The Reverse Markov Chain
What You'll Learn
You'll understand how the reverse diffusion process also follows a Markov chain structure, stepping backward through time from pure noise to clean data.
The Core Idea
Just as the forward process adds noise step-by-step in a Markov chain (where each step depends only on the previous one), the reverse process works the same way—but backward. Starting from pure Gaussian noise at time T, we denoise one small step at a time, moving from x_T → x_{T-1} → ... → x_1 → x_0, where each step depends only on the immediately previous (noisier) state.
The Markov Property in Reverse
The crucial insight: each denoising step is Markovian. To predict x_{t-1} from x_t, we don't need to know the entire history of all previous states—just the current noisy image x_t and the time step t. This makes the reverse process:
- Memoryless at each step (conditioned on current state)
- Sequential through time steps
- Tractable because we can model each transition independently
Why This Matters
Think of it like reversing a corrupted photograph restoration: you don't need to remember every version of the damage—just look at the current damaged state and know how much damage occurred (the time step) to remove one layer of corruption. Each denoising step uses the learned denoiser conditioned on time t to estimate the slightly cleaner x_{t-1}.
This Markov structure means we can learn a single neural network that handles all time steps, taking (x_t, t) as input and predicting how to move one step closer to clean data.
Key Takeaway: The reverse diffusion process is a Markov chain that steps backward from noise to data, where each denoising transition depends only on the current noisy state and time step, mirroring the forward process's Markovian structure.