertainty, 3. Selecting/allocating the next measurements for maximal expected error reduction. This adapts measurement strategy to the actual posterior structure, outperforming fixed or random sensing in empirical evaluations. ## 3. Diffusion Guidance: Adaptive Likelihood Scaling in Inverse Problems Recent advances deploy AdaPS in diffusion-based inverse problems where balancing the prior and likelihood is critical [2511.18471]. The challenge arises because explicit likelihood gradients $\nabla_{x_t} \log p_t(y \mid x_t)$ are intractable; practitioners use surrogates such as the DPS (Dirac) or $\Pi$GDM (Gaussian) approximations: - DPS: $g_1 = -\sigma_y^{-2} J_t^\top A^\top (y - A \hat{x}_0)$ - $\Pi$GDM: $g_2 = J_t^\top A^\top (r_t^2 A A^\top + \sigma_y^2 I)^{-1}(y - A \hat{x}_0)$ AdaPS derives an adaptive step size $\alpha_t$ per iteration by aligning the residual between predicted and MAP noise ($d_t$) with the chosen guidance direction $g_t$, yielding: \[ \alpha_t = 2 \gamma_t \frac{\langle d_t, g_t \rangle}{\|g_t\|_2^2} \] This data-dependent scaling eliminates the need for manual tuning, adapts naturally to the stochasticity parameter and number of diffusion steps, and empirically offers robust performance across varying observation noise and task settings [2511.18471]. The AdaPS algorithm in this context (see Section 4 table) iteratively updates: 1. Diffusion prior sample, 2. Estimate $\hat{x}_0$, 3. Compute guidance surrogate $g_t$ and residual $d_t$, 4. Adaptive guidance scaling $\alpha_t$, 5. Combined update of state with adaptive data term. | Step | Description | Details / Key Formula | |--------------|----------------------------------------------------------------------|--------------------------------------------------------| | 1 | Prior/prediction and residuals | $\hat{x}_0$, $d_t$ as in [2511.18471] | | 2 | Compute surrogate guidance $g_t$ | DPS/$\Pi$GDM as above | | 3 | Compute adaptive scale $\alpha_t$ | $2\gamma_t\langle d_t, g_t \rangle/\|g_t\|_2^2$ | | 4 | Update sample | $x_{t-1} \leftarrow$ prior step $- \alpha_t g_t$ | ## 4. Proximal and Candidate-Based Sampling Schemes AdaPS has also been instantiated as Diffusion Posterior Proximal Sampling (DPPS) for image restoration [2402.16907], where, instead of a single stochastic proposal per step, $n$ candidate latents are sampled at each reverse step, and the candidate most consistent with the measurement constraint is selected: \[ x_{t-1}^{i^*} = \arg\min_i \| A(x_{t-1}^i) - (C_1 x_t + C_2 y) \|_2^2 \] Aligned initialization further improves convergence by combining measurement signal and noise. Empirical analysis shows that this proximal selection mechanism reduces variance, accelerates measurement-convergence, and, beyond a moderate value of $n$, improvements become marginal while computational overhead grows sub-linearly [2402.16907]. ## 5. Adaptive Metric MCMC for Bayesian Posteriors The AdaPS concept also appears in the context of Bayesian neural network posterior sampling as adaptive-metric Langevin or preconditioned samplers [2403.08609]. These methods employ local, parameter-adaptive step-size matrices (e.g., as in RMSprop or Adam), updating the sample according to local geometry encoded by running gradient statistics. However, unless the required Itô correction term $\Gamma(\theta)$ is included, such samplers generally do not converge to the true posterior. In dimensions where the full correction is omitted or downscaled, the resulting Markov process converges to a biased (distorted) invariant distribution: \[ \pi(\theta) = Z p(\theta|D) G(\theta)^{-\alpha} \] where $G(\theta)$ is the (possibly running-averaged) preconditioner and $\alpha$ parameterizes the exponential averaging. This bias persists even as step size $\epsilon \to 0$, invalidating guarantees for exactness without full second-derivative computations [2403.08609]. ## 6. Practical Considerations and Empirical Performance Across instantiations: - **No retraining:** AdaPS typically requires only a pre-trained diffusion model, eschewing retraining or fine-tuning. - **Tuning:** Key hyperparameters, such as number of samples per adaptation $(s, n)$, block size $r$, and number of adaptation steps $N$, enable a trade-off between computational cost and fidelity [2407.08256, 2402.16907]. In practice, modest values ($s \approx 1.3 r$, $n=20$) suffice for near-optimal behavior. - **Computational cost:** Sampling-based approaches are more expensive than deterministic one-pass inversions; however, mechanisms such as proximal selection or accelerated DDPM variants (e.g., DDRM) keep runtimes tractable [2407.08256, 2402.16907]. - **Robustness and performance:** AdaPS methods outperform non-adaptive or heuristic sampling strategies across compressed sensing, MRI/CT, and standard image restoration, with statistically significant improvements in PSNR, SSIM, and LPIPS [2407.08256, 2402.16907, 2511.18471]. | Application Domain | Performance Gain (relative) | Additional Notes | |---------------------|-------------------------------|-----------------------------------------| | Faces (CS) | +1.6 dB vs. PCA, +6 dB vs. random | Unconstrained, mean-of-posterior improves further | | MRI (FastMRI) | +2 dB over Poisson-disk | Outperforms hand-engineered downsampling | | ImageNet SR/deblur | Lower LPIPS, sharper details | Hyperparameter-free guidance in adaptive schemes | ## 7. Limitations, Open Questions, and Prospects Prominent limitations across AdaPS frameworks include: - Dependence on the expressiveness and consistency of pre-trained diffusion samplers for accurate posterior characterization. - Sampling cost grows with posterior sample count. - Frameworks are currently best suited for linear inverse problems; extension to nonlinear or unknown operators remains an open research direction [2407.08256, 2511.18471]. - In adaptive-metric MCMC, computational feasibility of the required second-order correction hinders practical unbiasedness [2403.08609]. Promising directions involve leveraging distilled or parallelized samplers to amortize sampling cost, incorporating non-Gaussian likelihoods and non-linear operators, and deploying AdaPS in clinical and scientific scenarios for robust uncertainty quantification. --- **References**: - [2407.08256] Adaptive Compressed Sensing with Diffusion-Based Posterior Sampling - [2403.08609] On the Convergence of Locally Adaptive and Scalable Diffusion-Based Sampling Methods for Deep Bayesian Neural Network Posteriors - [2402.16907] Diffusion Posterior Proximal Sampling for Image Restoration - [2511.18471] Robust Posterior Diffusion-based Sampling via Adaptive Guidance Scale Source: https://www.emergentmind.com/topics/adaptive-posterior-diffusion-sampling-adaps