Posterior Mean Directly from Noise

mediumPROParameterizations and the Reverse PosteriorDiffusion ModelsDiffusion Models 1

Problem Statement

In practice the network predicts eps, and the posterior mean has a compact form that skips reconstructing x0 explicitly. Implement that direct eps-based mean.

Background

The DDPM reverse mean can be written directly in terms of the predicted noise:

$\mu_\theta = \frac{1}{\sqrt{\alpha_t}}\left(x_t - \frac{\beta_t}{\sqrt{1 - \bar{\alpha}_t}}\, \hat{\varepsilon}\right)$

with alpha_t = 1 - beta_t. This is the p_theta(x_{t-1}|x_t) mean used in the standard ancestral sampler (before adding the noise term).

Your Task

Implement:

def posterior_mean_from_eps(x_t, eps, beta_t, alpha_bar_t):

Return the mean as a list rounded to 4 decimals.

Input Format

x_t, eps: lists of length D.
beta_t, alpha_bar_t (float).

Output Format

A list of D floats rounded to 4 decimals.

Sample

print(posterior_mean_from_eps([1.0, 1.0], [0.5, -0.5], 0.2, 0.5))

Output:

[0.9599, 1.2761]

Example

Input

print(posterior_mean_from_eps([1.0, 1.0], [0.5, -0.5], 0.2, 0.5))

Output

[0.9599, 1.2761]

Reasoning

Compute the signal coefficient $\alpha_t$ by subtracting the noise level from 1: $\alpha_t = 1 - 0.2 = 0.8$ .
Determine the noise scaling factor by dividing $\beta_t$ by the square root of the remaining variance: $\text{coef} = \frac{0.2}{\sqrt{1 - 0.5}} = \frac{0.2}{\sqrt{0.5}} \approx 0.28284$ .
Calculate the numerator for each dimension by subtracting the scaled predicted noise from the current state $x_t$ $x_{t}$ :
- Dimension 1: $1.0 - (0.28284 \times 0.5) \approx 0.85858$
- Dimension 2: $1.0 - (0.28284 \times -0.5) \approx 1.14142$
Scale the results by the inverse square root of $\alpha_t$ $α_{t}$ to obtain the posterior mean:
- Dimension 1: $\frac{0.85858}{\sqrt{0.8}} \approx 0.9599$
- Dimension 2: $\frac{1.14142}{\sqrt{0.8}} \approx 1.2761$
The final output is [0.9599, 1.2761]

Constraints

len(x_t) == len(eps), 0 < beta_t < 1, 0 < alpha_bar_t < 1.
alpha_t = 1 - beta_t; apply the formula above.
Round to 4 decimals; avoid -0.0.

More from Diffusion Models 1

Classifier-Free Guidance: Uncond/Cond Mixingeasy DDPM Posterior Mean and Variance from x0medium Adaptive CFG Scheduling over the Trajectoryhard Classifier Gradient Guidance Stepmedium Negative Prompt as the Unconditional Branchmedium

Concept primerPREVIEW

1. Background Knowledge

In DDPM (Denoising Diffusion Probabilistic Models), the forward process gradually adds Gaussian noise to data, while the reverse process learns to denoise. The reverse step $p_\theta(x_{t-1} \mid x_t)$ is a Gaussian whose mean depends on the predicted noise $\hat{\varepsilon}$ rather than an explicit reconstruction of $x_0$ . This "direct-from-noise" parameterization is numerically stable and avoids an extra division by $\bar{\alpha}_t$ that can amplify floating-point error.

The key schedule variables are:

$\beta_t$ : the per-step noise variance, with $\alpha_t = 1 - \beta_t$ .
$\bar{\alpha}_t = \prod_{s=1}^{t} \alpha_s$ : the cumulative product, so $x_t = \sqrt{\bar{\alpha}_t}\, x_0 + \sqrt{1-\bar{\alpha}_t}\, \varepsilon$ .

The posterior mean formula combines two terms: a scaling of the noisy observation $x_t$ by $1/\sqrt{\alpha_t}$ , and a correction proportional to the predicted noise $\hat{\varepsilon}$ , weighted by $\beta_t / \sqrt{1 - \bar{\alpha}_t}$ . Intuitively, the first term "undoes" the single-step noise, while the second term nudges the estimate toward the learned denoised direction.

2. Algorithm Approach

This is a direct formula evaluation problem. There is no iterative or search-based algorithm; you simply plug the inputs into the closed-form expression element-wise. The pattern is:

Compute the scalar coefficients that depend only on $\beta_t$ and $\bar{\alpha}_t$ .
Apply the per-element affine transformation to each component of $x_t$ and $\hat{\varepsilon}$ .
Round and return.

Because the formula is element-wise, the implementation is a simple vectorized (or loop-based) operation over the $D$ -dimensional vectors.

3. Step-by-Step Strategy

Compute $\alpha_t$ : $\alpha_t = 1 - \beta_t$ .
Compute the scaling factor: $s = 1 / \sqrt{\alpha_t}$ .
Compute the noise-correction weight: $w = \beta_t / \sqrt{1 - \bar{\alpha}_t}$ .
For each dimension $i$ :
$\mu_i = s \cdot (x_t[i] - w \cdot \varepsilon[i])$
Round each $\mu_i$ to 4 decimal places.
Return the resulting list.

PRO

Continue the full explanation

You're reading the free preview. Unlock the complete walkthrough, the code editor, test runner and reference solution with Premium.

PRO

Editor locked

The code editor is locked for Pro problems. It is only available for free problems. Please upgrade to gain access to the code editor for all problems.

solution.py

Python 3.13

Test results